text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Using Rule-Based Labels for Weak Supervised Learning
A ChemNet for Transferable Chemical Property Prediction
arXiv:1712.02734v2 [stat.ML] 18 Mar 2018
Garrett B. Goh
Charles Siegel
PNNL
[email protected]
PNNL
[email protected]
Nathan Hodas
Abhinav Vishnu
PNNL
[email protected]
PNNL
[email protected]
ABSTRACT
1
With access to large datasets, deep neural networks (DNN) have
achieved human-level accuracy in image and speech recognition
tasks. However, in chemistry, data is inherently small and fragmented. In this work, we develop an approach of using rule-based
knowledge for training ChemNet, a transferable and generalizable deep neural network for chemical property prediction that
learns in a weak-supervised manner from large unlabeled chemical databases. When coupled with transfer learning approaches
to predict other smaller datasets for chemical properties that it
was not originally trained on, we show that ChemNet’s accuracy
outperforms contemporary DNN models that were trained using
conventional supervised learning. Furthermore, we demonstrate
that the ChemNet pre-training approach is equally effective on
both CNN (Chemception) and RNN (SMILES2vec) models, indicating that this approach is network architecture agnostic and is
effective across multiple data modalities. Our results indicate a pretrained ChemNet that incorporates chemistry domain knowledge,
enables the development of generalizable neural networks for more
accurate prediction of novel chemical properties.
In the chemical sciences, designing chemicals with desired characteristics, such as a drug that interacts specifically with its intended
target, or a material with specified physical performance ratings,
is, despite decades of research, still largely driven by serendipity
and chemical intuition. Over the decades, various machine learning (ML) algorithms have been developed to predict the activity or
property of chemicals, using engineered features developed using
domain knowledge. Recent work have also started using deep neural networks (DNN) [9, 19, 24, 27], that are on average, typically
more accurate than traditional ML models [12, 14].
CCS CONCEPTS
•Computing methodologies → Transfer learning; Neural networks; Natural language processing; Computer vision; •Applied
computing → Chemistry; Computational biology; •Theory of
computation → Semi-supervised learning;
KEYWORDS
Weak Supervised Learning, Transfer Learning, Computer Vision,
Natural Language Processing, Cheminformatics, Bioinformatics
ACM Reference format:
Garrett B. Goh, Charles Siegel, Nathan Hodas, and Abhinav Vishnu. 2018.
Using Rule-Based Labels for Weak Supervised Learning. In Proceedings of
ACM SIGKDD Conference, London, UK, Aug 2018 (KDD 2018), 9 pages.
DOI: 10.475/123 4
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
KDD 2018, London, UK
© 2018 Copyright held by the owner/author(s). 123-4567-24-567/08/06. . . $15.00
DOI: 10.475/123 4
1.1
INTRODUCTION
Limitations of Feature Engineering and
Data Challenges
Compared to modern deep learning research, the use of DNN models in chemistry relies heavily on engineered features. While such
an approach is advantageous because it utilizes existing knowledge,
using engineered features may limit the search space of potentially
developable representations. This is exacerbated in situations in
which engineered features are not appropriate or inadequate due
to the lack of well-developed domain knowledge.
With the growth of chemical data [14], it may be desirable to fully
leverage representation learning, which will enable one to predict
novel chemical properties for which little or no feature engineering
research has been performed. In computer vision research, this is
achieved by using raw data. For example, unaltered images are used
as the input in various CNN models [17, 30]. In chemistry, DNN
models that leverage representation learning from raw data are
starting to emerge. For example, with minimal feature engineering,
molecular graphs have been used to train DNN models [10, 20].
Other approaches use 2D or 3D images to train convolutional neural
network (CNN) models [15, 16, 32], or SMILES strings to train
recurrent neural network (RNN) models [3, 13].
One factor that complicates representation learning is the limited
amount of usable labeled data in chemistry, which is significantly
smaller than that available in modern deep learning research. For
example, having 100,000 labeled datapoints is considered a significant accomplishment in chemistry. In contrast, in computer
vision research, datasets like ImageNet [28] that includes over a
million images are typically the starting point. While sizable chemical databases like PubChem [21] and ChEMBL [11] do exist, their
labels are skewed towards biomedical data, and such databases
KDD 2018, Aug 2018, London, UK
are only sparsely labelled, where each labeled data (i.e. measurements) are only available for a small subset (typically under 10%)
of the entire database. Therefore, the current state of labeled chemical data is small and fragmented, which reduces the effectiveness of
representation learning when using conventional supervised training
approaches.
1.2
Contributions
Our work addresses the small and fragmented data landscape in
chemistry. This is achieved by leveraging rule-based knowledge
obtained from prior feature engineering research in chemistry [5]
to perform weak supervised learning, and combining it with transfer learning methods used in modern deep learning research [25].
Specifically, we develop ChemNet, the first deep neural network that
is pre-trained with chemistry-relevant representations, making it the
analogous counterpart of a ResNet or GoogleNet for use in the chemical
sciences. Our contributions are as follows.
• We demonstrate that ChemNet learns chemistry-relevant
internal representations, and when coupled with a transfer
learning approach, can be used to predict novel chemical
properties it was not originally trained on.
• We demonstrate the generalizability of ChemNet by predicting a broad range of chemical properties that are relevant to multiple chemical-affliated industries, including
pharmaceuticals, biotechnology, materials and consumer
goods.
• We demonstrate that the ChemNet pre-training approach is
network architecture agnostic and effective across multiple
data modalities, including both CNN and RNN models.
• We demonstrate that ChemNet-based models outperform
otherwise identical models that are trained using conventional supervised learning, and that ChemNet also matches
or exceeds the current state-of-the-art DNN models in the
chemistry literature.
The organization for the rest of the paper is as follows. In section
2, we outline the motivations in developing a chemistry-relevant
rule-based weak supervised learning approach, and the design principles behind ChemNet. In section 3, we examine the datasets, its
broad applicability to chemical-affliated industries, as well as the
training protocols used for pre-training ChemNet, and for evaluating its performance on unseen chemical tasks. Lastly, in section
4, we explore different ChemNet models, and the various factors
that affect model accuracy and generalization. The best ChemNet
model was then evaluated against other DNN models trained using
conventional supervised learning approaches.
1.3
Related Work
Transfer learning is an established technique in deep learning research [25]. This approach first trains a neural network on a larger
database, before fine-tuning it on a smaller dataset. For example,
using ResNet that has been pre-trained on ImageNet to classify
various common objects, may be used with transfer learning techniques to classify specific clothing type. In addition, as long as there
is sufficient overlap in the ”image space” on which the network
was trained on, seemingly unrelated outcomes can be achieved. For
example, a model pre-trained on ImageNet can also be fine-tuned
Goh G.B. et al.
to classify medical images [29]. While medical applications are
seemingly unrelated to conventional image recognition tasks, both
sets of data are natural photographs (i.e. in the same ”image space”)
and thus the lower-level basic representations can be utilized.
In the chemistry literature, because 2D molecular diagrams are
substantially different from natural photographs, and chemistryspecific information is encoded into the image channels, existing
pre-trained models in the computer vision literature on RGB images
for example, would not be fully applicable. In addition, there are
limited examples of using weak supervised learning to utilize large
chemical databases in traing neural networks. Thus, the challenge of
how to convert a sparsely labeled chemical database into a usable form
that can be used in a transfer learning approach for existing chemistry
DNN models is a non-trivial task. In our work, we will use molecular
descriptors to generate consistent and inexpensive rule-based labels,
combined with a weak supervised learning approach to train ChemNet
to develop chemically-relevant internal representations, which is an
approach that is conceptually unique relative to existing methods.
2
CHEMNET DESIGN
In this section, we provide a brief introduction to molecular descriptors, and its role as alternate labels for weak supervised learning
of ChemNet. Then, we document the design principles behind
ChemNet.
2.1
Alternate Labels for Weak Supervised
Learning
Much of the success in modern computer vision research comes
from the availability of large labeled datasets like ImageNet [28].
However, in the chemical sciences, generating labels are both
resource-intensive and time-intensive endeavors, and often the
goal in chemistry research is to predict the chemicals with desired
characteristics (i.e. the labels). Futhermore, because of the way
deep neural networks are trained to recognize patterns, they cannot
be easily programmed with specific rules for chemistry. Therefore,
in order to develop a generalizable ”chemistry-expert” DNN model
that incorporates existing domain knowledge, we need to train
a neural network in a manner such that it will learn basic and
universal chemical representations.
An analagous but hypothetical scenario that explains the data
challenges in chemistry in a traditional computer vision research
context would be as follows. In this fictitious example, the research
objective is to classify object’s names in ImageNet, but obtaining
the names for a large dataset is unfeasible. However, concepts
associated with images can be easily obtained. For example, in a
classification of 3 images [cats, trees, cars], let us assume alternative
labels that describe its texture [furry, rough, smooth] and whether
it is man-made [natural, natural, artificial] can be easily computed.
Therefore, when a CNN model is trained on these alternate labels
in a weak supervised learning approach, representations that are
developed to classify texture and whether an object is man-made,
can be leveraged during a second transfer learning phase when the
model is fine-tuned on smaller labeled dataset that has labels for the
object’s names. In this example, developing internal representations
to identify a furry surface from images, may assist in identifying
cats from trees and cars, as fur is a distinguishing characteristc that
Using Rule-Based Labels for Weak Supervised Learning
KDD 2018, Aug 2018, London, UK
Compute
Predict
Database
(~1 million entries)
Image
Chemception ChemNet
(Molecular Descriptors)
Transfer Learning
C\C=C(/C)C
(=O)OCC1=
CC=CC=C1
Predict
Small Dataset
(~1k - 10k entries)
Image
Chemception ChemNet
(Molecular Properties)
Figure 1: Schematic illustration of ChemNet pre-training on the ChEMBL database using rule-based molecular descriptors, followed by
fine-tuning on smaller labeled datasets on unseen chemical tasks.
cats have. Similarly, developing other representations to identify
artifical components such as wheels from the images, may assist in
identifying cars.
2.2
Using Rule-Based Features to Teach
Chemistry Representations
Moving back to the chemical sciences, our approach is therefore
to use molecular descriptors as alternate labels to perform weak
supervised learning, after which transfer learning methods can
be applied to fine-tune the pre-trained model directly on smaller
labeled datasets of other unseen chemical properties of interests.
Molecular descriptors are engineered features developed through
historical research that stretches back to the late 1940s [26], and over
5000 molecular descriptors have been developed from rule-based
chemistry knowledge [31]. Molecular descriptors are typically computable properties or rule-based descriptions of a chemical’s structure. Some molecular descriptors, such as hydrogen bond donor
count, correspond to intuitive chemical knowledge that chemists
use to conceptualize and understand more complex chemical phenomena. On other hand, there are descriptors such as the Balaban’s
J index [2] that may not be as intuitive, but nevertheless is an abstract topological description of a chemical that has been useful in
various modeling studies.
In the absence of copious amount of data, the representation
learning ability of deep neural networks may not learn optimal
features. Our solution as illustrated in Figure 1 uses molecular
descriptors to generate consistent and inexpensive rule-based labels for large chemical databases that are typically sparsely and
inconsistently labeled. Rule-based labels are then used to train
ChemNet in a supervised manner, using a multi-task learning configuration, where the network attempts to predict all molecular
descriptors simultaneously. In the process, the neural network develops ”chemistry-relevant” representations that we hypothesize
will serve as a better initialization of the network’s weights when
fine-tuning to other smaller and unrelated chemical tasks to be
predicted.
2.3
Hypothesis Behind ChemNet
In this work, we first evaluated the ChemNet approach, on the
Chemception CNN-based model [15, 16]. As there is a strong relationship between a chemical’s structure and its property, we
hypothesize that the use of images of molecular drawings with
a CNN-based model will therefore help facilitate the learning of
chemistry-relevant structural representations. Then, we evaulated
the ChemNet approach on SMILES2vec [13], an RNN-based model
that directly uses SMILES strings for predicting chemical properties. SMILES is a ”chemical language” [33] that encodes structural
information into a compact text representation. As there is a oneto-one mapping between specific characters in the string to specific
structural elements of the chemical, we hypothesize that ChemNet will also facilitate the learning of chemistry-relevant structural
representations even when using a text representation of the chemical. Lastly, by using multi-task learning, we anticipate that the
KDD 2018, Aug 2018, London, UK
Goh G.B. et al.
shared representation learned will be more generalizable, and it
will be used as building blocks to develop more sophisticated and
task-specific representations when fine-tuning on smaller datasets.
3
Dataset for Performance Evaluation
Once ChemNet has been pre-trained, we fine-tune and evaluate
ChemNet’s performance on smaller datasets. To ensure that our
results are comparable with contemporary DNN models reported
in the literature [34] and earlier work on Chemception CNN models [15, 16] and SMILES2vec RNN models [13], we used the Tox21,
HIV, and FreeSolv dataset from the MoleculeNet benchmark [34] for
predicting toxicity, activity and solvation free energy respectively.
The datasets used (Table 1), comprises of a mix of large vs small
datasets, physical vs non-physical properties and regression vs classification problems. None of the above-mentioned chemical tasks
are related to the molecular descriptors used to train ChemNet, and
thus also serve as a measure of ChemNet’s ability to generalize to
predict unseen chemical properties.
3.3
Industrial Application
Based on the datasets tested, we elaborate on applications to the
chemical industry. First, toxicity prediction is of high relevance,
most notably for chemicals that require FDA approval, which includes drugs and other therapeutics (pharmaceuticals) as well as
cosmetics (consumer goods). [22] Activity prediction is proxy to
how well-suited a chemical may be as a drug, and therefore is of relevance to both pharmaceuticals and biotechnology industries. [4]
Solvation free energy values are computable by physics-based simulations, and such methods are currently being employed by pharmaceuticals, consumer goods and materials industries. [7] Therefore,
using neural networks to predict such computable properties with
similar accuracy will potentially lead to several orders of magnitude speed up compared to traditional computational chemistry
simulations that typically take on the order of minutes to hours for
each calculation.
3.4
HIV
FreeSolv
Property
Task
Non-Physical
(Toxicity)
Non-Physical
(Activity)
Physical
(Solvation)
Multi-task
classification
Single-task
classification
Single-task
regression
Size
8014
41,193
643
Table 1: Characteristics of the 3 datasets used to evaluate the
performance of ChemNet.
Dataset for Pre-Training
ChemNet was first trained on the ChEMBL [11] database, which is
a manually curated database of bioactive molecules with drug-like
properties. In this work, after curation, approximately ˜1,700,000
compounds were used. In the initial pre-training stage, we compute
molecular descriptors that serve as inexpensive and consistent (i.e.
no missing) rule-based labels. Specifically, we used RDKit [23] to
compute a list of ˜100 2D descriptors that includes basic computable
properties (e.g MW, logP, etc.), connectivity, constitutional and
topological descriptors.
3.2
Tox21
METHODS
In this section, we provide details on the datasets used, data splitting
and data preparation steps. Then, we document the training and
transfer learning protocols, as well as the evaluation metrics used
in this work.
3.1
Dataset
Data Preparation
The preparation of chemical image data is identical to that reported
by earlier work [15]. Briefly, SMILES strings are converted to their
respective 2D molecular structures using RDKit [23]. The coordinates of the molecule is then used to map it onto a discretized image
of 80 x 80 pixels that corresponds to 0.5 A resolution per pixel. We
initially used the greyscale ”color-coding” scheme reported in the
earlier Chemception paper [15]. However, our subsequent experiments utilized the more sophisticated 4-channel ”color-coding”
scheme, where each atom and bond pixel is assigned a ”color”
based on its local (i.e. pixel-specific) atomic/bond properties, such
as atomic number, partial charge, valence and hybridization. Specifically, we used the ”EngD” augmented image representation ”colorcoding”, and further details about this data preparation protocol
can be obtained from published work [16].
The preparation of the chemical text data is identical to that
reported by earlier work [13]. The SMILES string were first canonicalized using RDKit [23], then unique characters in the string were
then mapped to one-hot vectors. Zero padding was also applied
to both the left and right of the string to construct uniform entries
that were 250 characters long.
3.5
Data Splitting
The dataset splitting steps are identical to that reported previously [15]. We used a 5-fold cross validation protocol for training
and evaluated the performance and early stopping criterion of the
model using the validation set. We also included the performance
on a separate test set as an indicator of generalizability. Specifically,
for the ChEMBL database, Tox21 and HIV dataset, 1/6th of the
dataset was separated out to form the test set, and for the Freesolv
dataset, 1/10th of the dataset was used to form the test set. The remaining 5/6th or 9/10th of the dataset was then used in the random
5-fold cross validation approach for training ChemNet.
For classification tasks (Tox21, HIV), we also over-sampled the
minority class to address the class imbalance observed in the dataset.
This was achieved by computing the imbalance ratio and appending
additional data from the minority class by that ratio. The oversampling step was performed after stratification, to ensure that the
same molecule is not repeated across training/validation/test sets.
3.6
Training the Neural Network
ChemNet was trained using a Tensorflow backend [1] with GPU
acceleration using NVIDIA CuDNN libraries[6]. The network was
created and executed using the Keras 2.0 functional API interface [8].
We use the RMSprop algorithm [18] to train for 50 epochs for
ChEMBL, or 500 epochs for Tox21, HIV, FreeSolv, using the standard
settings recommended (learning rate = 10−3 , ρ = 0.9, ϵ = 10−8 ). We
Using Rule-Based Labels for Weak Supervised Learning
used a batch size of 32, and also included an early stopping protocol
to reduce overfitting. This was done by monitoring the loss of the
validation set, and if there was no improvement in the validation
loss after 10 (ChEMBL) or 50 (Tox21, HIV, FreeSolv) epochs, the last
best model was saved as the final model. In addition, for images,
during the training of the ChemNet, we performed additional realtime data augmentation to the image using the ImageDataGenerator
function in the Keras API, where each image was randomly rotated
between 0 to 180 degrees.
Unless specified otherwise, we used the weights of the final
ChemNet model trained on molecular descriptors, as the initial
weights to initialize subsequent individual models for predicting
toxicity, activity and solvation energy. We also explored different
fine-tuning protocols (see Experiments), where segments (i.e. a
collection of convolutional layers) of ChemNet had its weights
fixed.
3.7
EXPERIMENTS
In this section, we first conduct several experiments to determine
the factors that may affect the performance and generalizability of
ChemNet CNN models. Next, we demonstrated that the ChemNet
approach on other data modalities by training a ChemNet RNN
model. After establishing the best ChemNet model, we compare its
performance against earlier Chemception/SMILES2vec models and
other contemporary DNN models in the literature.
4.1
Model
T1 F32 std
T3 F16 std
T3 F64 std
T1 F32 eng
T3 F16 eng
Architecture
T1
T3
T3
T1
T3
F32
F16
F64
F32
F16
Image
Params
Std
Std
Std
Eng
Eng
276,603
149,741
2,369,681
276,603
149,741
Table 2: Various pre-trained ChemNet models evaluated in
this work investigated different network architectures and
image representations.
Loss Functions and Performance Metrics
For classification tasks (Tox21, HIV) we used the binary crossentropy loss function, and for regression tasks (ChEMBL, FreeSolv)
we used the mean-squared-error loss function. For the initial ChemNet pre-training on the ChEMBL database, we performed min-max
normalization on the molecular descriptors, and these normalized
labels were used for training the neural network. This ensures that
each molecular descriptor is given equal emphasis during training.
For classification tasks (Tox21, HIV), the evaluation metric reported in our paper that determines model’s performance is the
area under the ROC-curve (AUC). For the FreeSolv dataset, the
evaluation metric is RMSE. The reported results in the paper are
the mean value of the evaluation metric, obtained from the 5 runs
in the 5-fold cross validation.
4
KDD 2018, Aug 2018, London, UK
ChemNet Model Exploration
In the absence of more data, network architecture has been a key
driver in increasing model accuracy [17, 30]. Therefore, we first
examine the network architecture and hyperparameters, followed
by an evaluation of the image representation used. The full list of
ChemNet CNN models explored is summarized in Table 2.
We first evaluated the effect of Chemception architecture on
the performance on the 3 chemical tasks: toxicity (Tox21), activity
(HIV) and free energy of solvation (FreeSolv). From earlier work
on optimizing the Chemception architecture, we evaluated both
the baseline T1 F32 and optimized T3 F16 architectures [15]. In the
nomenclature used, Tx refers to the general depth of the network,
and Fx refers to the number of filters in the convolutional layers.
In addition, we also tested a wider and deeper Chemception T3 F64
architecture.
Figure 2: The ChemNet T3 F16 architecture generally had consistently better performance on the validation AUC/RMSE for toxicity, activity and solvation energy predictions. For Tox21 and HIV,
higher AUC is better. For FreeSolv, lower RMSE is better.
The T3 F64 architecture with its wider layers that can accommodate more representations for the simultaneous prediction of ˜100
molecular descriptors attained the lowest normalized validation loss
of 5.39x10−4 during ChemNet pre-training. This is slightly lower
but still in the same order of magnitude as compared to that for
T1 F32 (5.80x10−4 ) and T3 F16 (5.56x10−4 ) architectures. As illustrated in Figure 2, in terms of the validation metrics on the 3 unseen
chemical tasks, the T3 F16 architecture generally had consistently
better performance than T1 F32 even though it has approximately
half the number of parameters. This implies that T3 F16 is not suffering from underfitting. At the same time, the T3 F16 architecture
also has better performance than T3 F64, which implies that adding
more parameters while retaining similar network architecture of
the same depth does not help to improve the generalizability of the
model. Therefore, our findings indicate that amongst the network architectures tested in this work, a deep and narrow T3 F16 architecture
provides the best performance when generalizing to unseen chemical
tasks.
4.2
Synergy of Image Representation and
Transfer Learning
Thus far, we have evaluated our results using the standard images
reported in earlier work [15]. However, subsequent improvements
has shown that augmenting the image channels with basic atom
and/or bond-specific chemical information improves overall performance [16]. We note that the addition of localized (i.e. pixel-specific)
chemical information to the image channels is complementary to
KDD 2018, Aug 2018, London, UK
Goh G.B. et al.
Beginning with a ChemNet model that has all its weights frozen,
we incrementally unfreeze (i.e. fine-tuned) the network starting
with the top segment. We used this segment-based approach instead of a more conventional layer-based approach as the network
architecture was designed with segments as the base unit in mind.
The resulting model performance across all 3 chemical tasks were
recorded as a function of number of segments fine-tuned.
Figure 3: Using augmented images results in consistently better
performance on the validation AUC/RMSE for toxicity, activity and
solvation energy predictions. For Tox21 and HIV, higher AUC is
better. For FreeSolv, lower RMSE is better.
ChemNet’s transfer learning approach, as in this work, we are training the neural network to reproduce global chemical properties
of the entire molecule (i.e. the entire image). As such, a combination of ChemNet with augmented images may lead to additional
performance improvement. The results as summarized in Figure 3
indicates that training ChemNet with augmented images consistently improved performance relative to standard images, and this
is independent of the network architecture and the chemical task.
Therefore, our findings indicate using augmented images of molecular drawings is a synergistic approach to ChemNet transfer learning
methods.
Having investigated various factors that may impact ChemNet
performance, we come to the following conclusions: (i) using the
Chemception T3 F16 architecture provided the best consistent performance, and (ii) training with augmented images consistently improved performance. Therefore, for the remainder of this work, we
will explore our results using the best model identified: T3 F16 eng.
4.3
Transferability of Learned Representations
The presented results in the preceding sections used the final weights
of ChemNet as an initialization scheme for the individually trained
networks for the smaller Tox21, HIV and FreeSolv datasets. We
anticipate that due to the hierarchical nature of deep neural networks, it will learn hierarchical chemical representations, and more
basic (lower-level) representations may not need to be re-trained.
In order to determine which layers of ChemNet needs to be finetuned, we systematically explore the freezing of weights for various
segments in the ChemNet model.
The T3 F16 architecture is constructed from 12 segments, where
each segment comprises of several convolutional layers that are
grouped together based on similarities in their function. Specifically, ChemNet starts with a stem segment that has a single 4x4
convolutional layer, which is used to define the basic spatial region
of the network. Following the stem segment is an alternating series of Inception-Resnet segments and Reduction segments. Each
Inception-Resnet segment is a group of 4 convolutional layers that
collectively perform inception-style operation with residual links,
and the reduction segment is a group of 3 convolutional layers
that downsamples the image. For further details of the network
architecture, we refer our readers to earlier work [15].
Figure 4: Fine-tuning beyond the 7th segment of the ChemNet
T3 F16 architecture yield diminishing returns in performance improvement (validation AUC/RMSE) for toxicity, activity and solvation energy predictions.
As illustrated in Figure 4, when less than 3 segments are finetuned, which is also when most of the network has its weights
frozen, the model performance is relatively poor. This is an expected
behavior as ChemNet was trained on a separate set of molecular
descriptors that are unrelated to toxicity, activity and solvation
energy predictions. As the number of segments that are fine-tuned
increases, so does the model’s performance. At the limit where all
12 segments are fine-tuned, we recover the results reported in the
previous section.
Using Rule-Based Labels for Weak Supervised Learning
We observed that there reaches a point of diminishing performance improvement at around the 7th segment, where additional
fine-tuning of more segments do not consistently improve the results. This indicates that almost half of the network does not need
to be re-trained, which suggest that the first half of ChemNet has
developed basic chemical representations that are eminently transferable to other chemical tasks, and the second half of ChemNet
develops more complex representations that needs to be fine-tuned
for the specific property to be predicted. Therefore, our findings
suggest that ChemNet, particularly for the lower layers have learned
universal chemical representations that are generalizable to the prediction of other chemical properties.
4.4
KDD 2018, Aug 2018, London, UK
an RNN-based model that uses SMILES strings (chemical text) as input. The network architecture is based on the original SMILES2vec
work [13]. The results as summarized in Figure 6, show that similar performance gain can be acheived with RNN models using
a chemical text representation. Specifically ChemNet achieves a
validation AUC of 0.81 and 0.80 for toxicity and activity predictions
respectively, and a validation RMSE of 1.23 kcal/mol for solvation
free energy. Therefore, our findings indicate that the ChemNet pretraining approach provides consistent performance improvement that
is independent of the network’s architecture and data modality.
Performance Gain from ChemNet Transfer
Learning
Having identified the best ChemNet model and fine-tuning protocol, we now evaluate the performance of ChemNet against earlier
Chemception models that do not utilize transfer learning. In Figure 5, we summarize the performance across various generations
of Chemception-based models: Chemception refers to the original
model that uses standard images [15], AugChemception refers to
the modification of using augmented images [16] and ChemNet
refers to the results of this work.
Figure 6: ChemNet (based on the SMILES2vec RNN model) provides consistently better performance on the validation AUC/RMSE
for toxicity, activity and solvation energy predictions, as compared
to earlier SMILES2vec models that do not utilize transfer learning.
4.6
Figure 5: ChemNet (based on the Chemception CNN model) provides consistently better performance on the validation AUC/RMSE
for toxicity, activity and solvation energy predictions, as compared
to earlier Chemception models that do not utilize transfer learning.
Across all 3 chemical tasks, we observed that ChemNet achieves
the best performance. Specifically, ChemNet achieves a validation
AUC of 0.81 and 0.78 for toxicity and activity predictions respectively, and a validation RMSE of 0.99 kcal/mol for solvation free
energy. Furthermore, we emphasize that there is no difference
between AugChemception and ChemNet in terms of the network
architecture and the images (data) used in the superivsed learning
step, which means that the performance improvement is solely
originating from the transfer learning techniques applied. Therefore,
our findings indicate that the transfer learning techniques used in
ChemNet provide a non-trivial improvement to model performance
even when all other factors are held constant.
4.5
ChemNet Against State-of-the-Art Models
Having established that ChemNet provides consistently better performance than its counterpart Chemception and SMILES2vec models, we now perform benchmarks relative to other contemporary
deep learning models in the literature. Specifically, we compare
it to the MLP DNN model that was trained on molecular fingerprints [34]. In addition, we also include the ConvGraph algorithm,
which is a novel graph-based method for representing chemical
data, and is the current state-of-the-art in many chemical tasks [34].
Figure 7: ChemNet consistently outperforms MLP models trained
on engineered features (molecular fingerprints), and matches the
performance of ConvGraph on the validation AUC/RMSE for toxicity, activity and solvation energy predictions.
ChemNet on Other Data Modalities
Next, to demonstrate that the ChemNet approach is not unique to
just CNN-based models and images, we pre-trained SMILES2vec,
As shown in Figure 7, compared to the MLP DNN model, which is
the current state-of-the-art model that uses engineered features, we
KDD 2018, Aug 2018, London, UK
observe that both ChemNet(CNN) and ChemNet(RNN) consistently
outperforms across all 3 chemical tasks. Relative to the ConvGraph
algorithm, the best ChemNet model matches the performance for
activity (val AUC 0.80 vs 0.80) and toxicity (val AUC 0.81 vs 0.81)
predictions, and ChemNet significantly outperforms for solvation
free energy (val RMSE 0.99 kcal/mol vs 1.30 kcal/mol) prediction.
4.7
Rule-Based Weak Supervised Learning in
Other Domains
While the rule-based weak supervised learning approach that we
have developed to train ChemNet is unique to chemistry, several
design principles can be generalized to other domains. Specifically,
the following factors were critical in enabling rule-based weak supervised learning in this work: (i) The availability of large datasets,
but with the inability to generate ground-truth labels on a largescale, and (ii) prior research in feature engineering and rule-based
models which can be used to generate relatively inexpensive labels. Therefore, it is likely that other scientific, engineering and
financial modeling applications, on which substantial research into
rule-based models have been historically invested will benefit from
this approach.
Furthermore, it should be emphasized that in this approach the
process of pre-training the network is arguably more important
than the accuracy of the initial ChemNet model in predicting the
various rule-based labels (molecular descriptors). Technically, the
rule-based labels on which ChemNet was pre-trained on are not
related to the subsequent chemical properties that ChemNet was
fine-tuned on. However, the hierarchical representations that deep
neural networks form, are a good parallel to the hierarchical nature
on which scientific concepts are built on top of one another. This
suggests that the process of using rule-based transfer learning
could potentially simulate the more conventional learning process
of a domain expert, but without the need to explicitly introduce
domain-specific rules. Thus, it is plausible that the lower layers of
the network will learn representations that are analogous to simpler
concepts in the technical domain, and such representations will
give it the ability to adapt to different and/or unseen data. In the
fine-tuning experiments in this work, which indicates that almost
half of ChemNet has developed universal representations that can
be re-used for predicting novel chemical properties, suggests that
at least for the example of chemistry, such an approach is possible.
5
CONCLUSIONS
In conclusion, we have developed an approach for integrating rulebased knowledge with deep neural networks through weak supervised learning. Using the chemistry domain as an example, we
demonstrate how rule-based knowledge (molecular descriptors),
can be adapted with transfer learning techniques, to train CNN and
RNN models on large unlabeled chemical databases of ˜1,700,000
chemicals. The resulting model, ChemNet, can be fine-tuned on
much smaller datasets of ˜1000 to ˜10,000 chemicals to predict unrelated and novel chemical properties, that are of relevance to
many chemistry-affliated industries. In addition, the ChemNet pretraining approach works effectively across network architectures
and data modalities - where both CNN models (Chemception) using
chemical images and RNN models (SMILES2vec) using chemical
Goh G.B. et al.
text had consistently better performance. For CNN models, we
show that a combination of using augmented chemical images with
the Chemception T3 F16 architecture, and fine-tuning about half
of the network provides the best and most generalizable performance, and RNN models also produce comparable results. ChemNet
consistently outperforms all earlier versions of Chemception and
SMILES2vec for toxicity, activity and solvation energy predictions,
achieving a validation AUC of 0.81, 0.80 and validation RMSE of 0.99
kcal/mol respectively. In addition, ChemNet consistently outperforms contemporary deep learning models trained on engineered
features like molecular fingerprints, and outperforms the current
state-of-the-art ConvGraph algorithm for certain tasks. Furthermore, our fine-tuning experiments suggest that the lower layers of
ChemNet have learned universal chemical representations inspired
from rule-based knowledge, which improves its generalizability to
the prediction of unseen chemical properties. Lastly, we anticipate
the design principles behind our rule-based weak supervised learning approach will be adaptable to other scientific and engineering
domains, where existing rule-based models can be used to generate
data for training ”domain-expert” neural networks in their specific
field of application.
ACKNOWLEDGMENTS
The authors would like to thank Dr. Nathan Baker for helpful
discussions. This work is supported by the following PNNL LDRD
programs: Pauling Postdoctoral Fellowship and Deep Learning for
Scientific Discovery Agile Investment.
REFERENCES
[1] Martı́n Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey
Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al.
2016. TensorFlow: A System for Large-Scale Machine Learning.. In OSDI, Vol. 16.
265–283.
[2] Alexandru T Balaban. 1982. Highly discriminating distance-based topological
index. Chemical Physics Letters 89, 5 (1982), 399–404.
[3] Esben Jannik Bjerrum. 2017. SMILES Enumeration as Data Augmentation for
Neural Network Modeling of Molecules. arXiv preprint arXiv:1703.07076 (2017).
[4] Peter Buchwald and Nicholas Bodor. 2002. Computer-aided drug design: the role
of quantitative structure–property, structure–activity and structure–metabolism
relationships (QSPR, QSAR, QSMR). Drugs Future 27, 6 (2002), 577–588.
[5] Artem Cherkasov, Eugene N Muratov, Denis Fourches, Alexandre Varnek, Igor I
Baskin, Mark Cronin, John Dearden, Paola Gramatica, Yvonne C Martin, Roberto
Todeschini, et al. 2014. QSAR modeling: where have you been? Where are you
going to? Journal of medicinal chemistry 57, 12 (2014), 4977–5010.
[6] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John
Tran, Bryan Catanzaro, and Evan Shelhamer. 2014. cudnn: Efficient primitives
for deep learning. arXiv preprint arXiv:1410.0759 (2014).
[7] John D Chodera, David L Mobley, Michael R Shirts, Richard W Dixon, Kim
Branson, and Vijay S Pande. 2011. Alchemical free energy methods for drug
discovery: progress and challenges. Current opinion in structural biology 21, 2
(2011), 150–160.
[8] François Chollet et al. 2015. Keras. (2015).
[9] George E Dahl, Navdeep Jaitly, and Ruslan Salakhutdinov. 2014. Multi-task
neural networks for QSAR predictions. arXiv preprint arXiv:1406.1231 (2014).
[10] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell,
Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. 2015. Convolutional
networks on graphs for learning molecular fingerprints. In Advances in neural
information processing systems. 2224–2232.
[11] Anna Gaulton, Louisa J Bellis, A Patricia Bento, Jon Chambers, Mark Davies,
Anne Hersey, Yvonne Light, Shaun McGlinchey, David Michalovich, Bissan
Al-Lazikani, et al. 2011. ChEMBL: a large-scale bioactivity database for drug
discovery. Nucleic acids research 40, D1 (2011), D1100–D1107.
[12] Erik Gawehn, Jan A Hiss, and Gisbert Schneider. 2016. Deep learning in drug
discovery. Molecular informatics 35, 1 (2016), 3–14.
[13] Garrett B Goh, Nathan O Hodas, Charles Siegel, and Abhinav Vishnu. 2017.
SMILES2Vec: An Interpretable General-Purpose Deep Neural Network for Predicting Chemical Properties. arXiv preprint arXiv:1712.02034 (2017).
Using Rule-Based Labels for Weak Supervised Learning
[14] Garrett B Goh, Nathan O Hodas, and Abhinav Vishnu. 2017. Deep learning for
computational chemistry. Journal of Computational Chemistry (2017).
[15] Garrett B Goh, Charles Siegel, Abhinav Vishnu, Nathan O Hodas, and Nathan
Baker. 2017. Chemception: A Deep Neural Network with Minimal Chemistry
Knowledge Matches the Performance of Expert-developed QSAR/QSPR Models.
arXiv preprint arXiv:1706.06689 (2017).
[16] Garrett B Goh, Charles Siegel, Abhinav Vishnu, Nathan O Hodas, and Nathan
Baker. 2017. How Much Chemistry Does a Deep Neural Network Need to Know
to Make Accurate Predictions? arXiv preprint arXiv:1710.02238 (2017).
[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Delving deep
into rectifiers: Surpassing human-level performance on imagenet classification.
In Proceedings of the IEEE international conference on computer vision. 1026–1034.
[18] G Hinton, N Srivastava, and K Swersky. 2012. RMSProp: Divide the gradient by
a running average of its recent magnitude. Neural networks for machine learning,
Coursera lecture 6e (2012).
[19] Tyler B Hughes, Na Le Dang, Grover P Miller, and S Joshua Swamidass. 2016.
Modeling Reactivity to Biological Macromolecules with a Deep Multitask Network. ACS central science 2, 8 (2016), 529–537.
[20] Steven Kearnes, Kevin McCloskey, Marc Berndl, Vijay Pande, and Patrick Riley.
2016. Molecular graph convolutions: moving beyond fingerprints. Journal of
computer-aided molecular design 30, 8 (2016), 595–608.
[21] Sunghwan Kim, Paul A Thiessen, Evan E Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A Shoemaker, et al. 2015.
PubChem substance and compound databases. Nucleic acids research 44, D1
(2015), D1202–D1213.
[22] Naomi L Kruhlak, Joseph F Contrera, R Daniel Benz, and Edwin J Matthews. 2007.
Progress in QSAR toxicity screening of pharmaceutical impurities and other
FDA regulated products. Advanced drug delivery reviews 59, 1 (2007), 43–55.
[23] G Landrum. 2016. RDKit: Open-Source Cheminformatics Software. (2016).
[24] Andreas Mayr, Günter Klambauer, Thomas Unterthiner, and Sepp Hochreiter.
2016. DeepTox: toxicity prediction using deep learning. Frontiers in Environmental Science 3 (2016), 80.
[25] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. 2014. Learning
and transferring mid-level image representations using convolutional neural
networks. In Proceedings of the IEEE conference on computer vision and pattern
recognition. 1717–1724.
[26] John R Platt. 1947. Influence of neighbor bonds on additive bond properties in
paraffins. The Journal of Chemical Physics 15, 6 (1947), 419–420.
[27] Bharath Ramsundar, Steven Kearnes, Patrick Riley, Dale Webster, David Konerding, and Vijay Pande. 2015. Massively multitask networks for drug discovery.
arXiv preprint arXiv:1502.02072 (2015).
[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean
Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al.
2015. Imagenet large scale visual recognition challenge. International Journal of
Computer Vision 115, 3 (2015), 211–252.
[29] Chuen-Kai Shie, Chung-Hisang Chuang, Chun-Nan Chou, Meng-Hsi Wu, and
Edward Y Chang. 2015. Transfer representation learning for medical image
analysis. In Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual
International Conference of the IEEE. IEEE, 711–714.
[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. 2015.
Going deeper with convolutions. In Proceedings of the IEEE conference on computer
vision and pattern recognition. 1–9.
[31] Roberto Todeschini and Viviana Consonni. 2008. Handbook of molecular descriptors. Vol. 11. John Wiley & Sons.
[32] Izhar Wallach, Michael Dzamba, and Abraham Heifets. 2015. AtomNet: a deep
convolutional neural network for bioactivity prediction in structure-based drug
discovery. arXiv preprint arXiv:1510.02855 (2015).
[33] David Weininger. 1988. SMILES, a chemical language and information system. 1.
Introduction to methodology and encoding rules. Journal of chemical information
and computer sciences 28, 1 (1988), 31–36.
[34] Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. 2017. MoleculeNet: A
Benchmark for Molecular Machine Learning. arXiv preprint arXiv:1703.00564
(2017).
KDD 2018, Aug 2018, London, UK
| 2cs.AI
|
Proceedings of Machine Learning for Healthcare 2017
JMLR W&C Track Volume 68
Generating Multi-label Discrete Patient Records
using Generative Adversarial Networks
Edward Choi1
MP 2893@ GATECH . EDU
1
Siddharth Biswal
SBISWAL 7@ GATECH . EDU
2
Bradley Malin
BRADLEY. MALIN @ VANDERBILT. EDU
arXiv:1703.06490v3 [cs.LG] 11 Jan 2018
Jon Duke1
JON . DUKE @ GATECH . EDU
Walter F. Stewart3
Jimeng Sun
1
STEWARWF @ SUTTERHEALTH . ORG
1
G EORGIA I NSTITUTE OF T ECHNOLOGY
JSUN @ CC . GATECH . EDU
2
VANDERBILT U NIVERSITY
3
S UTTER H EALTH
Abstract
Access to electronic health record (EHR) data has motivated computational advances in medical research.
However, various concerns, particularly over privacy, can limit access to and collaborative use of EHR data.
Sharing synthetic EHR data could mitigate risk.
In this paper, we propose a new approach, medical Generative Adversarial Network (medGAN), to generate
realistic synthetic patient records. Based on input real patient records, medGAN can generate high-dimensional
discrete variables (e.g., binary and count features) via a combination of an autoencoder and generative adversarial networks. We also propose minibatch averaging to efficiently avoid mode collapse, and increase the
learning efficiency with batch normalization and shortcut connections. To demonstrate feasibility, we showed
that medGAN generates synthetic patient records that achieve comparable performance to real data on many
experiments including distribution statistics, predictive modeling tasks and a medical expert review. We also
empirically observe a limited privacy risk in both identity and attribute disclosure using medGAN.
1. Introduction
The adoption of electronic health records (EHR) by healthcare organizations (HCOs), along with the large
quantity and quality of data now generated, has led to an explosion in computational health. However, the
wide adoption of EHR systems does not automatically lead to easy access to EHR data for researchers. One
reason behind limited access stems from the fact that EHR data are composed of personal identifiers, which
in combination with potentially sensitive medical information, induces privacy concerns. As a result, access
to such data for secondary purposes (e.g., research) is regulated, as well as controlled by the HCOs groups
that are at risk if data are misused or breached. The review process by legal departments and institutional
review boards can take months, with no guarantee of access (Hodge Jr et al., 1999). This process limits timely
opportunities to use data and may slow advances in biomedical knowledge and patient care (Gostin et al.,
2009).
HCOs often aim to mitigate privacy risks through the practice of de-identification (for Civil Rights, 2013),
typically through the perturbation of potentially identifiable attributes (e.g., dates of birth) via generalization,
suppression or randomization. (El Emam et al., 2015) However, this approach is not impregnable to attacks,
such as linkage via residual information to re-identify the individuals to whom the data corresponds (El Emam
et al., 2011b). An alternative approach to de-identification is to generate synthetic data (McLachlan et al.,
2016; Buczak et al., 2010; Lombardo and Moniz, 2008). However, realizing this goal in practice has been
challenging because the resulting synthetic data are often not sufficiently realistic for machine learning tasks.
Since many machine learning models for EHR data use aggregated discrete features derived from longitudinal
EHRs, we concentrate our effort on generating such aggregated data in this study. Although it is ultimately
c 2017.
desirable to generate longitudinal event sequences, in this work we focus on generating high-dimensional
discrete variables, which is an important and challenging problem on its own.
Generative adversarial networks (GANs) have recently been shown to achieve impressive performance in
generating high-quality synthetic images (Goodfellow et al., 2014; Radford et al., 2015; Goodfellow, 2016).
To understand how, it should first be recognized that a GAN consists of two components: a generator that
attempts to generate realistic, but fake, data and a discriminator that aims to distinguish between the generated
fake data and the real data. By playing an adversarial game against each other, the generator can learn
the distribution of the real samples - provided that both the generator and the discriminator are sufficiently
expressive. Empirically, a GAN outperforms other popular generative models such as variational autoencoders
(VAE) (Kingma and Welling, 2013) and PixelRNN/PixelCNN (van den Oord et al., 2016b,a) on the quality of
data (i.e., fake compared to real), in this case images, and on processing speed (Goodfellow, 2016). However,
GANs have not been used for learning the distribution of discrete variables.
To address this limitation, we introduce medGAN, a neural network model that generates high-dimensional,
multi-label discrete variables that represent the events in EHRs (e.g., diagnosis of a certain disease or treatment
of a certain medication). Using EHR source data, medGAN is designed to learn the distribution of discrete
features, such as diagnosis or medication codes via a combination of an autoencoder and the adversarial
framework. In this setting, the autoencoder assists the original GAN to learn the distribution of multi-label
discrete variables. The specific contributions of this work are as follows:
• We define an efficient algorithm to generate high-dimensional multi-label discrete samples by combining
an autoencoder with GAN, which we call medGAN. This algorithm is notable in that it handles both
binary and count variables.
• We propose a simple, yet effective, method called minibatch averaging to cope with the situation where
GAN overfits to a few training samples (i.e., mode collapse), which outperforms previous methods such
as minibatch discrimination.
• We demonstrate a close-to-real data performance of medGAN using real EHR datasets on a set of diverse
tasks, which include reporting distribution statistics, classification performance and medical expert
review.
• We empirically show that medGAN leads to acceptable privacy risks in both presence disclosure (i.e.,
discovery that a patient’s record contributed to the GAN) and attribute disclosure (i.e., discovery of a
patient’s sensitive medical data).
2. Related work
In this section, we begin with a discussion of existing methods for generating synthetic EHR data. This is
followed by a review recent advances in generative adversarial networks (GANs). Finally, we summarize
specific investigations into generating discrete variables using GANs.
Synthetic Data Generation for Health Data: De-identification of EHR data is currently the most generally
accepted technical method for protecting patient privacy when sharing EHR data for research in practice
(Johnson et al., 2016). However, de-identification does not guarantee that a system is devoid of risk. In certain
circumstances, re-identification of patients can be accomplished through residual distinguishable patterns in
various features (e.g., demographics (Sweeney, 1997; El Emam et al., 2011a), diagnoses (Loukides et al.,
2010), lab tests (Atreya et al., 2013), visits across healthcare providers (Malin and Sweeney, 2004), and
genomic variants (Erlich and Narayanan, 2014)) To mitigate re-identification vulnerabilities, researchers in the
statistical disclosure control community have investigated how to generate synthetic datasets. Yet, historically,
these approaches have been limited to summary statistics for only several variables at a time (e.g., (Dreschsler,
2011; Reiter, 2002). For instance McLachlan et al.(2016) used clinical practice guidelines and health incidence
statistics with a state transition machine to generate synthetic patient datasets.
There is some, but limited, work on synthetic data generation in the healthcare domain and, the majority
that has, tend to be disease specific. For example, Buczak et al. (2010) generated EHRs to explore questions
related to the outbreak of specific illnesses, where care patterns in the source EHRs were applied to generate
synthetic datasets. Many of these methods often rely heavily upon domain-specific knowledge along with
actual data to generate synthetic EHRs (Lombardo and Moniz, 2008). More recently, and most related to our
2
work, a privacy-preserving patient data generator was proposed based on a perturbed Gibbs sampler (Park
et al., 2013). Still, this approach can only handle binary variables and its utility was assessed with only a small,
low-dimensional dataset. By contrast, our proposed medGAN directly captures general EHR data without
focusing on a specific disease, which makes it suitable for a greater diversity of applications.
GAN and its Applications: Attempts to advance GANs (Goodfellow et al., 2014) include, but are not limited
to, using convolutional neural networks to improve image processing capacity (Radford et al., 2015), extending
GAN to a conditional architecture for higher quality image generation (Mirza and Osindero, 2014; Denton
et al., 2015; Odena et al., 2016), and text-to-image generation (Reed et al., 2016). We, in particular, pay
attention to the recent studies that attempted to handle discrete variables using GANs.
One way to generate discrete variables with GAN is to invoke reinforcement learning. SeqGAN (Yu
et al., 2016) trains GAN with REINFORCE (Williams, 1992) and Monte-Carlo search to generate word
sequences. Although REINFORCE enables an unbiased estimation of the gradients of the model via sampling,
the estimates come with a high variance. Moreover, SeqGAN focuses on sampling one word (i.e. one-hot) at
each timestep, whereas our goal is to generate multi-label binary/count variables. Alternatively, one could use
specialized distributions, such as the Gumbel-softmax (Jang et al., 2016; Kusner and Hernández-Lobato, 2016),
a concrete distribution (Maddison et al., 2016) or a soft-argmax function (Zhang et al., 2016) to approximate
the gradient of the model from discrete samples. However, since these approaches focus on the softmax
distribution, they cannot be directly invoked for multi-label discrete variables, especially in the count variable
case. Yet another way to handle discrete variables is to generate distributed representations, then decode them
into discrete outputs. For example, Glover (2016) generated document embeddings with a GAN, but did not
attempt to simulate actual documents.
To handle high-dimensional multi-label discrete variables, medGAN generates the distributed representations of patient records with a GAN. It then decodes them to simulated patient records with an autoencoder.
3. Method
This section begins with a formalization of the structure of EHR data and the corresponding mathematical
notation we adopt in this work, This is followed by a detailed description of the medGAN algorithm.
3.1 Description of EHR Data and Notations
We assume there are |C| discrete variables (e.g., diagnosis, medication or procedure codes) in the EHR data that
|C|
can be expressed as a fixed-size vector x ∈ Z+ , where the value of the ith dimension indicates the number of
occurrences (i.e., counts) of the i-th variable in the patient record. In addition to the count variables, a visit
can also be represented as a binary vector x ∈ {0, 1}|C| , where the ith dimension indicates the absence or
occurrence of the ith variable in the patient record. It should be noted that we can also represent demographic
information, such as age and gender, as count and binary variables, respectively.
Learning the distribution of count variables is generally more difficult than learning the distribution of
binary variables. This is because the model needs to learn more than simple co-occurrence relations between
the various dimensions. Moreover, in EHR data, certain clinical concepts tend to occur much more frequently
(e.g., essential hypertension) than others. This is problematic because it can skew a distribution around
different dimensions.
3.2 Preliminary: Generative Adversarial Network
In a GAN, the generator G(z; θg ) accepts a random prior z ∈ Rr and generates synthetic samples G(z) ∈ Rd ,
while the discriminator D(x; θd ) determines whether a given sample is real or fake. The optimal discriminator
D∗ would perfectly distinguish real samples from fake samples. The optimal generator G∗ would generate
fake samples that are indistinguishable from the real samples so that D is forced to make random guesses.
Formally, D and G play the following minimax game with the value function V (G, D):
min max V (G, D) = Ex∼pdata [log D(x)]
G
D
+ Ez∼pz [log(1 − D(G(z)))]
3
Real or Fake?
Figure 1: Architecture of medGAN: The discrete x comes from the source EHR data,
z is the random prior for the generator G;
G is a feedforward network with shortcut
connections (right-hand side figure); An autoencoder (i.e, the encoder Enc and decoder
Dec) is learned from x; The same decoder
Dec is used after the generator G to construct the discrete output. The discriminator
D tries to differentiate real input x and discrete synthetic output Dec(G(z)).
D
Dec(G(z))
Dec(Enc(x))
G(z)
Enc(x)
z
x
where pdata is the distribution of the real samples and pz is the distribution of the random prior, for which
N (0, 1) is generally used. Both G and D iterate in optimizing the respective parameters θg and θd as follows,
θd ← θd + α∇θd
m
1 X
log D(xi ) + log(1 − D(G(zi )))
m i=1
θg ← θg − α∇θg
m
1 X
log(1 − D(G(zi )))
m i=1
where m is the size of the minibatch and α the step size. In practice, however, G can be trained to maximize
log(D(G(z)) instead of minimizing log(1 − D(G(z)) to provide stronger gradients in the early stage of the
training (Goodfellow et al., 2014) as follows,
θg ← θg + α∇θg
m
1 X
log D(G(zi ))
m i=1
(1)
Henceforth, we use Eq.(1) as it showed significantly more stable performance in our investigation. We also
assume throughout the paper that both D and G are implemented with feedforward neural networks.
3.3 medGAN
Since the generator G is trained by the error signal from the discriminator D via backpropagation, the original
|C|
GAN can only learn to approximate discrete patient records x ∈ Z+ with continuous values. We alleviate
this limitation by leveraging the autoencoder. Autoencoders are trained to project given samples to a lower
dimensional space, then project them back to the original space. Such a mechanism leads the autoencoder
to learn salient features of the samples and has been successfully used in certain applications, such as image
processing (Goodfellow et al., 2016; Vincent et al., 2008).
In this work, We apply the autoencoder to learn the salient features of discrete variables that can be applied
to decode the continuous output of G. This allows the gradient flow from D to the decoder Dec to enable
the end-to-end fine-tuning. As depicted by Figure 1, an autoencoder consists of an encoder Enc(x; θenc )
|C|
that compresses the input x ∈ Z+ to Enc(x) ∈ Rh , and a decoder Dec(Enc(x); θdec ) that decompresses
Enc(x) to Dec(Enc(x)) as the reconstruction of the original input x. The objective of the autoencoder is to
minimize the reconstruction error:
m
1 X
||xi − x0i ||22
m i=0
(2)
m
1 X
xi log x0i + (1 − xi ) log(1 − x0i )
m i=0
(3)
where x0i = Dec(Enc(xi ))
where m is the size of the mini-batch. We use the mean squared loss (Eq.(2)) for count variables and cross
entropy loss (Eq.(3)) for binary variables. For count variables, we use rectified linear units (ReLU) as the
4
activation function in both Enc and Dec. For binary variables, we use tanh activation for Enc and the sigmoid
activation for Dec.1
With the pre-trained autoencoder, we can allow GAN to generate distributed representation of patient
records (i.e., the output of the encoder Enc), rather than generating patient records directly. Then the pretrained decoder Dec can pick up the right signals from G(z) to convert it to the patient record Dec(G(z)).
The discriminator D is trained to determine whether the given input is a synthetic sample Dec(G(z)) or a real
sample x. The architecture of the proposed model medGAN is depicted in Figure 1. medGAN is trained in a
similar fashion as the original GAN as follows,
θd ← θd + α∇θd
m
1 X
log D(xi ) + log(1 − D(xzi ))
m i=1
θg,dec ← θg,dec + α∇θg,dec
m
1 X
log D(xzi )
m i=1
where xzi = Dec(G(zi ))
It should be note that we can round the values of Dec(G(z)) to their nearest integers to ensure that the
discriminator D is trained on discrete values instead of continuous values. We experimented both with
and without rounding and empirically found that training D in the latter scenario led to better predictive
performance in section 4.2. Therefore, we assume, for the remainder of this paper, that D is trained without
explicit rounding.
We fine-tune the pre-trained parameters of the decoder θdec while optimizing for G. Therefore, the
generator G can be viewed as a neural network with an extra hidden layer pre-trained to map continuous
samples to discrete samples. We used ReLU for all of G’s activation functions, except for the output layer,
where we used the tanh function2 . For D, we used ReLU for all activation functions except for the output
layer, where we used the sigmoid function for binary classification.
3.4 Minibatch Averaging
Since the objective of the generator G is to produce samples that can fool the discriminator D, G could learn
to map different random priors z to the same synthetic output, rather than producing diverse synthetic outputs.
This problem is denoted as mode collapse, which arises most likely due to the GAN’s optimization strategy
often solving the max-min problem instead of the min-max problem (Goodfellow, 2016). Some methods
have been proposed to cope with mode collapse (e.g., minibatch discrimination and unrolled GANs), but they
require ad hoc fine-tuning of the hyperparameters and scalability is often neglected (Salimans et al., 2016;
Metz et al., 2016).
By contrast, medGAN offers a simple and efficient method to cope with mode collapse when generating
discrete outputs. Our method, minibatch averaging, is motivated by the philosophy behind minibatch
discrimination. It allows the discriminator D to view the minibatch of real samples x1 , x2 , . . . and the
minibatch of the fake samples G(z1 ), G(z2 ), . . ., respectively, while classifying a real sample and a fake
sample. Given a sample to discriminate, minibatch discrimination calculates the distance between the given
sample and every sample in the minibatch in the latent space. Minibatch averaging, by contrast, provides the
average of the minibatch samples to D, modifying the objective as follows:
θd ← θd + α∇θd
m
1 X
log D(xi , x̄) + log(1 − D(xzi , x̄z ))
m i=1
θg,dec ← θg,dec + α∇θg,dec
where x̄ =
m
1 X
xi ,
m i=1
m
1 X
log D(xzi , x̄z )
m i=1
xzi = Dec(G(zi )),
x̄z =
m
1 X
xz
m i=1 i
where m denotes the size of the minibatch. Specifically, the average of the minibatch is concatenated on the
sample and provided to the discriminator D.
1. We considered a denoising autoencoder (dAE) (Vincent et al., 2008) as well, but there was no discernible improvement in performance.
2. We also applied tanh activation for the encoder Enc for consistency.
5
Binary variables: When processing binary variables x ∈ {0, 1}|C| , the average of minibatch samples x̄
and x̄z are equivalent to the maximum likelihood estimate of the Bernoulli success probability p̂k of each
dimension k. This information makes it easier for D to ascertain whether a given sample is real or fake, if p̂k ’s
of fake samples are considerably different from those of real samples. This is especially likely when mode
collapse occurs because the p̂k ’s for most dimensions of the fake samples become dichotomized (either 0 or 1),
whereas the p̂k ’s of real samples generally take on a value between 0 and 1. Therefore, if G wants to fool D, it
will have to generate more diverse examples within the minibatch Dec(G(z1 , z2 , . . .)).
Count variables: Count variables are a more accurate description of clinical events. They can indicate the
number of times a certain diagnosis was made or a certain medication was prescribed over multiple hospital
|C|
visits. For count variables x ∈ Z+ , the average of minibatch samples x̄ and x̄z can be viewed as the estimate
of the binomial distribution mean npbk of each dimension k, where n is the number of hospital visits. Hence
minibatch averaging for the count variables also provides helpful statistics to the discriminator D, guiding the
generator G to generate more diverse and realistic samples. As our experiments show, minibatch averaging
works surprisingly well and does not require additional parameters like minibatch discrimination. As a
consequence, it has minimal impact to the training time. It is further worth mentioning that, for both binary
and count variables, a minibatch that is larger than usual is recommended to properly capture the statistics of
the real data. We use 1,000 records for a minibatch in this investigation.
3.5 Enhanced Generator Training
Similar to image processing GANs, we observed that balancing the power of D and G in the multi-label
discrete variable setting was quite challenging (Goodfellow, 2016). Empirically, we observed that training
medGAN with minibatch averaging demonstrated D consistently overpowering G after several iterations.
While G still managed to learn under such situation, the performance seemed suboptimal, and updating θg
and θdec more often than θd in each iteration only degraded performance. Considering the importance of an
optimal D (Goodfellow, 2016), we chose not to limit the discriminative power of D, but rather improve the
learning efficiency of G by applying batch normalization (Ioffe and Szegedy, 2015) and shortcut connection
(He et al., 2016). G’s k th layer is now formulated as follows:
xk = ReLU(BNk (Wk xk−1 )) + xk−1
where ReLU is the rectified linear unit, BNk is the batch normalization at the k-th layer, Wk is the weight
matrix of the k-th layer, and xk−1 is the input from the previous layer. The right-hand side of Figure 1
depicts the first two layers of G. Note that we do not incorporate the bias variable into each layer because
batch normalization negates the necessity of the bias term. Additionally, batch normalization and shortcut
connections could be applied to the discriminator D, but the experiments showed that D was consistently
overpowering G without such techniques, and we empirically found that a simple feedforward network was
sufficient for D. We describe the overall optimization algorithm in the Appendix A.
3.6 Privacy Consideration
When EHRs are de-identfied via methods such generalization or randomization, there often remains a 1-to-1
mapping to the underlying records from where they were derived. However, in our case, the mapping between
the generated data from medGAN and the training data of specific patients is not explicit. Intuitively, this
seems to imply that the privacy of the patients can be better preserved with medGAN; however, it also begs the
question of how to evaluate the privacy in the system. We perform a formal assessment of medGAN’s privacy
risks based on two definitions of privacy.
Presence disclosure occurs when an attacker can determine that medGAN was trained with a dataset including
the record from patient x. (Nergiz and Clifton, 2010) Presence disclosure for medGAN happens when a
powerful attacker, one who already possesses the complete records of a set of patients P , can determine
whether anyone from P are in the training set by observing the generated patient records. More recently, for
machine learned models, this has been referred to as an membership inference attack (Shokri et al., 2017).
the knowledge gained by the attacker may be limited, if the dataset is well balanced in its clinical concepts.
Attribute disclosure occurs when attackers can derive additional attributes such as diagnoses and medications
about patient x based on a subset of attributes they already know about x. (Matwin et al., 2015) We believe
6
Table 1: Basic statistics of datasets A, B and C
Dataset
# of patients
# of unique codes
Avg. # of codes per patient
Max # of codes for a patient
Min # of codes for a patient
(A) Sutter PAMF
(B) MIMIC-III
(C) Sutter Heart Failure
258,559
615
38.37
198
1
46,520
1071
11.27
90
1
30,738
569
53.02
871
2
that attribute disclosure for medGAN could be a more prominent issue because the attacker only needs to know
a subset of attributes of a patient. Moreover, the goal of the attacker is to gain knowledge of the unknown
attributes by observing similar patients generated by medGAN.
Considering the difficulty of deriving analytic proof of privacy for GANs and simulated data, we report
the empirical analysis of both risks to understand the extent to which privacy can be achieved, as commonly
practiced in the statistical disclosure control community. (Domingo-Ferrer and Torra, 2003)
4. Experiments
We evaluated medGAN with three distinct EHR datasets. First, we describe the datasets and baseline models.
Next, we report the quantitative evaluation results using both binary and count variables. We then perform a
qualitative analysis through medical expert review. Finally, we address the privacy aspect of medGAN. The
source code of medGAN is publicly available at https://github.com/mp2893/medgan.
4.1 Experimental Setup
Source data: The datasets in this study were from A) Sutter Palo Alto Medical Foundation (PAMF), which
consists of 10-years of longitudinal medical records of 258K patients, B) the MIMIC-III dataset (Johnson
et al., 2016; Goldberger et al., 2000), which is a publicly available dataset consisting of the medical records of
46K intensive care unit (ICU) patients over 11 years old and C) a heart failure study dataset from Sutter, which
consists of 18-months observation period of 30K patients. From dataset A and C, we extracted diagnoses,
medications and procedure codes, which were then respectively grouped by Clinical Classifications Software
(CCS) for ICD-93 , Generic Product Identifier Drug Group4 and for CPT5 . From dataset B, we extracted ICD9
codes only and grouped them by generalizing up to their first 3 digits. Finally, we aggregate a patient’s
|C|
longitudinal record into a single fixed-size vector x ∈ Z+ , where |C| equals 615, 1071 and 569 for dataset A,
B and C respectively. Note that datasets A and B are binarized for experiments regarding binary variables
while dataset C is used for experiments regarding count variables. A summary of the datasets are in Table 1.
Models for comparison: To assess the effectiveness of our methods, we tested multiple versions of medGAN:
• GAN: We use the same architecture as medGAN with the standard training strategy, but do not pre-train
the autoencoder.
• GANP : We pre-train the autoencoder (in addition to the GAN).
• GANP D : We pre-train the autoencoder and use minibatch discrimination (Salimans et al., 2016).
• GANP A : We pre-train the autoencoder and use minibatch averaging.
• medGAN: We pre-train the autoencoder and use minibatch averaging. We also use batch normalization
and a shortcut connection for the generator G.
We also compare the performance of medGAN with several popular generative methods as below.
• Random Noise (RN): Given a real patient record x, we invert the binary value of each code (i.e.,
dimension) with probability 0.1. This is not strictly a generative method, but rather it is a simple
implementation of a privacy protection method based on randomization.
• Independent Sampling (IS): For the binary variable case, we calculate the Bernoulli success probability
of each code in the real dataset, based on which we sample binary values to generate the synthetic
dataset. For the count variable case, we use the kernel density estimator (KDE) for each code then
sample from that distribution.
3. https://www.hcup-us.ahrq.gov/toolssoftware/ccs/ccs.jsp
4. http://www.wolterskluwercdi.com/drug-data/medi-span-electronic-drug-file/
5. https://www.hcup-us.ahrq.gov/toolssoftware/ccs svcsproc/ ccssvcproc.jsp
7
GAN
GANP
GANPD
GANPA
medGAN
(a) Dimension-wise probability performance of various versions of medGAN.
RN
IS
DBM
VAE
medGAN
(b) Dimension-wise probability performance of baseline models and medGAN.
Figure 2: Scatterplots of dimension-wise probability results. Each dot represents one of 615 codes. The x-axis
represents the Bernoulli success probability for the real dataset A, and y-axis the probability for the synthetic
counterpart generated by each model. The diagonal line indicates the ideal performance where the real and
synthetic data show identical quality.
• Stacked RBM (DBM): We train a stacked Restricted Boltzmann Machines (Hinton and Salakhutdinov,
2006), then, using Gibbs sampling, we can generate synthetic binary samples. There are studies that
extend RBMs beyond binary variables (Hinton and Salakhutdinov, 2009; Gehler et al., 2006; Tran et al.,
2011). In this work, however, as our goal is to study medGAN’s performance in various aspects, we use
the original RBM only.
• Variational Autoencoder (VAE): We train a variational autoencoder (Kingma and Welling, 2013)
where the encoder and the decoder are constructed with feed-forward neural networks.
Implementation details: We implemented medGAN with TensorFlow 0.12 (Team, 2015). For training models,
we used Adam (Kingma and Ba, 2014) with the learning rate set to 0.001, and a mini-batch of 1,000 patients
on a machine equipped with Intel Xeon E5-2630, 256GB RAM, four Nvidia Pascal Titan X’s and CUDA 8.0.
The hyperparameter details are provided in Appendix B.
4.2 Quantitative Evaluation for Binary Variables
We evaluate the model performance for binary variables in this section, and provide the evaluation results of
count variables in Appendix D. For all evaluations, we divide the dataset into a training set R ∈ {0, 1}N ×|C|
and a test set T ∈ {0, 1}n×|C| by 4:1 ratio. We use R to train the models, then generate synthetic samples
S ∈ {0, 1}N ×|C| that are assessed in various tasks. For medGAN and VAE, we round the values of the
generated dataset to the nearest integer values.
• Dimension-wise probability: This is a basic sanity check to confirm the model has learned each
dimension’s distribution correctly. We use the training set R to train the models, then generate the same
number of synthetic samples S. Using R and S, we compare the Bernoulli success probability pk of each
dimension k.
• Dimension-wise prediction: This task indirectly measures how well the model captures the interdimensional relationships of the real samples. After training the models with R to generate S, we choose
one dimension k to be the label yRk ∈ {0, 1}N and ySk ∈ {0, 1}N . The remaining R\k ∈ {0, 1}N ×|C|−1
and S\k ∈ {0, 1}N ×|C|−1 are used as features to train two logistic regression classifiers LRRk and LRSk to
predict yRk and ySk , respectively. Then, we use the model LRRk and LRSk to predict label yTk ∈ {0, 1}n
of the test set T . We can assume that the closer the performance of LRSk to that of LRRk , the better
the quality of the synthetic dataset S. We use F1-score to measure the prediction performance, with the
threshold set to 0.5.
To mitigate the repetition of results, we present our evaluation of dataset A in this section and direct the reader
to Appendix C for the results from dataset B.
8
GAN
GANP
GANPD
GANPA
medGAN
(a) Dimension-wise prediction performance of various versions of medGAN.
RN
IS
DBM
VAE
medGAN
(b) Dimension-wise prediction performance of baseline models and medGAN.
Figure 3: Scatterplots of dimension-wise prediction results. Each dot represents one of 615 codes. The x-axis
represents the F1-score of the logistic regression classifier trained on the real dataset A. The y-axis represents
the F1-score of the classifier trained on the synthetic counterpart generated by each model. The diagonal line
indicates the ideal performance where the real and synthetic data show identical quality.
4.2.1 D IMENSIONS - WISE PROBABILITY
There are several notable findings that are worth highlighting. The dimension-wise probability performance
increased as we used more advanced versions of medGAN, where the full medGAN shows the best performance
as depicted by figure 2a. Note that minibatch averaging significantly increases the performance. Since
minibatch averaging provides Bernoulli success probability information of real data to the model during
training, it is natural that the generator learns to output synthetic data that follow a similar distribution.
Minibatch discrimination does not seem to improve the results. This is most likely due to the discrete nature
of the datasets. Improving the learning efficiency of the generator G with batch normalization and shortcut
connection clearly helped improve the results.
Figure 2b compares the dimension-wise probability performance of baseline models with medGAN.
Independent sampling (IS) naturally shows great performance as expected. DBM, given its stochastic binary
nature, shows comparable performance as medGAN. VAE, although slightly inferior to DBM and medGAN,
seems to capture the dimension-wise distribution relatively well, showing specific weakness at processing
codes with low probability. Overall, we can see that medGAN clearly captures the independent distribution of
each code.
4.2.2 D IMENSIONS - WISE PREDICTION
Figure 3a shows the dimension-wise prediction performance of various versions of medGAN. The full medGAN
again shows the best performance as it did in the dimension-wise probability task. Although the advanced
versions of medGAN do not seem to dramatically increase the performance as they did for the previous task,
this is due to the complex nature of inter-dimensional relationship compared to the independent dimension-wise
probability. Figure 3b shows the dimension-wise prediction performance of baseline models compared to
medGAN. As expected, IS is incapable of capturing the inter-dimensional relationship, given its naive sampling
method. VAE shows similar behavior as it did in the previous task, showing weakness at predicting codes with
low occurrence probability. Again, DBM shows comparable, if not slightly better performance to medGAN,
which seems to come from its binary nature.
4.3 Qualitative Evaluation for Count Variables
We conducted a qualitative evaluation of medGAN with the help from a medical doctor. A discussion with the
doctor taught us that count data are easier to assess its realistic-ness than binary data. Therefore we use dataset
C to train medGAN and generate synthetic count samples. In this experiment, we randomly pick 50 records
from real data and 50 records from synthetic data, randomly shuffle the order, present them to a medical
doctor (specialized in internal medicine) who is asked to score how realistic each record is using scale 1 to 10
(10 being most realistic). Here the human doctor is served as the role of discriminator to provide the quality
assessment of the synthetic data generated by medGAN.
9
The results of this assessment is shown in Figure 4. The findings suggest
that medGAN’s synthetic data are generally indistinguishable to a human doctor
8
except for several outliers. In those cases, the fake records identified by the doctor
either lacked appropriate medication codes, or had both male-related codes (e.g.
6
prostate cancer) and female-related codes (e.g. menopausal disorders) in the
4
same record. The former issue also existed in some of the real records due to
2
missing data, but the latter issue demonstrates a current limitation in medGAN
0
which could potentially be alleviated by domain specific heuristics. In addition
Figure 4: Boxplot of the im- to medGAN’s impressive performance in statistical aspects, this medical review
pression scores from a med- lends credibility to the qualitative aspect of medGAN.
ical expert.
4.4 Privacy Risk Evaluation
Real
medGAN
Sensitivity
1.0
(a)
.75
.50
.25
1.0
Precision
10
.50
.25
0.0 0 1 2 3 4 5 6 7 8
2 2 2 2 2 2 2 2 2
0.0 0 1 2 3 4 5 6 7 8
2 2 2 2 2 2 2 2 2
1.0
1.0
Number of known attributes
(c)
.75
.50
.25
0.0
3k
9k
15k 21k 27k 33k 37k
Size of the synthetic dataset
Precision
Number of known attributes
Sensitivity
(b)
.75
(d)
.75
.50
.25
0.0
3k
9k
15k 21k 27k 33k 37k
Size of the synthetic dataset
Figure 5: a,b: Sensitivity and precision when varying the number of known attributes. The total number
of attributes (i.e. codes) of dataset B is 1,071. c,d: Sensitivity and precision when varying the size of the
synthetic dataset. The maximum size of the synthetic dataset S ∈ {0, 1}N ×|C| is matched to the size of the
training set R ∈ {0, 1}N ×|C| .
We evaluate both presence and attribute disclosure using dataset B with binary variables. Due to the space
constraint, we present the results of the attribute disclosure in the main paper and leave out the results of
presence disclosure in Appendix F.
Experiment setup: We randomly sample 1% of the training set R as the compromised records, which is
approximately 370 records. For each record r, we randomly choose s attributes as those which are known to
the attacker. Next, the attacker performs k-nearest neighbor classifications to estimate the values of unknown
attributes based on the synthetic records. More specifically, based on the known attributes, k-nearest neighbors
in the synthetic dataset S are retrieved for each compromised record. Then, |C| − s unknown attributes are
estimated based on the majority vote of the k nearest neighbors. Finally, for each unknown attribute, we
calculate classification metrics in the form of precision and sensitivity. We repeat this process for all records of
the 1% samples and obtain the mean precision and mean sensitivity. We vary the number of known attributes s
and the number of neighbors k to study the attribute disclosure risk of medGAN. Note that the s attributes are
randomly sampled across patients, so the attacker may know different s attributes for different patients.
Impact of attacker’s knowledge: Figures 5a and 5b depict the sensitivity (i.e., recall) and the precision of
the attribute disclosure test when varying the number of attributes known to the attacker. In this case, x%
sensitivity means the attacker, using the known attributes of the compromised record and the synthetic data, can
correctly estimate x% of the positive unknown attributes (i.e., attribute values are 1). Likewise, x% precision
means the positive unknown attributes estimated by the attacker are on average x% accurate. Both figures
show that an attacker who knows approximately 1% of the target patient’s attributes (8 to 16 attributes) will
estimate the target’s unknown attributes with less than 10% sensitivity and 20% precision.
Impact of synthetic data size: Next, we fixed the number of known attributes to 16 and varied the number
of records in the synthetic dataset S. Figures 5c and 5d show that the size of the synthetic dataset has little
10
influence on the effectiveness of the attack. In general, 1 nearest neighbor seems to be the most effective attack,
although the sensitivity is still below 25% at best.
Overall, our privacy experiments indicate that medGAN does not simply remember the training samples
and reproduce them. Rather, medGAN generates diverse synthetic samples that reveal little information to
potential attackers unless they already possess significant amount of knowledge about the target patient.
5. Conclusion
In this work, we proposed medGAN, which uses generative adversarial framework to learn the distribution
of real-world multi-label discrete electronic health records (EHR). Through rigorous evaluation using real
datasets, medGAN showed impressive results for both binary variables and count variables. Considering
the difficult accessibility of EHRs, we expect medGAN to make a contribution for healthcare research. We
also provided empirical evaluation of privacy, which demonstrates very limited risks of medGAN in attribute
disclosure. For future directions, we plan to explore the sequential version of medGAN, and also try to include
other modalities such as lab measures, patient demographics, and free-text medical notes.
Acknowledgments
This work was supported by the National Science Foundation, award IIS-#1418511 and CCF-#1533768, Children’s
Healthcare of Atlanta, Google Faculty Award, UCB and Samsung Scholarship. Dr. Malin was supported by the National
Science Foundation, award IIS-#1418504.
11
References
R. V. Atreya, J. C. Smith, A. B. McCoy, B. Malin, and R. A. Miller. Reducing patient re-identification risk for
laboratory results within research datasets. Journal of the American Medical Informatics Association, 20(1):
95–101, 2013.
Anna Buczak, Steven Babin, and Linda Moniz. Data-driven approach for creating synthetic electronic medical
records. BMC Medical Informatics and Decision Making, 10(1):59, 2010.
Emily Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid
of adversarial networks. In NIPS, pages 1486–1494, 2015.
Josep Domingo-Ferrer and Vicenç Torra. Disclosure risk assessment in statistical microdata protection
via advanced record linkage. Statistics and Computing, 13(4):343–354, 2003. ISSN 1573-1375. doi:
10.1023/A:1025666923033. URL http://dx.doi.org/10.1023/A:1025666923033.
J Dreschsler. Synthetic datasets for statistical disclosure control. Springer Press, 2011.
K. El Emam, D. Buckeridge, R. Tamblyn, A. Neisa, E. Jonker, and A. Verma. The re-identification risk of
canadians from longitudinal demographics. BMC Medical Informatics and Decision Making, 11:46, 2011a.
K. El Emam, E. Jonker, L. Arbuckle, and B. Malin. A systematic review of re-identification attacks on health
data. PLoS ONE, 6(12):e28071, 2011b.
K. El Emam, S. Rodgers, and B. Malin. Anonymising and sharing individual patient data. British Medical
Journal, 350:h1139, 2015.
Y. Erlich and A. Narayanan. Routes for breaching and protecting genetic privacy. Nature Reviews Genetics,
15(6):409–421, 2014.
Office for Civil Rights. Guidance Regarding Methods for De-identification of Protected Health Information
in Accordance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule. U.S.
Department of Health and Human Services, 2013.
Peter V Gehler, Alex D Holub, and Max Welling. The rate adapting poisson model for information retrieval
and object recognition. In Proceedings of the 23rd international conference on Machine learning, pages
337–344. ACM, 2006.
John Glover. Modeling documents with generative adversarial networks. arXiv:1612.09122, 2016.
Ary Goldberger et al. Physiobank, physiotoolkit, and physionet components of a new research resource for
complex physiologic signals. Circulation, 2000.
Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv:1701.00160, 2016.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, pages 2672–2680, 2014.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
Lawrence Gostin, Laura Levit, Sharyl Nass, et al. Beyond the HIPAA Privacy Rule: Enhancing Privacy,
Improving Health Through Research. National Academies Press, 2009.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In
CVPR, pages 770–778, 2016.
Geoffrey Hinton and Ruslan Salakhutdinov. Reducing the dimensionality of data with neural networks.
Science, 313(5786):504–507, 2006.
12
Geoffrey E Hinton and Ruslan R Salakhutdinov. Replicated softmax: an undirected topic model. In Advances
in neural information processing systems, pages 1607–1614, 2009.
James Hodge Jr, Lawrence O Gostin, and Peter Jacobson. Legal issues concerning electronic health information:
privacy, quality, and liability. Jama, 282(15):1466–1471, 1999.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv:1502.03167, 2015.
Eric Jang, Shixiang Gu, and Ben Poole.
arXiv:1611.01144, 2016.
Categorical reparameterization with gumbel-softmax.
Alistair Johnson et al. Mimic-iii, a freely accessible critical care database. Scientific Data, 3, 2016.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014.
Diederik Kingma and Max Welling. Auto-encoding variational bayes. arXiv:1312.6114, 2013.
Matt J Kusner and José Miguel Hernández-Lobato. Gans for sequences of discrete elements with the gumbelsoftmax distribution. arXiv:1611.04051, 2016.
Joseph S Lombardo and Linda J Moniz. Ta method for generation and distribution. Johns Hopkins APL
Technical Digest, 27(4):356, 2008.
G. Loukides, J. C. Denny, and B. Malin. The disclosure of diagnosis codes can breach research participants’
privacy. J Am Med Inform Assoc, 17(3):322–327, 2010.
Chris Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of
discrete random variables. arXiv:1611.00712, 2016.
B. Malin and L. Sweeney. How (not) to protect genomic data privacy in a distributed network: using trail
re-identification to evaluate and design anonymity protection systems. Journal of Biomedical Informatics,
37(3):179–192, 2004.
Stan Matwin, Jordi Nin, Morvarid Sehatkar, and Tomasz Szapiro. A Review of Attribute Disclosure Control,
pages 41–61. Springer International Publishing, Cham, 2015. ISBN 978-3-319-09885-2. doi: 10.1007/
978-3-319-09885-2 4. URL http://dx.doi.org/10.1007/978-3-319-09885-2_4.
Scott McLachlan, Kudakwashe Dube, and Thomas Gallagher. Using the caremap with health incidents
statistics for generating the realistic synthetic electronic healthcare record. In Healthcare Informatics (ICHI),
2016 IEEE International Conference on, pages 439–448. IEEE, 2016.
Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks.
arXiv:1611.02163, 2016.
Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv:1411.1784, 2014.
Mehmet Nergiz and Chris Clifton. δ-presence with complete world knowledge. IEEE Transactions on
Knowledge Engineering, 22:868–883, 2010.
Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier
gans. arXiv:1610.09585, 2016.
Yubin Park, Joydeep Ghosh, and Mallikarjun Shankar. Perturbed gibbs samplers for generating large-scale
privacy-safe synthetic health data. In Healthcare Informatics (ICHI), 2013 IEEE International Conference
on, pages 493–498. IEEE, 2013.
13
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional
generative adversarial networks. arXiv:1511.06434, 2015.
Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative
adversarial text to image synthesis. In ICML, 2016.
J. Reiter. Satisfying disclosure restrictions with synthetic datasets. Journal of Official Statistics, 18(4):531–543,
2002.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved
techniques for training gans. In NIPS, pages 2226–2234, 2016.
Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against
machine learning models. In IEEE Security & Privacy Conference, page in press, 2017.
L. Sweeney. Weaving technology and policy together to maintain confidentiality. Journal of Law, Medicine,
and Ethics, 25(2-3):98–110, 1997.
TensorFlow Team. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL
http://tensorflow.org/. Software available from tensorflow.org.
Truyen Tran, Dinh Phung, and Svetha Venkatesh. Mixed-variate restricted boltzmann machines. In Asian
Conference on Machine Learning, pages 213–229, 2011.
Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image
generation with pixelcnn decoders. In NIPS, pages 4790–4798, 2016a.
Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu.
arXiv:1601.06759, 2016b.
Pixel recurrent neural networks.
Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing
robust features with denoising autoencoders. In ICML, pages 1096–1103, 2008.
Ronald Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning.
Machine learning, 8(3-4):229–256, 1992.
Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets with policy
gradient. arXiv:1609.05473, 2016.
Yizhe Zhang, Zhe Gan, and Lawrence Carin. Generating text via adversarial training. NIPS Workshop on
Adversarial Training, 2016.
14
Algorithm 1 medGAN Optimization
θd , θg , θenc , θdec ← Initialize with random values.
repeat // Pre-train the autoencoder
Randomly sample x1 , x2 , . . . , xm from X
Update θenc , θdec by minimizing Eq.(2) (or Eq.(3))
until convergence or fixed iterations
repeat
for k steps do // Update the discriminator.
Randomly sample z1 , z2 , . . . , zm from pz
Randomly sample x1 , x2 , . . . , xm from X
xzi ← Dec(G(z
Pm i ))
1
x̄z ← m
Pmi=1 xzi
1
x̄ ← m
i=1 xi
Ascend θd by the gradient:
1
∇θd m
Pm
i=1
log D(xi , x̄) + log(1 − D(xzi , x̄z ))
end for
// Update the generator and the decoder.
Randomly sample z1 , z2 , . . . , zm from pz
xzi ← Dec(G(z
Pm i ))
1
x̄z ← m
i=1 xzi
Ascend θg , θdec by the gradient:
1
∇θg,dec m
Pm
i=1
log D(xzi , x̄z )
until convergence or fixed iterations
Appendices
Appendix A. medGAN training algorithm
Algorithm 1 describes the overall optimization process of medGAN. Note that θd is updated k times per
iteration, while θg and θdec are updated once per iteration to ensure optimality of D. However, typically, a
larger k has not shown a clear improvement (Goodfellow, 2016). And we set k = 2 in our experiments.
Appendix B. Hyperparameter details
We describe the architecture and the hyper-parameter values used for each model. We tested all models by
varying the number of hidden layers (while matching the number of parameters used for generating synthetic
data), the size of the minibatch, the learning rate, the number of training epochs, and we report the best
performing configuration for each model.
• medGAN: Both the encoder Enc and the decoder Dec are single layer feedforward networks, where
the original input x is compressed to a 128 dimensional vector. The generator G is implemented as a
feedforward network with two hidden layers, each having 128 dimensions. For the batch normalization in
the generator G, we use both the scale parameter γ and the shift parameter β, and set the moving average
decay to 0.99. The discriminator D is also a feedforward network with two hidden layers where the first
layer has 256 dimensions and the second layer has 128 dimensions. medGAN is trained for 1,000 epochs
with the minibatch of 1,000 records.
• DBM: In order to match the number of parameters used for data generation in medGAN (G + Dec), we
used four layers of Restricted Boltzmann Machines where the first layer is the input layer. All hidden
layers used 128 dimensions. We performed layer-wise greedy persistent contrastive divergence (20-step
15
Gibbs sampling) to train DBM. We used 0.01 for learning rate and 100 samples per minibatch. All layers
were separately trained for 100 epochs. Synthetic samples were generated by performing Gibbs sampling
at the two two layers then propagating the values down to the input layer. We ran Gibbs sampling for 1000
iterations per sample. Using three stacks showed small performance degradation.
• VAE: In order to match the number of parameters used for data generation in medGAN (G + Dec), both the
encoder and the decoder were implemented with feedforward networks, each having 3 hidden layers. The
encoder accepts the input x and compresses it to a 128 dimensional vector and the decoder reconstructs it
to the original dimension space. VAE was trained with Adam for 1,000 iterations with the minibatch of
1,000 records. Using two hidden layers for the encoder and the decoder showed similar performance.
Appendix C. Quantitative evaluation results for binary dataset B
GAN
GANP
GANPD
GANPA
medGAN
(a) Dimension-wise probability performance of various versions of medGAN.
RN
IS
DBM
VAE
medGAN
(b) Dimension-wise probability performance of baseline models and medGAN.
Figure 6: Scatterplots of dimension-wise probability results. Each dot represents one of 1,071 codes. The
x-axis represents the Bernoulli success probability for the real dataset B, and y-axis the probability for the
synthetic counterpart generated by each model. The diagonal line indicates the ideal performance where the
real and synthetic data show identical quality.
GAN
GANP
GANPD
GANPA
medGAN
(a) Dimension-wise prediction performance of various versions of medGAN.
RN
IS
DBM
VAE
medGAN
(b) Dimension-wise prediction performance of baseline models and medGAN.
Figure 7: Scatterplots of dimension-wise prediction results. Each dot represents one of 1,071 codes. The
x-axis represents the F1-score of the logistic regression classifier trained on the real dataset B. The y-axis
represents the F1-score of the classifier trained on the synthetic counterpart generated by each model. The
diagonal line indicates the ideal performance where the real and synthetic data show identical quality.
16
C.1 Dimension-wise probability
Figure 6a shows the consistent superiority of the full version of medGAN compared other versions. The
effect of minibatch averaging is even more dramatic for dataset B. Figure 6b shows that VAE has some
difficulty capturing the dimension-wise distribution of dataset B. Again, DBM shows comparable performance
to medGAN, slightly outperforming medGAN for low-probability codes, but slightly underperforming for
high-probability codes. Overall, dimension-wise probability performance is somewhat weaker for dataset B
than for dataset A, most likely due to smaller data volume and sparser code distribution.
C.2 Dimension-wise prediction
Figure 7a shows the dimension-wise predictive performance for different versions of medGAN where the full
version outperforms others. Figure 7b shows similar pattern as Figure 3b. Independent sampling completely
fails to make any meaningful prediction. VAE demonstrates weakness at predicting low-probability codes.
DBM seems to slightly outperform medGAN, especially for highly predictable codes. Again, due to the nature
of the dataset, all models show weaker predictive performance for dataset B than they did for dataset A.
Appendix D. Quantitative results for count variables
In order to evaluate for count variables, we use dataset C, consisting of 30,738 patients whose records were
taken for exactly 18 months. The same subset was used to perform qualitative evaluation in section 4.3. The
details of constructing dataset C for heart failure studies are described in Appendix E. Note that each patient’s
number of hospital visits within the 18 months period can vary, which is a perfect test case for count variables.
N ×|C|
Again, we aggregate the dataset into a fixed-size vector and divide it into the training set R ∈ Z+
and the
n×|C|
test set T ∈ Z+
in 4:1 ratio. Since we have confirmed the superior performance of full medGAN compared
to other versions of GANs in binary variables evaluation, we focus on the comparison with baseline models
in this section. Note that, to generate count variables, we replaced all activation functions in both VAE and
medGAN (except the discriminator’s output) to ReLU. We also use kernel density estimator with Gaussian
kernel (bandwidth=0.75) to perform the independent sampling (IS) baseline. We no longer test random noise
(RN) method in this section as it is difficult to determine how much noise should be injected to count variables
to keep them sufficiently realistic but different enough from the training set.
For count variables, we conduct similar quantitative evaluations as binary variables with slight modifications. We first calculate dimension-wise average count instead of dimension-wise probability. For
dimension-wise prediction, we use the binary labels yRk ∈ {0, 1}N and ySk ∈ {0, 1}N as before, but we train
N ×|C|−1
N ×|C|−1
the logistic regression classifier with count samples R\k ∈ Z+
and S\k ∈ Z+
. The classifiers
use count features as oppose to binary features while the evaluation metric is still F1-score.
D.0.1 D IMENSIONS - WISE AVERAGE COUNT
5k
2.5k
Essention
hypertension
Connective
tissue disease
Disorder of
lipid metabolism
Cardiac
dysrhythmias
Skin
disorders
0
5k
2.5k
0
0
25
50
0
25
50 0
25
50
0
25
50
0
25
50
Figure 8: Histogram of counts of five most frequent codes from dataset C. The top row was plotted using the
training dataset, the bottom row using medGAN’s synthetic dataset.
Figure 9 shows the performance of baseline models and medGAN. The discontinuous behavior of VAE is
due to its extremely low-variance synthetic samples. We found that, on average, VAE’s synthetic samples had
17
IS-KDE
VAE
medGAN
Figure 9: Scatterplot of dimension-wise average count of the training dataset (x-axis) versus the synthetic
counterpart (y-axis).
IS-KDE
VAE
medGAN
Figure 10: Scatterplot of dimension-wise prediction F1-score of logistic regression trained on the training
dataset (x-axis) versus the classifier trained on the synthetic counterpart (y-axis).
nine orders of magnitude smaller standard deviation than medGAN’s synthetic samples. medGAN, on the other
hand, shows good performance with just a simple substitution of the activation functions.
Figure 8 shows the count histograms of five most frequent codes from the count dataset, where the top
row was plotted with the training dataset and the bottom row with medGAN’s synthetic dataset. We can see
that medGAN’s synthetic counterpart has very similar distribution as the real data. This tells us that medGAN
is not just trying to match the average count of codes (i.e. binomial distribution mean), but learns the actual
distribution of the data.
D.0.2 D IMENSIONS - WISE PREDICTION
Figure 10 shows the performance of baseline models and medGAN. We can clearly see that medGAN shows
superior performance. The experiments on count variables is especially interesting, as medGAN seems to make
a smooth transition from binary variables to count variables, with just a replacement of the activation function.
We also speculate that the medGAN’s dimension-wise prediction performance will increase with more training
data, as the count dataset used in this section consists of only 30,738 samples.
Appendix E. Dataset construction for heart failure studies
Case patients were 40 to 85 years of age at the time of HF diagnosis. HF diagnosis (HFDx) is defined as: 1)
Qualifying ICD-9 codes for HF appeared in the encounter records or medication orders. Qualifying ICD-9
codes are displayed in Table 2. 2) a minimum of three clinical encounters with qualifying ICD-9 codes had to
occur within 12 months of each other, where the date of diagnosis was assigned to the earliest of the three
dates. If the time span between the first and second appearances of the HF diagnostic code was greater than 12
months, the date of the second encounter was used as the first qualifying encounter. The date at which HF
diagnosis was given to the case is denoted as HFDx. Up to ten eligible controls (in terms of sex, age, location)
were selected for each case, yielding an overall ratio of 9 controls per case. Each control was also assigned
an index date, which is the HFDx of the matched case. Controls are selected such that they did not meet the
operational criteria for HF diagnosis prior to the HFDx plus 182 days of their corresponding case. Control
subjects were required to have their first office encounter within one year of the matching HF case patients
first office visit, and have at least one office encounter 30 days before or any time after the cases HF diagnosis
date to ensure similar duration of observations among cases and controls.
Appendix F. Presence disclosure
We performed a series of experiments to assess the extent to which medGAN leaks the presence of a patient.
To do so, we randomly sample r patient records from each of the training set R ∈ {0, 1}N ×|C| and the test set
18
Table 2: Qualifying ICD-9 codes for heart failure
19
T ∈ {0, 1}n×|C| . We assume the attacker has complete knowledge on those 2r records. Then for each record,
we calculate its hamming distance to each sample from the synthetic dataset S ∈ {0, 1}N ×|C| . If there is at
least one synthetic sample within a certain distance, we treat that as its claimed match. Now, since we sample
from both R and T , the match could be a true positive (i.e., attacker correctly claims their targeted record is
in the GAN training set), false positive (i.e., attacker incorrectly claims their targeted record is in the GAN
training set), true negative (i.e., attacker correctly claims their targeted record is not in the GAN training set),
or false negative (i.e., attacker incorrectly claims their targeted record is not in the GAN training set).
We varied the number of patients r and the hamming distance threshold and calculated the sensitivity and
precision.
(a)
2k
(b)
4k
6k
8k
10k
Number of patients known to attacker
(c)
5k
10k
15k
20k
25k
30k
35k
2k
6k
4k
8k
10k
Number of patients known to attacker
(d)
5k
10k
15k
20k
25k
30k
Number of synthetic patients
Number of synthetic patients
Number of synthetic patients
Number of synthetic patients
35k
Figure 11: a,b: Sensitivity and precision while varying the number of patients known to the attacker. c,d:
Sensitivity and precision while varying the number of synthetic patients.
Impact of attacker’s knowledge: Figures 11a and 11b depict the sensitivity (i.e. recall) and the precision of
the presence disclosure test when varying the number of real patient the attacker knows. In this setting, x%
sensitivity means the attacker has successfully discovered that x% of the records that he/she already knows
were used to train medGAN. Similarly, x% precision means, when an attacker claims that a certain number of
patients were used for training medGAN, only x% of them were actually used. Figure 11a shows that with low
threshold of hamming distance (e.g. hamming distance of 0) attacker can only discover 10% percent of the
known patients to attacker were used to train medGAN. Figure 11b shows that, the precision is mostly 50%
except when the number of known patients are small. This indicates that the attacker’s knowledge is basically
useless for presence disclosure attack unless the attacker is focusing on a small number of patients (less than a
hundred), in which case the precision is approximately 80%.
We conducted an additional experiment to evaluate the impact of the size of the synthetic data on presence
disclosure risk. In this experiment, we fix the number of known real patients to 100 and varied the number of
records in the synthetic dataset S. Figures 11c and 11d show that the size of the generated synthetic dataset
has almost no impact on presence disclosure.
20
| 9cs.NE
|
arXiv:1602.05281v1 [cs.SY] 17 Feb 2016
Generalized Jensen Inequalities with Application to Stability
Analysis of Systems with Distributed Delays over Infinite
Time-Horizons ⋆
Kun Liu a Emilia Fridman b Karl Henrik Johansson c Yuanqing Xia a
a
School of Automation, Beijing Institute of Technology, 100081 Beijing, China (e-mails: kunliubit, xia [email protected]).
b
c
School of Electrical Engineering, Tel Aviv University, 69978 Tel Aviv, Israel (e-mail: [email protected]).
ACCESS Linnaeus Centre and School of Electrical Engineering, KTH Royal Institute of Technology, SE-100 44 Stockholm,
Sweden (e-mail: [email protected]).
Abstract
The Jensen inequality has been recognized as a powerful tool to deal with the stability of time-delay systems. Recently, a new
inequality that encompasses the Jensen inequality was proposed for the stability analysis of systems with finite delays. In this
paper, we first present a generalized integral inequality and its double integral extension. It is shown how these inequalities
can be applied to improve the stability result for linear continuous-time systems with gamma-distributed delays. Then, for the
discrete-time counterpart we provide an extended Jensen summation inequality with infinite sequences, which leads to less
conservative stability conditions for linear discrete-time systems with poisson-distributed delays. The improvements obtained
thanks to the introduced generalized inequalities are demonstrated by examples.
Key words: new integral and summation inequalities, gamma-distributed delays, poisson-distributed delays, Lyapunov
method.
1
Introduction
Time-delay often appears in many control systems either in the state, the control input, or the measurements.
During the last two decades, the stability of time-delay
systems has received considerable attention (e.g., [3], [8],
[15], [17] and references therein). One of the most popular approaches is the use of Lyapunov-Krasovskii functionals (LKF) to derive stability conditions (e.g., [1], [5],
[9], [26]). The choice of the Lyapunov functional and the
method of bounding an integral term in the derivative of
the LKF are important ways to reduce conservativeness
of the stability results. The Jensen inequality [8], has
been widely used as an efficient bounding technique, although at a price of an unavoidable conservativeness [7],
[12]. The Jensen inequality claims that for any continuous function ω : [a, b] → Rn and n × n positive definite
⋆ This work was partially supported by the National Natural
Science Foundation of China (grant no. 61503026, 61440058),
the Knut and Alice Wallenberg Foundation, the Swedish
Research Council, and the Israel Science Foundation (grant
no. 754/10 and 1128/14).
Preprint submitted to *****
matrix R,
Rb
a
ω T (s)Rω(s)ds ≥
1
b−a
Rb
a
ω T (s)dsR
Rb
a
ω(s)ds
holds. There is a discrete counterpart, which involves
sums instead of integrals [3], [4].
Some recent efforts have been made to overcome the
conservativeness induced by the Jensen inequality when
applied to the stability analysis of time-delay systems.
The bound on the gap of the Jensen inequality was analyzed in [2] by using the Grüss inequality. Based on the
Wirtinger inequality [11], Seuret and Gouaisbaut [19]
derived an extended integral inequality, which encompasses Jensen inequality as a particular case. Recently,
the inequality they proposed was further refined in [20].
By combining the newly developed integral inequality
and an augmented Lyapunov functional, a remarkable
result was obtained for systems with constant discrete
and distributed delays. Let us recall the inequality provided in [20] (see [21] for the discrete counterpart): for
any continuous function ω : [a, b] → Rn and n × n posi-
13 March 2018
Question 1 Is it possible to derive more accurate lower
bounds to reduce the conservativeness of integral inequalities (4) and (5)? If so, how much improvements can we
obtain by applying the generalized inequalities to the stability analysis of continuous-time systems with gammadistributed delays?
tive definite matrix R, the inequality
Rb
a
ω T (s)Rω(s)ds ≥
Rb T
Rb
1
b−a a ω (s)dsR a
3
+ b−a
ΩT RΩ
ω(s)ds
(1)
holds, where
Ω=
Rb
a
ω(s)ds −
2
b−a
RbRs
a
a
ω(r)drds.
(2)
We further analyze the discrete-time case. Poissondistribution is widespread in queuing theory [6]. In
[18], the experimental data on the arrivals of pulses in
indoor environments revealed that each cluster’s timedelay is poisson-distributed (see also [10]). Therefore,
we study the stability of linear discrete-time systems
with poisson-distributed delays via appropriate Lyapunov functionals. The Lyapunov-based analysis uses
the discrete counterpart of integral inequalities (4) and
(5), i.e., Jensen inequalities with infinite sequences [13],
[25]. The following question corresponds to Question 1
in the discrete case:
To prove (1), a function f (u), u ∈ [a, b], was introduced
in [20] as follows:
f (u) = ż(u) = ω(u) −
1
b−a
Rb
a
ω(s)ds −
a+b−2u
(b−a)2 Θ,
(3)
where Θ ∈ R Rn is a constant
be defined
R b vector to(b−u)(u−a)
u
and z(u) = a ω(s)ds − u−a
ω(s)ds
−
Θ,
b−a a
(b−a)2
Rb
u ∈ [a, b]. It is noted that a (a + b − 2u)du = 0 plays
an important role in the utilization of (3). Since Θ is
a constant vector, it is obvious that in (3), a + b − 2u
could be replaced by c(a + b − 2u), c ∈ R\{0}, because
Rb
c(a + b − 2u)du = 0. By using a more general auxa
Rb
iliary function g(u) with a g(u)du = 0, an extended
integral inequality, which included the one proposed
in [20] as a particular case, was provided in [16].
Question 2 Is it possible to generalize Jensen summation inequalities with infinite sequences? If so, how much
improvements can be achieved by applying the generalized inequalities to the stability analysis of discrete-time
systems with poisson-distributed delays?
The central aim of the present paper is to answer the
above questions. First, we present generalized Jensen
integral inequality and its double integral extension,
which are over infinite intervals of integration. We
show how they can be applied to improve the stability
result for linear continuous-time systems with gammadistributed delays. Then, for the discrete counterpart
we provide extended Jensen summation inequality with
infinite sequences, which leads to less conservative stability conditions for linear discrete-time systems with
poisson-distributed delays. In both the continuous-time
and discrete-time cases, the considered infinite distributed delays are shown to have stabilizing effects.
Following [22], we derive the results via augmented
Lyapunov functionals.
Recently, the stability analysis of systems with gammadistributed delays was studied [22]. The Lyapunov-based
analysis was based on two kinds of integral inequalities
with infinite intervals of integration: given an n × n positive definite matrix R, a scalar h ≥ 0, a vector function
ω : [0, +∞) → Rn and a scalar function K : [0, +∞) →
R+ such that the integrations concerned are well defined,
the following inequalities
R +∞
K(s)ω T (s)Rω(s)ds
R +∞
R +∞
≥ K0−1 0 K(s)ω T (s)dsR 0 K(s)ω(s)ds
0
and
R +∞ R t
T
t−θ−h K(θ)ω (s)Rω(s)dsdθ
R
R
−1 +∞ t
≥ K1h
K(θ)ω T (s)dsdθR
R +∞0 R t t−θ−h
× 0
t−θ−h K(θ)ω(s)dsdθ,
R +∞
R +∞
hold, where K0 = 0 K(s)ds and K1h = 0
(4)
0
(5)
The structure of this paper is as follows. In Section 2 we
derive generalized Jensen integral inequalities. Section 3
presents stability results for linear continuous-time systems with gamma-distributed delays to illustrate the efficiency of the proposed inequalities. Sections 4 and 5
discuss the corresponding extended Jensen summation
inequality with infinite sequences and its application to
the stability analysis of linear discrete-time systems with
poisson-distributed delays, respectively. The conclusions
and the future work will be stated in Section 6.
K(s)(s+
h)ds. The inequalities (4) and (5) were used in [23] to
the stability and passivity analysis for diffusion partial
differential equations with infinite distributed delays.
To obtain more accurate lower bounds of integral inequalities (4) and (5) over infinite intervals of integration, the method developed in [20] for the integral inequality over finite intervals of integration seems not to
be applicable, since the function f of (3) is directly dependent on both the lower limit a and the upper limit b.
Therefore, an interesting question arises:
Notations: The notations used throughout the paper
are standard. The superscript ‘T ’ stands for matrix
transposition, Rn denotes the n dimensional Euclidean
space with vector norm | · |, Rn×m is the set of all n × m
2
real matrices, and the notation P ≻ 0, for P ∈ Rn×n
means that P is symmetric and positive definite. The
symmetric term in a symmetric matrix is denoted by ∗.
The symbols R, R+ , Z+ and N denote the set of real
numbers, non-negative real numbers, non-negative integers and positive integers, respectively.
2
Rewriting the last two terms as sum of squares yields
Rb T
Rb
1
ω (s)dsR a ω(s)ds
ω T (s)Rω(s)ds ≥ b−a
a
hR
i−1 R
Rb
b
b
T
+ a g 2 (s)ds
a g(s)ω (s)dsR a g(s)ω(s)ds
hR
i
b
− a g 2 (s)ds [Θ − Υ]T R[Θ − Υ],
(8)
where
Rb
a
Extended Jensen integral inequalities
The objective of this section is to provide extended
Jensen integral inequalities over infinite intervals. To
do so, we first prove the generalized Jensen integral
inequality introduced in [16] over finite intervals in a
simpler way. Then we extend the method to prove the
inequality over infinite intervals.
2.1
Υ=
hZ
a
b
g 2 (s)ds
i−1 Z
b
g(s)ω(s)ds.
a
Since (8) holds independently of the choice of Θ, we may
choose Θ = Υ, which leads to the maximum of the righthand side of (8) and thus, (6) holds. This concludes the
proof.
Extended Jensen integral inequality over finite intervals
Remark 1 In [16], the proof was more complicated as
the corresponding construction of (7) relied on an auxiliary function ḡ(u), where ḡ(u) satisfies g(u) = ḡ(u) −
Rb
1
b−a a ḡ(s)ds.
By changing a + b − 2u of (3) to a more general scalar
Rb
function g(u) with a g(u)du = 0 and g(u) not identically zero, we first present the extended Jensen inequality over finite intervals of integration.
Lemma 1 [16] If there exist an n × n matrix R ≻ 0, a
scalar function g : [a, b] → R and a vector function ω :
[a, b] → Rn such that the integrations concerned are well
Rb
defined and a g(s)ds = 0, where g(s) is not identically
zero, then the following inequality holds:
2.2
Generalized Jensen integral inequalities over infinite intervals
We extend the method used for proving Lemma 1 from
finite intervals of integration to infinite ones in the following result.
Rb
Rb T
Rb
1
ω T (s)Rω(s)ds ≥ b−a
ω (s)dsR a ω(s)ds
a
hR
i−1 R
Rb
b
b
T
+ a g 2 (s)ds
a g(s)ω (s)dsR a g(s)ω(s)ds.
(6)
a
Theorem 1 For a given n×n matrix R ≻ 0, scalar functions g : [0, +∞) → R, K : [0, +∞) → R+ and a vector
function ω : [0, +∞) → Rn , assume that the integrations
R +∞
concerned are well defined and 0 K(s)g(s)ds = 0 with
g(s) not identically zero. Then the following inequality
holds:
Proof: Define a function f (u) for all u ∈ [a, b] by
f (u) = ω(u) −
1
b−a
Rb
a
ω(s)ds − g(u)Θ,
(7)
R +∞
K(s)ω T (s)Rω(s)ds
R +∞
R +∞
≥ K0−1 0 K(s)ω T (s)dsR 0 K(s)ω(s)ds
hR
i−1
+∞
+ 0 K(s)g 2 (s)ds
Ω̄T RΩ̄,
0
where Θ ∈ Rn is a constant vector to be defined. Then,
since R ≻ 0 it follows that
0≤
=
Rb
f T (s)Rf (s)ds
a
h
iT h
i
Rb
Rb
Rb
1
1
ω(s)−
ω(θ)dθ
ω(θ)dθ
ds
R
ω(s)−
b−a a
b−a a
ah
iR
Rb T
b
2
+ b−a a ω (s)dsRΘ a g(s)ds
i
hR
Rb
b
+ a g 2 (s)ds ΘT RΘ − 2ΘT R a g(s)ω(s)ds.
Noting that
Rb
a
where
R +∞
K0 = 0 K(s)ds,
R +∞
Ω̄ = 0 K(s)g(s)ω(s)ds.
(10)
Proof:
g(s)ds = 0, we obtain
Define a function f¯(u) for all u ∈ [0, +∞) by
Rb T
Rb
1
ω (s)dsR a ω(s)ds
ω T (s)Rω(s)ds ≥ b−a
a
a h
i
Rb
Rb
− a g 2 (s)ds ΘT RΘ + 2ΘT R a g(s)ω(s)ds.
Rb
(9)
f¯(u) =
3
i
h
p
R +∞
K(u) ω(u)−K0−1 0 K(s)ω(s)ds− g(u)Θ̄ ,
where Θ̄ ∈ Rn is a constant vector to be defined. Because
R ≻ 0 we have
Corollary 1 For a given n × n matrix R ≻ 0, a scalar
function K : [0, +∞) → R+ and a vector function ω :
[0, +∞) → Rn , assume that the integrations concerned
are well defined. Then the following inequality holds:
R +∞
0 ≤ 0 f¯T (s)Rf¯(s)ds
iT
p
R +∞hp
R +∞
= 0
K(s)ω(s)−K0−1 K(s) 0 K(θ)ω(θ)dθ R
i
hp
p
R +∞
× K(s)ω(s)−K0−1 K(s) 0 K(θ)ω(θ)dθ ds
h
iR
R +∞
+∞
+ 2K0−1 Θ̄T R 0 K(s)ω(s)ds 0 K(s)g(s)ds
i
hR
+∞
+ 0 K(s)g 2 (s)ds Θ̄T RΘ̄
R +∞
−2Θ̄T R 0 K(s)g(s)ω(s)ds.
R +∞
(16)
where K0 , K1 and K2 are given by (10), (12) and (15),
respectively, and
RepresentingR the last two terms as sum of squares to+∞
gether with 0 K(s)g(s)ds = 0 yields
R +∞
K(s)ω T (s)Rω(s)ds
R +∞
R +∞
≥ K0−1 0 K(s)ω T (s)dsR 0 K(s)ω(s)ds
i−1
hR
+∞
Ω̄T RΩ̄
+ 0 K(s)g 2 (s)ds
hR
i
+∞
− 0 K(s)g 2 (s)ds [Θ̄ − Ῡ]T R[Θ̄ − Ῡ],
Ω̃ =
(11)
+∞
K(s)g 2 (s)ds
0
i−1 Z
+∞
K(s)g(s)ω(s)ds.
0
0
Note that the choice of g(s) plays a crucial role in the
application of Theorem 1. Given K0 in (10) and
let
R +∞
0
sK(s)ds,
g(u) = c(K0 u − K1 ), c ∈ R\{0}, u ≥ 0,
such that
R +∞
0
0
K(s)ω(s)ds −
R +∞
0
sK(s)ω(s)ds.
R +∞ R t
Then, the same arguments in the proof of Lemma 1 and
the choice Θ̄ = Ῡ lead to the maximum of the right-hand
side of (11) and thus, (9) holds. This concludes the proof.
K1 =
R +∞
Theorem 2 If there exist an n × n matrix R ≻ 0, scalar
functions g : [t − θ − h, t] → R, K : [0, +∞) → R+ , a
scalar h ≥ 0 and a vector function ω : [t − θ − h, t] →
Rn such
R +∞that
R t the integrations concerned are well defined
and 0
t−θ−h K(θ)g(s)dsdθ = 0, where g(s) is not
identically zero, then the following inequality holds:
where Ω̄ is given in (10) and
hZ
K1
K0
The same methodology to prove Lemma 1 and Theorem 1 can be applied to generalize the inequality (5). We
have the following result:
0
Ῡ =
K(s)ω T (s)Rω(s)ds
R +∞
R +∞
≥ K0−1 0 K(s)ω T (s)dsR 0 K(s)ω(s)ds
−1
K2
+ K2 − K10
Ω̃T RΩ̃,
0
T
t−θ−h K(θ)ω (s)Rω(s)dsdθ
R
R
+∞
t
−1
≥ K1h
K(θ)ω T (s)dsdθR
0
R +∞
R t t−θ−h
× 0
t−θ−h K(θ)ω(s)dsdθ
hR
i−1
+∞ R t
+ 0
K(θ)g 2 (s)dsdθ
ΣT RΣ,
t−θ−h
(17)
where
(12)
R +∞
K1h = 0 K(s)(s + h)ds = hK0 + K1 ,
R +∞ R t
Σ= 0
t−θ−h K(θ)g(s)ω(s)dsdθ.
(13)
(18)
Proof: See Appendix A.
K(s)g(s)ds = 0 holds. Then, we find that
(14)
Remark 2 Theorems 1 and 2 refine the inequalities of
[22], in which the last terms of the right-hand-side of (9)
and (17) are zero. Hence, our new inequalities develop
R +∞
more accurate lower bounds of 0 K(s)ω T (s)Rω(s)ds
R +∞ R t
and 0
K(θ)ω T (s)Rω(s)dsdθ than the ones
t−θ−h
provided in [22].
(15)
We choose a scalar function
From (13), (14) and Theorem 1, we have the following
corollary:
g(u) = −u + t −
R +∞
0
K(s)g 2 (s)ds = c2
R +∞
0
K(s)(K0 s − K1 )2 ds
= c2 (K02 K2 − K0 K12 ),
R +∞
Ω̄ = 0 K(s)g(s)ω(s)ds
i
h R
R +∞
+∞
= c K0 0 sK(s)ω(s)ds − K1 0 K(s)ω(s)ds ,
where
K2 =
R +∞
0
s2 K(s)ds.
4
hK1h +hK1 +K2
,
2K1h
(19)
R +∞ R t
such that 0
K(θ)g(s)dsdθ = 0, where K1 , K2
t−θ−h
and K1h are given by (12), (15) and (18), respectively.
Hence, we have
R +∞ R t
0
=
=
K(θ)g 2 (s)dsdθ
t−θ−h K(θ) − s + t −
t−θ−h
R +∞ R t
0
3
h
2
K0 + 2K3 +
hK1h +hK1 +K2
2K1h
3h2 K0 (hK1 +2K2 )−3K22 ∆
= K̃1 ,
2K1h
2
where x(t) ∈ Rn is the state vector, A, A1 ∈ Rn×n
are constant system matrices, and h ≥ 0 represents
a fixed time gap. The smooth kernel Γ is given by
N −1 − θ
T
Γ(θ) = Tθ N (Ne−1)!
, where N ≥ 2, N ∈ N, is a shape
parameter of the distribution and T > 0 is a scale
parameter. The matrices A and A + A1 are not allowed to be Hurwitz. The initial condition is given by
φ ∈ C 1 (−∞, 0], where C 1 (−∞, 0] denotes the space of
continuously differentiable functions φ : (−∞, 0] → Rn
with the norm kφkC 1 = kφkC + kφ̇kC < +∞,
kφkC = sups∈(−∞,0] |φ(s)| < +∞.
dsdθ
(20)
and
R +∞ R t
Σ= 0
K(θ)g(s)ω(s)dsdθ
R +∞ R t t−θ−h
+hK1 +K2
= 0
K(θ) −s+t− hK1h 2K
ω(s)dsdθ
t−θ−h
1h
R +∞R t
Rt
= 0
K(θ)ω(s)drdsdθ
t−θ−h s R
+∞ R t
hK1+K2
h
− 2 + 2K1h
K(θ)ω(s)dsdθ
0
t−θ−h
R +∞ R t
Rr
= 0
K(θ)ω(s)dsdrdθ
R
t−θ−h t−θ−h
Rt
∆
+∞
1+K2
− h2 + hK
2K1h
0
t−θ−h K(θ)ω(s)dsdθ = Σ̃,
(21)
where
R +∞
K3 = 0 s3 K(s)ds.
Following [22] and introducing
y(t) =
0
ρ(t) =
R +∞
R0+∞
R0+∞
R0+∞
(24)
−∞ Ψ(t− s)x(s−
θ
θ N −2 e− T
T N (N −2)! .
h)ds =
R +∞
0
Ψ(θ)x(t −θ− h)dθ,
Ψ(θ)dθ =
1 ∆
T =
Ψ0 ,
∆
θΨ(θ)dθ = N − 1 = Ψ1 ,
∆
θ2 Ψ(θ)dθ = N (N − 1)T = Ψ2 ,
∆
θ3 Ψ(θ)dθ = (N + 1)N (N − 1)T 2 = Ψ3 ,
0
R +∞
∆
(θ + h)Ψ(θ)dθ = hΨ0 + Ψ1 = N − 1 + Th =
0
(22)
Ψ1h .
(25)
In the following, we provide two sufficient conditions for
the stability of system (24); one is derived by applying
(16) and (5), the other is obtained from (16) and (22).
3.1
Stability result I
We consider the following augmented LKF:
V (t) = V1 (t) + VG (t) + VH (t),
V1 (t) = η T (t)W η(t),
R +∞ R t
VG (t) = 0
Ψ(θ)xT (s)Gx(s)dsdθ,
R +∞ Rt−θ−h
R
θ+h t
T
VH (t) = 0
0
t−λ Ψ(θ)ẋ (s)H ẋ(s)dsdλdθ,
(26)
where W ≻ 0, G ≻ 0, H ≻ 0, η(t) = col{x(t), y(t)}.
Since A and A+A1 are not allowed to be Hurwitz, we use
Consider the linear continuous-time systems with
gamma-distributed delays:
Γ(θ)x(t − θ − h)dθ,
Γ(t− s)x(s− h)ds,
It follows readily that
Stability analysis of continuous-time systems
with gamma-distributed delays
0
Rt
Ψ(θ) =
The generalized integral inequality (16) and its double
integral extension (22) will be employed for the stability analysis of continuous-time systems with gammadistributed delays in the next section.
R +∞
−∞
where
where K̃1 and Σ̃ are given by (20) and (21), respectively.
ẋ(t) = Ax(t) + A1
Rt
ẏ(t) = − T1 y(t) + ρ(t),
+K̃1−1 Σ̃T RΣ̃,
3
Γ(θ)x(t −θ− h)dθ =
ẋ(t) = Ax(t) + A1 y(t),
Corollary 2 If there exist an n × n matrix R ≻ 0, a
scalar function K : [0, +∞) → R+ , a scalar h ≥ 0 and
a vector function ω : [t − θ − h, t] → Rn such that the
integrations concerned are well defined, then the following
inequality holds:
T
t−θ−h K(θ)ω (s)Rω(s)dsdθ
R
R
−1 +∞ t
≥ K1h
K(θ)ω T (s)dsdθR
0
R +∞ R t t−θ−h
× 0
t−θ−h K(θ)ω(s)dsdθ
0
the system (23) can be transformed into
From (19)–(21) and Theorem 2, we arrive at the following result:
R +∞ R t
R +∞
(23)
5
with ξ(t) = col{x(t), y(t), ρ(t)}. By applying Corollary 1 we obtain
augmented Lyapunov functionals. The term VH extends
the triple integral of [24] for finite delay to infinite delay
[22].
R +∞
V̇G (t) = Ψ0 xT (t)Gx(t)− 0 Ψ(θ)xT (t−θ−h)Gx(t−θ−h)dθ
≤ Ψ0 xT (t)Gx(t) − Ψ−1
ρT (t)Gρ(t)
−1 h 0
iT
2
Ψ
Ψ1
− Ψ2 − Ψ10
ρ(t) − (N − 1)y(t) G
Ψ
0
i
h
Ψ1
ρ(t)
−
(N
−
1)y(t)
× Ψ
0
Remark 3 The recent method of [22] for the stability
analysis of system (24) is based on a functional of the
form
V (t) = V1 (t) + VG (t) + VH (t) + VE (t) + VF (t),
R +∞ R t
(27)
VE (t) = 0
Γ(θ)xT (s)Ex(s)dsdθ,
R +∞ Rt−θ−h
R
θ+h t
T
VF (t) = 0
Γ(θ)ẋ (s)F ẋ(s)dsdλdθ
0
t−λ
T
T
= ξ T (t)[Ψ1h F01
HF01 − Ψ−1
1h F13 HF13 ]ξ(t).
Therefore, (28) guarantees that V̇ (t) ≤ ξ T (t)Ξξ(t) ≤
−β|x(t)|2 for some β > 0, which proves the asymptotic
stability.
Proposition 1 If there exist 2n × 2n positive definite
matrix W and n × n positive definite matrices G, H such
that the following LMI is feasible:
F0 =
A A1 0
0 − T1 I
#
, F1 =
"
I 0 0
0 I 0
#
Stability result II
The stability of system (24) can be alternatively analyzed via a LKF given by
V̄ (t) = V̄1 (t) + VG (t) + VH (t), V̄1 (t) = η̄ T (t)W̄ η̄(t),
(28)
where W̄ ≻ 0, η̄(t) = col{x(t), y(t), ζ(t)}, ζ(t) =
R +∞ R t
Ψ(θ)x(s)dsdθ, VG (t) and VH (t) are given by
0
t−θ−h
(26). Noting that ζ̇(t) = T1 x(t) − ρ(t) and differentiating
V1 (t) along (24), we have
where Ψ1h is given by (25), Σ = diag{ T1 G, 0, −T G} and
"
1 T
T
T x (t)Gx(t) − T ρ (t)Gρ(t)
T
GF23 ξ(t).
− NT−1 ξ T (t)F23
V̇H (t) ≤ Ψ1h ẋT (t)H ẋ(t)
h
iT h
i
H Ψ0 x(t) − ρ(t)
−Ψ−1
1h Ψ0 x(t) − ρ(t)
3.2
T
T
+Ψ1h F01
HF01 − Ψ−1
1h F13 HF13 ≺ 0,
=
− T ρT (t)Gρ(t)
h
iT h
i
− NT−1 T ρ(t) − y(t) G T ρ(t) − y(t)
Furthermore, applying (5) we find that
The following proposition is provided for the asymptotic
stability of system (24).
N −1 T
T F23 GF23
1 T
T x (t)Gx(t)
(29)
together with the utilization of the integral inequalities
(4) and (5). Compared to (26), the functional (27) has
two additional terms VE (t) and VF (t). In the example
below, we will show the advantages of our proposed approach (larger stability region in the (T, h) plane and less
number of scalar decision variables). The improvement is
achieved due to that the application of Corollary 1 leads
to one more negative term in the derivative of the LKF.
Ξ = Σ + F1T W F0 + F0T W F1 −
=
,
Ax(t) + A1 y(t)
1
V̄˙ 1 (t) = 2η̄ T (t)W̄
− T y(t) + ρ(t)
1
T x(t) − ρ(t)
= 2ξ¯T (t)F̄ T W̄ F̄0 ξ̄(t)
F01 = [A A1 0], F13 = [ T1 I 0 − I], F23 = [0 − I T I],
(30)
1
then system (24) is asymptotically stable.
with ξ̄(t) = col{x(t), y(t), ρ(t), ζ(t)} and
Proof: Differentiating V1 (t) along (24), we have
I 0 0 0
A
A1
0 0
1
F̄1 =
0 I 0 0 , F̄0 = 0 − T I I 0 .
1
0 −I 0
0 0 0 I
TI
V̇1 (t) = 2η T (t)W η̇(t) = 2ξ T (t)F1T W F0 ξ(t)
6
(31)
Table 1
Example 1: maximum allowable value of T for different h
Furthermore, by applying (22) we find that
[max T ]
10−5
0.01
0.15
0.34
0.35
0.36
\ h
V̇H (t) ≤ Ψ1h ẋT (t)H ẋ(t)
h
iT h
i
−Ψ−1
H Ψ0 x(t) − ρ(t)
1h Ψ0 x(t) − ρ(t)
T
−Ψ̃−1
1 ϕ (t)Hϕ(t)
= ξ̄
T
(32)
where Ψ1h is given in (25) and
F̄01 = [A A1 0 0], F̄13 = [ T1 I 0 − I 0],
ϕ(t) = ~[Ψ0 x(t) − ρ(t)] − ζ(t) + hρ(t) + (N − 1)y(t)
= F̄33 ξ̄(t),
F̄33 = [ T~ I (N − 1)I (h − ~)I − I],
~=
h3
2T
h
2
+ 2Ψ3 +
+
3h2 Ψ0 (hΨ1 +2Ψ2 )−3Ψ22
,
2Ψ1h
hΨ1 +Ψ2
2Ψ1h
.
Ξ̄ = Σ̄ + F̄1T W̄ F̄0 + F̄0T W̄ F̄1 −
T
H F̄01 −
+Ψ1h F̄01
0.274
0.265
0.141
0.005
-
-
22
0.305
0.296
0.158
0.008
0.002
-
16
Prop. 2
0.322
0.312
0.168
0.014
0.008
0.003
31
Let us now compare the number of scalar decision variables in the LMIs. The LMIs of [22] have
{4n2 + 3n}n=2 = 22 variables. Proposition 1 in this paper not only possess a fewer number {3n2 +2n}n=2 = 16
of variables but also lead to less conservative results. In comparison with Proposition 1, Proposition 2 slightly improves the results at the price of
{2.5n2 + 2.5n}n=2 = 15 additional decision variables.
(33)
Therefore, by combining (29), (30) and (32) we obtain
¯ ≤ −β̄|x(t)|2 for some β̄ > 0, if
V̄˙ (t) ≤ ξ̄ T (t)Ξ̄ξ(t)
N −1 T
T F̄23 GF̄23
−1 T
T
Ψ1h F̄13 H F̄13 − Ψ̃−1
1 F̄33 H F̄33
[22]
Prop. 1
[22] and using Propositions 1 and 2, we obtain the maximum allowable values of T that achieve the stability. Fig.
1 presents tradeoff curves between maximal allowable T
and h by applying the above three methods. Furthermore, the stability region in the (T, h) plane that preserves the asymptotic stability is depicted in Figs. 2–4
by using the condition in [22], Proposition 1 and Proposition 2, respectively. From Figs. 2–4 we can see that
Proposition 1 induces a more dense stability region than
[22], but guarantees a little sparser stability region than
Proposition 2. Therefore, Figs. 1–4 show that Proposition 1 improves the results in [22] and that the conditions can be further enhanced by Proposition 2.
−1 T
T
T
¯
H F̄01−Ψ−1
(t)[Ψ1h F̄01
1h F̄13 H F̄13−Ψ̃1 F̄33 H F̄33 ]ξ(t),
Ψ̃1 =
Decision
variables
≺ 0,
(34)
where
0.4
Σ̄
=
diag{ T1 G, 0, −T G, 0},
F̄23 = [0 − I T I 0].
[22]
Proposition 1
Proposition 2
0.35
(35)
0.3
0.25
h
We have thus proved the following proposition:
0.2
0.15
Proposition 2 If there exist 3n × 3n positive definite
matrix W̄ and n × n positive definite matrices G, H such
that LMI (34) with notations given by (25), (31), (33)
and (35) is feasible, then system (24) is asymptotically
stable.
0.1
0.05
0
Example 1
4
We illustrate the efficiency of the presented results
through an example of two cars on a ring, see [14] and
[22] for details. In this example,
A = 0 and A1 =
"
−2 2
2 −2
#
0.05
0.1
0.15
0.2
T
0.25
0.3
0.35
0.4
Fig. 1. Example 1: tradeoff curve between maximal allowable
T and h for Propositions 1 and 2 compared with the result
of [22]
Next we present an example to illustrate the applicability of the theoretical results.
3.3
0
Extended Jensen summation inequalities
with infinite sequences
The objective of this section is to present the discrete
counterpart of the results obtained in Section 2 and to
provide extended Jensen summation inequality with infinite sequences. We first introduce the following lemma
for the discrete counterpart of the integral inequalities
(4) and (5):
,
Lemma 2 Assume that there exist an n × n matrix R ≻
0, a scalar function M (i) ∈ R+ and a vector function
so neither A nor A + A1 is Hurwitz. For the values of h
given in Table I and N = 2, by applying the method in
7
hold, where
0.4
condition of [22]
0.35
M1h
h
0.25
0.2
0.15
0.05
0
0.05
0.1
0.15
0.2
T
0.25
0.3
0.35
(38)
The proof of (36) and (37) follows from [22] by using
sums instead of integrals and is therefore omitted here.
By applying the arguments of Theorem 1 to the discrete
case, we obtain the following theorem for the extended
Jensen summation inequality with infinite sequences.
Note that this result includes (36) as a special case and
that the generalization of (37) can be done by the same
approach as exploited in Theorem 2.
0.1
0
P+∞
i=0 M (i),
P+∞
= i=0 (i + h)M (i).
M0 =
0.3
0.4
Fig. 2. Example 1: stability region by the condition of [22]
0.4
Proposition 1
0.35
Theorem 3 For a given n × n matrix R ≻ 0, scalar
functions M (i) : Z+ → R+ , g(i) : Z+ → R and a vector
function x(i) : Z+ → Rn , assume that the series conP+∞
cerned are well defined and i=0 M (i)g(i) = 0 with g(i)
not identically zero. Then the following inequality holds:
0.3
h
0.25
0.2
0.15
0.1
P+∞
M (i)xT (i)Rx(i)
hP
iT h P
i
+∞
+∞
M
(i)x(i)
≥ M0−1
R
M
(i)x(i)
i=0
i=0
hP
i−1
+∞
2
T
+
Π RΠ
i=0 M (i)g (i)
(39)
with M0 given by (38) and
0.05
0
i=0
0
0.05
0.1
0.15
0.2
T
0.25
0.3
0.35
0.4
Fig. 3. Example 1: stability region by Proposition 1
0.4
Proposition 2
0.35
Π=
0.3
h
0.25
0.2
P+∞
i=0
M (i)g(i)x(i).
(40)
Proof: See Appendix B.
0.15
In order to apply Theorem 3 to the stability analysis of
time-delay systems, we take
0.1
0.05
0
0
0.05
0.1
0.15
0.2
T
0.25
0.3
0.35
0.4
g(v) = c(M0 v − M1 ), c ∈ R\{0}, v ∈ Z+ ,
Fig. 4. Example 1: stability region by Proposition 2
such that
n
x(i) ∈ R such that the series concerned are convergent.
Then the inequality
M (i)xT (i)Rx(i)
hP
iT h P
i
+∞
+∞
≥ M0−1
R
i=0 M (i)x(i)
i=0 M (i)x(i) ,
i=0
P+∞
i=0
(36)
M (i)xT (j)Rx(j)
iT
+∞ Pk−1
−1
≥ M1h
M
(i)x(j)
R
j=k−i−h
h P i=0
i
+∞ Pk−1
×
i=0
j=k−i−h M (i)x(j) ,
j=k−i−h
hP
M (i)g(i) = 0 is satisfied, where
M1 =
P+∞
M (i)g 2 (i) = c2
P+∞
i=0
iM (i).
(42)
2
i=0 M (i)(M0 i − M1 )
c2 (M02 M2 − M0 M12 ),
=
P+∞
Π = i=0 M (i)g(i)x(i)
i
h
P+∞
P
M
(i)x(i)
,
iM
(i)x(i)
−
M
= c M0 +∞
1
i=0
i=0
and its double summation extension
i=0
i=0
Hence, we have
P+∞
P+∞ Pk−1
P+∞
(41)
(43)
(37)
where
M2 =
8
P+∞
i=0
i2 M (i).
(44)
τ
where P(τ ) = e τ !λ . We next derive LMI conditions for
the asymptotic stability of (47) via a direct Lyapunov
method.
−λ
From (41) and (43), Theorem 3 is reduced to the following corollary, which will be employed in the next section
for the stability analysis of discrete-time systems with
poisson-distributed delays.
Denoting
Corollary 3 Given an n × n matrix R ≻ 0, a scalar
function M (i) : Z+ → R+ and a vector function x(i) :
Z+ → Rn such that the series concerned are well defined,
the following inequality holds:
P+∞
f (k) =
M (i)x (i)Rx(i)
hP
iT h P
i
+∞
+∞
≥ M0−1
M
(i)x(i)
R
M
(i)x(i)
i=0
i=0
−1
M12
T
+ M2 − M0
Π̃ RΠ̃,
(45)
where M0 , M1 and M2 are given by (38), (42) and (44),
respectively, and
5
M1
M0
P+∞
i=0
M (i)x(i) −
P+∞
i=0
= e−λ Ax(k − h) + e−λ A1 f (k − h)
P+∞
+ τ =0 Q(τ )x(k − τ − h),
iM (i)x(i).
(48)
where Q(τ ) =
It is noted that the augmented
system (48) has not only distributed but also discrete
delays. This is different from augmented system (24) for
the case of gamma-distributed delays. Moreover, we find
that
e−λ λτ +1
(τ +1)! .
In this section, we will demonstrate the efficiency of the
extended Jensen summation inequality (45) through the
stability analysis of linear discrete-time systems with
poisson-distributed delays. Consider the following system:
P+∞
i=0 (i
P+∞
i=0
P+∞
P+∞
− τ ), k ∈ Z+ ,
(46)
where x(k) ∈ Rn is the state vector, the system matrices
A and A1 are constant with appropriate dimensions. We
do not allow A and A + A1 to be Schur stable. The initial condition is given as col{x(0), x(−1), x(−2), . . . } =
col{φ(0), φ(−1), φ(−2), . . . }. The function p(v), v ∈ Z+ ,
is a poisson distribution with a fixed time gap h ∈ Z+ :
p(v) =
(
τ =0 p(τ )x(k
e−λ λv−h
(v−h)!
v ≥ h,
0
v < h,
i=0
P+∞
i=0
P+∞
τ =0 p(τ )x(k
− τ) =
=
P+∞
τ =h p(τ )x(k
P+∞
θ=0 p(θ
P+∞
2
2
i Q(i) = λ − λ + 1 − e
(49)
−λ ∆
= Q̄2 ,
∆
+ h)Q(i) = λ+(1 − e−λ )(h−1) = Q̄1h .
i
P2 h
V (k) = x̂T (k)Ŵ x̂(k)+ i=1 VGi (k) + VHi (k) + VSi (k) ,
P+∞ Pk−1
VG1 (k) = i=0 s=k−i−h P(i)xT (s)G1 x(s),
P+∞ Pi+h Pk−1
VH1 (k) = i=0 j=1 s=k−j P(i)η1T (s)H1 η1 (s),
P+∞ Pk−1
VG2 (k) = i=0 s=k−i−h Q(i)xT (s)G2 x(s),
P+∞ Pi+h Pk−1
VH2 (k) = i=0 j=1 s=k−j Q(i)η1T (s)H2 η1 (s),
Pk−1
VS1 (k) = s=k−h xT (s)S1 x(s)
P−1 Pk−1
+h j=−h s=k+j η1T (s)R1 η1 (s),
Pk−1
VS2 (k) = s=k−h f T (s)S2 f (s)
P−1 Pk−1
+h j=−h s=k+j η2T (s)R2 η2 (s),
− τ)
+ h)x(k − θ − h),
τ =0 P(τ )x(k
∆
iQ(i) = λ − 1 + e−λ = Q̄1 ,
Consider system (48) with both distributed and discrete
delays. The stability analysis will be based on the following discrete-time LKF:
we arrive at the equivalent system:
x(k + 1) = Ax(k)+A1
+ h)P(i) = λ + h,
Q(i) = 1 − e−λ ,
i=0 (i
where λ > 0 is a parameter of the distribution. The mean
value of p is h + λ. Due to the fact that
P+∞
− τ − h), k ∈ Z+ ,
x(k + 1) = Ax(k) + A1 f (k),
P+∞ −λ τ
f (k + 1) = τ =0 e τ !λ x(k + 1 − τ − h)
P+∞ −λ τ
= e−λ x(k + 1 − h) + τ =1 e τ !λ x(k + 1 − τ − h)
P+∞ −λ λτ +1
x(k − τ − h)
= e−λ x(k + 1 − h) + τ =0 e (τ +1)!
Stability analysis of discrete-time systems
with poisson-distributed delays
x(k + 1) = Ax(k) + A1
τ =0 P(τ )x(k
the system (47) can be transformed into the following
augmented form
T
i=0
Π̃ =
P+∞
−τ − h), k ∈ Z+ ,
(47)
where Ŵ ≻ 0, Gi ≻ 0, Hi ≻ 0, Si ≻ 0, Ri ≻ 0, i = 1, 2,
9
and
and
x̂(k) = col{x(k), f (k)},
h
i=1 VGi (k + 1) + VHi (k + 1) + VSi (k + 1)
i
−VGi (k) − VHi (k) − VSi (k)
h
T
= ξ T (k) Σ̂ + F̂01
[(λ + h)H1 + Q̄1h H2 + h2 R1 ]F̂01
P2
η1 (k) = x(k + 1) − x(k),
η2 (k) = f (k + 1) − f (k).
Here the last two terms VS1 (k) and VS2 (k) are added to
compensate the delayed terms x(k − h) and f (k − h) of
(48), respectively. Therefore, for system (48) with h = 0,
the terms VS1 (k) and VS2 (k) are not necessary. From
standard arguments, we arrive at the following result for
the asymptotic stability of (48):
T
T
R2 F̂02
H1 F̂12 + h2 F̂02
−(λ + h)−1 F̂12
i
T
T
−F̂13 R1 F̂13 − F̂24 R2 F̂24 ξ(k)
+(1 − e−λ )−1 q T (k)G2 q(k)
P+∞
− i=0 Q(i)xT (k − i − h)G2 x(k − i − h)
P+∞ Pk−1
− i=0 s=k−i−h Q(i)η1T (s)H2 η1 (s).
Proposition 3 Given a real scalar λ > 0 and an integer
h ≥ 0, assume that there exist 2n × 2n positive definite
matrix Ŵ and n × n positive definite matrices Gi , Hi ,
Si , Ri , i = 1, 2, such that the following LMI is satisfied:
(52)
Applying the generalized Jensen inequality (45) with infinite sequences, we obtain
−
T
Ξ̂ = Σ̂+ F̂0T Ŵ F̂0 − F̂1T Ŵ F̂1 − (λ + h)−1 F̂12
H1 F̂12
T
T
R1 F̂13
F̂ T H2 F̂15 − F̂13
+h2 F̂02
R2 F̂02 − Q̄−1
1h 15 2 −1
Q̄1
T
T
−F̂24 R2 F̂24 − Q̄2 − 1−e−λ
F̂25 G2 F̂25 ≺ 0,
F̂0 =
A A1
F̂1 =
"
I 0 0 00
0 0 e
0
0
−λ
−λ
0 I 0 00
Ae
#
0
A1 I
,
−λ
−λ
= −Q̄−1
)x(k)−q(k)]T H2 [(1−e−λ )x(k)−q(k)]
1h [(1−e
)I 0 0 0 −I],
T
T
= −Q̄−1
1h ξ (k)F̂15 H2 F̂15 ξ(k).
Q̄1
F̂24 = [0 I 0 −I 0], F̂25 = [0 − λI 0 0 ( 1−e
−λ + 1)I].
(54)
Therefore, (51)–(54) yield ∆V (k) = V (k + 1) − V (k) ≤
ξ T (k)Ξ̂ξ(k). Then if (50) holds for given scalars λ > 0
and h ≥ 0, the system (48) is asymptotically stable.
Then the system (48) is asymptotically stable.
Remark 4 The LMI condition in Proposition 3 is derived by employing the generalized Jensen inequality (45).
The system (48) can be alternatively analyzed by inequality (36). In this case, (53) is reduced to:
Proof: Define
τ =0
P+∞ Pk−1
T
s=k−i−h Q(i)η1 (s)H2 η1 (s)
hi=0
i
h
iT
−1
P+∞
P+∞ Pk−1
≤−
i=0 (i+h)Q(i)
i=0
s=k−i−h Q(i)η1 (s) H2
i
hP
+∞ Pk−1
Q(i)η
(s)
×
1
s=k−i−h
i=0
−
, F̂01 = [A−I A1 0 0 0],
F̂13 = [I 0 −I 0 0], F̂15 = [(1 − e
P+∞
(53)
Furthermore, the application of (37) leads to
F̂02 = [0 −I e−λ A e−λ A1 I], F̂12 = [I −I 0 0 0],
q(k) =
Q(i)xT (k − i − h)G2 x(k − i − h)
= −(1 − e−λ )−1 q T (k)G2 q(k)
−1
Q̄2
T
ξ T (k)F̂25
G2 F̂25 ξ(k).
− Q̄2 − 1−e1−λ
(50)
where Q̄1 , Q̄2 and Q̄1h are given by (49), Σ̂ = diag{S1 +
G1 +(1−e−λ )G2 , −G1 +S2 , −S1 , −S2 , −(1−e−λ )−1 G2 },
#
i=0
≤ −(1 − e−λ )−1 q T (k)G2 q(k)
−1 h
iT
Q̄2
Q̄1
− Q̄2 − 1−e1−λ
+1
q(k)−λf
(k)
G2
−λ
1−e
i
h
Q̄1
× 1−e−λ + 1 q(k) − λf (k)
T
+F̂01
[(λ+h)H1 + Q̄1h H2 + h2 R1 ]F̂01
"
P+∞
Q(τ )x(k − τ − h),
ξ(k) = col{x(k), f (k), x(k − h), f (k − h), q(k)}.
−
By taking difference of V (k) along (48) and applying
Jensen inequalities with finite sequences (see e.g., Chapter 6 of [3]), we have
x̂T (k + 1)Ŵ x̂(k + 1) − x̂T (k)Ŵ x̂(k)
= ξ T (k)[F̂0T Ŵ F̂0 − F̂1T Ŵ F̂1 ]ξ(k)
P+∞
i=0
Q(i)xT (k − i − h)G2 x(k − i − h)
≤ −(1 − e−λ )−1 q T (k)G2 q(k).
It yields Ξ̂|F̂25 =0 ≺ 0, which is more conservative than
the condition proposed in Proposition 3 since the matrix
−1
Q̄2
T
− Q̄2 − 1−e1−λ
F̂25
G2 F̂25 of (50) is negative definite.
(51)
10
Θ̂ ∈ Rn his a constant vector to be defined and fˆ(r,
i u) =
p
R
R
−1 +∞ t
K(u) ω(r) − K1h 0
with
t−θ−h K(θ)ω(s)dsdθ
K1h given by (17). Since R ≻ 0 we have
Remark 5 Both conditions in Proposition 3 and Remark 4 are derived by the use of inequality (37). It is
worth noting that the results could be further improved
(in the (λ, h) plane preserving the stability) by the discrete counterpart of Theorem 2.
5.1
Example 2
Consider the linear discrete-time system (46) with
A=
"
−0.5 0
0
1
#
and A1 =
"
−0.5 0.8
0.5 −0.2
#
R +∞ R t
f T (s, θ)Rf (s, θ)dsdθ
R0+∞ Rt−θ−h
t
ˆT
ˆ
= 0
t−θ−h f (s, θ)Rf (s, θ)dsdθ
h
i
R +∞ R t
−1 T
+ 2K1h Θ̂ R 0
t−θ−h K(θ)ω(s)dsdθ
R +∞ R t
×
K(θ)g(s)dsdθ
i
h R 0 R t−θ−h
+∞ t
2
Θ̂T RΘ̂
K(θ)g
(s)dsdθ
+ 0
t−θ−h
R
R
+∞ t
−2Θ̂T R 0
t−θ−h K(θ)g(s)ω(s)dsdθ,
0 ≤
.
Here neither A nor A + A1 is Schur stable. For h = 0
the values of λ that guarantee the asymptotic stability
of the system by Remark 4 and Proposition 3 are shown
in Fig. 5, where we can see that the results achieved by
Proposition 3 are less conservative than those obtained
by Remark 4. It is noted that Proposition 3 and Remark
4 possess the same number {5n2 + 4n}n=2 = 28 of variables.
Rewriting
last two terms as sum of squares together
R +∞ Rthe
t
with 0
K(θ)g(s)dsdθ = 0 leads to
t−θ−h
R +∞ R t
0
Remark 4
Proposition 3
T
t−θ−h K(θ)ω (s)Rω(s)dsdθ
R
R
−1 +∞ t
T
≥ K1h
0
t−θ−h K(θ)ω (s)dsdθR
R +∞ R t
×
K(θ)ω(s)dsdθ
h R 0 R t−θ−h
i−1
+∞ t
2
+ 0
K(θ)g
(s)dsdθ
ΣT RΣ
t−θ−h
hR
i
R
+∞ t
2
T
− 0
t−θ−h K(θ)g (s)dsdθ [Θ̂ − Υ̂] R[Θ̂
− Υ̂],
(55)
where Σ is given by (18) and
Υ̂ =
0
0.5
1
1.5
λ
2
2.5
3
Then the inequality (55) with Θ̂ = Υ̂ implies (17). This
concludes the proof.
Fig. 5. Example 2: stabilizing values of λ
6
i−1
+∞ R t
2
K(θ)g
(s)dsdθ
0
t−θ−h
R +∞ R t
× 0
t−θ−h K(θ)g(s)ω(s)dsdθ.
hR
Conclusions
Appendix B
In this paper, we have provided extended Jensen integral
inequalities. For the discrete counterpart we have generalized Jensen summation inequality. Applications to
the stability analysis of linear continuous-time systems
with gamma-distributed delays and linear discrete-time
systems with poisson-distributed delays demonstrated
the advantages of these generalized inequalities. In both
cases, the considered infinite distributed delays with a
gap have stabilizing effects. The future research may include other applications of these developed inequalities.
Proof of Theorem 3p
Define a function f (v) for all v ∈
Z+ by f (v) = f˜(v) − M (v)g(v)Θ̃, where Θ̃ ∈ hRn is a
p
constant vector to be defined and f˜(v) = M (v) x(v)−
i
P+∞
M0−1 i=0 M (i)x(i) with M0 given by (38) . Then since
R ≻ 0 it follows that
P+∞ T
f (i)Rf (i)
iP
h i=0
P+∞
+∞
= 2M0−1 Θ̃T R i=0 M (i)x(i)
i=0 M (i)g(i)
P+∞ ˜T
˜
+ i=0 f (i)Rf (i)
hP
i
+∞
2
+
M
(i)g
(i)
Θ̃T RΘ̃
i=0
P+∞
−2Θ̃T R i=0 M (i)g(i)x(i).
0 ≤
Appendix A
Proof of Theorem 2 Following the proof of Theorem 1,
we define a function f (r, u) for allpu ∈ [0, +∞), r ∈
[t − u − h, t] by f (r, u) = fˆ(r, u) − K(u)g(r)Θ̂, where
11
RepresentingPthe last two terms as sum of squares to+∞
gether with i=0 M (i)g(i) = 0 yields
[13] Y.R. Liu, Z.D. Wang, J.L. Liang, and X.H. Liu.
Synchronization and state estimation for discrete-time
complex networks with distributed delays.
IEEE
Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics, 38(5):1314–1325, 2008.
P+∞
M (i)xT (i)Rx(i)
hP
iT h P
i
+∞
+∞
M
(i)x(i)
≥ M0−1
R
M
(i)x(i)
i=0
i=0
hP
i−1
+∞
2
T
+
M (i)g (i)
Π RΠ
h Pi=0
i
+∞
2
T
−
i=0 M (i)g (i) [Θ̃ − Υ̃] R[Θ̃ − Υ̃],
i=0
[14] C.I. Morarescu, S.I. Niculescu, and K. Gu. Stability crossing
curves of shifted gamma-distributed delay systems. SIAM
Journal on Applied Dynamical Systems, 6(2):475–493, 2007.
[15] S.I. Niculescu. Delay Effects on Stability: A Robust Control
Approach. Springer-Verlag, Berlin, 2001.
[16] P.G. Park, W.I. Lee, and S.Y. Lee. Auxiliary functionbased integral inequalities for quadratic functions and their
applications to time-delay systems. Journal of the Franklin
Institute, 352(4):1378–1396, 2015.
(56)
where Π is given by (40) and
Υ̃ =
+∞
hX
i=0
+∞
i−1 X
M (i)g(i)x(i).
M (i)g 2 (i)
[17] J.P. Richard.
Time-delay systems: An overview of
some recent advances and open problems. Automatica,
39(10):1667–1694, 2003.
i=0
[18] A. Saleh and R. Valenzuela. A statistical model for indoor
multipath propagation. IEEE Journal on Selected Areas in
Communications, 5(2):128–137, 1987.
The choice of Θ̃ = Υ̃ results in the maximum of the
right-hand side of (56) and thus (39). This concludes the
proof.
[19] A. Seuret and F. Gouaisbaut. On the use of the Wirtinger
inequalities for time-delay systems. In Proceedings of the 10th
IFAC workshop on time-delay systems, Boston, MA, USA,
2012.
References
[20] A. Seuret and F. Gouaisbaut. Wirtinger-based integral
inequality: application to time-delay systems. Automatica,
49(9):2860–2866, 2013.
[1] Y. Ariba and F. Gouaisbaut. Construction of LyapunovKrasovskii functional for time-varying delay systems. In
Proceedings of the 47th IEEE Conference on Decision and
Control, Cancun, Mexico, December 2008.
[21] A. Seuret, F. Gouaisbaut, and E. Fridman. Stability of
discrete-time systems with time-varying delays via a novel
summation inequality. IEEE Transactions on Automatic
Control, 60(10):2740–2745, 2015.
[2] C. Briat. Convergence and equivalence results for the
Jensen’s inequality–application to time-delay and sampleddata systems. IEEE Transactions on Automatic Control,
56(7):1660–1665, 2011.
[3] E. Fridman.
Introduction
Birkhäuser, 2014.
to
Time-Delay
[22] O. Solomon and E. Fridman. New stability conditions for
systems with distributed delays. Automatica, 49(11):3467–
3475, 2013.
[23] O. Solomon and E. Fridman. Stability and passivity analysis
of semilinear diffusion PDEs with time-delays. International
Journal of Control, 88(1):180–192, 2015.
Systems.
[4] E. Fridman and U. Shaked. Delay-dependent stability and
H∞ control: constant and time-varying delays. International
Journal of Control, 76(1):48–60, 2003.
[24] J. Sun, G. Liu, and J. Chen. Delay-dependent stability and
stabilization of neutral time-delay systems. International
Journal of Robust and Nonlinear Control, 19(12):1364–1375,
2009.
[5] E. Fridman, U. Shaked, and K. Liu. New conditions for delayderivative-dependent stability. Automatica, 45(11):2723–
2727, 2009.
[25] Z.G. Wu, P. Shi, H.Y. Su, and J. Chu. Reliable control for
discrete-time fuzzy systems with infinite-distributed delay.
IEEE Transactions on Fuzzy Systems, 20(1):22–31, 2012.
[6] D. Gross, J. Shortle, J. Thompson, and C. Harris.
Fundamentals of queueing theory. John Wiley & Sons, 2008.
[26] Y.Q. Xia and Y.M. Jia. Robust control of state delayed
systems with polytopic type uncertainties via parameterdependent Lyapunov functionals. Systems & Control Letters,
50(3):183–193, 2003.
[7] K. Gu. An improved stability criterion for systems with
distributed delays. International Journal of Robust and
nonlinear control, 13(9):819–831, 2003.
[8] K. Gu, V. Kharitonov, and J. Chen. Stability of Time-Delay
Systems. Birkhäuser, Boston, 2003.
[9] V.B. Kolmanovskii and J.P. Richard. Stability of some
linear systems with delays. IEEE Transactions on Automatic
Control, 44(5):984–989, 1999.
[10] S.A. Kotsopoulos and K. Loannou. Handbook of Research
on Heterogeneous Next Generation Networking: Innovations
and Platforms. IGI Global, 2008.
[11] K. Liu and E. Fridman. Wirtinger’s inequality and Lyapunovbased sampled-data stabilization. Automatica, 48(1):102–
108, 2012.
[12] K. Liu and E. Fridman. Delay-dependent methods and the
first delay interval. Systems & Control Letters, 64(1):57–63,
2014.
12
| 3cs.SY
|
arXiv:1403.2000v1 [cs.CE] 8 Mar 2014
A Galois-Connection between Myers-Briggs’
Type Indicators and Szondi’s Personality Profiles
Simon Kramer
[email protected]
May 8, 2014
Abstract
We propose a computable Galois-connection between Myers-Briggs’
Type Indicators (MBTIs), the most widely-used personality measure for
non-psychiatric populations (based on C.G. Jung’s personality types), and
Szondi’s personality profiles (SPPs), a less well-known but, as we show,
finer personality measure for psychiatric as well as non-psychiatric populations (conceived as a unification of the depth psychology of S. Freud,
C.G. Jung, and A. Adler). The practical significance of our result is that
our Galois-connection provides a pair of computable, interpreting translations between the two personality spaces of MBTIs and SPPs: one concrete from MBTI-space to SPP-space (because SPPs are finer) and one
abstract from SPP-space to MBTI-space (because MBTIs are coarser).
Thus Myers-Briggs’ and Szondi’s personality-test results are mutually interpretable and inter-translatable, even automatically by computers.
Keywords: applied order theory, computational and mathematical psychology, depth psychology, machine translation, MBTI, personality tests.
1
Introduction
According to [8, Page xxi and 210], the Myers-Briggs Type Indicator (MBTI) [7],
based on C.G. Jung’s personality types [3], has become “the most widely-used
personality measure for non-psychiatric populations” and “the most extensively
used personality instrument in history” with over two million tests taken per
year. In this paper, we propose a computable Galois-connection [2] between
MBTIs and Szondi’s personality profiles (SPPs) [9], a less well-known but, as
we show, finer personality measure for psychiatric as well as non-psychiatric
populations, and conceived as a unification [10] of the depth psychology of S.
Freud, C.G. Jung, and A. Adler.
Our result is a contribution to mathematical psychology in the area of depth
psychology, which does not yet seem to have been explored with mathematical
means besides those of statistics (often not part of mathematics departments). It
is also meant as a contribution towards practicing psychological research with
1
the methods of the exact sciences, for obvious ethical reasons. The practical
significance of our result is that our Galois-connection provides a pair of efficiently computable, interpreting translations between the two personality spaces
of MBTIs and SPPs (and thus hopefully also between their respective academic
and non-academic communities): one concrete translation from MBTI-space to
SPP-space (because SPPs are finer than MBTIs) and one abstract translation
from SPP-space to MBTI-space (because MBTIs are coarser than SPPs). Thus
Myers-Briggs’ and Szondi’s personality-test results are mutually interpretable
and inter-translatable, even automatically by computers. The only restriction to
this mutuality is the subjective interpretation of the faithfulness of these translations. In our interpretation, we intentionally restrict the translation from
SPP-space to MBTI-space, and only that one, in order to preserve (our perception of) its faithfulness. More precisely, we choose to map some SPPs to the
empty set in MBTI-space (but every MBTI to a non-empty set in SPP-space).
Our readers can experiment with their own interpretations, as we explain below.
We stress that our Galois-connection between the spaces of MBTIs and SPPs
is independent of their respective test, which evaluate their testees in terms of
structured result values—the MBTIs and SPPs—in the respective space. Both
tests are preference-based, more precisely, test evaluation is based on choices of
preferred questions in the case of the MBTI-test [7] and on choices of preferred
portraits in the case of the Szondi-test [9, 6]. Due to the independence of our
Galois-connection from these tests, their exact nature need not concern us here.
All what we need to be concerned about is the nature of the structured result
values that these tests generate. (Other test forms can generate the same form
of result values, e.g. [5].) We also stress that our proposed Galois-connection is
what we believe to be an interesting candidate brain child for adoption by the
community, but that there are other possible candidates, which our readers are
empowered to explore themselves. In fact, not only do we propose a candidate
Galois-connection between MBTI-space and SPP-space, but also do we propose
a whole methodology for generating such candidates. All what readers interested
in generating such connections themselves need to do is map their own intuition
about the meaning of MBTIs to a standard interlingua, called Logical Pivot
Language (LPL) here, and check that their mapping has a single simple property,
namely the one stated as Fact 2.1 about our mapping i in Figure 1. Their desired
Galois-connection is then automatically induced jointly by their chosen mapping
and a mapping, called p here, from SPP-space to LPL that we choose once and
for all possible Galois-connections of interest. What is more, our methodology is
applicable even more generally to the generation of Galois-connections between
pairs of result spaces of other personality tests. SPPs just happen to have a
finer structure than other personality-test values that we are aware of, and so
are perhaps best suited to play the distinguished role of explanatory semantics
for result values of other personality tests. Of course our readers are still free
to choose their own preferred semantic space.
An SPP can be conceived as a tuple of eight, so-called signed factors whose
signatures can in turn take twelve values. So SPPs live in an eight-dimensional
space. On the other hand, an MBTI can be conceived as a quadruple of two2
Figure 1: Mappings between personality spaces and interlingua
.
MBT I
SPP
/
p
i
LPL
valued components, namely, first, extro-/introversion, second, perception, being
either sensing or intuition, third, judgment, being either thinking or feeling, and
fourth, a dominance flag, indicating either a dominance of perception or judgment. So MBTIs live in a coarser, four dimensional space. Hence the translation
from SPPs to MBTIs must be a projection (and thus surjection) of SPP-space
onto MBTI-space. An insight gained in the finer referential system of SPPs is
that MBTIs turn actually out to be non-orthogonal or not independent, contrary
to common belief [7, 8]. Of course our readers are still free to disagree on the
value of this insight by giving a convincing argument for why SPP-space would
be an inappropriate semantics for MBTI-space. After all, Szondi conceived his
theory of human personality as a unifying theory that also includes Jung’s theory, on which MBTI-theory is based. We now put forward our own argument
for why we believe SPP-space is indeed an appropriate—though surely not the
only—semantics for MBTI-space. In Section 2.1, we present the defining mathematical structures for each space, and in Section 2.2, the defining mathematical
mappings for their translation. No prior knowledge of either MBTIs or SPPs is
required to appreciate the results of this paper.
2
The connection
In this section, we present the defining mathematical structures for MBTI-space,
the interlingua LPL, and SPP-space, as well as the defining mathematical mappings for the concrete translation of MBTI-space to SPP-space and the abstract
translation of SPP-space back to MBTI-space, both via LPL, see Figure 1.
2.1
Structures
In this section, we present the defining mathematical structures for MBTI-space,
the interlingua LPL, and SPP-space. We start with defining MBTI-space.
Definition 1 (The Myers-Briggs Type Indicator Space). Let
MB = {E, I, F, T, N, S, J, P}
be the set of basic type indicators, with E meaning “extroversion,” I “introversion,” F “feeling,” T “thinking,” N “intuition,” S “sensing,” J “judging,” and P
3
“perceiving.” Further let
MBTI = { ISTJ, ISFJ, INFJ, INTJ, ISTP, ISFP, INFP, INTP,
ESTP, ESFP, ENFP, ENTP, ESTJ, ESFJ, ENFJ, ENTJ }
be the set of Myers-Briggs Type Indicators (MBTIs) [7, 8].
Then,
MBT I = h 2MBTI , ∅, ∩, ∪, MBTI, · , ⊆ i
defines our Myers-Briggs Type Indicator Space, that is, the (inclusion-ordered,
Boolean) powerset algebra [2] on MBTI (the set of all subsets of MBTI).
Note that we do need to define MBT I as the set of all subsets of MBTI and
not simply as the set of all elements of MBTI. The reason is the aforementioned
fact that in the finer referential system of SPP-space (see Definition 2), MBTIs
turn out to be non-orthogonal or not independent, and thus an MBTI may have
to be mapped to a proper set of SPPs (see Table 2). So the proper setting for
SPP-space is a set of subsets of SPPs, which in turn, via the backward translation from SPP-space to MBT I, means that the proper setting for MBT I,
as the target of a mapping of subsets, is also a set of subsets. Further, notice
that the MBTI-test [7], which as previously mentioned requires answering to
questions, actually requires the P-faculty (perception of the question), the Nfaculty (intuition in the sense of textual, and thus symbolic understanding), and
the J-faculty (judgment in the sense of choice of and decision about an answer)
as its own prerequisites. Incidentally, the concept of choice is the key concept
in Szondi’s depth-psychological fate analysis [11, 10], which is the background
theory for his test [9] and the SPPs that it generates.
We continue to define SPP-space.
Definition 2 (The Szondi Personality Profile Space). Let us consider the Hassediagram [2] in Figure 2 of the partially ordered set of Szondi’s twelve signatures
[9] of human reactions, which are:
• approval: from strong +!!! , +!! , and +! to weak + ;
• indifference/neutrality: 0 ;
• rejection: from weak − , −! , and −!! to strong −!!! ; and
• ambivalence: ±! (approval bias), ± (no bias), and ±! (rejection bias).
(Szondi calls the exclamation marks in his signatures quanta.)
Further let us call this set of signatures S, that is,
S = { −!!!, −!!, −!, −, 0, +, +!, +!!, +!!!, ±! , ±, ±! }.
Now let us consider Szondi’s eight factors and four vectors of human personality [9] as summarised in Table 1. (Their names are of clinical origin and
need not concern us here.) And let us call the set of factors F, that is,
F = { h, s, e, hy, k, p, d, m }.
4
Figure 2: Hasse-diagram of Szondi’s signatures
+!!!
+!!
+!
+
±!
0
±
−
±!
−!
−!!
−!!!
Then,
SPP = { ((h, s1 ), (s, s2 ), (e, s3 ), (hy, s4 ), (k, s5 ), (p, s6 ), (d, s7 ), (m, s8 )) |
s1 , . . . , s8 ∈ S }
is the set of Szondi’s personality profiles, and
SPP = h 2SPP , ∅, ∩, ∪, SPP, · , ⊆ i
defines our Szondi Personality Profile Space, that is, the (inclusion-ordered,
Boolean) powerset algebra [2] on SPP (the set of all subsets of SPP).
As an example of an SPP, consider the norm profile for the Szondi-test [9]:
((h, +), (s, +), (e, −), (hy, −), (k, −), (p, −), (d, +), (m, +))
Spelled out, this norm profile describes the personality of a human being who
approves of physical love, has a proactive attitude, has unethical but moral
behaviour, wants to have and be less, and is unfaithful and dependent.
We conclude this subsection with the definition of our interlingua LPL.
Definition 3 (The Logical Pivot Language). Let
A = { hs1 , ss2 , es3 , hys4 , ks5 , ps6 , ds7 , ms8 | s1 , . . . , s8 ∈ S }
5
Table 1: Szondi’s factors and vectors
Vector
S (Id)
P
(SuperEgo)
Sch (Ego)
C (Id)
Signature
Factor
h (love)
s (attitude)
e (ethics)
hy (morality)
k (having)
p (being)
d (relations)
m (bindings)
+
−
physical love
(proactive) activity
ethical behaviour
immoral behaviour
having more
being more
unfaithfulness
dependence
platonic love
(receptive) passivity
unethical behaviour
moral behaviour
having less
being less
faithfulness
independence
be our set of atomic logical formulas, and LPL(A) the classical propositional
language over A, that is, the set of sentences constructed from the elements in
A and the classical propositional connectives ¬ (negation, pronounced “not”),
∧ (conjunction, pronounced “and”), ∨ (disjunction, pronounced “or”), etc.
Then,
LPL = h LPL(A), ⇒ i
defines our logical pivot language, with ⇒ being logical consequence.
Logical equivalence ≡ is defined in terms of ⇒ such that for every φ, ϕ ∈
LPL(A), φ ≡ ϕ by definition if and only if φ ⇒ ϕ and ϕ ⇒ φ.
2.2
Mappings between structures
In this section, we present the defining mathematical mappings for the concrete
translation . of MBT I to SPP via LPL and the abstract translation / of SPP
back to MBT I again via LPL by means of the auxiliary mappings i and p. We
also prove that the ordered pair ( . , / ) is a Galois-connection, as promised.
Definition 4 (Mappings). Let the mapping (total function)
• i be defined in the function space (MBTI → LPL(A)) as in Table 2 and
in the function space (2MBTI → LPL(A)) such that for every I ∈ 2MBTI ,
^
i(I) = { i(i) | i ∈ I } ;
• p be defined in the function space (SPP → LPL(A)) such that
p(((h, s1 ), (s, s2 ), (e, s3 ), (hy, s4 ), (k, s5 ), (p, s6 ), (d, s7 ), (m, s8 ))) =
hs1 ∧ ss2 ∧ es3 ∧ hys4 ∧ ks5 ∧ ps6 ∧ ds7 ∧ ms8
and in the function space (2SPP → LPL(A)) such that for every P ∈ 2SPP ,
_
p(P ) = { p(p) | p ∈ P } .
6
Table 2: Translating MB and MBTI to LPL(A)
i(E) = hy+ ∨ hy+! ∨ hy+!! ∨ hy+!!! ∨ hy±!
i(I) = hy− ∨ hy−! ∨ hy−!! ∨ hy−!!! ∨ hy±!
i(F) = (h+ ∨ h± ∨ h±! ) ∧ (p− ∨ p± ∨ p±! )
i! (F) = (h+! ∨ h+!! ∨ h+!!! ∨ h±! ) ∧
(p−! ∨ p−!! ∨ p−!!! ∨ p±! )
i(T) = k− ∨ k± ∨ k±!
i! (T) = k−! ∨ k−!! ∨ k−!!! ∨ k±!
i(N) = (k+ ∨ k± ∨ k±! ) ∧ (p+ ∨ p± ∨ p±! )
i! (N) = (k+! ∨ k+!! ∨ k+!!! ∨ k±! ) ∧
(p+! ∨ p+!! ∨ p+!!! ∨ p±! )
i(S) = (k+ ∨ k± ∨ k±! ) ∧
((h+ ∨ e− ∨ hy− ∨ d+ ∨ m+) ∨
(h± ∨ e± ∨ hy± ∨ d± ∨ m±) ∨
(h±! ∨ e±! ∨ hy±! ∨ d±! ∨ m±! ))
i! (S) = (k+! ∨ k+!! ∨ k+!!! ∨ k±! ) ∧
((h+! ∨ e−! ∨ hy−! ∨ d+! ∨ m+!) ∨
(h+!! ∨ e−!! ∨ hy−!! ∨ d+!! ∨ m+!!) ∨
(h+!!! ∨ e−!!! ∨ hy−!!! ∨ d+!!! ∨ m+!!!) ∨
(h±! ∨ e±! ∨ hy±! ∨ d±! ∨ m±! ))
i(ISTJ) = i(I) ∧ i! (S) ∧ i(T)
i(ISFJ) = i(I) ∧ i! (S) ∧ i(F)
i(INFJ) = i(I) ∧ i! (N) ∧ i(F)
i(INTJ) = i(I) ∧ i! (N) ∧ i(T)
i(ISTP) = i(I) ∧ i(S) ∧ i! (T)
i(ISFP) = i(I) ∧ i(S) ∧ i! (F)
i(INFP) = i(I) ∧ i(N) ∧ i! (F)
i(INTP) = i(I) ∧ i(N) ∧ i! (T)
i(ESTP) = i(E) ∧ i! (S) ∧ i(T)
i(ESFP) = i(E) ∧ i! (S) ∧ i(F)
i(ENFP) = i(E) ∧ i! (N) ∧ i(F)
i(ENTP) = i(E) ∧ i! (N) ∧ i(T)
i(ESTJ) = i(E) ∧ i(S) ∧ i! (T)
i(ESFJ) = i(E) ∧ i(S) ∧ i! (F)
i(ENFJ) = i(E) ∧ i(N) ∧ i! (F)
i(ENTJ) = i(E) ∧ i(N) ∧ i! (T)
7
Then, the mapping
•
.
: MBT I → SPP defined such that for every I ∈ 2MBTI ,
I . = { p ∈ SPP | p(p) ⇒ i(I) }
is the so-called right polarity and
•
/
: SPP → MBT I defined such that for every P ∈ 2SPP ,
P / = { i ∈ MBTI | p(P ) ⇒ i(i) }
is the so-called left polarity of the ordered pair ( . , / ).
Spelled out, (1) the result of applying the mapping i to a set I of MBTIs i as
defined in Definition 4 is the conjunction of the results of applying i to each
one of these i as defined in Table 2; (2) the result of applying the mapping p
to a set P of SPPs p as defined in Definition 4 is the disjunction of the results
of applying p to each one of these p, which simply is the conjunction of all
signed factors in p taken each one as an atomic proposition; (3) the result of
applying the mapping . to a set I of MBTIs is the set of all those SPPs p whose
mapping under p implies the mapping of I under i; (4) the result of applying
the mapping / to a set P of SPPs is the set of all those MBTIs i whose mapping
under i is implied by the mapping of P under p. Thus from a computer science
perspective [2, Section 7.35], MBTIs are specifications of SPPs and SPPs are
implementations or refinements of MBTIs. The Galois-connection then connects
correct implementations to their respective specification by stipulating
that a
V
correct
implementation
imply
its
specification.
By
convention,
∅
=
>
and
W
∅ = ⊥ , that is, the conjunction over the empty set ∅ is tautological truth > ,
and the disjunction over ∅ is tautological falsehood ⊥ , respectively.
Note that an example of an SPP that maps to the empty set under / happens
to be the Szondi norm profile mentioned before, because its mapping under p
p(((h, +), (s, +), (e, −), (hy, −), (k, −), (p, −), (d, +), (m, +))) =
h+ ∧ s+ ∧ e− ∧ hy− ∧ k− ∧ p− ∧ d+ ∧ m+ ,
does not contain any (individually) dominant factor (all factors are simply either positive or negative, and thus are without quanta). So it does not imply
the mapping of any MBTI under i, which requires a dominant factor, as we
have alluded to in the introduction with the dominance flag, and are going to
explain now by spelling out the translation of MBTIs to SPPs, see Table 2.
There, dominance is indicated by a quantum subscript of the translation i. As
can be seen, J and P merely act as flag values, which indicate which one of the
judgment or perception faculties is the dominant faculty for dealing with the
outer world, (that is, the observably dominant faculty) and this as a function of
extro-/introversion. The rule is that extroverts do show their dominant faculty
for dealing with the outer world, whereas introverts do not [8, Page 14]. So with
judging (J) introverts (I), a perceptive faculty shows up as dominant, with perceiving (P) introverts (I), a judging faculty, with perceiving (P) extroverts (E), a
8
perceptive faculty, and with judging (J) extroverts (E), a judging faculty. As can
be seen in Table 2, our interpretation of extroversion is a positive tendency of
Szondi’s factor hy (immoral behaviour: being seen, showing off). Whereas our
interpretation of introversion is a negative tendency thereof (moral behaviour:
seeing, hiding). Our readers might want to stipulate additional constraints for
their own interpretation, but must take care not to create conflicts. Consistency
is a necessary condition for the faithfulness of the translation i ! (It is also one
for the faithfulness of the translation p, but there it is obvious.)
Fact 1 (Consistency of the translation i).
1. For every b ∈ {E, I} and b0 ∈ MB \ {E, I}, i(b) ∧ i(b0 ) 6≡ ⊥ .
2. For every b ∈ {F, T},
• i(b) ∧ i! (N ) 6≡ ⊥ and i(b) ∧ i! (S) 6≡ ⊥ ;
• i! (b) ∧ i(N ) 6≡ ⊥ and i! (b) ∧ i(S) 6≡ ⊥ .
Proof. By inspection of Table 2.
Our interpretations of the remaining faculties follow a simple generating pattern:
• for a non-dominant positive factor f , the pattern is f + ∨ f ± ∨ f ±! ;
• for a non-dominant negative factor f , it is f − ∨ f ± ∨ f ±! ;
• for a dominant positive factor f , it is f +! ∨ f +!! ∨ f +!!! ∨ f ±! ;
• for a dominant negative factor f , it is f −! ∨ f −!! ∨ f −!!! ∨ f ±! .
As can be seen in Table 2, our interpretation of feeling is the conjunction of
personal warmth (h+) and empathy (p−), which [7] stipulate as the characteristic properties of this faculty. Our interpretation of thinking is simply its
corresponding factor (k−) in Szondi’s system. Our interpretation of the two
perceptive faculties contains as a conjunct the factor (k+), which corresponds
to perception in Szondi’s system. Our interpretation of intuition then further
contains its corresponding factor (p+) in Szondi’s system. Finally, our interpretation of sensing further contains the disjunction of all those factors that
correspond to the human senses in Szondi’s system, namely: touching (h+),
hearing (e−), seeing (hy−), smelling (d+), and tasting (m+).
Our readers might be interested in comparing our interpreting mapping of
MBTIs with D.W. Keirsey’s [4]: for him, ISTJ maps to the Inspector, ISFJ to
the Protector, INFJ to the Counselor, INTJ to the Mastermind, ISTP to the
Crafter, ISFP to the Composer, INFP to the Healer, INTP to the Architect,
ESTP to the Promoter, ESFP to the Performer, ENFP to the Champion, ENTP
to the Inventor, ESTJ to the Supervisor, ESFJ to the Provider, ENFJ to the
Teacher, and ENTJ to the Fieldmarshal.
We now prove in two intermediate steps that the ordered pair ( . , / ) is indeed a Galois-connection. The first step is the following announced fact, from
9
which the second step, Lemma 1, follows, from which in turn the desired result,
Theorem 1, then follows—easily. As announced, all that our readers need to
check on their own analog of our mapping i is that it has the property stated
as Fact 2.1. Their own Galois-connection is then automatically induced.
Fact 2 (Some facts about i and p).
1. if I ⊆ I 0 then i(I 0 ) ⇒ i(I)
2. if P ⊆ P 0 then p(P ) ⇒ p(P 0 )
3. The functions i and p are injective but not surjective.
Proof. By inspection of Definition 4 and Table 2.
We need Fact 2.1 and 2.2 but not Fact 2.3 in the following development. Therefor, note the two macro-definitions ./ := . ◦ / and /. := / ◦ . with ◦ being
function composition, as usual (from right to left, as usual too).
Lemma 1 (Some useful properties of
1. if I ⊆ I 0 then I 0. ⊆ I .
2. if P ⊆ P 0 then P 0/ ⊆ P /
3. P ⊆ (P / ).
4. I ⊆ (I . )/
.
and / ).
( . is antitone)
( / is antitone)
( ./ is inflationary)
( /. is inflationary)
Proof. For (1), let I, I 0 ∈ 2MBTI and suppose that I ⊆ I 0 . Hence i(I 0 ) ⇒ i(I) by
Fact 2.1. Further suppose that p ∈ I 0. . By definition, I 0. = { p ∈ SPP | p(p) ⇒
i(I 0 ) }. Hence p(p) ⇒ i(I 0 ). Hence p(p) ⇒ i(I) by transitivity. By definition,
I . = { p ∈ SPP | p(p) ⇒ i(I) }. Hence p ∈ I . . Thus I 0. ⊆ I . .
For (2), let P, P 0 ∈ 2SPP and suppose that P ⊆ P 0 . Hence p(P ) ⇒ p(P 0 )
by Fact 2.2. Further suppose that i ∈ P 0/ . By definition, P 0/ = { i ∈ MBTI |
p(P 0 ) ⇒ i(i) }. Hence p(P 0 ) ⇒ i(i). Hence p(P ) ⇒ i(i) by transitivity. By
definition, P / = { i ∈ MBTI | p(P ) ⇒ i(i) }. Hence i ∈ P / . Thus P 0/ ⊆ P / .
For (3), consider:
1.
p∈P
2.
{p} ⊆ P
3.
p(p) ⇒ p(P )
hypothesis
1
2, Fact 2.2
4.
p(p) is true
hypothesis
5.
p(P ) is true
3, 4
6.
φ ∈ { i(i) | p(P ) ⇒ i(i) }
7.
there is i s.t. φ = i(i) and p(P ) ⇒ i(i)
hypothesis
8.
φ = i(i) and p(P ) ⇒ i(i)
9.
p(P ) ⇒ i(i)
6
hypothesis
8
10
10.
i(i) is true
11.
φ = i(i)
12.
φ is true
13.
14.
15.
16.
17.
18.
5, 9
8
10, 11
φ is true
7, 8–12
for every φ ∈ { i(i) | p(P ) ⇒ i(i) }, φ is true
V
{ i(i) | p(P ) ⇒ i(i) } is true
V
{ i(i) | i ∈ { i ∈ MBTI | p(P ) ⇒ i(i) }} is true
V
{ i(i) | i ∈ P / } is true
/
i(P ) is true
14
15
16
17
/
19.
p(p) ⇒ i(P )
20.
p ∈ { p ∈ SPP | p(p) ⇒ i(P / ) }
21.
6–13
4–18
19
/ .
p ∈ (P )
20
/ .
1–21.
22. P ⊆ (P )
For (4), consider:
1.
i∈I
2.
{i} ⊆ I
3.
i(I) ⇒ i(i)
hypothesis
1
2, Fact 2.1
7.
p(I . ) is true
W
{ p(p) | p ∈ I . } is true
W
{ p(p) | p ∈ { p ∈ SPP | p(p) ⇒ i(I) } } is true
W
{ p(p) | p(p) ⇒ i(I) } is true
8.
there is p s.t. p(p) ⇒ i(I) and p(p) is true
4.
5.
6.
p(p) ⇒ i(I) and p(p) is true
9.
hypothesis
4
5
6
7
hypothesis
10.
i(I) is true
9
11.
i(i) is true
3, 10
12.
13.
14.
15.
i(i) is true
8, 9–11
.
p(I ) ⇒ i(i)
4–12
.
i ∈ { i ∈ MBTI | p(I ) ⇒ i(i) }
. /
i ∈ (I )
13
14
16. I ⊆ (I . )/
1–15.
We are ready for making the final step.
11
Theorem 1 (The Galois-connection property of ( . , / )). The ordered pair ( . , / )
is an antitone or order-reversing Galois-connection between MBT I and SPP.
That is, for every I ∈ 2MBTI and P ∈ 2SPP ,
P ⊆ I . if and only if I ⊆ P / .
Proof. Let I ∈ 2MBTI and P ∈ 2SPP . Suppose that P ⊆ I . . Hence (I . )/ ⊆ P /
by Lemma 1.2. Further, I ⊆ (I . )/ by Lemma 1.4. Hence I ⊆ P / by transitivity.
Conversely suppose that I ⊆ P / . Hence (P / ). ⊆ I . by Lemma 1.1. Further,
P ⊆ (P / ). by Lemma 1.3. Hence P ⊆ I . .
Thus from a computer science perspective [2, Section 7.35], smaller (larger) sets
of MBTIs and thus less (more) restrictive specifications correspond to larger
(smaller) sets of SPPs and thus more (fewer) possible implementations.
Note that Galois-connections are connected to residuated mappings [1]. Further, natural notions of equivalence on MBT I and SPP are given by the kernels
of . and / , respectively, which are, by definition:
I ≡ I0
if and only if I . = I 0. ;
P ≡ P0
if and only if P / = P 0/ .
Proposition 1 (The efficient computability of ( . , / )).
1. Given I ∈ 2MBTI , I . is efficiently computable.
2. Given P ∈ 2SPP , P / is efficiently computable.
Proof. Even a relatively brute-force approach is feasible with today’s laptop
computers, which have a (high-speed) random access memory of several giga
(109 ) bytes and a processor power of several giga instructions per second. Moreover in psychological practice, the sets I and P will usually be singletons of either
only one type indicator or only one personality profile, respectively.
• Given I ∈ 2MBTI (thus 0 ≤ |I| ≤ 16), we precompute the list of all
SPPs, which contains 168 entries, once and for all I, and store the list
on a computer hard disk and then load it into the (fast) random access
memory of the computer. Then we model-check the formula i(I), which
can also be precomputed, against each entry p in the list. That is, we
check whether a given model p satisfies (makes true) the given i(I), and
collect up into the result set all those p that do satisfy i(I). It is wellknown that model-checking a propositional formula, such as i(I), takes
only a polynomial number of computation steps in the size of the formula,
and thus only a polynomial number in the size of I (being at most 16).
Of course, this computation can be done once and for all of the 216 possible
sets I, and the results stored on a hard disk for faster, later look-up.
• Given P ∈ 2SPP (thus 0 ≤ |P | ≤ 168 ), we model-check each p in P against
the (pre-computable) mapping i(i) of each i of the 16 MBTIs, and collect
up into the result set all those i whose mapping i(i) satisfies p.
12
Of course, optimisations of the computation procedure given in the previous
proof are possible, but we consider them as not sufficiently interesting implementation details.
3
Conclusion
We have proposed a computable Galois-connection between Myers-Briggs Type
Indicators and Szondi’s personality profiles as promised in the abstract. In
addition, we have proposed a simple methodology for generating other such
Galois-connections, including Galois-connections not only between this pair of
spaces of personality-test result values but also between other such pairs.
Acknowledgements I thank Danilo Diedrichs for proof-reading this article.
The LATEX-package TikZ was helpful for graph drawing.
References
[1] T.S. Blyth. Lattices and Ordered Algebraic Structures. Springer, 2005.
[2] B.A. Davey and H.A. Priestley. Introduction to Lattices and Order. Cambridge University Press, 2nd edition, 1990 (2002).
[3] C.G. Jung. Psychological Types, volume 6 of The Collected Works of C.G.
Jung. Princeton University Press, 1971.
[4] D. Keirsey. Please Understand Me II. Prometheus Nemesis Book Company,
1984 (1998).
[5] R. Kenmo. Let the Personality Bloom. Humankonsult, 2009.
[6] S. Kramer. www.szondi-test.ch, 2014. forthcoming.
[7] I. Briggs Myers and M.H. McCaulley. Manual: A Guide to the Development and Use of the Myers-Briggs Type Indicator. Consulting Psychologists
Press, 2nd edition, 1985.
[8] I. Briggs Myers and P.B. Myers. Gifts Differing: Understanding Personality
Type. Consulting Psychologists Press, 1980 (1995).
[9] L. Szondi. Lehrbuch der Experimentellen Triebdiagnostik, volume I: TextBand. Hans Huber, 3rd edition, 1972.
[10] L. Szondi. Ich-Analyse: Die Grundlage zur Vereinigung der Tiefenpsychologie. Hans Huber, 1999. English translation: https://sites.google.
com/site/ajohnstontranslationsofszondi/.
13
[11] L. Szondi.
Schicksalsanalyse: Wahl in Liebe, Freundschaft, Beruf,
Krankheit und Tod. Schwabe, 4th edition, 2004.
14
| 5cs.CE
|
Published as a conference paper at ICLR 2018
C OMPOSITIONAL ATTENTION N ETWORKS
FOR M ACHINE R EASONING
Christopher D. Manning
Department of Computer Science
Stanford University
[email protected]
arXiv:1803.03067v1 [cs.AI] 8 Mar 2018
Drew A. Hudson
Department of Computer Science
Stanford University
[email protected]
A BSTRACT
We present the MAC network, a novel fully differentiable neural network architecture, designed to facilitate explicit and expressive reasoning. Drawing inspiration
from first principles of computer organization, MAC moves away from monolithic
black-box neural architectures towards a design that encourages both transparency
and versatility. The model approaches problems by decomposing them into a
series of attention-based reasoning steps, each performed by a novel recurrent
Memory, Attention, and Composition (MAC) cell that maintains a separation between control and memory. By stringing the cells together and imposing structural
constraints that regulate their interaction, MAC effectively learns to perform iterative reasoning processes that are directly inferred from the data in an end-to-end
approach. We demonstrate the model’s strength, robustness and interpretability on
the challenging CLEVR dataset for visual reasoning, achieving a new state-of-theart 98.9% accuracy, halving the error rate of the previous best model. More importantly, we show that the model is computationally-efficient and data-efficient,
in particular requiring 5x less data than existing models to achieve strong results.
1
I NTRODUCTION
Reasoning, the ability to manipulate previously acquired knowledge to form novel inferences or
answer new questions, is one of the fundamental building blocks of the intelligent mind. As we
seek to advance neural networks beyond their current great success with sensory perception towards
tasks that require higher-order and more deliberate thinking, conferring them with the ability to
move from facts to conclusions is thus of crucial importance. To this end, we consider here how best
to design a neural network to perform the structured and iterative reasoning necessary for complex
problem solving.
Concretely, we develop a novel model that we apply to the CLEVR
task (Johnson et al., 2017a) of visual question answering (VQA).
VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal
task that requires responding to natural language questions about images. However, Agrawal et al. (2016) show how the first generation
of successful VQA models tends to acquire only superficial comprehension of both the image and the question, exploiting dataset biases
rather than capturing a sound perception and reasoning process that
would lead to the correct answer (Sturm, 2014). CLEVR was created to address this problem. As illustrated in figure 1, the dataset
features unbiased, highly compositional questions that require an array of challenging reasoning skills, such as transitive and logical relations, counting and comparisons, without allowing any shortcuts
around such reasoning.
Q: Do the block in front of the
tiny yellow cylinder and the
tiny thing that is to the right
of the large green shiny object
have the same color? A: No
Figure 1: A CLEVR example.
Color added for illustration.
However, deep learning approaches often struggle to perform well on tasks with a compositional and
structured nature (Garnelo et al., 2016; Lake et al., 2017). Most neural networks are essentially very
large correlation engines that will hone in on any statistical, potentially spurious pattern that allows
them to model the observed data more accurately. The depth, size and statistical nature that allows
1
Published as a conference paper at ICLR 2018
Figure 2: Model Overview. The MAC network consists of an input unit, a core recurrent network and an output unit. (1) The input unit transforms the raw image and question into distributed vector representations.
(2) The core recurrent network reasons sequentially over the question by decomposing it into a series of operations (control)
that retrieve information from the image
(knowledge base) and aggregate the results
into a recurrent memory. (3) The output
classifier computes the final answer using
the question and the final memory state.
them to cope with noisy and diverse data often limits their interpretability and hinders their capacity
to perform explicit and sound inference procedures that are vital for problem solving tasks. To mitigate this issue, some recent approaches adopt symbolic structures, resembling the expression trees of
programming languages, that compose neural modules from a fixed predefined collection (Andreas
et al., 2016a; Johnson et al., 2017b). However, they consequently rely on externally provided structured representations and functional programs, brittle handcrafted parsers or expert demonstrations,
and require relatively complex multi-stage reinforcement learning training schemes. The rigidity
of these models’ structure and the use of an inventory of specialized operation-specific modules
ultimately undermines their robustness and generalization capacities.
Seeking a balance between the versatility and robustness of end-to-end neural approaches on the
one hand and the need to support more explicit and structured reasoning on the other, we propose
the MAC network, a novel fully differentiable architecture for reasoning tasks. Our model performs
structured and explicit reasoning by sequencing a new recurrent Memory, Attention and Composition (MAC) cell. The MAC cell was deliberately designed to capture the inner workings of an
elementary, yet general-purpose reasoning step, drawing inspiration from the design principles of
computer architectures. The cell explicitly separates out memory from control, both represented
recurrently, and consists of three operational units that work in tandem to perform a reasoning step:
the control unit updates the control state to attend at each iteration to some aspect of a given question; the read unit extracts information out of a knowledge base, guided by the control and memory
states; and the write unit integrates the retrieved information into the memory state, iteratively computing the answer. This universal design of the MAC cell serves as a structural prior that encourages
the network to solve problems by decomposing them into a sequence of attention-based reasoning
operations that are directly inferred from the data, without resorting to any strong supervision. With
self-attention connections between the cells, the MAC network is capable of representing arbitrarily complex acyclic reasoning graphs in a soft manner, while still featuring a physically sequential
structure and end-to-end differentiabillity, amenable to training simply by backpropagation.
We demonstrate the model’s quantitative and qualitative performance on the CLEVR task and its
associated datasets. The model achieves state-of-the-art accuracy across a variety of reasoning tasks
and settings, both for the primary dataset as well as the more difficult human-authored questions.
Notably, it performs particularly well on questions that involve counting and aggregation skills,
which tend to be remarkably challenging for other VQA models (Santoro et al., 2017; Hu et al.,
2017; Johnson et al., 2017b). Moreover, we show that the MAC network learns rapidly and generalizes effectively from an order of magnitude less data than other approaches. Finally, extensive
ablation studies and error analysis demonstrate MAC’s robustness, versatility and generalization
capacity. These results highlight the significance and value of imposing strong structural priors
to guide the network towards compositional reasoning. Realizing the vision proposed by Bottou
(2014), enriching the model with structures that encourage it to explicitly perform a chain of operations that build upon each other allows MAC to develop reasoning skills from the ground up.
Although each cell’s functionality has only a limited range of possible continuous behaviors, geared
to perform a simple reasoning operation, when chained together in a MAC network, the whole system becomes expressive and powerful.
2
Published as a conference paper at ICLR 2018
Figure 3: The MAC cell architecture. The MAC recurrent cell consists of a control unit, read unit, and
write unit, that operate over dual control and memory hidden states. The control unit successively attends to
different parts of the task description (question), updating the control state to represent at each timestep the
reasoning operation the cell intends to perform. The read unit extracts information out of a knowledge base
(image), guided by the control state. The write unit integrates the retrieved information into the memory state,
yielding the new intermediate result that follows from applying the current reasoning operation.
2
T HE MAC N ETWORK
A MAC network is an end-to-end diffrentiable architecture primed to perform an explicit multi-step
reasoning process, by stringing together p recurrent MAC cells, each responsible for performing one
reasoning step. Given a knowledge base K (for VQA, an image) and a task description q (for VQA,
a question), the model infers a decomposition into a series of p reasoning operations that interact
with the knowledge base, iteratively aggregating and manipulating information to perform the task
at hand. It consists of three components: (1) an input unit, (2) the core recurrent network, composed
out of p MAC cells, and (3) an output unit, all described below.
2.1
T HE I NPUT U NIT
The input unit transforms the raw inputs given to the model into distributed vector representations.
Naturally, this unit is tied to the specifics of the task we seek to perform. For the particular case of
VQA, it receives a question and an image and processes each of them respectively:
The question string, of length S, is converted into a sequence of learned word embeddings that is
further processed by a d-dimensional biLSTM yielding: (1) contextual words: a series of output
states cw1 , . . . , cwS that represent
each word in the context of the question, and (2) the question
−−, −
−→
representation: q = ←
cw
1 cw S , the concatenation of the final hidden states from the backward and
forward LSTM passes. Subsequently, for each i = 1, . . . , p, the question q is transformed through
a learned linear transformation into a position-aware vector qi = Wid×2d q + bdi , representing the
aspects of the question that are relevant to the ith reasoning step.
The image is first processed by a fixed feature extractor pre-trained on ImageNet (Russakovsky
et al., 2015) that outputs conv4 features from ResNet101 (He et al., 2016), matching prior work
for CLEVR (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017). The resulting tensor is then
passed through two CNN layers with d output channels to obtain a final image representation, the
d
knowledge base K H×W ×d = {kh,w
|H,W
h,w=1,1 }, where H = W = 14 are the height and width of the
processed image, corresponding to each of its regions.
2.2
T HE MAC CELL
The MAC cell is a recurrent cell designed to capture the notion of an atomic and universal reasoning
operation and formulate its mechanics. For each step i = 1, . . . , p in the reasoning process, the ith
cell maintains dual hidden states: control ci and memory mi , of dimension d, initialized to learned
parameters m0 and c0 , respectively.
The control ci represents the reasoning operation the cell should accomplish in the ith step, selectively focusing on some aspect of the question. Concretely, it is represented by a soft attention-based
weighted average of the question words cws ; s = 1, . . . , S.
3
Published as a conference paper at ICLR 2018
cq i = W d×2d [ci−1 , qi ] + bd
ca i,s = W
1×d
(cq i
cw s ) + b
cv i,s = softmax(cai,s )
ci =
S
X
cv i,s · cw s
(c1)
1
(c2.1)
(c2.2)
(c2.3)
s=1
Figure 4: The Control Unit (CU) architecture. The control unit attends at each iteration to some part of the
question, by applying soft attention over the question words, and updates the control state accordingly. The
unit’s inputs and outputs are in bold. See section 2.2.1 for details.
The memory mi holds the intermediate result obtained from the reasoning process up to the ith
step, computed recurrently by integrating the preceding hidden state mi−1 with new information
ri retrieved from the image in order to perform the ith reasoning operation ci . Analogously, ri is a
weighted average over its regions {kh,w |H,W
h,w=1,1 }.
Building on the design principles of computer organization, the MAC cell consists of three operational units: control unit CU, read unit RU and write unit WU, that work together to accomplish
tasks by performing an iterative reasoning process: The control unit identifies a series of operations,
represented by a recurrent control state; the read unit extracts relevant information from a given
knowledge base to perform each operation, and the write unit iteratively integrates the information
into the cell’s memory state, producing a new intermediate result.
Through their operation, the three units impose together an interface that regulates the interaction
between the control and memory states. Specifically, the control state, which is a function of the
question, guides the integration of content from the image into the memory state only through indirect means: soft-attention maps and sigmoidal gating mechanisms. Consequently, the interaction
between these two modalities – visual and textual, or knowledge base and query – is mediated
through probability distributions only. This stands in stark contrast to common approaches that fuse
the question and image together into the same vector space through linear combinations, multiplication, or concatenation. As we will see in section 3, maintaining a strict separation between the
representational spaces of question and image, which can interact only through interpretable discrete
distributions, greatly enhances the generalizability of the network and improves its transparency.
In the following, we describe the cell’s three components: control, read and write units, and detail
their formal specification. Unless otherwise stated, all the vectors are of dimension d.
2.2.1
T HE C ONTROL U NIT
The control unit (see figure 4) determines the reasoning operation that should be performed at each
step i, attending to some part of the question and updating the control state ci accordingly. It
receives the contextual question words cw 1 , . . . , cw S , the question position-aware representation
qi , and the control state from the preceding step ci−1 and consists of two stages:
1. First, we combine qi and ci−1 through a linear transformation into cq i , taking into account
both the overall question representation qi , biased towards the ith reasoning step, as well
as the preceding reasoning operation ci−1 . This allows the cell to base its decision for the
ith reasoning operation ci on the previously performed operation ci−1 .
2. Subsequently, we cast cq i onto the space of the question words. Specifically, this is
achieved by measuring the similarity between cq i and each question word cw s and passing the result through a softmax layer, yielding an attention distribution over the question
words cw 1 , . . . , cw S . Finally, we sum the words according to the distribution to produce
the reasoning operation ci , represented in terms of the question words.
The casting of cq i onto question words serves as a form of regularization that restricts the space of
the valid reasoning operations by anchoring them back in the original question words, and due to
the use of soft attention may also improve the cell transparency, since we can interpret the control
state content and the cell’s consequent behavior based on the words it attends to.
4
Published as a conference paper at ICLR 2018
d×d
Ii,h,w =[Wm
mi−1 + bdm ]
(r1)
[Wkd×d kh,w + bdk ]
0
Ii,h,w
= W d×2d [Ii,h,w , kh,w ] + bd
ra i,h,w = W
d×d
(ci
0
Ii,h,w
)
rv i,h,w = softmax(ra i,h,w )
+b
d
(r2)
(r3.1)
(r3.2)
H,W
X
ri =
rv i,h,w · kh,w
(r3.3)
h,w=1,1
Figure 5: The Read Unit (RU) architecture. The read unit retrieves information from the knowledge base that
is necessary for performing the current reasoning operation (control) and potentially related to previously obtained intermediate results (memory). It extracts the information by performing a two-stages attention process
over the knowledge base elements. See section 2.2.2 for details.
2.2.2
T HE R EAD U NIT
For the ith step, the read unit (see figure 5) inspects the knowledge base (the image) and retrieves
the information ri that is required for performing the ith reasoning operation ci . The content’s
relevance is measured by an attention distribution rv i that assigns a probability to each element
in the knowledge base kdh,w , taking into account the current reasoning operation ci and the prior
memory mi−1 , the intermediate result produced by the preceding reasoning step. The attention
distribution is computed in several stages:
1. First, we compute the direct interaction between the knowledge-base element kh,w and the
memory mi−1 , resulting in Ii,h,w . This term measures the relevance of the element to
the preceding intermediate result, allowing the model to perform transitive reasoning by
considering content that now seems important in light of information obtained from the
prior computation step.
2. Then, we concatenate the element kh,w to Ii,h,w and pass the result through a linear trans0
formation, yielding Ii,h,w
. This allows us to also consider new information that is not
directly related to the prior intermediate result, as sometimes a cogent reasoning process
has to combine together independent facts to arrive at the answer (e.g., for a logical OR
operation, set union and counting).
3. Finally, aiming to retrieve information that is relevant for the reasoning operation ci , we
measure its similarity to each of the interactions Ii,h,w and pass the result through a softmax
layer. This produces an attention distribution over the knowledge base elements, which we
then use to compute a weighted average over them – ri .
To give an example of the read unit operation, consider the question in figure 6, which refers to the purple cylinder in the image.
Initially, no cue is provided to the model to attend to the cylinder, since no direct mention of it is given in the question. Instead,
the model approaches the question in steps: in the first iteration
it attends to the “tiny blue block”, updating m1 accordingly to
the visual representation of the block. At the following step, the
control unit realizes it should now look for “the sphere in front”
of the block, storing that in c2 . Then, when considering both m1
and c2 , the read unit realizes it should look for “the sphere in
front” (c2 ) of the blue block (stored in m1 ), thus finding the cyan
sphere and updating m2 . Finally, a similar process repeats in the
next iteration, allowing the model to traverse from the cyan ball to
the final objective – the purple cylinder, and answer the question
correctly.
Figure 6: Attention maps produced
by a MAC network of length 3.
5
Published as a conference paper at ICLR 2018
= W d×2d [ri , mi−1 ] + bd
mprev
i
sa ij = softmax W 1×d (ci cj ) + b1
msa
i =
i−1
X
sa ij · mj
(w1)
(w2.1)
(w2.2)
j=1
d×d prev
+ bd
m0i = Wsd×d msa
mi
i + Wp
(w2.3)
c0i = W 1×d ci + b1
mi = σ c0i mi−1 + 1 − σ c0i m0i
(w3.1)
(w3.2)
Figure 7: The Write Unit (WU) architecture. The write unit intergreates the information retrieved from the
knowledge base into the recurrent memory state, producing a new intermediate result mi that corresponds to
the reasoning operation ci . See section 2.2.3 for details.
2.2.3
T HE W RITE U NIT
The write unit (see figure 7) is responsible for computing the ith intermediate result of the reasoning
process and storing it in the memory state mi . Specifically, it integrates the information retrieved
from the read unit ri with the preceding intermediate result mi−1 , guided by the ith reasoning operation ci . The integration proceeds in three steps, the first mandatory while the others are optional1 :
1. First, we combine the new information ri with the prior intermediate result mi−1 by a
linear transformation, resulting in mprev
.
i
2. Self-Attention. To support non-sequential reasoning processes, such as trees or graphs, we
allow each cell to consider all previous intermediate results, rather than just the preceding
one mi−1 : We compute the similarity between the ith operation ci and the previous ones
c1 , . . . , ci−1 and use it to derive an attention distribution over the prior reasoning steps sa i,j
for j = 0, . . . , i − 1. The distribution represents the relevance of each previous step j to
the current one i, and is used to compute a weighted average of the memory states, yielding
prev
msa
to produce m0i . Note that while we compute the
i , which is then combined with mi
attention based on the control states, we use it to average over the memory states, in a way
that resembles Key-Value Memory Networks (Miller et al., 2016).
3. Memory Gate. Not all questions are equally complex – some are simpler while others
are more difficult. To allow the model to dynamically adjust the reasoning process length
to the given question, we add a sigmoidal gate over the memory state that interpolates
between the previous memory state mi−1 and the new candidate m0i , conditioned on the
reasoning operation ci . The gate allows the cell to skip a reasoning step if necessary,
passing the previous memory value further along the network, dynamically reducing the
effective length of the reasoning process as demanded by the question.
2.3
T HE O UTPUT U NIT
The output unit predicts the final answer to the question based
on the question representation q and the final memory mp ,
which represents the final intermediate result of the reasoning process, holding relevant information from the knowledge base.2 For CLEVR, where there is a fixed set of possible answers, the unit processes the concatenation of q and
mp through a 2-layers fully-connected softmax classifier that
produces a distribution over the candidate answers.
1
Figure 8: The output unit. A classifier that predicts an answer based on the
question and the final memory state.
Both self-attention connections as well as the memory gate serve to reduce long-term dependencies. However, note that for the CLEVR dataset we were able to maintain almost the same performance with the first step
only, and so we propose the second and third ones as optional extensions of the write unit, and explore their
impact on the model’s performance in section 3.3.
2
Note that some questions refer to important aspects that do not have counterpart information in the knowledge base, and thus considering both the question and the memory is critical to answer them.
6
Published as a conference paper at ICLR 2018
Table 1: CLEVR and CLEVR-Humans Accuracy by baseline methods, previous methods, and our method
(MAC). For CLEVR-Humans, we show results before and after fine-tuning. (*) denotes use of extra supervisory
information through program labels. († ) denotes use of data augmentation. (‡ ) denotes training from raw pixels.
Model
Human (Johnson et al., 2017b)
Q-type baseline (Johnson et al., 2017b)
LSTM (Johnson et al., 2017b)
CNN+LSTM (Johnson et al., 2017b)
CNN+LSTM+SA+MLP (Johnson et al., 2017a)
N2NMN* (Hu et al., 2017)
PG+EE (9K prog.)* (Johnson et al., 2017b)
PG+EE (18K prog.)* (Johnson et al., 2017b)
PG+EE (700K prog.)* (Johnson et al., 2017b)
CNN+LSTM+RN†‡ (Santoro et al., 2017)
CNN+GRU+FiLM (Perez et al., 2017)
CNN+GRU+FiLM‡ (Perez et al., 2017)
MAC
3
CLEVR
Overall
92.6
41.8
46.8
52.3
73.2
83.7
88.6
95.4
96.9
95.5
97.7
97.6
Count
Exist
96.6
50.2
61.1
65.2
77.9
85.7
89.7
97.3
97.1
97.8
99.1
99.3
Compare
Numbers
86.5
51.0
69.8
67.1
75.1
84.9
79.1
96.5
98.7
93.6
96.8
93.4
Query
Attribute
95.0
36.0
36.8
49.3
80.9
90.0
92.6
97.4
98.1
97.9
99.1
99.3
Compare
Attribute
96.0
51.3
51.8
53.0
70.8
88.7
96.0
98.0
98.9
97.1
99.1
99.3
Humans
before FT
27.5
37.7
50.4
54.0
56.6
-
Humans
after FT
36.5
43.2
57.6
66.6
75.9
-
86.7
34.6
41.7
43.7
59.7
68.5
79.7
90.1
92.7
90.1
94.3
94.3
98.9
97.1
99.5
99.1
99.5
99.5
57.4
81.5
E XPERIMENTS
We evaluate our model on the recent CLEVR task for visual reasoning (Johnson et al., 2017a). The
dataset consists of rendered images featuring 3D-objects of various shapes, materials, colors and
sizes, coupled with compositional multi-step questions that measure performance on an array of
challenging reasoning skills such as following transitive relations, counting objects and comparing
their properties. Each question is also associated with a tree-structured functional program, specifying the reasoning operations that should be performed to compute the answer.
In the following experiments, our model’s training is cast as a supervised classification problem
to minimize the cross-entropy loss of the predicted candidate answer out of the 28 possibilities.
The model is set to hidden state size of d = 512 and, unless otherwise stated, length of p = 12
MAC cells.3 While some prior work uses the functional programs associated with each question as
additional supervisory information at training time (see table 1), note that we intentionally do not
use these structured representations to train our model, aiming to infer coherent reasoning strategies
directly from the question and answer pairs in an end-to-end approach.
We first perform experiments on the primary 700k dataset. As shown in table 1, our model outperforms all prior work both in overall accuracy, as well as in each of the categories of specific
reasoning skills. In particular, for the overall performance, we achieve 98.94% accuracy, more than
halving the error rate of the prior best model, FiLM (Perez et al., 2017).
Counting and Numerical Comparison. In particular, our performance on questions about counting
and numerical comparisons is significantly higher than existing models, which consistently struggle
on these question types. Again, we nearly halve the corresponding error rate. These are significant
results, as counting and aggregations are known to be particularly challenging in the area of VQA
(Chattopadhyay et al., 2017). In contrast to CNNs, using attention enhances our model’s ability to
perform reasoning operations such as counting that pertain to the global aggregation of information
across different regions of the image.
3.1
C OMPUTATIONAL AND DATA E FFICIENCY
We examine the learning curves of MAC and compare them to previous models4,5 : specifically,
FiLM (Perez et al., 2017), the strongly-supervised PG+EE (Johnson et al., 2017b), and stackedattention networks (Johnson et al., 2017b; Yang et al., 2016). As shown in figure 9, our model
learns significantly faster than other approaches. While we do not have learning curves for the
3
We initialize the word embeddings of our model to random vectors using uniform distribution. In an
earlier version of this work, we have used pretrained GloVe vectors, but found that they did not improve the
performance for CLEVR and led to only a marginal improvement for CLEVR-Humans.
4
For previous models, we use the author’s original publicly available implementations. All the models were
trained with an equal batch size of 64 (as in the original implementations) and using the same hardware – a
single Maxwell Titan X GPU per model.
5
To make sure the results are statistically significant, we run each model multiple (10) times, and plot the
averages and confidence intervals.
7
Published as a conference paper at ICLR 2018
Model Ablations
1
0.98
Accuracy (Val)
0.9
Accuracy (Val)
Model Variations
1
0.8
standard model
unshared weights
question vector as control
w/o word attention (c2)
w/o memory-control
0.7
0.6
0
5
10
Epoch
15
0.96
standard model
prediction w/o question (o1)
w/o direct KB elements (r2)
w/ self-attention (w2)
w/ gating (w3)
0.94
0.92
0.9
20
0
5
10
Epoch
15
20
Figure 9: From left to right: (1) Learning curve of MAC and alternative approaches (accuracy / epoch). (2)
Models’ performance as a function of the CLEVR subset size used for training, ranging from 1% to 100%.
(3),(4) Learning curves for ablated MAC variants. See section 3.3 for details.
recent Relation Network model, Santoro et al. (2017) report 1.4 million iterations (equivalent to 125
epochs) to achieve 95.5% accuracy, whereas our model achieves a comparable accuracy after only 3
epochs, yielding a 40x reduction in the length of the training process. Likewise, Perez et al. (2017)
report a training time of 4 days, equivalent to 80 epochs, to reach accuracy of 97.7%. In contrast, we
achieve higher accuracy in 6 epochs, 9.5 hours overall, leading to a 10x reduction in training time.
In order to study the ability of MAC to generalize from a smaller amount of data, we explore its
performance on subsets of CLEVR, sampled at random from the original 700k dataset. As shown
in figure 9, MAC outperforms the other models by a wide margin: For 50% of the data, equivalent
to 350k samples, other models obtain accuracies ranging between 70% and 93%, while our model
achieves 97.6%. The gap becomes larger as the dataset size reduces: for 25% of the data, equivalent
to 175k samples, the performance of other models is between 50% and 77%, while MAC maintains
a high 94.3% accuracy.
Finally, for just 10% of the data, amounting to 70k samples, our model is the only one to generalize well, with performance of 85.5% on average, whereas the other leading models fail, achieving
49.0%-54.9%. Note that, as pointed out by Johnson et al. (2017a), a simple baseline that predicts
the most frequent answer for each question type already achieves 42.1%, suggesting that answering only half of the questions correctly means that the other models barely learn to generalize from
the smaller subsets. These results demonstrate the robustness and generalization capacity of our
architecture and its key role as a structural prior guiding MAC to learn the intended reasoning skills.
3.2
CLEVR H UMANS AND E RROR A NALYSIS
We analyze our model’s performance on the CLEVR-Humans dataset
(Johnson et al., 2017b), consisting of natural language questions collected through crowdsourcing. As such, the dataset has a diverse vocabulary and linguistic variations, and it also demands more varied
reasoning skills. Since the training set is relatively small, comprising
18k samples, we use it to finetune a model pre-trained on the primary
CLEVR dataset, following prior work.
As shown in table 1, our model achieves state-of-the-art performance
on CLEVR-Humans both before and after fine-tuning. It surpasses
the next-best FiLM model by 5.6% percent, achieving 81.5%. The results substantiate the model’s robustness against linguistic variations
and noise as well as its ability to adapt to new and more diverse vocabulary and varied reasoning skills. The soft attention performed over
the question allows the model to focus on the words that are most
critical to answer the question while paying less attention to irrelevant linguistic variations. See figure 12, and figures 16 and 17 in the
appendix for examples.
In order to gain insight into the nature of the mistakes our model
makes, we perform an error analysis for the CLEVR and CLEVRHumans datasets (See figure 10). Overall, we see that most of the
errors in the CLEVR dataset either are off-by-one counting mistakes,
8
Figure 10: Error distribution for CLEVR and CLEVRHumans.
Published as a conference paper at ICLR 2018
Q: What is the shape of
the large item, mostly
occluded by the metallic
cube? A: sphere 3
Q: What color is the
object that is a different
size? A: purple 3
Q: What color ball is
close to the small purple
cylinder? A: gray 3
Q: What color block is
farthest front? A: purple
3
Q: Are any objects gold?
A: yes 3
Figure 12: CLEVR-Humans examples showing the model performs novel reasoning skills that do not appear
in CLEVR, including: obstructions, object uniqueness, relative distances, superlatives and new concepts.
or result from heavy object occlusions. For CLEVR-Humans, we observe many errors that involve
new reasoning skills that the model has not been trained for, such as ones that relate to physical
properties (such as stability and reflections), relative distances and amounts, commonalities and
uniqueness of objects, or negative questions. See appendix B for further details. Nevertheless, note
that the model does respond correctly to many of the questions that fall under these reasoning skills,
as illustrated in figures 12 and 16, and so we speculate that the errors the model makes stem from
the small size of the CLEVR-Human dataset.
A BLATIONS
To gain better insight into the relative contribution of the design choices
we made, we perform extensive ablation studies. See figure 9 and appendix C for accuracies and learning curves. The experiments demonstrate the robustness of the model to hyperparameter variations such as
network dimension and length, and shed light on the significance of various aspects and components of the model, as discussed below:
Network Length
1
0.99
Accuracy (Val)
3.3
0.98
0.97
0.96
0.95
2
8
0.94
3
12
4
16
5
20
6
Question Attention. The ablations show that using attention over the
Epoch
question words (see section 2.2.1) is highly effective in accelerating learnFigure 11: Model perforing and enhancing generalization capacity. Using the complete question mance as a function of the
q instead of the attention-based control state leads to a significant drop of network length.
18.5% in the overall accuracy. Likewise, using unconstrained recurrent
control states, without casting them back onto the question words space (step (3) in section 2.2.1)
leads to a 6x slowdown in the model convergence rate. These results illustrate the importance and
usefulness of decomposing the question to a series of “simple” operations, such that a single cell
is faced with learning the semantics of one or a few words at a time, rather than grasping all the
question at once. They provide evidence for the efficacy of using attention as a regularization
mechanism, by restricting the input and output spaces of each MAC cell.
0
5
10
15
20
Control and Memory. Maintaining separation between control and memory proves to be another
key property that contributes significantly to the model’s accuracy, learning rate and data efficiency.
We perform experiments for a variant of the MAC cell in which we maintain one hidden state that
plays both the roles of the control and memory, iteratively attending and integrating information
from both the question and the image. While this approach achieves a final accuracy of 93.75%, it
leads to a sharp drop in the convergence rate, as shown in figure 9, and a 20.2% reduction in the
final accuracy for a smaller 10% subset of CLEVR. The results make a strong case for our model’s
main design choice, namely, splitting the computation into two dual paths: one that decomposes the
linguistic information and another that reconstructs the corresponding visual information.
The design choices discussed above were found to be the most significant to the model’s overall
accuracy, convergence rate and generalization. Other design choices that were found beneficial include (1) predicting the final answer based on both the final memory state and the question (see
section 2.3), and (2) considering knowledge base elements directly (step (2) in section 2.2.2), resulting in 19.8% and 11.1% improvement for a 10% subset of CLEVR, respectively. Please refer to
appendix C for further discussion and results.
9
Published as a conference paper at ICLR 2018
Figure 14: Attention maps produced by MAC, providing some evidence for the ability of the model to perform
counting and summation of small numbers. Note how the first iterations focus on the key structural question
words “many” and “or” that inform the model of the required reasoning operation it has to perform.
3.4
I NTERPRETABILITY
To obtain better insight into the underlying reasoning processes MAC
learns to perform, we study visualizations of the attention distributions
produced by the model during its iterative computation, and provide
examples in figures 13, 14, 17, and 18. Examining the sequence of attention maps over the image and the question reveals several qualitative
patterns and properties that characterize MAC’s mode of operation.
First, we observe that both the linguistic and visual attentions of the
model are very focused on specific terms or regions in the image, and
commonly refer to concrete objects (“the shiny red cube” or the “metallic cylinder”) or question structural keywords (“or”, “and” or “how
many”). More importantly, the attetnion maps give evidence for the
ability of the model to capture the underlying semantic structure of the
question, traversing the correct transitive relations between the objects
it refers to. For instance, we see in figure 13 how the model explicitly
decomposes the question into the correct reasoning steps: first identifying the green ball, then focusing on the red cylinder that is located left Figure 13: Attention maps
of the ball, and finally attending to the yellow cylinder. In the second produced by MAC, showing
step, note how the model attends only to the relevant red cylinder and how it tracks transitive relanot to other red rubber things, correctly resolving the indirect reference tions between objects.
in the question. This shows strong evidence for the ability of the model
to perform transitive reasoning, integrating information from prior steps that allows it to focus only
on the relevant objects, even when they are not mentioned explicitly.
In figure 14, we further see how the model interprets a multi-step counting question, apparently
summing up the amounts of two referenced object groups to produce the correct overall count. These
observations suggest that the model infers and effectively performs complex reasoning processes in
a transparent manner.
4
C ONCLUSION
We have introduced the Memory, Attention and Composition (MAC) network, an end-to-end differentiable architecture for machine reasoning. The model solves problems by decomposing them into
a series of inferred reasoning steps that are performed successively to accomplish the task at hand.
It uses a novel recurrent MAC cell that aims to formulate the inner workings of a single universal reasoning operation by maintaining a separation between memory and control, chained together
to produce explicit and structured multi-step reasoning processes. We demonstrate the versatility,
robustness and transparency of the model through quantitative and qualitative studies, achieving
state-of-the-art results on the CLEVR task for visual reasoning, and generalizing well even from a
10% subset of the data. The experimental results further show that the model can adapt to novel
situations and diverse language, and generate interpretable attention-based rationales that reveal the
underlying reasoning it performs. While CLEVR provides a natural testbed for our approach, we
believe that the architecture will prove beneficial for other multi-step reasoning and inference tasks,
including reading comprehension, textual question answering, and real-world VQA.
10
Published as a conference paper at ICLR 2018
R EFERENCES
Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question
answering models. In Proceedings of the 2016 Conference on Empirical Methods in Natural
Language Processing, pp. 1955–1960, 2016.
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48,
2016a.
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural
networks for question answering. In Proceedings of NAACL-HLT, pp. 1545–1554, 2016b.
Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In Proceedings of the IEEE International
Conference on Computer Vision, pp. 2425–2433, 2015.
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint
arXiv:1607.06450, 2016.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. 2015.
Léon Bottou. From machine learning to machine reasoning. Machine learning, 94(2):133–149,
2014.
Prithvijit Chattopadhyay, Ramakrishna Vedantam, Ramprasaath R Selvaraju, Dhruv Batra, and Devi
Parikh. Counting everyday objects in everyday scenes. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 1135–1144, 2017.
Marta Garnelo, Kai Arulkumaran, and Murray Shanahan. Towards deep symbolic reinforcement
learning. arXiv preprint arXiv:1609.05518, 2016.
Alex Graves, Greg Wayne, and Ivo Danihelka.
arXiv:1410.5401, 2014.
Neural turing machines.
arXiv preprint
Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538
(7626):471–476, 2016.
Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint
arXiv:1705.03865, 2017.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770–778, 2016.
Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to
reason: End-to-end module networks for visual question answering. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp. 804–813, 2017.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,
2015.
Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and
Ross Girshick. CLEVR: A diagnostic dataset for compositional language and elementary visual
reasoning. In Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on, pp.
1988–1997. IEEE, 2017a.
Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei,
C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.
2989–2998, 2017b.
11
Published as a conference paper at ICLR 2018
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor
Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for
natural language processing. In International Conference on Machine Learning, pp. 1378–1387,
2016.
Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building
machines that learn and think like people. Behavioral and Brain Sciences, 40, 2017.
Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention
for visual question answering. In Advances In Neural Information Processing Systems, pp. 289–
297, 2016.
Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason
Weston. Key-value memory networks for directly reading documents. In Proceedings of the 2016
Conference on Empirical Methods in Natural Language Processing, pp. 1400–1409, 2016.
Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual
reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual
recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
Adam Santoro, David Raposo, David G Barrett, Mateusz Malinowski, Razvan Pascanu, Peter
Battaglia, and Tim Lillicrap. A simple neural network module for relational reasoning. In Advances in neural information processing systems, pp. 4974–4983, 2017.
Bob L Sturm. A simple method to determine if a music information retrieval system is a horse.
IEEE Transactions on Multimedia, 16(6):1636–1644, 2014.
Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and
textual question answering. In International Conference on Machine Learning, pp. 2397–2406,
2016.
Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks
for image question answering. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 21–29, 2016.
12
Published as a conference paper at ICLR 2018
Control Unit
1
Write Unit (1)
1
256d
0
5
10
Epoch
15
w/o word attention (c2)
w/o control
0.6
0.99
Accuracy (Val)
attention over word vectors
0.7
standard writing unit
0.98
w/ self-attention
0.98
0.96
0.94
standard write unit
W · retrieved + b -> memory
0.92
w/ gating
0.5
20
0.4
5
10
Epoch
15
20
0
5
10
15
0.94
1
-1
0
only gating
0.9
0.9
0.97
0
0.96
0.92
retrieved -> memory
w/ self-attention + gating
512d
0.6
Accuracy (Val)
Accuracy (Val)
Accuracy (Val)
128d
0.7
standard control unit
0.8
Gate Bias Initialization
1
0.98
0.9
0.8
Write Unit (2)
1
0.9
Accuracy (Val)
Network Dimension
1
20
Epoch
0
5
10
Epoch
15
20
0
5
10
15
20
Epoch
Figure 15: Learning curves for ablated MAC variants (accuracy / epoch). See appendix C for details.
S UPPLEMENTARY M ATERIAL
A
I MPLEMENTATION AND T RAINING DETAILS
We train our model using Adam (Kingma & Ba, 2014), with a learning rate of 10−4 and a batch
size of 64. We employ early stopping based on the validation accuracy, resulting in training process
of 10–20 epochs, equivalent to roughly 15–30 hours on a single Maxwell Titan X GPU. Word
vectors have dimension 300 and were initialized randomly using standard uniform distribution. The
exponential moving averages of the model weights are maintained during training, with a decay rate
of 0.999, and used at test time instead of the raw weights. We use variational dropout of 0.85 across
the network, along with ELU as non-linearity, which, in our experience, accelerates training and
performs favorably compared to the more standard ReLU.
B
E RROR A NALYSIS
To gain insight into the model’s failure cases, we perform error analysis for the CLEVR and CLEVRHumans datasets. For CLEVR, we see that many of the errors arise from object occlusions, which
may make it harder for the model to recognize the objects’ material or shape. Most of the other
errors are off-by-one counting mistakes, oftentimes for questions that ask to sum up two groups of
objects (see examples in figure 16). Interestingly, we noticed that when the model is required to
count objects that are heavily occluded, it may lead the model to slightly underestimate the correct
number of objects, suggesting that it may perform some sort of “continuous” counting rather than
discrete.
For CLEVR-Humans, the errors made by the model are more diverse. About half of the errors result from questions about reasoning skills that the model has not been exposed to while training on
CLEVR. These include questions about physical properties: lighting and shadows, reflections and
objects stability (“How many of the items are casting a shadow?”, “Can a ball stay still on top of
one another?”); relative distances (“How many objects. . . are almost touching?”, “What color is the
sphere positioned closest to. . . ”, “What object is in between. . . ”); relative amounts (“Are half the
items. . . ”, “Are the objects mainly. . . ”); commonalities (“What color are the identical objects?”,
“What shape do the items that . . . have in common?”); and negative questions, which refer to objects
that do not maintain some property (“How many items do not. . . ”). In addition, we observed some
cases where the model misinterprets unseen words, capturing plausible but incorrect semantics: for
instance, in some cases it interpreted a “caramel” object as yellow whereas the original question
referred to a brown one, or considered a cylinder to be “circle” while the question referred to the
spheres only. We believe that these errors may arise from diversity in the semantics of these words
across the dataset, making the model learn those potentially valid but incorrect interpretations. Similarly to CLEVR, we observed some off-by-one errors in CLEVR-Humans. Finally, in fifth of the
cases, the errors result from faulty or ambiguous questions, which may mistakenly regard a cyan
object as blue or mention references that cannot be uniquely resolved to a specific object.
C
A BLATION STUDIES
Based on the validation set, we have conducted an ablation study for MAC to better understand the
impact of each of its components on the overall performance. We have tested each setting for the
primary 700K CLEVR dataset as well as on a 10% subset of it. See table 2, figure 9 and figure 15 for
13
Published as a conference paper at ICLR 2018
final accuracies and training curves. The following discussion complements the main conclusions
presented in section 3.3:
Network Length. We observe a positive correlation between the network length and its performance, with significant improvements up to length p = 8. These results stand out from other
multi-hop architectures that tend to benefit from a lower number of iterations, commonly 2–3 only
(Yang et al., 2016; Kumar et al., 2016), and suggest that MAC makes effective use of the recurrent
cells to perform compositional reasoning.
Weight Sharing. Weight sharing across the p cell instances has also proven to be useful both for
the primary CLEVR task as well as for settings of limited data. Note that in contrast to alternative
approaches that apply specialized modules for different parts of the reasoning, these results provide
some evidence for the ability of the same MAC cell to adapt its behavior to the task at hand and
demonstrate different behaviors as inferred from the context.
Control Unit. We have performed several ablations in the control unit architecture to identify its
contribution to the model behavior and performance. First, we observe that, as would be expected,
an ablated model that reasons over the image alone with no access to the question, performs poorly,
achieving accuracy of 51.1%. As discussed in section 3.3, the ablations further show the importance
of applying attention over the question words to decompose it into an explicit sequence of steps.
Finally, we find that using the “contextual words” – the output states of a biLSTM processing the
question – results in better performance and faster learning than attending directly to the learned
word vectors. This implies that the model benefits from interpreting the word semantics and entailed
behaviors in the broader context of the question rather then as a sequence of independent entities.
Write Unit. The standard MAC write unit integrates new information ri with the previous memory state mi−1 through a linear transformation (step (1) in section 2.2.3). In this experiment, we
explore other variants of the unit. We begin by measuring the impact of self-attention and gating
mechanisms, both aiming to reduce long-range dependencies in the reasoning process. Compared
to the standard MAC model, which achieves 98.94% on the validation set, self-attention yields accuracy of 99.23%, memory gating – 99.36%, and adding both results in 99.48%. While we can see
some gain from using these components for CLEVR, we speculate that they may prove more useful
for tasks that necessitate longer or more complex reasoning processes over larger knowledge bases.
Next, we examine ablated write unit variants that assign the newly retrieved information ri (or its
linear transformation) to mi directly, ignoring the prior memory content mi−1 . Notably, the results
show that in fact such variants are only slightly worse than our standard model, reducing accuracy
by 0.4% only. We perform further experiments in which we compute the new memory state mi by
averaging the retrieved information ri with the previous memory state mi−1 using a sigmoidal gate
alone. This architecture results in equivalent performance to that of the standard write unit variant.
Gate Bias Initialization. Finally, we test the impact of the gate bias (step (3) in section 2.2.3), initializing it to either −1, 0 or 1. Intuitively, initialization of −1 amounts to biasing the model to retain
previous memory states and thereby shortening the effective reasoning process, while initialization
of 1 is equivalent to using all new intermediate results derived by each of the cells. The experiments
show that a bias of 1 is optimal for training on the full dataset while 0 is ideal for settings of limited
data (training on 10% of the data). These results demonstrate that when enough data is available, the
MAC network benefits from utilizing its full capacity, whereas biasing the model towards applying
less cells helps to mitigate overfitting when data is more scarce.
D
R ELATED W ORK
There have been several prominent models that address the CLEVR task. By and large they can be
partitioned into two groups: module networks, which in practice have all used the strong supervision
provided in the form of tree-structured functional programs that accompany each data instance, and
large, relatively unstructured end-to-end differentiable networks that complement a fairly standard
stack of CNNs with components that aid them in performing reasoning tasks. In contrast to modular
approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017b), our model is fully
differentiable and does not require additional supervision, making use of a single computational cell
chained in sequence (like an LSTM) rather than a collection of custom modules deployed in a rigid
tree structure. In contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017),
14
Published as a conference paper at ICLR 2018
Table 2: Accuracies for ablated MAC models, measured for the validation set after training on the full CLEVR
dataset (left) and 10% subset of it (right).
Model
MAC standard model
state dimension 256
state dimension 128
unshared weights
attention over word vectors
w/o word-attention
question vector as control
w/o control
w/o memory-control separation
w/o direct KB elements
retrieved → memory
W · retrieved + b → memory
only memory gate
w/ self-attention
w/ memory gate
w/ self-attention and memory gate
gate bias 0
gate bias 1
gate bias −1
prediction w/o question
Standard CLEVR
98.9
98.4
97.6
97.8
98.3
95.3
80.7
55.6
93.9
98.4
98.2
98.5
99.3
99.2
99.4
99.5
98.7
99.4
99.0
97.8
10% CLEVR
84.5
76.3
77.0
67.5
61.4
63.2
65.0
51.5
64.7
73.4
84.5
83.7
83.1
83.2
83.1
85.5
84.9
68.5
77.1
64.7
we suggest that our approach provides an ability for relational reasoning with better generalization
capacity, higher computational efficiency and enhanced transparency.
D.1
M ODULE N ETWORKS
The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017b) first translates a given question into a tree-structured action plan, aiming to imitate the question underlying
structural representation externally provided as strong supervision. Then, it constructs a tailor-made
network that progressively executes the plan over the image. The network is composed of discrete
units selected out of a fixed collection of predefined “modules”, each responsible for an elementary
reasoning operation, such as identifying an object’s color, filtering them for their shape, or comparing their amounts. Each module has its own set of learned parameters (Johnson et al., 2017b),
or even a hand-crafted design (Andreas et al., 2016a) that guides it towards its intended behavior.
Overall, this approach makes discrete choices at two levels: the identity of each module – the behavior it should learn among a fixed set of possible behavior types, and the network layout – the way
in which the modules are wired together to compute the answer. The model differentiability is thus
confined to the boundaries of a single module.
Several key differences exist between our approaches. First, MAC replaces the fixed and specialized
modules inventory with one universal cell that adapts its operation to the task at hand, selected from
a continuous range of reasoning behaviors. Therefore, in contrast to module networks, our cell can
be applied across all the reasoning steps, sharing both its parameters and architecture. Second, we
replace the dynamic recursive tree structures with a sequential topology, augmented by soft attention
mechanisms, inspired by Bahdanau et al. (2015). This confers the network with the capacity to
represent arbitrarily complex Directed Acyclic Graphs (DAGs) in a soft way, while still having
efficient and readily deployed physical sequential structure. Together, these relaxations allow us to
effectively train our model end-to-end by backpropagation alone, whereas module networks demand
more involved training schemes that rely on the strongly-supervised programs at the first stage, and
on various reinforcement learning (RL) techniques at the second. Finally, since the only source of
supervision in training our model arises from the answer to each question, MAC is free to acquire
more robust and adaptive reasoning strategies from the bottom up – inferred directly form the data,
rather than trying to imitate the behaviors dictated by brittle parsers or closed domain functional
programs, and is thus more applicable to real-world settings.
D.2
AUGMENTED C ONVOLUTIONAL N EURAL N ETWORKS
Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong
supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard
multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling
compositional and relational questions.
15
Published as a conference paper at ICLR 2018
Relation Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN. This
layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason over
binary relations between objects. While this approach is very simple and elegant conceptually, it
suffers from quadratic computational complexity, in contrast to our approach, which is linear. But
beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary.
Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to
objects and their pairwise interactions to relations, an RN layer attempts to grasp the induced graph
between objects all at once in one shallow and broad layer. Conversely, our attention-based model
proceeds in steps, iteratively comparing the image to a memory state that had aggregated information
from the image in prior iterations. By the same analogy, MAC traverses a narrow and deep reasoning
“path” that progressively follows transitive relations. Consequently, our model exhibits a relational
capacity while circumventing the computational inefficiency.
FiLM. Perez et al. (2017) proposes a model for visual reasoning that interleaves standard CNN
layers with linear layers, reminiscent of layer normalization techniques (Ba et al., 2016; Ioffe &
Szegedy, 2015). Each of these layers, called FiLM, is conditioned on the question, which is translated into matching bias and variance terms that tilt the layer’s activations to reflect the specifics of
the given question, thus influencing the computation done over the image. Similarly to our model,
this approach features distant modulation between the question and the image, where the former can
affect the latter only through constrained means. However, since the same normalization is applied
across all the activations homogeneously, agnostic to both their spatial location as well as their feature values, FiLM does not allow the question to differentiate between regions in the image based
on their semantics – the objects or concepts they represent.
This stands in stark contrast to our attention-based model, which readily allows and actually encourages the question to inform the model about relevant regions to focus on. As supported by section 3,
this more selective interaction between the question and the image facilitates learning and increases
the model’s generalizability. Indeed, since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta, 2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to
assume that it would be beneficial to incorporate such methods into visual reasoning systems for the
CLEVR task as well. In fact, attention mechanisms should be especially useful for multi-step reasoning questions such as those present in CLEVR. Such questions refer to several relations between
different objects in the image and feature compositional structure that may be approached one step
at a time. Thus, it should be beneficial for a cogent responder to have the capacity to selectively
focus on one or some objects at each step, traversing the relevant relational links one after the other,
both at the image level, and at the question level.
D.3
M EMORY AND ATTENTION
Our architecture draws inspiration from recent research on mechanisms for neural memory and
attention (Kumar et al., 2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016)
and Xiong et al. (2016) propose the Dynamic Memory Network (DMN) model that proceeds in
an iterative process, attending to relevant information from a given knowledge base, which is then
successively accumulated into the model’s memory state. However, DMN views the question as one
atomic unit, whereas our model decomposes it into a multi-step action plan that informs each cell of
its specific objective. Another key difference is the distant interaction between the question and the
knowledge base that characterizes our model. Conversely, DMN fuses their representations together
into the same vector space.
Graves et al. (2014; 2016) complement a neural network with external memory it can interact with
through the means of soft attention. Similarly to our approach, the model consists of a controller that
performs read and write operations over a fixed-size memory array. However, in contrast to Graves
et al. (2014; 2016), we employ a recurrent memory structure, where each MAC cell is associated
with its own memory state. Rather than reading and writing iteratively into multiple slots in a shared
memory resource, each cell creates a new memory, building upon the contents of the prior ones.
This allows us to avoid potential issues of content blurring due to multiple global write operations,
while still supporting the emergence of complex reasoning processes that progressively interact with
preceding memories and intermediate results to accomplish the task at hand.
16
Published as a conference paper at ICLR 2018
Q: What color is the metallic
cylinder in front of the silver
cylinder? A: cyan 3
Q: What is the object made of
hiding behind the green cube?
A: rubber 3
Q: What is the color of the ball
that is farthest away? A: blue 3
Q: How many matte cubes are
there? A: 2 3
Q: How many spheres are
pictured? A: 4 3
Q: how many square objects are
in the picture? A: 4 3
Q: What object is to the far
right? A: cube 3
Q: Are the yellow blocks the
same? A: no 3
Q: What shape is the smallestt
Q: What object looks like a
object in this image? A: sphere 3 caramel? A: cube 3
Q: Can a ball stay still on top of
one another? A: yes (no) 7
Q: What color is the center
object? A: blue 3
Q: How many gray objects are
shown? A: 3 3
Q: How many small objects are
rubber? A: 2 3
Q: What color is the largest
cube? A: yellow 3
Q: What shape are most of the
shiny items? A: sphere 3
Q: What is the tan object made
of? A: rubber 3
Q: Are half the items shown
green? A: yes (no) 7
Q: What color object is biggest?
A: blue 3
Q: Which shape is a different
color from the others? A:
cylinder 3
Q: How many other objects are
Q: What is the shape of the object Q: How many tiny objects are green things or red objects? A: 4 3
the same size as the blue ball? A: that is to the left of the red rubber
73
cube and behind the metallic
cylinder? A: sphere (cylinder) 7
Figure 16: The first five rows show examples of the final attention map produced by the model for CLEVRHuman questions, demonstrating the ability of the model to perform novel reasoning skills and cope with
new concepts that have not been introduced in CLEVR. These include in particular: obstructions, object
uniqueness, relative distances, superlatives and new terms. The final row shows examples from CLEVR
with object occlusions and summation.
17
Published as a conference paper at ICLR 2018
Figure 17: Attention maps produced by MAC network of lengths 4 and 6, providing evidence for the ability
of the model to track transitive relations and perform logical operations. Note how the model tends to proceed
from the end of the question backwards, tracking the relevant objects iteratively.
18
Published as a conference paper at ICLR 2018
Figure 18: Attention maps produced by MAC network of length 6, providing evidence for the ability of the
model to track transitive relations, and perform logical operations, counting and summation. Note how the first
iterations focus on the key structural question words “many” and “or” that serve as indicators for the model
of the required reasoning operation it has to perform. Also note how the model correctly sums up two object
groups in the second example, while correctly accounting for the intersection between them.
19
| 2cs.AI
|
Published as a conference paper at ICLR 2017
A S TRUCTURED S ELF - ATTENTIVE
S ENTENCE E MBEDDING
arXiv:1703.03130v1 [cs.CL] 9 Mar 2017
Zhouhan Lin‡∗, Minwei Feng , Cicero Nogueira dos Santos , Mo Yu ,
Bing Xiang , Bowen Zhou & Yoshua Bengio‡†
IBM Watson
‡
Montreal Institute for Learning Algorithms (MILA), Université de Montréal
†
CIFAR Senior Fellow
[email protected]
{mfeng, cicerons, yum, bingxia, zhou}@us.ibm.com
A BSTRACT
This paper proposes a new model for extracting an interpretable sentence embedding by introducing self-attention. Instead of using a vector, we use a 2-D matrix
to represent the embedding, with each row of the matrix attending on a different
part of the sentence. We also propose a self-attention mechanism and a special
regularization term for the model. As a side effect, the embedding comes with an
easy way of visualizing what specific parts of the sentence are encoded into the
embedding. We evaluate our model on 3 different tasks: author profiling, sentiment classification and textual entailment. Results show that our model yields a
significant performance gain compared to other sentence embedding methods in
all of the 3 tasks.
1
I NTRODUCTION
Much progress has been made in learning semantically meaningful distributed representations of
individual words, also known as word embeddings (Bengio et al., 2001; Mikolov et al., 2013).
On the other hand, much remains to be done to obtain satisfying representations of phrases and
sentences. Those methods generally fall into two categories. The first consists of universal sentence
embeddings usually trained by unsupervised learning (Hill et al., 2016). This includes SkipThought
vectors (Kiros et al., 2015), ParagraphVector (Le & Mikolov, 2014), recursive auto-encoders (Socher
et al., 2011; 2013), Sequential Denoising Autoencoders (SDAE), FastSent (Hill et al., 2016), etc.
The other category consists of models trained specifically for a certain task. They are usually
combined with downstream applications and trained by supervised learning. One generally finds
that specifically trained sentence embeddings perform better than generic ones, although generic
ones can be used in a semi-supervised setting, exploiting large unlabeled corpora. Several models
have been proposed along this line, by using recurrent networks (Hochreiter & Schmidhuber, 1997;
Chung et al., 2014), recursive networks (Socher et al., 2013) and convolutional networks (Kalchbrenner et al., 2014; dos Santos & Gatti, 2014; Kim, 2014) as an intermediate step in creating sentence
representations to solve a wide variety of tasks including classification and ranking (Yin & Schütze,
2015; Palangi et al., 2016; Tan et al., 2016; Feng et al., 2015). A common approach in previous
methods consists in creating a simple vector representation by using the final hidden state of the
RNN or the max (or average) pooling from either RNNs hidden states or convolved n-grams. Additional works have also been done in exploiting linguistic structures such as parse and dependence
trees to improve sentence representations (Ma et al., 2015; Mou et al., 2015b; Tai et al., 2015).
For some tasks people propose to use attention mechanism on top of the CNN or LSTM model to
introduce extra source of information to guide the extraction of sentence embedding (dos Santos
et al., 2016). However, for some other tasks like sentiment classification, this is not directly applicable since there is no such extra information: the model is only given one single sentence as input.
In those cases, the most common way is to add a max pooling or averaging step across all time steps
∗
This work has been done during the 1st author’s internship with IBM Watson.
1
Published as a conference paper at ICLR 2017
...
......
n
mr
2u
mi
m1 m2
da
M
Ai2
Ai3
h1
h2
h3
...
Ai4
...
h4
hn
tanh
Ain
...
...
Ws1
hn
r
Ai1
h1 h2
Ws2
softmax
ran
(b)
res
tau
A
n
.
t
t
wn
r
...
tha
e
eri
en
c
ex
p
in
...
w4
nic
e
w3
a
d
w2
ha
we
w1
(a)
Figure 1: A sample model structure showing the sentence embedding model combined with a fully
connected and softmax layer for sentiment analysis (a). The sentence embedding M is computed as
multiple weighted sums of hidden states from a bidirectional LSTM (h1 , ..., hn ), where the summation weights (Ai1 , ..., Ain ) are computed in a way illustrated in (b). Blue colored shapes stand for
hidden representations, and red colored shapes stand for weights, annotations, or input/output.
(Lee & Dernoncourt, 2016), or just pick up the hidden representation at the last time step as the
encoded embedding (Margarit & Subramaniam, 2016).
A common approach in many of the aforementioned methods consists of creating a simple vector
representation by using the final hidden state of the RNN or the max (or average) pooling from
either RNNs hidden states or convolved n-grams. We hypothesize that carrying the semantics along
all time steps of a recurrent model is relatively hard and not necessary. We propose a self-attention
mechanism for these sequential models to replace the max pooling or averaging step. Different from
previous approaches, the proposed self-attention mechanism allows extracting different aspects of
the sentence into multiple vector representations. It is performed on top of an LSTM in our sentence
embedding model. This enables attention to be used in those cases when there are no extra inputs. In
addition, due to its direct access to hidden representations from previous time steps, it relieves some
long-term memorization burden from LSTM. As a side effect coming together with our proposed
self-attentive sentence embedding, interpreting the extracted embedding becomes very easy and
explicit.
Section 2 details on our proposed self-attentive sentence embedding model, as well as a regularization term we proposed for this model, which is described in Section 2.2. We also provide a
visualization method for this sentence embedding in section 2.3. We then evaluate our model in
author profiling, sentiment classification and textual entailment tasks in Section 4.
2
2.1
A PPROACH
M ODEL
The proposed sentence embedding model consists of two parts. The first part is a bidirectional
LSTM, and the second part is the self-attention mechanism, which provides a set of summation
weight vectors for the LSTM hidden states. These set of summation weight vectors are dotted
with the LSTM hidden states, and the resulting weighted LSTM hidden states are considered as
an embedding for the sentence. It can be combined with, for example, a multilayer perceptron to
2
Published as a conference paper at ICLR 2017
be applied on a downstream application. Figure 1 shows an example when the proposed sentence
embedding model is applied to sentiment analysis, combined with a fully connected layer and a
softmax layer. Besides using a fully connected layer, we also proposes an approach that prunes
weight connections by utilizing the 2-D structure of matrix sentence embedding, which is detailed
in Appendix A. For this section, we will use Figure 1 to describe our model.
Suppose we have a sentence, which has n tokens, represented in a sequence of word embeddings.
S = (w1 , w2 , · · · wn )
(1)
Here wi is a vector standing for a d dimentional word embedding for the i-th word in the sentence.
S is thus a sequence represented as a 2-D matrix, which concatenates all the word embeddings
together. S should have the shape n-by-d.
Now each entry in the sequence S are independent with each other. To gain some dependency between adjacent words within a single sentence, we use a bidirectional LSTM to process the sentence:
→
−
−−→
−−−−→
ht = LST M (wt , ht−1 )
(2)
←
− ←−−−−
←−−
ht = LST M (wt , ht+1 )
(3)
→
−
←
−
And we concatenate each ht with ht to obtain a hidden state ht . Let the hidden unit number for each
unidirectional LSTM be u. For simplicity, we note all the n ht s as H, who have the size n-by-2u.
H = (h1 , h2 , · · · hn )
(4)
Our aim is to encode a variable length sentence into a fixed size embedding. We achieve that by
choosing a linear combination of the n LSTM hidden vectors in H. Computing the linear combination requires the self-attention mechanism. The attention mechanism takes the whole LSTM hidden
states H as input, and outputs a vector of weights a:
a = sof tmax ws2 tanh Ws1 H T
(5)
Here Ws1 is a weight matrix with a shape of da -by-2u. and ws2 is a vector of parameters with
size da , where da is a hyperparameter we can set arbitrarily. Since H is sized n-by-2u, the annotation vector a will have a size n. the sof tmax()˙ ensures all the computed weights sum up to 1.
Then we sum up the LSTM hidden states H according to the weight provided by a to get a vector
representation m of the input sentence.
This vector representation usually focuses on a specific component of the sentence, like a special set
of related words or phrases. So it is expected to reflect an aspect, or component of the semantics in
a sentence. However, there can be multiple components in a sentence that together forms the overall
semantics of the whole sentence, especially for long sentences. (For example, two clauses linked
together by an ”and.”) Thus, to represent the overall semantics of the sentence, we need multiple m’s
that focus on different parts of the sentence. Thus we need to perform multiple hops of attention.
Say we want r different parts to be extracted from the sentence, with regard to this, we extend the
ws2 into a r-by-da matrix, note it as Ws2 , and the resulting annotation vector a becomes annotation
matrix A. Formally,
A = sof tmax Ws2 tanh Ws1 H T
(6)
Here the sof tmax()˙ is performed along the second dimension of its input. We can deem Equation
6 as a 2-layer MLP without bias, whose hidden unit numbers is da , and parameters are {Ws2 , Ws1 }.
The embedding vector m then becomes an r-by-2u embedding matrix M . We compute the r
weighted sums by multiplying the annotation matrix A and LSTM hidden states H, the resulting
matrix is the sentence embedding:
M = AH
(7)
2.2
P ENALIZATION TERM
The embedding matrix M can suffer from redundancy problems if the attention mechanism always
provides similar summation weights for all the r hops. Thus we need a penalization term to encourage the diversity of summation weight vectors across different hops of attention.
3
Published as a conference paper at ICLR 2017
The best way to evaluate the diversity is definitely the Kullback Leibler divergence between any 2
of the summation weight vectors. However, we found that not very stable in our case. We conjecture
it is because we are maximizing a set of KL divergence (instead of minimizing only one, which is
the usual case), we are optimizing the annotation matrix A to have a lot of sufficiently small or
even zero values at different softmax output units, and these vast amount of zeros is making the
training unstable. There is another feature that KL doesn’t provide but we want, which is, we want
each individual row to focus on a single aspect of semantics, so we want the probability mass in the
annotation softmax output to be more focused. but with KL penalty we cant encourage that.
We hereby introduce a new penalization term which overcomes the aforementioned shortcomings.
Compared to the KL divergence penalization, this term consumes only one third of the computation.
We use the dot product of A and its transpose, subtracted by an identity matrix, as a measure of
redundancy.
2
P = AAT − I F
(8)
Here k•kF stands for the Frobenius norm of a matrix. Similar to adding an L2 regularization term,
this penalization term P will be multiplied by a coefficient, and we minimize it together with the
original loss, which is dependent on the downstream application.
Let’s consider two different summation vectors ai and aj in A. Because of the softmax, all entries
within any summation vector in A should sum up to 1. Thus they can be deemed as probability
masses in a discrete probability distribution. For any non-diagonal elements aij (i 6= j) in the AAT
matrix, it corresponds to a summation over elementwise product of two distributions:
0 < aij =
n
X
aik ajk < 1
(9)
k=1
where aik and ajk are the k-th element in the ai and aj vectors, respectively. In the most extreme case,
where there is no overlap between the two probability distributions ai and aj , the correspond aij will
be 0. Otherwise, it will have a positive value. On the other extreme end, if the two distributions are
identical and all concentrates on one single word, it will have a maximum value of 1. We subtract
an identity matrix from AAT so that forces the elements on the diagonal of AAT to approximate
1, which encourages each summation vector ai to focus on as few number of words as possible,
forcing each vector to be focused on a single aspect, and all other elements to 0, which punishes
redundancy between different summation vectors.
2.3
V ISUALIZATION
The interpretation of the sentence embedding is quite straight forward because of the existence of
annotation matrix A. For each row in the sentence embedding matrix M , we have its corresponding
annotation vector ai . Each element in this vector corresponds to how much contribution the LSTM
hidden state of a token on that position contributes to. We can thus draw a heat map for each row of
the embedding matrix M This way of visualization gives hints on what is encoded in each part of
the embedding, adding an extra layer of interpretation. (See Figure 3a and 3b).
The second way of visualization can be achieved by summing up over all the annotation vectors,
and then normalizing the resulting weight vector to sum up to 1. Since it sums up all aspects of
semantics of a sentence, it yields a general view of what the embedding mostly focuses on. We can
figure out which words the embedding takes into account a lot, and which ones are skipped by the
embedding. See Figure 3c and 3d.
3
R ELATED WORK
Various supervised and unsupervised sentence embedding models have been mentioned in Section
1. Different from those models, our proposed method uses a new self-attention mechanism that
allows it to extract different aspects of the sentence into multiple vector-representations. The matrix
structure together with the penalization term gives our model a greater capacity to disentangle the
latent information from the input sentence. We also do not use linguistic structures to guide our
sentence representation model. Additionally, using our method we can easily create visualizations
that can help in the interpretation of the learned representations.
4
Published as a conference paper at ICLR 2017
Some recent work have also proposed supervised methods that use intra/self-sentence attention. Ling
et al. (2015) proposed an attention based model for word embedding, which calculates an attention
weight for each word at each possible position in the context window. However this method cannot
be extended to sentence level embeddings since one cannot exhaustively enumerate all possible
sentences. Liu et al. (2016a) proposes a sentence level attention which has a similar motivation but
done differently. They utilize the mean pooling over LSTM states as the attention source, and use
that to re-weight the pooled vector representation of the sentence.
Apart from the previous 2 variants, we want to note that Li et al. (2016) proposed a same self
attention mechanism for question encoding in their factoid QA model, which is concurrent to our
work. The difference lies in that their encoding is still presented as a vector, but our attention
produces a matrix representation instead, with a specially designed penalty term. We applied the
model for sentiment anaysis and entailment, and their model is for factoid QA.
The LSTMN model (Cheng et al., 2016) also proposed a very successful intra-sentence level attention mechanism, which is later used by Parikh et al. (2016). We see our attention and theirs as having
different granularities. LSTMN produces an attention vector for each of its hidden states during the
recurrent iteration, which is sort of an ”online updating” attention. It’s more fine-grained, targeting
at discovering lexical correlations between a certain word and its previous words. On the contrary,
our attention mechanism is only performed once, focuses directly on the semantics that makes sense
for discriminating the targets. It is less focused on relations between words, but more on the semantics of the whole sentence that each word contributes to. Computationally, our method also scales up
with the sentence length better, since it doesn’t require the LSTM to compute an annotation vector
over all of its previous words each time when the LSTMN computes its next step.
4
E XPERIMENTAL RESULTS
We first evaluate our sentence embedding model by applying it to 3 different datasets: the Age
dataset, the Yelp dataset, and the Stanford Natural Language Inference (SNLI) Corpus. These 3
datasets fall into 3 different tasks, corresponding to author profiling, sentiment analysis, and textual entailment, respectively. Then we also perform a set of exploratory experiments to validate
properties of various aspects for our sentence embedding model.
4.1
AUTHOR PROFILING
The Author Profiling dataset1 consists of Twitter tweets in English, Spanish, and Dutch. For some of
the tweets, it also provides an age and gender of the user when writing the tweet. The age range are
split into 5 classes: 18-24, 25-34, 35-49, 50-64, 65+. We use English tweets as input, and use those
tweets to predict the age range of the user. Since we are predicting the age of users, we refer to it
as Age dataset in the rest of our paper. We randomly selected 68485 tweets as training set, 4000 for
development set, and 4000 for test set. Performances are also chosen to be classification accuracy.
Table 1: Performance Comparision of Different Models on Yelp and Age Dataset
Models
BiLSTM + Max Pooling + MLP
CNN + Max Pooling + MLP
Our Model
Yelp
61.99%
62.05%
64.21%
Age
77.40%
78.15%
80.45%
We compare our model with two baseline models: biLSTM and CNN. For the two baseline models.
The biLSTM model uses a bidirectional LSTM with 300 dimensions in each direction, and use max
pooling across all LSTM hidden states to get the sentence embedding vector, then use a 2-layer
ReLU output MLP with 3000 hidden states to output the classification result. The CNN model
uses the same scheme, but substituting biLSTM with 1 layer of 1-D convolutional network. During
training we use 0.5 dropout on the MLP and 0.0001 L2 regularization. We use stochastic gradient
descent as the optimizer, with a learning rate of 0.06, batch size 16. For biLSTM, we also clip the
1
http://pan.webis.de/clef16/pan16-web/author-profiling.html
5
Published as a conference paper at ICLR 2017
(a) 1 star reviews
(b) 5 star reviews
Figure 2: Heatmap of Yelp reviews with the two extreme score.
norm of gradients to be between -0.5 and 0.5. We searched hyperparameters in a wide range and
find the aforementioned set of hyperparameters yields the highest accuracy.
For our model, we use the same settings as what we did in biLSTM. We also use a 2-layer ReLU
output MLP, but with 2000 hidden units. In addition, our self-attention MLP has a hidden layer with
350 units (the da in Section 2), we choose the matrix embedding to have 30 rows (the r), and a
coefficient of 1 for the penalization term.
We train all the three models until convergence and select the corresponding test set performance
according to the best development set performance. Our results show that the model outperforms
both of the biLSTM and CNN baselines by a significant margin.
4.2
S ENTIMENT ANALYSIS
We choose the Yelp dataset2 for sentiment analysis task. It consists of 2.7M yelp reviews, we take
the review as input and predict the number of stars the user who wrote that review assigned to the
corresponding business store. We randomly select 500K review-star pairs as training set, and 2000
for development set, 2000 for test set. We tokenize the review texts by Stanford tokenizer. We use
2
https://www.yelp.com/dataset challenge
6
Published as a conference paper at ICLR 2017
100 dimensional word2vec as initialization for word embeddings, and tune the embedding during
training across all of our experiments. The target number of stars is an integer number in the range
of [1, 5], inclusive. We are treating the task as a classification task, i.e., classify a review text into
one of the 5 classes. We use classification accuracy as a measurement.
For the two baseline models, we use the same setting as what we used for Author Profiling dataset,
except that we are using a batch size of 32 instead. For our model, we are also using the same
setting, except that we choose the hidden unit numbers in the output MLP to be 3000 instead. We
also observe a significant performance gain comparining to the two baselines. (Table 1)
As an interpretation of the learned sentence embedding, we use the second way of visualization
described in Section 2.3 to plot heat maps for some of the reviews in the dataset. We randomly
select 5 examples of negative (1 star) and positive (5 stars) reviews from the test set, when the model
has a high confidence (> 0.8) in predicting the label. As shown in Figure 2, we find that the model
majorly learns to capture some key factors in the review that indicate strongly on the sentiment
behind the sentence. For most of the short reviews, the model manages to capture all the key factors
that contribute to an extreme score, but for longer reviews, the model is still not able to capture all
related factors. For example, in the 3rd review in Figure 2b), it seems that a lot of focus is spent on
one single factor, i.e., the ”so much fun”, and the model puts a little amount of attention on other
key points like ”highly recommend”, ”amazing food”, etc.
4.3
T EXTUAL ENTAILMENT
We use the biggest dataset in textual entailment, the SNLI corpus (Bowman et al., 2015) for our
evaluation on this task. SNLI is a collection of 570k human-written English sentence pairs manually
labeled for balanced classification with the labels entailment, contradiction, and neutral. The model
will be given a pair of sentences, called hypothesis and premise respectively, and asked to tell if the
semantics in the two sentences are contradicting with each other or not. It is also a classification
task, so we measure the performance by accuracy.
We process the hypothesis and premise independently, and then extract the relation between the two
sentence embeddings by using multiplicative interactions proposed in Memisevic (2013) (see Appendix B for details), and use a 2-layer ReLU output MLP with 4000 hidden units to map the hidden
representation into classification results. Parameters of biLSTM and attention MLP are shared across
hypothesis and premise. The biLSTM is 300 dimension in each direction, the attention MLP has
150 hidden units instead, and both sentence embeddings for hypothesis and premise have 30 rows
(the r). The penalization term coefficient is set to 0.3. We use 300 dimensional GloVe (Pennington
et al., 2014) word embedding to initialize word embeddings. We use AdaGrad as the optimizer,
with a learning rate of 0.01. We don’t use any extra regularization methods, like dropout or L2
normalization. Training converges after 4 epochs, which is relatively fast.
This task is a bit different from previous two tasks, in that it has 2 sentences as input. There are
a bunch of ways to add inter-sentence level attention, and those attentions bring a lot of benefits.
To make the comparison focused and fair, we only compare methods that fall into the sentence
encoding-based models. i.e., there is no information exchanged between the hypothesis and premise
before they are encoded into some distributed encoding.
Table 2: Test Set Performance Compared to other Sentence Encoding Based Methods in SNLI Datset
Model
300D LSTM encoders (Bowman et al., 2016)
600D (300+300) BiLSTM encoders (Liu et al., 2016b)
300D Tree-based CNN encoders (Mou et al., 2015a)
300D SPINN-PI encoders (Bowman et al., 2016)
300D NTI-SLSTM-LSTM encoders (Munkhdalai & Yu, 2016a)
1024D GRU encoders with SkipThoughts pre-training (Vendrov et al., 2015)
300D NSE encoders (Munkhdalai & Yu, 2016b)
Our method
7
Test Accuracy
80.6%
83.3%
82.1%
83.2%
83.4%
81.4%
84.6%
84.4%
Published as a conference paper at ICLR 2017
We find that compared to other published approaches, our method shows a significant gain (≥ 1%)
to them, except for the 300D NSE encoders, which is the state-of-the-art in this category. However,
the 0.2% different is relatively small compared to the differences between other methods.
4.4
E XPLORATORY EXPERIMENTS
In this subsection we are going to do a set of exploratory experiments to study the relative effect of
each component in our model.
4.4.1
E FFECT OF PENALIZATION TERM
Since the purpose of introducing the penalization term P is majorly to discourage the redundancy
in the embedding, we first directly visualize the heat maps of each row when the model is presented
with a sentence. We compare two identical models with the same size as detailed in Section 4.1
trained separately on Age dataset, one with this penalization term (where the penalization coefficient
is set to 1.0) and the other with no penalty. We randomly select one tweet from the test set and
compare the two models by plotting a heat map for each hop of attention on that single tweet. Since
there are 30 hops of attention for each model, which makes plotting all of them quite redundant, we
only plot 6 of them. These 6 hops already reflect the situation in all of the 30 hops.
(a)
(b)
(c) without penalization
(d) with 1.0 penalization
Figure 3: Heat maps for 2 models trained on Age dataset. The left column is trained without the
penalization term, and the right column is trained with 1.0 penalization. (a) and (b) shows detailed
attentions taken by 6 out of 30 rows of the matrix embedding, while (c) and (d) shows the overall
attention by summing up all 30 attention weight vectors.
(a) Yelp without penalization
(b) Yelp with penalization
Figure 4: Attention of sentence embedding on 3 different Yelp reviews. The left one is trained
without penalization, and the right one is trained with 1.0 penalization.
8
Published as a conference paper at ICLR 2017
Table 3: Performance comparision regarding the penalization term
Penalization coefficient
1.0
0.0
Yelp
64.21%
61.74%
Age
80.45%
79.27%
From the figure we can tell that the model trained without the penalization term have lots of redundancies between different hops of attention (Figure 3a), resulting in putting lot of focus on the word
”it” (Figure 3c), which is not so relevant to the age of the author. However in the right column, the
model shows more variations between different hops, and as a result, the overall embedding focuses
on ”mail-replies spam” instead. (Figure 3d)
For the Yelp dataset, we also observe a similar phenomenon. To make the experiments more explorative, we choose to plot heat maps of overall attention heat maps for more samples, instead of
plotting detailed heat maps for a single sample again. Figure 4 shows overall focus of the sentence
embedding on three different reviews. We observe that with the penalization term, the model tends
to be more focused on important parts of the review. We think it is because that we are encouraging
it to be focused, in the diagonals of matrix AAT (Equation 8).
To validate if these differences result in performance difference, we evaluate four models trained
on Yelp and Age datasets, both with and without the penalization term. Results are shown in Table
3. Consistent with what expected, models trained with the penalization term outperforms their
counterpart trained without.
In SNLI dataset, although we observe that introducing the penalization term still contributes to encouraging the diversity of different rows in the matrix sentence embedding, and forcing the network
to be more focused on the sentences, the quantitative effect of this penalization term is not so obvious
on SNLI dataset. Both models yield similar test set accuracies.
4.4.2
E FFECT OF MULTIPLE VECTORS
Having multiple rows in the sentence embedding is expected to provide more abundant information
about the encoded content. It makes sence to evaluate how significant the improvement can be
brought by r. Taking the models we used for Age and SNLI dataset as an example, we vary r from
1 to 30 for each task, and train the resulting 10 models independently (Figure 5). Note that when
r = 1, the sentence embedding reduces to a normal vector form.
From this figure we can find that, without having multiple rows, the model performs on-par with
its competitiors which use other forms of vector sentence embeddings. But there is significant
(a)
(b)
Figure 5: Effect of the number of rows (r) in matrix sentence embedding. The vertical axes indicates
test set accuracy and the horizontal axes indicates training epoches. Numbers in the legends stand
for the corresponding values of r. (a) is conducted in Age dataset and (b) is conducted in SNLI
dataset.
9
Published as a conference paper at ICLR 2017
difference between having only one vector for the sentence embedding and multiple vectors. The
models are also quite invariant with respect to r, since in the two figures a wide range of values
between 10 to 30 are all generating comparable curves.
5
C ONCLUSION AND DISCUSSION
In this paper, we introduced a fixed size, matrix sentence embedding with a self-attention mechanism. Because of this attention mechanism, there is a way to interpret the sentence embedding in
depth in our model. Experimental results over 3 different tasks show that the model outperforms
other sentence embedding models by a significant margin.
Introducing attention mechanism allows the final sentence embedding to directly access previous
LSTM hidden states via the attention summation. Thus the LSTM doesn’t need to carry every piece
of information towards its last hidden state. Instead, each LSTM hidden state is only expected to
provide shorter term context information around each word, while the higher level semantics, which
requires longer term dependency, can be picked up directly by the attention mechanism. This setting
reliefs the burden of LSTM to carry on long term dependencies. Our experiments also support that,
as we observed that our model has a bigger advantage when the contents are longer. Further more,
the notion of summing up elements in the attention mechanism is very primitive, it can be something
more complex than that, which will allow more operations on the hidden states of LSTM.
The model is able to encode any sequence with variable length into a fixed size representation,
without suffering from long-term dependency problems. This brings a lot of scalability to the model:
without any modification, it can be applied directly to longer contents like paragraphs, articles, etc.
Though this is beyond the focus of this paper, it remains an interesting direction to explore as a
future work.
As a downside of our proposed model, the current training method heavily relies on downstream
applications, thus we are not able to train it in an unsupervised way. The major obstacle towards
enabling unsupervised learning in this model is that during decoding, we don’t know as prior how
the different rows in the embedding should be divided and reorganized. Exploring all those possible
divisions by using a neural network could easily end up with overfitting. Although we can still do
unsupervised learning on the proposed model by using a sequential decoder on top of the sentence
embedding, it merits more to find some other structures as a decoder.
ACKNOWLEDGMENTS
The authors would like to acknowledge the developers of Theano (Theano Development Team,
2016) and Lasagne. The first author would also like to thank IBM Watson for providing resources,
fundings and valuable discussions to make this project possible, and Caglar Gulcehre for helpful
discussions.
R EFERENCES
Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. In
Advances in Neural Information Processing Systems, pp. 932–938, 2001.
Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015.
Samuel R Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D Manning, and
Christopher Potts. A fast unified model for parsing and sentence understanding. arXiv preprint
arXiv:1603.06021, 2016.
Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine
reading. In Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 2016.
Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of
gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
10
Published as a conference paper at ICLR 2017
Cicero dos Santos and Maira Gatti. Deep convolutional neural networks for sentiment analysis of
short texts. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pp. 69–78, 2014.
Cicero dos Santos, Ming Tan, Bing Xiang, and Bowen Zhou. Attentive pooling networks. arXiv
preprint arXiv:1602.03609, 2016.
Minwei Feng, Bing Xiang, Michael R. Glass, Lidan Wang, and Bowen Zhou. Applying deep learning to answer selection: a study and an open task. In 2015 IEEE Workshop on Automatic Speech
Recognition and Understanding, ASRU 2015, Scottsdale, AZ, USA, December 13-17, 2015, pp.
813–820, 2015.
Felix Hill, Kyunghyun Cho, and Anna Korhonen. Learning distributed representations of sentences
from unlabelled data. In Proceedings of the 2016 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies, pp. 1367–
1377, San Diego, California, June 2016. Association for Computational Linguistics. URL http:
//www.aclweb.org/anthology/N16-1162.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network for
modelling sentences. arXiv preprint arXiv:1404.2188, 2014.
Yoon Kim.
Convolutional neural networks for sentence classification.
arXiv:1408.5882, 2014.
arXiv preprint
Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In Advances in neural information processing
systems, pp. 3294–3302, 2015.
Quoc V Le and Tomas Mikolov. Distributed representations of sentences and documents. In ICML,
volume 14, pp. 1188–1196, 2014.
Ji Young Lee and Franck Dernoncourt. Sequential short-text classification with recurrent and convolutional neural networks. arXiv preprint arXiv:1603.03827, 2016.
Peng Li, Wei Li, Zhengyan He, Xuguang Wang, Ying Cao, Jie Zhou, and Wei Xu. Dataset and neural
recurrent sequence labeling model for open-domain factoid question answering. arXiv preprint
arXiv:1607.06275, 2016.
Wang Ling, Lin Chu-Cheng, Yulia Tsvetkov, and Silvio Amir. Not all contexts are created equal:
Better word representations with variable attention. In Proceedings of the 2015 Conference on
Empirical Methods in Natural Language Processing, pp. 1367–1372, Lisbon, Portugal, September 2015. Association for Computational Linguistics.
Yang Liu, Chengjie Sun, Lei Lin, and Xiaolong Wang. Learning natural language inference using
bidirectional LSTM model and inner-attention. CoRR, abs/1605.09090, 2016a.
Yang Liu, Chengjie Sun, Lei Lin, and Xiaolong Wang. Learning natural language inference using
bidirectional lstm model and inner-attention. arXiv preprint arXiv:1605.09090, 2016b.
Mingbo Ma, Liang Huang, Bing Xiang, and Bowen Zhou. Dependency-based convolutional neural
networks for sentence embedding. In Proceedings of the 53rd Annual Meeting of the Association
for Computational Linguistics and the 7th International Joint Conference on Natural Language
Processing, volume 2, pp. 174–179, 2015.
Horia Margarit and Raghav Subramaniam. A batch-normalized recurrent network for sentiment
classification. In Advances in Neural Information Processing Systems, 2016.
Roland Memisevic. Learning to relate images. IEEE transactions on pattern analysis and machine
intelligence, 35(8):1829–1846, 2013.
11
Published as a conference paper at ICLR 2017
Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013.
Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan, and Zhi Jin. Natural language inference by
tree-based convolution and heuristic matching. arXiv preprint arXiv:1512.08422, 2015a.
Lili Mou, Hao Peng, Ge Li, Yan Xu, Lu Zhang, and Zhi Jin. Discriminative neural sentence modeling by tree-based convolution. In Proceedings of the 2015 Conference on Empirical Methods in
Natural Language Processing, pp. 2315–2325, Lisbon, Portugal, September 2015b. Association
for Computational Linguistics. URL http://aclweb.org/anthology/D15-1279.
Tsendsuren Munkhdalai and Hong Yu. Neural tree indexers for text understanding. arXiv preprint
arXiv:1607.04492, 2016a.
Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv preprint arXiv:1607.04315,
2016b.
Hamid Palangi, Li Deng, Yelong Shen, Jianfeng Gao, Xiaodong He, Jianshu Chen, Xinying Song,
and Rabab Ward. Deep sentence embedding using long short-term memory networks: Analysis
and application to information retrieval. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(4):694–707, 2016.
Ankur P. Parikh, Oscar Tackstrom, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention
model for natural language inference. In Proceedings of EMNLP, 2016.
Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word
representation. In EMNLP, volume 14, pp. 1532–43, 2014.
Richard Socher, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning.
Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings
of the 2011 Conference on Empirical Methods in Natural Language Processing, pp. 151–161,
Edinburgh, Scotland, UK., July 2011. Association for Computational Linguistics. URL http:
//www.aclweb.org/anthology/D11-1014.
Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng,
and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment
treebank. In Proceedings of the conference on empirical methods in natural language processing
(EMNLP), volume 1631, pp. 1642. Citeseer, 2013.
Kai Sheng Tai, Richard Socher, and Christopher D. Manning. Improved semantic representations
from tree-structured long short-term memory networks. In Proceedings of ACL, pp. 1556–1566,
2015.
Ming Tan, Cicero dos Santos, Bing Xiang, and Bowen Zhou. Improved representation learning for
question answer matching. In Proceedings of ACL, pp. 464–473, Berlin, Germany, August 2016.
Association for Computational Linguistics. URL http://www.aclweb.org/anthology/
P16-1044.
Theano Development Team. Theano: A {Python} framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.0, 2016. URL http://arxiv.org/abs/1605.
02688.
Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and
language. arXiv preprint arXiv:1511.06361, 2015.
Wenpeng Yin and Hinrich Schütze. Convolutional neural network for paraphrase identification.
In Proceedings of the 2015 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, pp. 901–911, 2015.
12
Published as a conference paper at ICLR 2017
APPENDIX
A
P RUNED MLP
FOR
S TRUCTURED M ATRIX S ENTENCE E MBEDDING
As a side effect of having multiple vectors to represent a sentence, the matrix sentence embedding
is usually several times larger than vector sentence embeddings. This results in needing more parameters in the subsequent fully connected layer, which connects every hidden units to every units
in the matrix sentence embedding. Actually in the example shown in Figure 1, this fully connected
layer takes around 90% percent of the parameters. See Table 4. In this appendix we are going to
introduce a weight pruning method which, by utilizing the 2D structure of matrix embedding, is able
to drastically reduce the number of parameters in the fully connected hidden layer.
Inheriting the notation used in the main paper, let the matrix embedding M has a shape of r by u,
and let the fully connected hidden layer has b units. The normal fully connected hidden layer will
require each hidden unit to be connected to every unit in the matrix embedding, as shown in Figure
1. This ends up with r × u × b parameters in total.
However there are 2-D structures in the matrix embedding, which we should make use of. Each
row (mi in Figure 1) in the matrix is computed from a weighted sum of LSTM hidden states, which
means they share some similarities
To reflect these similarity in the fully connected layer, we split the hidden states into r equally sized
groups, with each group having p units. The i-th group is only fully connected to the i-th row in
the matrix representation. All connections that connects the i-th group hidden units to other rows
of the matrix are pruned away. In this way, Simillarity between different rows of matrix embedding
are reflected as symmetry of connecting type in the hidden layer. As a result, the hidden layer can
be interperated as also having a 2-D structute, with the number (r) and size (p) of groups as its
two dimensions (The M v in Figure 6). When the total number of hidden units are the same (i.e.,
q
p
Mv
u
Mh
r
u m1 m2
mi
M
r
Figure 6: Hidden layer with pruned weight connections. M is the matrix sentence embedding, M v
and M h are the structured hidden representation computed by pruned weights.
Table 4: Model Size Comparison Before and After Pruning
Yelp, Original, b=3000
Yelp, Pruned, p=150, q=10
Age, Original, b=4000
Age, Pruned, p=25, q=20
SNLI, Original, b=4000
SNLI, Pruned, p=300, q=10
Hidden layer
54M
2.7M
72M
822K
72M
5.6M
Softmax
15K
52.5K
20K
63.75K
12K
45K
13
Other Parts
1.3M
1.3M
1.3M
1.3M
22.9M
22.9M
Total
55.3M
4.1M
73.2M
2.1M
95.0M
28.6M
Accuracy
64.21%
63.86%
80.45%
77.32%
84.43%
83.16%
Published as a conference paper at ICLR 2017
r × p = b), this process prunes away (r − 1)/r of weight values, which is a fairly large portion when
r is large.
On the other dimension, another form of similarity exists too. For each vector representation mi in
M , the j-th element mij is a weighted sum of an LSTM hidden unit at different time steps. And for
a certain j-th element in all vector representations, they are summed up from a same LSTM hidden
unit. We can also reflect this similarity into the symmetry of weight connections by using the same
pruning method we did above. Thus we will have another 2-D structured hidden states sized u-by-q,
noted as M h in Figure 6.
Table 4 takes the model we use for yelp dataset as a concrete example, and compared the number of
parameters in each part of the model, both before and after pruning. We can see the above pruning
method drastically reduces the model size. Note that the p and q in this structure can be adjusted
freely as hyperparameters. Also, we can continue the corresponding pruning process on top of M v
and M h over and over again, and end up with having a stack of structured hidden layers, just like
stacking fully connected layers.
The subsequent softmax layer will be fully connected to both Mv and Mh , i.e., each unit in the
softmax layer is connected to all units in Mv and Mh . This is not a problem since the speed of
softmax is largely dependent of the number of softmax units, which is not changed.In addition, for
applications like sentiment analysis and textural entailment, the softmax layer is so tiny that only
contains several units.
Experimental results in the three datasets has shown that, this pruning mechanism lowers performances a bit, but still allows all three models to perform comparable or better than other models
compared in the paper.
B
D ETAILED S TRUCTURE OF THE M ODEL FOR SNLI DATASET
In Section 4.3 we tested our matrix sentence embedding model for the textual entailment task on
the SNLI dataset. Different from the former two tasks, the textual entailment task consists of a pair
of sentences as input. We propose to use a set of multiplicative interactions to combine the two
...
......
Gated Encoder
Fr
Fh
Fp
Premise
Hypothesis
Mh
w1
w2
w3
w4
...
...
Mp
wn
w1
w2
w3
w4
...
...
Figure 7: Model structure used for textual entailment task.
14
wn
Published as a conference paper at ICLR 2017
matrix embeddings extracted for each sentence. The form of multiplicative interaction is inspired
by Factored Gated Autoencoder (Memisevic, 2013).
The overall structure of our model for SNLI is dipicted in Figure 7. For both hypothesis and premise,
we extract their embeddings (Mh and Mp in the figure) independently, with a same LSTM and
attention mechanism. The parameters of this part of model are shared (rectangles with dashed orange
line in the figure).
Comparing the two matrix embeddings corresponds to the green dashed rectangle part in the figure,
which computes a single matrix embedding (Fr ) as the factor of semantic relation between the two
sentences. To represent the relation between Mh and Mp , Fr can be connected to Mh and Mp
through a three-way multiplicative interaction. In a three-way multiplicative interaction, the value
of anyone of Fr , Mh and Mp is a function of the product of the others. This type of connection is
originally introduced to extract relation between images (Memisevic, 2013). Since here we are just
computing the factor of relations (Fr ) from Mh and Mp , it corresponds to the encoder part in the
Factored Gated Autoencoder in Memisevic (2013). We call it Gated Encoder in Figure 7.
First we multiply each row in the matrix embedding by a different weight matrix. Repeating it
over all rows, corresponds to a batched dot product between a 2-D matrix and a 3-D weight tensor.
Inheriting the name in (Memisevic, 2013), we call the resulting matrix as factor. Doing the batched
dot for both hypothesis embedding and premise embedding, we have Fh and Fp , respectively.
Fh = batcheddot(Mh , Wf h )
(10)
Fp = batcheddot(Mp , Wf p )
(11)
Here Wf h and Wf p are the two weight tensors for hypothesis embedding and premise embedding.
The factor of the relation (Fr ) is just an element-wise product of Fh and Fp (the triangle in the
middle of Figure 7):
Fr = Fh
Fp
(12)
Here stands for element-wise product. After the Fr layer, we then use an MLP with softmax
output to classify the relation into different categlories.
15
| 9cs.NE
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
A Functional Regression approach to Facial
Landmark Tracking
arXiv:1612.02203v2 [cs.CV] 20 Sep 2017
Enrique Sánchez-Lozano, Georgios Tzimiropoulos*, Brais Martinez*,
Fernando De la Torre and Michel Valstar
Abstract—Linear regression is a fundamental building block in many face detection and tracking algorithms, typically used to predict
shape displacements from image features through a linear mapping. This paper presents a Functional Regression solution to the least
squares problem, which we coin Continuous Regression, resulting in the first real-time incremental face tracker. Contrary to prior work
in Functional Regression, in which B-splines or Fourier series were used, we propose to approximate the input space by its first-order
Taylor expansion, yielding a closed-form solution for the continuous domain of displacements. We then extend the continuous least
squares problem to correlated variables, and demonstrate the generalisation of our approach. We incorporate Continuous Regression
into the cascaded regression framework, and show its computational benefits for both training and testing. We then present a fast
approach for incremental learning within Cascaded Continuous Regression, coined iCCR, and show that its complexity allows real-time
face tracking, being 20 times faster than the state of the art. To the best of our knowledge, this is the first incremental face tracker that
is shown to operate in real-time. We show that iCCR achieves state-of-the-art performance on the 300-VW dataset, the most recent,
large-scale benchmark for face tracking.
Index Terms—Continuous Regression, Face Tracking, Functional Regression, Functional Data Analysis.
F
1
I NTRODUCTION
L
INEAR R EGRESSION is a standard tool in many Computer Vision problems, such as pose estimation [16],
and object tracking [49], and it is the foremost approach for
facial landmark detection, especially after the introduction
of the so-called Cascaded Regression method [11], [44], [47],
[53], [55]. The goal of facial landmark detection is to locate a
sparse set of facial landmarks in still images or videos. It is a
problem of wide interest to the Computer Vision community
because many higher level face analysis tasks, such as facial
expression recognition [15], and face recognition [60], are
critically affected by the performance of facial landmark
detection systems.
Arguably, the most popular method for facial landmark
detection is the Supervised Descent Method (SDM) [53], in
which a set of Linear Regressors are used in a cascade, each
approximating the average descent direction for a specific
input space. SDM generates samples by randomly perturbing the training data, i.e., each regressor is learnt by minimising the least squares error on a training set of annotated
images and known displacements. Many extensions to SDM
have been proposed [2], [23], [39], [47], yet little attention has
been paid to the limitations of the sampling-based approach
to linear regression, inherent to all aforementioned methods.
More specifically, it has been shown that the samplingbased approach is directly related to the following three
limitations:
•
•
•
E. Sánchez-Lozano, G. Tzimiropoulos, B.Martinez and M. Valstar are
with the School of Computer Science. University of Nottingham.
E-mail: [email protected]
F. De la Torre is with the Robotics Institute. Carnegie Mellon University.
*Both authors contributed equally
Manuscript received April 19, 2005; revised August 26, 2015.
•
•
•
The least squares formulation theoretically needs an
exponential number of samples with respect to the
dimensionality of the input vector in order to avoid
biased models.
Training an SDM model is computationally expensive as it requires the data to be sampled per cascade
level, with both memory and time needed increasing
dramatically with the number of perturbations.
Incremental learning for SDM has been reported to
be extremely slow [2], [54]. This limits the capacity
of SDM and its extensions for learning on-line and in
real-time, which for an application such as face tracking is impractical and hence highly unsatisfactory.
In order to overcome the aforementioned limitations, in
this paper we propose Continuous Regression, a Functional
Regression solution to the Least-Squares problem, and show
its application to real-time face tracking. Contrary to SDM,
the proposed Continuous Regression solution only requires
the data to be sampled at the ground-truth landmarks, i.e.,
no sampling at perturbations is required. This way, one
can sample and store the ground-truth data only once, and
then train each cascade level, or even a new model under
a different configuration extremely quickly. While the SDM
can take hours to train, our proposed Continuous Regression can train different models in seconds, once the data
is extracted. Moreover, and contrary to existing cascaded
regression methods, our continuous formulation allows for
real-time incremental learning, which as we demonstrate is
crucial for attaining state-of-the-art tracking performance on
the most challenging subset of the 300-VW dataset [42].
Continuous Regression approaches the standard least
squares formulation from a continuous perspective where
0162-8828 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/
publications/rights/index.html for more information.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
2
Fig. 1. Left: Difference between sampling-based regression and continuous regression. The continuous regression accounts for all the samples
within a neighbourhood, whereas sampling-based needs to sample the data from a given distribution. Right: Our new approach to Continuous
Regression can be seen as the inverse of a Monte Carlo sampling estimation. We will see that the probability density function defined by p(δs)
defines the volume within which samples are taken.
the target variable is treated as a continuum. Unlike existing
approaches to Functional Regression, which approximate
the input space by means of B-splines [27], Fourier Series
[34], or Radial Basis Functions [32], in this paper, we propose
a first-order Taylor expansion of the feature space. This way,
the input space becomes linear with respect to shape displacements yielding a closed-form solution. We go beyond
prior work in Functional Regression, and study the case of
correlated variables which naturally arises within landmark
localisation and tracking. To this end, we introduce a “data
term”, tasked with correlating the different dimensions over
which the problem is solved.
We then extend Continuous Regression within the cascaded regression framework, and demonstrate that training
our method has notable computational benefits compared to
that of standard SDM, without sacrificing accuracy. Finally,
we devise an approach for incremental learning in the
continuous domain, and show that its complexity allows for
real-time tracking. To the best of our knowledge, our Incremental Cascaded Continuous Regression method (iCCR), is
the first cascaded regression method that incorporates realtime incremental learning capabilities. Crucially, we show
the importance of incremental learning in achieving stateof-the-art results on the 300-VW dataset.
This manuscript re-formulates the Continuous Regression over uncorrelated variables, which has been presented
in [38], and further extends it to the domain of correlated
variables, presented in [40]. Building on [38], [40], we
present a complete formulation for the Continuous Regression problem, and then provide a geometric interpretation
and link to previous works. An overview of our work is
depicted in Fig. 1.
1.1
Contributions
Our main contributions are as follows:
•
•
We propose a complete formulation for Continuous
Regression including a closed-form solution based
on a first-order Taylor expansion of the input space.
Notably, prior work in Functional Regression has
only considered approximating the input space with
pre-defined basis functions.
We then go one step further in the classical Functional Regression formulation and propose a novel
optimisation problem for Continuous Regression
•
•
•
2
which allows the target variable to be correlated. We
also analytically derive the solution to this problem.
We incorporate Continuous Regression into the
Cascaded Regression framework, demonstrating its
computational benefits over SDM. We coin our
method Cascaded Continuous Regression (CCR).
We derive the incremental learning updates for CCR,
coined iCCR. We show that iCCR is capable of realtime incremental face tracking, being also an order of
magnitude less complex than previous works based
on incremental SDM.
We evaluate iCCR on the 300VW data set [42] and
show the importance of incremental learning in
achieving state-of-the-art performance.
R ELATED W ORK
Our work applies and extends concepts from Functional
Regression to the problem of face tracking. This section provides a brief description of Functional Regression methods
and how they are related to our work. We follow with a
review of related face alignment and tracking methods.
2.1
Prior Work in Functional Regression Analysis
Functional Data Analysis. Functional Data Analysis (FDA)
[33] is a branch of statistics that aims to model realisations
of stochastic processes as continuous functions [33]. FDA
assumes that observations, and/or responses, are outcomes
of continuous processes [29], and attempts to parameterise
functions by means of basis functions, rather than as a set
of samples. Several basis functions have been proposed
to date for different FDA domains, such as Radial Basis
Functions [32], Fourier Series [34], or B-splines [27].
A typical use of Functional Data Analysis is in the field
of Linear Regression, called Functional Regression [58], in
which either the responses or the observed data, or both, are
modelled as continuous functions. Functional Regression on
longitudinal data can be seen as an extension of Multivariate
Polynomial Regression to complex time series. Typically,
basis functions are used to approximate the observed data,
although some elegant solutions have also been proposed
to model responses as well [29]. However, in practice it is
not computationally feasible to approximate image features
by means of Radial Basis Functions or Fourier Series. We
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
propose to approximate the input features by their Taylor
expansion with respect to the variable of interest (variation
on shapes) to address this. To the best of our knowledge,
this is the first time that this approximation is applied in the
context of Functional Regression.
FDA in Computer Vision. To the best of our knowledge,
there are very few works that apply ideas from FDA to
Computer Vision problems, and none address the problem
of facial landmark localisation. The work that is perhaps
closest to ours (in some of its preliminary assumptions)
is PCA applied over continuous subspaces presented by
Levin and Shashua [22]. The main idea of [22] is to extend
the covariance matrix of the input data, necessary for the
computation of PCA, to all possible linear combinations
of the training samples. That is, given a set of training
examples (images) aRi , i = 1...d, the covariance is defined
as Cov(W ) = V1 a∈W aaT da, where W is the set of
convex combinations of training examples, and V is the
volume of the polytope defined by W . The inverse of the
volume outside the integral indicates that a uniform density
function when sampling the points a ∈ W is assumed.
In this work, we will set out how one can formulate the
Least-Squares problem for Continuous Regression using a
covariance matrix, providing a link between our work and
[22]. Our solution, however, does not necessarily need to be
formulated using a uniform distribution.
2.2
Prior Work on Face Tracking
Facial landmark tracking methods have often been adaptations of facial landmark detection methods. For example,
Active Shape Models [9], Active Appearance Models (AAM)
[11], [28], Constrained Local Models (CLM) [41], or the
Supervised Descent Method (SDM) [53] were all presented
as detection algorithms. It is thus natural to group facial
landmark tracking algorithms in the same way as detection
algorithms, i.e. into discriminative and generative methods.
On the generative side, AAMs have often been used for
tracking. Because model fitting relies on gradient descent,
it suffices to start the fitting from the previously tracked
shape1 . AAMs have been particularly regarded as very reliable for person specific tracking, but not for generic tracking
(i.e., tracking faces unseen during training) [17]. Recently
[45], [48] showed however that an improved optimisation
procedure and the use of in-the-wild images for training can
lead to satisfactorily person independent AAM. Eliminating
the piecewise-affine representation and adopting a partbased model led to the Gauss-Newton Deformable Part
Model (GN-DPM) [46] which is the state-of-the-art AAM.
Discriminative methods directly learn a mapping from
image features to shape displacements, effectively eliminating the need to minimise a reconstruction error. A
certain family of discriminative methods rely on training
local classifier-based models of appearance, with the local
responses being then constrained by a shape model [9], [13],
[41]. These algorithms are typically cast into the Constrained
Local Model (CLM) formulation [13], [26], [41].
One of the simplest tools for discriminative methods is
Linear Regression, which was firstly used to bypass the need
1. “Implementation tricks” can be found in [44], which provides a
very detailed account of how to optimise an AAM tracker
3
of an appearance model when fitting AAMs to images [11].
Herein, we refer to linear regression as a sampling-based
method, in which the training is performed by generating
random perturbations of the target variable (the shape displacements). Despite its simplicity and its fast computation,
a single regressor has been shown to be a poor choice to
account for all input variance. In order to overcome this
limitation, boosted regression methods were proposed in
[12], [43] where a set of weak regressors are added to a
master model, resulting in stronger regressors [12], [25].
Furthermore, in order to reduce the input variance of each
of the weak regressors, [16] proposed the use of weakly
invariant features. The use of such features requires boosted
regression be split into different stages, in which a different
regressor has to be learnt for each feature extraction step.
This novel approach was firstly exploited for the task of
pose estimation [16] and then was successfully applied to
face alignment in [6], [7]. However, the most successful
form of Cascaded Regression is the Supervised Descent
Method (SDM) [53], which applies a sampling-based linear
regression for each cascade level, and in which each level
can be seen as an average descent direction for the full shape
displacement estimation. The use of SIFT [24] features, and
the direct estimation of the full shape variation, made the
SDM the state-of-the-art method for face alignment. Successive works have further shown impressive efficiency [19],
[35] and reliable performance [47], [56].
However, how to best exploit discriminative cascaded regression for tracking and, in particular, how to efficiently integrate incremental learning, is still an open problem. More
importantly, SDM has not yet overcome one of the main
intrinsic limitations of sampling-based linear regression, to
wit, the generation of biased models. This is especially
important when it comes to tracking, since the initialisation
is given by the points of the previous frame, and therefore
there is no guarantee that the output of the model will
serve as a good initialisation point for the subsequent frame.
Eventually, this might cause the tracker to drift. Also, the
SDM is prone to fail for large out-of-plane head poses.
In order to overcome this limitation, recent methods have
proposed to pay special attention to the way the fitting
is initialised. In [55], a Global Supervised Descent Method
applies a partition to the shape space, in order to generate
a different cascaded regression model for each possible
domain of homogeneous descent directions.
Regarding evaluation and benchmarking, the only largescale face tracking benchmark that exists to date is the 300
Videos in the Wild (300-VW) challenge [42]. Two methods
were shown to clearly outperform all others. In [57], a
multi-view cascaded regression is used in which a different
model is trained for each different partition of the space of
poses. This approach, along with an accurate re-initialisation
system achieved the best results in the challenge. In [51], a
multi-stage regression approach is used, which guides the
initialisation process based on the localisation of a subset of
landmarks with a strong semantic meaning. The resulting
method achieved results very close to those of [57].
Finally, it is worth mentioning the advent of facial
landmark localisation using Deep Learning [5], [30], [52].
However, despite achieving impressive accuracy, these techniques are far from being able to operate at real-time speed.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
4
L INEAR R EGRESSION FOR FACE A LIGNMENT
This section reviews Cascaded Regression training using
Linear Regression, which is the learning problem Continuous Regression builds upon. In the following, we will
represent a face image by I. A face shape s ∈ R2n is a vector
describing the location of the n landmarks considered. We
will represent vector shapes as s = (x1 , . . . , xn , y1 , . . . yn )T .
We also define x = f (I, s) ∈ Rd as the feature vector
representing shape s. An asterisk represents the ground
truth, e.g., s∗j is the ground truth shape for image j .
Cascaded Regression [6], [16], [53] is an iterative regression method in which the output of regression at level l
is used as input for level l + 1, and each level uses image
features depending on the current shape estimate. The most
widely used form of Cascaded Regression is the Supervised
Descent Method (SDM, [53]), which learns a cascade of
linear regressors on SIFT features.
Training SDM for face alignment requires a set of M
images {Ij }j=1...M and their corresponding K perturbed
(0)
shapes {sj,k }j=1...M,k=1...K . When the fitting is carried out
from the output of a face detector, the different perturbations, k = 1 . . . K , are generated by applying a random
perturbation to the bounding boxes, capturing the variance
of the face detector. When the fitting is initialised from the
fitting of the previous frame, as is the case of tracking, the
training perturbations are generated by applying a random
perturbation to the ground-truth points, ideally capturing
the variance of facial shapes between consecutive frames.
For each cascade level l (starting with l = 0), a regressor
is trained and applied to the training data, resulting in an
update to the input shapes. Therefore, a new set of training
(l+1)
shapes {sj,k }j=1...M,k=1...K is generated for level l + 1,
in which the updated perturbed shapes are on average
closer to their corresponding ground-truth. This way, the
input variance and the training error are decreased as one
descends through the cascade.
Mathematically speaking, if we are given a set of pertur(l)
(l)
bations δsj,k = sj,k − s∗j between the initial shapes for level
l, and their corresponding ground-truth, a regressor R(l) is
learnt that aims to minimise the least-squares error:
arg min
R
M X
K
X
(l)
(l)
kδsj,k − Rf (Ij , s∗j + δsj,k )k22 ,
(1)
j=1 k=1
where the bias term is implicitly included by appending a 1 to the feature vector. If we store the features
(l)
xj,k = f (Ij , sj,k ) into matrix X(l) , and the corresponding
(l)
(l)
displacements δsj,k = sj,k −s∗j into matrix Y (l) , the solution
to Eq. (1) can be written in closed form:
T
T −1
R(l) = Y(l) (X(l) ) X(l) (X(l) )
.
(2)
The regressor R(l) is then applied to each of the extracted
features xj,k , and a new training set is generated, in which
the training shapes are now given as:
(l+1)
sj,k
=
(l)
sj,k
(l)
−R
(l)
f (Ij , sj,k ).
(3)
The process is repeated until the average of the differences
(l)
(l)
δsj,k = sj,k − s∗j no longer decreases. Typically, the training
error is minimised after 4 or 5 cascade levels.
During testing, the initial shape s(0) is forwarded to the
first regressor R(0) , to generate a new shape s(1) , which
is subsequently forwarded to R(1) . For each level l, a new
shape s(l+1) is generated by applying regressor R(l) to the
features extracted at s(l) :
s(l+1) = s(l) − R(l) f (I, s(l) ).
(4)
Provided the prediction statistics of R(l) are known
for all cascade levels, the training of Cascaded Regression
can be done independently for each of them (par-SDM,
[2]). As pointed out in [2], for a sufficiently large and
general training set, we can first train SDM, following the
sequential algorithm shown above. After SDM training, we
can compute the mean µ(l) and covariance Σ(l) of the
differences that result from updating the training shapes:
(l)
(l)
δsj,k = sj,k − s∗j . Then, for each cascade level, we can
(l)
generate a set of random perturbations δsj,k , drawn from a
Gaussian distribution N (µ(l) , Σ(l) ), and then compute each
R(l) using Eq. (2). This way, computing each regressor is
independent from the other levels, meaning that the training
can be performed in parallel.
As empirically shown by [2], such a model achieves
similar accuracy as a model trained in a sequential manner.
In addition, [2] showed that one can use these statistics to
train a par-SDM using a larger training set, without compromising accuracy with respect to a sequentially trained SDM.
This implies that, for a sufficiently large original training set,
the computed statistics properly model the displacements
for each level. We will see that our Continuous Regression
can be straightforwardly introduced within the Cascaded
Regression framework, with the advantage that our proposed method only needs the data to be sampled once,
whereas both the SDM and the par-SDM require a sampling
step for each Cascade Level.
4
C ONTINUOUS R EGRESSION
In this section we present a complete formulation and
solution for Continuous Regression, in which the shape displacements are treated as continuous variables. We propose
the use of a first-order Taylor expansion approximation of
the feature space, which yields a closed-form solution in
the continuous domain for the Least-Squares problem. We
will see that our Taylor-based closed-form solution depends
only on the features extracted from the ground-truth shapes
and their derivatives. Contrary to existing approaches, our
Continuous Regression solution does not need to sample
over perturbations.
Let us consider the Linear Regression problem shown in
Eq. (1). The extension of the Least-Squares problem to the
continuous domain of shape displacements is defined as:
Z a2n
M Z a1
X
arg min
...
kδs − Rf (Ij , s∗j + δs)k22 dδs, (5)
R
j=1 −a1
−a2n
where2 the line element is defined as δs
=
δx1 . . . δxn , δy1 . . . δyn , the limits are bounded, and
defined independently for each dimension. Without loss of
generality, we will assume that the limits are symmetric,
2. For the sake of clarity, we drop the dependence on the level
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
5
and the displacement assumed for each of the points is
therefore unbiased.
Unlike previous works on Functional Regression, which
approximate the input feature space by B-splines or Fourier
series, we propose to approximate the input feature space
by its first-order Taylor expansion:
f (Ij , s∗j + δs) ≈ f (Ij , s∗j ) + J∗j δs,
(6)
∂f (I ,s)
j
where J∗j =
|(s=s∗j ) ∈ Rd×2n , evaluated at s = s∗j ,
∂s
is the Jacobian of the feature representation of image Ij ,
with respect to shape coordinates s, at s∗j . To compute the
Jacobian of the image features (pixels) with respect to the xpositions, we have to use some approximation, e.g. the Sobel
filter. We can also compute the empirical derivative of other
complex features, such as SIFT or HOG. Indeed, gradientbased features, such as SIFT or HOG, are smoother than
pixels, and are therefore more feasible to be approximated
by a first-order Taylor expansion. Experiments will show
that the use of HOG features suffices to attain state of the
art performance, reinforcing the idea that a first-order Taylor
expansion is sufficient. Herein, we will define the empirical
derivative of the image features with respect to the x (or y )
coordinate, as follows:
∇fx =
∂f (I, x)
f (I, x + ∆x) − f (I, x − ∆x)
≈
∂x
2∆x
(7)
Then, the i-th column of the Jacobian J∗ , for i = 1, . . . , n, is
given by ∇fxi , whereas the (i+n)-th column of the Jacobian
is given by ∇fyi . In practice, ∆x is set to the minimum
displacement, i.e., ∆x = 1. Given the approximation of
Eq. (6), the original problem can be expressed as:
Z
kδs − R f (Ij , s∗j ) + J∗j δs k22 dδs ≈
(8)
Z
T
δs Aj δs + 2δsT bj + x∗j T RT Rx∗j dδs,
(9)
where Aj = (I − RJ∗j )T (I − RJ∗j ) and bj = J∗j T RT Rx∗j −
Rx∗j . Given the independence between each dimension, the
linear term equals to zero for symmetric limits. For the
quadratic term, the solution stems from the fact that
Z
XZ
δsT Aj δsdδs =
Auv
(10)
j δsu δsv dδs,
u,v
j=1
j=1
The solution shown in Eq. (13) only accounts for the
ground-truth features and their corresponding Jacobians.
We can see that, once the data has been sampled, training a
new regressor for different chosen limits is straightforward,
as it does not require recalculation of perturbations. This
implies that, for the Cascaded Regression framework, we
only need to sample the images once. This is a huge time
saving compared to the sampling-based SDM.
5 C ONTINUOUS R EGRESSION ON C ORRELATED
VARIABLES
This section extends Continuous Regression to the space
of correlated variables. We will show that it is possible
to find a closed-form solution for a general case, by assuming a different measure (Section 5.1). We then study
the validity of the Taylor expansion to approximate the
input features (Section 5.2). We later introduce our solution into the Cascaded Regression framework, and show
how our Cascaded Continuous Regression (CCR) implies
a huge training time saving (Section 5.3). We also study
the importance of the new solution (Section 5.4). Then, we
provide a theoretical link to previous work (Section 5.5) and
a geometrical interpretation (Section 5.6).
5.1
Reformulating Continuous Regression
The main problem of existing Functional Regression approaches is that the integral is defined with respect to
the Lebesgue measure, which implicitly assumes that an
(unnormalised) uniform distribution, in which samples are
uncorrelated, is used. Thus, it is not possible to solve it for
correlated dimensions. To overcome this limitation, we need
to solve the integral with respect to a different measure µ.
We can readily see that the proper measure is the probability
measure, i.e., µ = Pr. Therefore, the problem becomes
M Z
X
j=1 δs
where Auv
is the {u, v}-th entry of Aj , and su , sv refer
j
indistinctly to any pair of elements of s. We can see that
(
Z
a2u Q
if u = v
Auu
uv
j 3
k 2ak
(11)
Aj δsu δsv dδs =
0
otherwise
Q
2n
Let V =
be a diagonal
k 2ak , and let Σ ∈ R
a2
matrix the entries of which are defined by 3u . We can see
R
P
a2u
that δsT Aj δsdδs = V u Auu
j 3 . We can further observe
P
a2
that
Auu u = T r(Aj Σ). Similarly, we can see that
R ∗ T uT j ∗ 3
xj R Rxj dδs = x∗j T RT Rx∗j V . Then, the solution to
Eq. (9) is given by:
T r(Aj Σ)V + x∗j T RT Rx∗j V.
The minimisation of Eq. (13) w.r.t. R has a closed-form
as follows:
−1
M
M
X
X
T
T
T
R = Σ
J∗j
x∗j x∗j + J∗j ΣJ∗j .
(13)
(12)
kδs − Rf (Ij , s∗j + δs)k22 dPr(δs).
(14)
Applying the Riemannian form, we can write Eq. (14) as
M Z
X
j=1 δs
kδs − Rf (Ij , s∗j + δs)k22 p(δs)dδs,
(15)
where now p(δs) accounts for the pdf of the sampling
distribution. Interestingly, we can see that by formulating
the integral with respect to a probability measure, we are actually formulating Continuous Regression by means of the
average expected loss function, rather than from the classical
Functional Regression perspective, which minimises the empirical loss function. The expected loss is actually the function
from which the empirical loss is derived, for example, in
[53], the expected loss function is reduced to the empirical
loss by applying a Monte Carlo sampling approximation.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
Again, we will approximate the feature space by its firstorder Taylor expansion. The integrals are now generally
unbounded (they become bounded for e.g. a uniform distribution). Following the steps carried out in Section 4,
we expand the features, and group linear, quadratic, and
independent terms, with respect to δs.
Z
p(δs)kδs − Rf (Ij , s∗j + δs)k22 dδs ≈
δs
Z
≈
p(δs) δsT Aj δs + 2δsT bj + x∗j T RT Rx∗j dδs, (16)
δs
where, as previously, Aj = (I − RJ∗j )T (I − RJ∗j ) and bj =
J∗j T RT Rx∗j − Rx∗j . We only require the pdf p(δs) to be
parameterised by its mean µ and covariance Σ, so that, for
any symmetric matrix A, the following properties hold [4]:
Z
Z
p(δs)dδs = 1,
δsp(δs)dδs = µ,
δs
Z δs
p(δs)δsT Aδsdδs = T r(AΣ) + µT Aµ.
(17)
δs
It is then straightforward to compute the error, for the j -th
training example:
Z
p(δs)kδs − Rf (Ij , s∗j + δs)k22 dδs ≈
δs
T r(Aj Σ) + µT Aj µ + 2µT bj + x∗j T RT Rx∗j .
(18)
Again, R is obtained after minimising Eq. (18) w.r.t. R, in
which derivatives are obtained as follows:
∂
T r(Aj Σ) = 2RJ∗j ΣJ∗j T − 2ΣJ∗j T
∂R
∂ T
µ Aj µ = 2RJ∗j µµT J∗j T − 2µµT J∗j T
∂R
∂
2µT bj = 4Rx∗j µT J∗j T − 2µx∗j T
∂R
∂ ∗T T
x R Rx∗j = 2Rx∗j x∗j T .
∂R j
This leads to the closed-form solution:
M
X
R=
µx∗j T + (Σ + µµT )J∗j T ·
(19)
j=1
M
X
−1
x∗j x∗j T
+
2x∗j µT J∗j T
+ J∗j (Σ +
µµT )J∗j T
. (20)
j=1
The similarities between the closed form error in Eq. (18)
and that of Eq. (12) are clear. If p(δs) is defined as a
zero-mean uniform distribution with diagonal covariance
2
matrix with entries defined as a3 (i.e., defined for the limits
a1 , . . . , a2n ), then Eq. (20) and Eq. (13) would be the same.
It is worth highlighting that the solution does not depend
on the actual sampling pdf, but rather on its first and
second order moments, i.e. Continuous Regression bypasses
the question of which distribution should be used when
sampling the data.
The proposed formulation has important computational
advantages besides the theoretical differences between Continuous Regression and sampling-based methods like SDM.
In particular, once the training images have been sampled,
training a new regressor under a different configuration
requires very little computation as opposed to the samplingbased approach. The reason for this is that Continuous
Regression does not require the sampling process to be
repeated, instead it only changes the sampling statistics. For
instance, in the context of Cascaded Regression, given the
statistics for each level of the cascade with respect to the
ground truth, there is no need to sample the images more
than once, which yields a huge time saving. Furthermore,
we can see that there is a matrix form that can help compute
Eq. (20). Let us introduce the following shorthand notation:
µT
∗
∗
∗
M = [µ, Σ + µµT ], B = µ1 Σ+µµ
T , Dj = [xj , Jj ] and
∗
∗
∗
D̄ = [D1 , . . . , DM ]. Then:
T
M
−1
X
T
(21)
R = M
D∗j D̄∗ B̂D̄∗
j=1
where B̂ = B ⊗ IM . Since B̂ is sparse, computing Eq. (21),
once the data is given, is done in a matter of seconds.
5.2
Validity of Taylor expansion
In order to validate the Taylor approximation to represent
the image features, we have designed the following experiment. For a given set of initial statistics Σl , a set of
random perturbations f (I, s∗ + δs) was created, where δs
was drawn from N (0, Σl ). Then, we measured the distance
between the Taylor approximation to the ground-truth features, as well as to the features extracted from the same
image at different locations. Comparing these two distances
allows to distinguish whether the Taylor expansion gives a
reasonable estimation or a random estimation. The distance
between the Taylor approximation at δsi and the features
collected at δsj (including i = j ) is given as:
disti,j =
kf (I, s∗ + δsj ) − (f (I, s∗ ) + f 0 (I, s∗ )δsi )k
, (22)
kf (I, s∗ + δsj ) + (f (I, s∗ ) + f 0 (I, s∗ )δsi )k
This distance is expected to increase with δs. However, despite this increase, we expect samples to be distinguishable
from samples taken from other images.
We have used a random subset of 1000 images from
the training set (see Section 7 for further details). We have
evaluated the distances at the locations defined by a set
of 5 different Σl . More specifically, for l = 3 . . . 7, Σl is
defined as a diagonal matrix the elements of which are set
to 2l−1 (which represents the variance of the displacement
per landmark, in pixels). The results are shown in Fig. 2.
It can be seen that even when the variance is fairly high,
the approximated samples are still distinguishable from
samples collected at other locations (i.e. red lines have
lower error than corresponding blue lines). This validates
the Taylor approximation for a far long margin.
In order to also check the capabilities of a regressor
trained using a Taylor approximation, we have trained two
regressors for each of the Σl described above (one using
the sampling-based approach, and one using Continuous
Regression). For each of the images, a new set of random
locations was generated, and the regressors were used to
predict the shape displacements. Results are shown in Fig. 3.
It can be seen that both regressors have similar accuracy,
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Fig. 2. Results obtained for the Taylor approximation. Red curves correspond to the distances described in Eq. (22), when i = j , for a varying
set of initial perturbations ranging from l = 3 to l = 7. Blue curves
correspond to the distances described in Eq. (22) when i 6= j .
thus validating the assumptions made through the paper.
It has to be noted that these validations hold for such a
large range thanks to the smoothness and the constrained
environment that face tracking generally offers. When the
displacement is far from the ground-truth, the distance
between the approximation and the sampled data increases,
but so it does when collecting the data somewhere else.
Therefore, there is still the possibility of training a regressor
capable of reducing the error, as shown in Fig. 3.
7
advantage of our CCR with respect to SDM is that we only
need to sample the training images once, given that Eq. (20)
only needs to account for the ground-truth features and
their corresponding Jacobians. That is to say, the sampling
process in the CCR training is done only once.
Denoting the cost of sampling an image as O(q), for a
set of M images, K perturbations, and L cascade levels, the
total sampling cost of SDM is O(LKM q). In our proposed
CCR, the sampling is done only once, which is O(5qM ),
given that extracting the ground-truth features and Jacobians is O(5q). Therefore, our CCR training presents a computational advantage, with respect to SDM, of LK/5. In our
setting, L = 4 and K = 10, meaning that the Continuous
Regression sampling is 8 times faster (in FLOPS). However,
we have to note that, if we were to train a different model
for a different set of statistics, we would not need to
sample any data again. This means that, under a different
configuration, we only need Eq. (21) to be computed, which
can be done in seconds. In this case, we can see that the total
sampling cost of SDM training would remain O(LKM q),
while retraining a CCR does not entail any sampling, thus
having a null cost. In our configuration, M ≈ 7000 images,
meaning that the computational saving of retraining a
CCR, with respect to SDM, is O(106 ).
In the case that no prior SDM has been trained, Algorithm 1 summarises the training of CCR. Given the
training images and ground-truth points, CCR needs the
data to be pre-computed, by computing all D∗j . Then, the
process of CCR training basically consists of updating the
initial statistics µ0 and Σ0 (computed for a training set
of annotated videos), to generate a new regressor for each
cascade level. The statistics for each level are generated by
computing the difference of the outputs of the previous level
with respect to the corresponding ground-truth points. For
the first level, a set of initial shapes s0j are generated by
randomly perturbing the ground-truth with µ0 and Σ0 .
Algorithm 1: CCR training
1
2
3
4
5
6
Fig. 3. Results attained by a sampling-based linear regression (green
curves) and a regressor trained using Continuous Regression (red
curves) for the given initial perturbations (blue curves)
7
8
9
5.3
Data: {Ij , s∗j , s0j }j=1:M , µ0 and Σ0 ; L levels
Pre-compute: Extract D̄ = {D∗j }j=1:M
P
T
M
∗
Pre-compute: D̂ =
j=1 Dj
for l = 0 : L − 1 do
Compute Ml and Bl
−1
Rl = Ml D̂ D̄B̂l D̄T
(l)
(l+1)
Apply Rl to sj to generate sj
(l+1)
Compute distances δsj = sj
− s∗j
l+1
Update µ
= mean({δs}) and
Σl+1 = cov({δs})
end
Cascaded Continuous Regression
We can readily extend Continuous Regression to the Cascaded Regression framework. To do so, we point out that
the statistics defined in Eq. (20) correspond to those used
in the parallel SDM training proposed in [2], described
in Section 3. In this context, we can train each cascade
level using the corresponding µ(l) and Σ(l) . We coin our
method Cascaded Continuous Regression (CCR). The main
5.4
The impact of the data term
To empirically demonstrate the theoretical influence of the
data term which allows for correlated variables, we compared the performance of a CCR model trained using the
“correlated” solution in Eq. (20) (cor-CCR), with a CCR
model trained using the “uncorrelated” version described
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
8
the polytope approximation would not fit our problem. We
can compute the functional covariance matrix as follows:
XZ
Cov(X, X) =
p(δs)f (I, s∗j + δs)f (I, s∗j + δs)T dδs.
j
δs
(24)
If we approximate the input features by its first-order
Taylor expansion, we can see that:
f (I, s∗j + δs)f (I, s∗j + δs)T ≈
T
T
T
T
x∗j x∗j + x∗j δsT J∗j + J∗j δsx∗j + J∗j δsδsT J∗j .
(25)
R
T
We can use Eq. (17), and the fact that p(δs)δsδs =
Σ + µµT to further expand the covariance as:
X
Cov(X, X) =
x∗j x∗j T + 2x∗j µT J∗j T + J∗j (Σ + µµT )J∗j T ,
j
Fig. 4. Cumulative Error Distribution (CED) curve for both the uncor-CR
(red) and cor-CR (blue). Results are shown for the 49 points configuration. The contribution of a full covariance matrix is clear
(26)
which is, exactly, the invertible part of Eq. (18). We can
readily see that the covariance can be expressed as:
Cov(X, X) = D̄∗ B̂(D̄∗ )T .
in Section 4 (uncor-CCR). Both models were trained on the
Helen database [21], using Algorithm 1. In both cases, the
initial training set is generated by perturbing the groundtruth shapes according to the statistics measured at the
training partition of 300-VW dataset [42] (see Section 7 for
further details). In uncor-CCR, the statistics for each level
were forced to have zero-mean and diagonal covariance.
Otherwise the two versions are the same. We evaluated
both models in the most challenging subset of the 300VW test partition. Results in Fig. 4 show the Cumulative
Error Distribution (CED). It is immediately clear that the
contribution of cor-CCR is significant, and thus that shape
dimensions are significantly correlated. The result of this
experiment clearly demonstrates the importance of solving
Continuous Regression in spaces of correlated variables.
Similarly, we can see that:
XZ
Cov(X, Y ) =
δsf (I, s∗j + δs)T dδs ≈
Connection to related work
In this section, we link our work to that of [22]. More specifically, we show that one can first compute the functional
covariance in a similar fashion to that of [22], and then link
it to the Least-Squares solution. To do so, we first observe
that the normal equations shown in Eq. (13) can actually be
written by means of a covariance matrix as [50]3 :
Cov(X, Y ) = Cov(X, X)R,
(23)
where we used the term Cov to refer to the data covariance, as opposite to Σ referring to the covariance of the
sampling pdf. While [22] applies a polytope approximation
to estimate the covariance matrix, we rely on the first-order
Taylor expansion of the feature space. We have to note
that [22] approximates the covariance matrix by considering
combinations of training samples, while in our case we want
to generate perturbations on the training data, and therefore
3. Without loss of generality, we can assume that the input data and
the shape displacements are zero-mean
δs
j
≈
X
µx∗j T + (Σ + µµT )J∗j T .
(28)
j
Obtaining Eq. (20) from Eq. (23) is straightforward given
Eq. (28) and Eq. (27). Interestingly, we can take advantage of
this approach to generate our PCA models for each cascade
level. These models are used to reduce the dimensionality
of the feature space, and therefore need to be computed in
the perturbed image space. This way, we can easily generate
each of the models just by applying an eigendecomposition
of the functional covariance of Eq. (27).
5.6
5.5
(27)
Geometric interpretation
Finally, we want to elaborate on the meaning of the “data
term” in the Continuous Regression framework, i.e., the role
of a pdf given that an infinite set of samples is taken. To do
so, we can analyse the solution presented in this paper from
the Information Geometry perspective [1], [31], in which a
geometrical interpretation can be applied to the probability
measure. In the uncorrelated case, i.e. when using a diagonal
covariance matrix, the manifold of shape displacements is
a 2n-dimensional parallelepiped, spanned by the Cartesian
basis, defined for each of the shape dimensions. When we
extend the Continuous Regression to full covariance matrices, we are rotating the manifold of shape displacements
according to the eigenvectors of Σ, and the origin of the
coordinates axes is displaced to µ. Fig. 5 illustrates this
geometrical interpretation graphically.
6
I NCREMENTAL C ASCADED C ONTINUOUS R E -
GRESSION
This section introduces the process of updating a CCR
model using a new set of images and estimated shapes.
We will show that our incremental CCR (which we coin
iCCR) has a complexity that yields real-time performance.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Fig. 5. Difference between classical Functional Regression (Left) and
CR in correlated variables (Right). The green area represents the
volume within which samples are taken. Left: a diagonal covariance
2
matrix, with entries defined as a3 . Right: a full covariance matrix and
a non-zero mean vector. The sampling space is translated to the centre,
defined by µ, and rotated according to the eigenvectors of Σ.
9
turn to be tracked. Such affordances are not present when
tracking from a live video stream, e.g. a webcam.
Let us devise the incremental learning update rule for the
Continuous Regression. To do so, let us assume that we have
a regressor RT , trained using Eq. (21), on a training set T .
Also, we denote the covariance matrix for the training data
as VT := Cov(X, X) = D̄∗T B̂(D̄∗T )T . The incremental (also
known as online) learning aims to update RT with a set of S
new images, for which the ground-truth image features and
Jacobians are extracted. Let D∗S be the data corresponding
to the updating samples. We can define a forgetting factor
λ, which sets the influence of a given frame (the lower the
λ, the more influence a frame has on the update), and define
Wλ = λI2n+1 , with I2n+1 the (2n + 1)-dimensional matrix.
Then, the new regressor would be computed as:
T
M
X
−1
−1
∗
∗
RT ∪S = M
Dj + Wλ DS (VT ∪S ) ,
(29)
j=1
where
−1
(VT ∪S )
−1
T
= VT + D∗S BWλ−1 D∗S
.
(30)
We can apply the Woodbury identity [4] to Eq. (30):
VT ∪S −1 = VT −1 −
−1
VT −1 D∗S Wλ B−1 + D∗S T VT −1 D∗S
D∗S T VT −1 . (31)
Fig. 6. Overview of our incremental cascaded continuous regression
algorithm (iCCR). The originally model RT learnt offline is updated with
each new frame, thus sequentially adapting to the target face.
To the best of our knowledge, the iCCR update is the first
Cascaded Regression method achieving real-time capabilities. Results shown in Section 7 demonstrate the importance
of incremental learning to achieve state-of-the-art results,
thus illustrating the advantages of a real-time incremental
learning algorithm.
Once a regressor has been trained, we might want to
incorporate new images to the model, without the need of
recomputing Eq. (20). This is very important for the task of
face tracking, since it has been reported that generic models
perform worse than person-specific models. Since training a
person-specific model a priori is not an option in most use
cases, adding samples to a generic model as the face is being
tracked may help the model to track a specific person better
in future frames. An overview of the incremental learning
procedure is depicted in Fig. 6.
However, we can not retrain the full models by incorporating new images to the original training set, as it would be
far too slow, and therefore is not computationally tractable.
Some previous works [2], [54] have attempted to incorporate
the online learning to their current models, by applying
the recursive least squares to a trained model. However,
both methods are very slow, and thus are impractical for
real-time tracking. It should be noted that the good results
reported by [2], [54] rely on the fact that videos were tracked
offline, meaning that the update methods could take their
time to learn while following frames politely waited for their
This way, obtaining the inverse of the covariance matrix is computationally feasible. During tracking, the set S
consists of a tracked image with its estimated landmarks,
assuming these to have been correctly fitted. We can readily
see that the sampling cost of the iCCR update is fixed to
O(5q), and does not depend on the number of cascade levels, whereas the incremental SDM (iSDM), requires the sampling process to be carried out for each cascade level. We can
see that Eq. (31) needs to compute the inverse of the matrix
Wλ B−1 + D∗S T VT −1 D∗S , which is (2n + 1)-dimensional.
This operation has a complexity of O((2n + 1)3 ). That is, the
computational complexity of inverting that matrix is cubic
with respect to the number of points. However, we can alleviate this computation by working with a PDM [10], instead
of predicting over the points in a direct way. In a PDM, a
shape s is parameterised in terms of p = [q, c] ∈ Rm , where
q ∈ R4 represents the rigid parameters and c represents
the flexible shape parameters, so that s = tq (s0 + Bs c),
where t is a Procrustes transformation parameterised by q.
Bs ∈ R2n×m and s0 ∈ R2n are learnt during training and
represent the linear subspace of flexible shape variations.
This way, the Continuous Regression formulation would be
simply transformed into:
M Z
X
p(δp)kδp − Rf (Ij , p∗j + δp)k22 dδp,
(32)
j=1 δp
where the Jacobians now need to be computed with respect
to the PDM. To do so, it suffices to apply the chain rule:
∂f (I, p + δp) ∂s
∂f (I, p + δp)
=
,
∂p
∂s
∂p
where
∂s
∂p
(33)
can be analytically computed from the shape
model, and
∂f (I,p+δp)
∂s
is computed as in Eq. (7). This
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
10
way, the Continuous Regression solution is exactly the
same as shown in Eq. (20), with the Jacobians now being computed with respect to the PDM. Now, the matrix
Wλ B−1 + D∗S T VT −1 D∗S is m dimensional, and thus its
inversion is O(m3 ). In our experimental set-up, 2n = 132,
and m = 24. The computational saving is obvious.
Now, we can further analyse the update complexity,
noting that the most expensive step in the updating process
is not computing the inverse of Wλ B−1 + D∗S T VT −1 D∗S ,
−1
but rather computing DS T VTC , which has a complexity
2
cost of O(d m) (assuming a PDM is used). The overall
computational cost of updating a regressor is
O(3md2 ) + O(3m2 d) + O(m3 ).
(34)
We compare this updating cost with the sampling-based
SDM update proposed in [2], which was formulated as:
RT ∪S
Q
U
VT ∪S
= RT − RT Q + YS XTS VT ∪S
= XS UXTS VT
−1
=
IK + XTS VT XS
= VT − VT Q
(35)
E XPERIMENTAL RESULTS
To evaluate our proposed approach, we have tested our
CCR and iCCR methods on the most extensive benchmark
that exists to date: 300VW [42]. To compare our proposed
methods against top participants, we develop a fully automated system, which includes automated initialisation, as
well as a tool to detect that tracking is lost. This way, we can
compare our methods in exactly the same conditions that
were set up for the benchmark. We will show that our fully
automated system achieves state of the art results, as well
as the importance of incremental learning. Code for training
and testing is available on the lead author’s website.
7.1
where douter is the Euclidean distance between the points
defined for the outer corner of the eyes, measured on the
ground-truth. The results are summarised in the form of
Cumulative Error Distribution curves (CED), along with the
Area Under the Curve (AUC).
7.2
where IK is the K -dimensional identity matrix. We can
readily see that the main bottleneck in this update comes
from computing VT Q, which is O(d3 ), i.e., the iCCR update
is an order of magnitude faster than the sampling-based SDM
update proposed in [2]. This brings the cost of updating from
the ∼ 4 seconds reported in [2] down to ∼ 0.2 seconds,
which is the total update time in our sequential Matlab
implementation of iCCR.
7
[47]. All frames in which the face appears beyond profileview have been removed from the challenge evaluation, and
therefore are not considered in our experiments either.
Error measure: The error measure is that defined for 300VW.
It is computed for each frame by dividing the average
point-to-point Euclidean error by the inter-ocular distance,
understood to be the distance between the two outer eye
corners. More specifically, if ŝ is the estimated shape, and s∗
is the ground-truth shape, the RMSE is given as:
Pn p
(x̂i − x∗i )2 + (ŷi − yi∗ )2
RM SE = i=1
,
(36)
douter n
Experimental set-up
Test Data: All methods are tested on the 300VW [42] dataset,
which, to the best of our knowledge, is the most extensive
and recent benchmark in Face Tracking. The dataset consists
of 114 videos, each ∼ 1 minute long, divided into different
categories. 50 videos are used for training, whereas the
64 remaining videos are subdivided into three categories,
intended to represent increasingly unconstrained scenarios.
Category 1 contains 31 videos recorded in controlled conditions, whereas Category 2 includes 19 videos recorded under severe changes in illumination. Category 3 contains 14
videos captured in totally unconstrained scenarios. All the
videos are of a single-person, and have been annotated in
a semi-supervised manner using two different methods [8],
Training
Data: We use data from different datasets of static images
to construct our training set. Specifically, we use a total of
∼7000 images taken from Helen [21], LFPW [3], AFW [61],
IBUG [37], and a subset of MultiPIE [18]. We have used the
facial landmark annotations provided by the 300 faces in
the wild challenge [36], [42], as they are consistent across
datasets. Our models are trained for a 66-point configuration. In order to ensure the consistency of the annotations
with respect to the test set, the Shape Model is constructed
using the training partition of 300VW, and has 20 non-rigid
parameters and 4 rigid parameters [28].
Training: As shown in Algorithm 1, the training of CCR
starts from a set of given statistics (µ0 and Σ0 ). In our
experiments, the statistics are computed across the training
sequences, by computing the differences between consecutive frames. That is to say, µ0 and Σ0 are meant to
model how the shapes vary from frame to frame. The main
idea of this approach is to replicate a real scenario when
“perturbing” the ground-truth. Given that the training set
shows smoother variations w.r.t. the test set, we compute the
differences between frames separated by two and three time
steps. This way, higher displacements are also captured.
Features: We use a HOG [14] implementation, with a blocksize of 24 pixels, subdivided in 4 blocks, and 9 bins, resulting in a 9504-dimensional vector. We apply PCA on
the output, retaining 2000 dimensions, i.e., d = 2000. To
generate a PCA model per cascade level, we compute the
functional covariance shown in Eq. (27).
Cascade levels: In our experimental set-up, we have fixed
the number of Cascade levels to L = 4.
7.3
Testing
Initialisation: The tracker is initialised at the beginning of
each sequence, as well as each time the tracker is detected
to have lost a fitting (see details below). The initialisation
utilises the open-source dlib face detection to locate the face
bounding box (dlib.net), and then predicts the shape with
a Context-based SDM [39]. Then, a single CCR iteration is
carried out.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Fig. 7. CED’s for a set of different forgetting factors, along with the corresponding AUCs (Category 3). Red curve shows the Oracle. Magenta
shows the best factor found for that Category. Green curve shows the
results attained by the forgetting factor found on the training set. Blue
curves are results attained by other learning rates
Incremental learning update: The incremental learning
needs to filter frames to decide whether a fitting is suitable
or harmful to be used for the model updates. We use
a simple heuristic with a linear SVM, trained to decide
whether a particular fitting is “correct”, understood as being
below a threshold error. In order to avoid redundancy, once
a frame has been used to update the models, we consider
no more frames within a window of 3 consecutive frames.
Given that videos are of a single person, we do not need
to perform any face recognition to guarantee that updated
models keep tracking the same person with updated, and
thus personalised, models. For the iCCR update, we collect
person-specific statistics on a sliding window of 50 frames,
so that the person-specific variations are encoded in the
statistics used to update the model. The forgetting factor λ,
has been found through validation on the training set, and is
set to λ1 = 0.01, λ2 = 0.025, λ3 = 0.05 and λ4 = 0.1, where
λl is the learning rate for the l-th cascade level, and the
lower the λ, the more influence a frame has on the update.
We evaluate the sensitiveness of iCCR w.r.t. λ in Section 7.4.
Loss of tracking detection: Sometimes the tracker fails
to track a face. This occurs in particular when there are
significant occlusions or head-poses beyond profile. In these
cases, we need to detect that the tracker has lost a face, in
order to reinitialise it in the next frame from the face detector
output. We used the same SVM learnt for the incremental
learning update to also detect whether the tracker is lost
or not, by empirically selecting a suitable threshold for the
score returned by the SVM model.
7.4
Influence of the forgetting factor
The forgetting factor has been found through validation on
the training partition, and set up as shown above. In general,
it has been found that the top levels of the cascade are less
sensitive to an error in the update, and therefore a smaller
factor can be used (meaning a higher impact of the given
frame). However, for the last levels of the cascade, a low
11
factor can have a big impact if the frame to be updated is
not “entirely” correct. Since the score of an SVM is used to
detect whether a frame is suitable or not for the update, it is
likely that eventually a partially correct frame will be used.
To measure how well the found values generalise to the test
set, we have included an experiment in Category 3 showing
the performance of other forgetting factors. The overall
results are shown in Fig. 7. It can be seen that other values
would produce similar results. In addition, we include the
error given by an “Oracle” (red curve), understood as the
aggregated error for a scenario in which the best forgetting
factor is chosen for each video independently. That is to say,
red curve represents the results that would be attained if we
could choose the best forgetting factor per video (fixed for
the whole video). It can be seen that the red curve does not
show a significant improvement over the best blue curve,
meaning that in general the learning rate might have a
limited impact in the results.
7.5 Equivalence with SDM
In order to demonstrate the performance of our CCR
method compared to SDM, we have trained an SDM model
under the same conditions as for CCR. More specifically,
we have used the same training set, and the same statistics
measured for the first level. The training was then done in
a sequential manner, in a similar fashion to that of Algorithm 1. The number of cascade levels was again set to 4.
The PCA models were directly computed over the sampled
images, and comprise 2000 components. The SDM model
was tested in the test partition of 300-VW under exactly the
same conditions than those of CCR and iCCR. The results
attained by our SDM model are shown in Fig. 8.
7.6 Comparison with state of the art
We evaluate our fully automated tracking system in exactly
the same conditions as those used for the 300VW benchmark. We compare our method with respect to the top two
300VW results [51], [57]. We report results for the configurations of 49 points. In order to compare our method against
[51], [57], we asked these authors for their results, which
were kindly provided. For the fairest comparison with respect to the 300VW evaluation, we removed the frames that
were not considered for the evaluation (i.e. those including
faces beyond profile). To show the benefits of incremental
learning, we include the results of both our CCR and iCCR
systems in all curves, as well as our trained SDM. CED
results are shown in Fig. 8, and the AUC for each method is
shown in Table 1. It is interesting to remark that the methods
we compare with made contributions to the initialisation
of the tracker in each frame, or to failure detection. In our
case, even with a simple heuristic for both initialisation and
failure detection, we are capable of attaining comparable or
even superior performance over [51], [57]. The initialisation
for each frame in our tracking system is simply the output
of the previous frame. We do not include any multi-view
models or progressive initialisation, but instead ensure that
the statistics used to generate our models generalise well to
unseen scenarios.
It is unlikely that algorithms with slower running times
than iCCR, such as [51], [57], would attain their high accuracies if forced to process challenging videos in real-time. This
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
12
has important consequences for tracking streaming video,
in which frames must be tracked as fast as they arrive in
order to make use of the limited movement of facial points
between adjacent frames, something that is beneficial to
trackers of any ilk.
It is worth highlighting that CCR and SDM models
were reinitialised in ∼ 0.51% and ∼ 0.56% of total frames
(> 121000), respectively, whilst the iCCR was reinitialised
in only ∼ 0.34% of the frames. This illustrates the stability of our models, resulting in a simple yet effective
tracking system. Specially remarkable is the importance of
the Incremental Learning in challenging sequences, such
as those shown in Category 3, to achieve state of the art
results. As shown throughout the paper, the complexity of
our incremental learning allows for real-time implementation, something that could not be achieved by previous works on Cascaded Regression. Some visual examples
are depicted in Fig. 9. Code and videos can be found at
http://continuousregression.wordpress.com.
Method
iCCR
CCR
SDM
Yang et al. [57]
Xiao et al. [51]
Category 1
0.5978
0.5657
0.5617
0.5981
0.5814
Category 2
0.5918
0.5539
0.5522
0.6025
0.6093
Category 3
0.5141
0.4410
0.4380
0.4996
0.4865
TABLE 1
AUC for 49 points configuration for the different categories.
8
C ONCLUSION
We have proposed a novel formulation for the problem of
Continuous Least Squares in spaces of correlated variables,
namely Continuous Regression. We have shown the relation
that exists between the sampling distribution and the space
within which samples are taken, and have proposed a solution for general feature descriptors. We then incorporated
the Continuous Regression within the Cascaded Regression framework, and showed its performance is similar to
sampling-based methods. We devised an incremental learning approach using Continuous Regression, and showed
its complexity allows for real-time performance. Results
demonstrate the state-of-the-art performance in challenging
sequences, whilst being able to run in real-time. We also
want to note that while we tackle the facial landmark
tracking problem, cascaded regression has also been applied
to a wider range of problems such as pose estimation [16],
model-free tracking [49] or object localisation [59], thus
meaning the contributions of this paper can be extended
to other problems. Future work will include also the optimisation of the learning process, given that a better classifier
than a SVM would probably select better frames, and we
will also investigate the effect of the forgetting factor under
different configurations.
Fig. 8. CED’s for the 49-points configuration. Better seen in colour
also supported by the Vice-Chancellors Scholarship for Research Excellence of the University of Nottingham. The
work of Tzimiropoulos was supported in part by the EPSRC project EP/M02153X/1 Facial Deformable Models of
Animals. We are also grateful for access to the University of
Nottingham High Performance Computing Facility.
ACKNOWLEDGMENTS
R EFERENCES
The work of Sánchez-Lozano, Martinez and Valstar was
supported by the European Union Horizon 2020 research
and innovation programme under grant agreement No
645378, ARIA-VALUSPA. The work of Sánchez-Lozano was
[1]
[2]
Amari, S.: Differential-geometrical methods in statistics Springer
Science and Business Media (1985)
Asthana, A., Zafeiriou, S., Cheng, S., Pantic, M.: Incremental face
alignment in the wild. In: IEEE Conf. on Computer Vision and
Pattern Recognition. (2014)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
13
Fig. 9. Visual examples from Category 3 on 300VW. Upper rows show the CCR fitting (red points), whereas the lower rows show the iCCR fitting
(green points). The sequences on the left show a clear case where iCCR outperforms CCR thanks to its incremental learning capabilities (fast
movement in the upper sequence, low resolution in the sequence at the bottom). The sequences on the right show a failure case where the learned
features are not capable to outperform CCR. All qualitative videos and results can be found at https://continuousregression.wordpress.com/code.
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
Belhumeur, P.N., Jacobs, D.W., Kriegman, D.J., Kumar, N.: Localizing parts of faces using a consensus of exemplars. In: IEEE Conf.
on Computer Vision and Pattern Recognition. (2011) 545–552
Brookes, M.: The matrix reference manual (2011)
Bulat, A., Tzimiropoulos, G.: Convolutional aggregation of local
evidence for large pose face alignment. In: British Machine Vision
Conf. (2016)
Cao, X., Wei, Y., Wen, F., Sun, J.: Face alignment by explicit
shape regression. In: IEEE Conf. on Computer Vision and Pattern
Recognition. (2012) 2887–2894
Cao, X., Wei, Y., Wen, F., Sun, J.: Face alignment by explicit shape
regression. Int’l Journal of Computer Vision 107(2) (2014) 177–190
Chrysos, G.S., Antonakos, E., Zafeiriou, S., Snape, P.: Offline deformable face tracking in arbitrary videos. In: Int’l Conf. Computer
Vision - Workshop. (2015)
Cootes, T.F., Taylor, C.J., Cooper, D.H., Graham, J.: Active shape
models-their training and application. Comp. Vision and Image
Understanding 61(1) (1995) 38–59
Cootes, T.F., Taylor, C.J.: Statistical models of appearance for
computer vision (2004)
Cootes, T.F., Edwards, G.J., Taylor, C.J.: Active appearance models.
Trans. on Pattern Analysis and Machine Intelligence 23(6) (2001)
681–685
Cootes, T.F., Ionita, M.C., Lindner, C., Sauer, P.: Robust and accurate shape model fitting using random forest regression voting.
In: European Conf. on Computer Vision. (2012) 278–291
Cristinacce, D., Cootes, T.: Feature detection and tracking with
constrained local models. In: British Machine Vision Conf. (2006)
929–938
Dalal, N., Triggs, B.: Histograms of Oriented Gradients for human detection In: IEEE Conf. on Computer Vision and Pattern
Recognition. (2005)
Dhall, A., Goecke, R., Joshi, J., Sikka, K., Gedeon, T.: Emotion
recognition in the wild challenge 2014: Baseline, data and protocol.
In: Int’l Conf. on Multimodal Interaction. (2014) 461–466
Dollár, P., Welinder, P., Perona, P.: Cascaded pose regression. In:
IEEE Conf. on Computer Vision and Pattern Recognition. (2010)
1078–1085
Gross, R., Matthews, I., Baker, S.: Generic vs. person specific active
appearance models. Image and Vision Computing 23(11) (2005)
1080–1093
Gross, R., Matthews, I., Cohn, J., Kanade, T., Baker, S.: Multi-pie.
Image and Vision Computing 28(5) (2010) 807–813
Kazemi, V., Sullivan, J.: One millisecond face alignment with an
ensemble of regression trees. In: IEEE Conf. on Computer Vision
and Pattern Recognition. (2014)
[20] Kristan, M., Matas, J., Leonardis, A., Vojir, T., Pflugfelder, R.,
Fernandez, G., Nebehay, G., Porikli, F., Čehovin, L.: A novel
performance evaluation methodology for single-target trackers.
arXiv (2015)
[21] Le, V., Brandt, J., Lin, Z., Bourdev, L.D., Huang, T.S.: Interactive
facial feature localization. In: European Conf. on Computer Vision.
(2012) 679–692
[22] Levin, A., Shashua A.: Principal Component Analysis Over Continuous Subspaces and Intersection of Half-spaces In: European
Conference on Computer Vision (2002)
[23] Liu, L., Hu, J., Zhang, S., Deng, W.: Extended supervised descent
method for robust face alignment In: Asian Conf. on Computer
Vision Workshops. (2014) 71–84
[24] Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision 60(2) (2004) 91–
110
[25] Martinez, B., Valstar, M., Binefa, X., Pantic, M.: Local Evidence
Aggregation for Regression Based Facial Point Detection IEEE
Trans.Pattern Analysis and Machine Intelligence 35(5) (2013) 1149–
1163
[26] Martins, P., Henriques, J.F., Caseiro, R., Batista, J.: Bayesian
constrained local models revisited IEEE Trans. Pattern Analysis
and Machine Intelligence 38(4) (2016) 704–716
[27] Marx, B.D., Eilers, P.H.C.: Generalized Linear Regression on
Sampled Signals and Curves: A P-spline approach Techno. 41(1)
1–13
[28] Matthews, I., Baker, S.: Active appearance models revisited. Int’l
Journal of Computer Vision 60(2) (2004) 135–164
[29] Morris, J.S.: Functional Regression Annual Review of Statistics
and Its Applications 2 (2015) 321–359
[30] Peng, X., Feris, R., Wang, X., Metaxas, D.: A recurrent encoderdecoder for sequential face alignment In: European Conf. on
Computer Vision. (2016)
[31] Pennec, X.: Intrinsic statistics on Riemannian Manifolds: basic
tools for geometrical measurements J. Math Imagin Vis. 25, 127–
154. (2006)
[32] Quak, E., Sivakumar, N., Ward, J.D.: Least squares approximation
by radial functions SIAM J. Math. Anal. 24(4) (1993) 1043–1066
[33] Ramsay, J., Silverman, B.: Functional Data Analysis Springer
(2005).
[34] Ratliffe, S.J., Heller, G.Z., Leader, L.R. : Functional data analysis
with application to periodically stimulated foetal heart rate data.
I: Functional regression. Stat. in Med. 21(8) 1115–1127
[35] Ren, S., Cao, X., Wei, Y., Sun, J.: Face alignment at 3000 FPS
via regressing local binary features. In: IEEE Conf. on Computer
Vision and Pattern Recognition. (2014) 1685–1692
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[36] Sagonas, C., Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: 300
faces in-the-wild challenge: the first facial landmark localization
challenge. In: Int’l Conf. Computer Vision - Workshop. (2013)
[37] Sagonas, C., Tzimiropoulos, G., Zafeiriou, S., Pantic, M.: A semiautomatic methodology for facial landmark annotation. In: IEEE
Conf. on Computer Vision and Pattern Recognition - Workshops.
(2013)
[38] Sánchez-Lozano, E., De la Torre, F., González-Jiménez, D.: Continuous regression for non-rigid image alignment. In: European
Conf. on Computer Vision. (2012) 250–263
[39] Sánchez-Lozano, E., Martinez, B., Valstar, M.F.: Cascaded regression with sparsified feature covariance matrix for facial landmark
detection Pattern Recognition Letters 73 (2016) 19–25
[40] Sánchez-Lozano, E., Martinez, B., Tzimiropoulos, G., Valstar, M.:
Cascaded continuous regression for real-time incremental face
tracking In: European Conf. on Computer Vision. (2016) 645–661
[41] Saragih, J.M., Lucey, S., Cohn, J.F.: Deformable model fitting by
regularized landmark mean-shift. Int’l Journal of Computer Vision
91(2) (2011) 200–215
[42] Shen, J., Zafeiriou, S., Chrysos, G.S., Kossaifi, J., Tzimiropoulos,
G., Pantic, M.: The first facial landmark tracking in-the-wild
challenge: Benchmark and results. In: Int’l Conf. Computer Vision
- Workshop. (2015)
[43] Tresadern, P., Sauer, P., Cootes, T.: Additive update predictors in
Active Appearance Models In: British Machine Vision Conf. (2010)
[44] Tresadern, P., Ionita, M., Cootes, T.: Real-time facial feature
tracking on a mobile device. Int’l Journal of Computer Vision
96(3) (2012) 280–289
[45] Tzimiropoulos, G., Pantic, M.: Optimization problems for fast
AAM fitting in-the-wild. In: Int’l Conf. Computer Vision. (2013)
[46] Tzimiropoulos, G., Pantic, M.: Gauss-newton deformable part
models for face alignment in-the-wild. In: IEEE Conf. on Computer Vision and Pattern Recognition. (2014) 1851–1858
[47] Tzimiropoulos, G.: Project-out cascaded regression with an application to face alignment. In: IEEE Conf. on Computer Vision and
Pattern Recognition. (2015) 3659–3667
[48] Tzimiropoulos, G., Pantic, M.: Fast algorithms for fitting Active
Appearance Models to unconstrained images Int’l Journal of
Computer Vision (in press)
[49] Wang, X., Valstar, M., Martinez, B., Khan, M.H., Pridmore, T.: Trictrack: Tracking by regression with incrementally learned cascades.
In: Int’l Conf. Computer Vision. (2015)
[50] Wang, J.L., Chiou, J.M., Muller, H.G.: Review of functional data
analysis arXiv (2015)
[51] Xiao, S., Yan, S., Kassim, A.: Facial landmark detection via progressive initialization. In: ICCV-W 1st Facial Landmark Tracking
in-the-wild challenge and workshop (2015)
[52] Xiao, S., Feng, J., Xing, J., Lai, H., Yan, S., Kassim, A.: Robust facial
landmark detection via recurrent attentive-refinement networks
In: European Conf. on Computer Vision. (2016)
[53] Xiong, X., De la Torre, F.: Supervised descent method and its
applications to face alignment. In: IEEE Conf. on Computer Vision
and Pattern Recognition. (2013)
[54] Xiong, X., la Torre, F.D.: Supervised descent method for solving nonlinear least squares problems in computer vision. arXiv
abs/1405.0601 (2014)
[55] Xiong, X., De la Torre, F.: Global supervised descent method In:
IEEE Conf. on Computer Vision and Pattern Recognition (2015)
[56] Yan, J., Lei, Z., Yi, D., Li, S.: Learn to combine multiple hypotheses
for accurate face alignment. In: Int’l Conf. Computer Vision Workshop. (2013) 392–396
[57] Yang, J., Deng, J., Zhang, K., Liu, Q.: Facial shape tracking via
spatio-temporal cascade shape regression In: ICCV-W, 1st Facial
Landmark Tracking in-the-wild challenge and workshop (2015)
[58] Yao, F., Muller, H.G., Wang, J.L.: Functional Linear regression
analysis for longitudinal data The Annals of Statistics 33(2) (2005)
2873–2903
[59] Yan, J., Lei, Z., Yang, Y., Li, S.: Stacked deformable part model with
shape regression for object part localization. In: European Conf.
on Computer Vision. (2014) 568–583
[60] Zhou, S., Krueger, V., Chellappa, R.: Probabilistic recognition of
human faces from video. Comp. Vision and Image Understanding
91(12) (2003) 214 – 245
[61] Zhu, X., Ramanan, D.: Face detection, pose estimation, and
landmark localization in the wild. In: IEEE Conf. on Computer
Vision and Pattern Recognition. (2012) 2879–2886
14
Enrique Sánchez-Lozano is a Research Fellow
in the Computer Vision Lab at the University of
Nottingham, where he received his PhD in Computer Science (2017). He received his M.Eng.
in Telecommunication Engineering (2009) and
M.Sc. in Signal Theory and Communications
(2011) from the University of Vigo (Spain). He
spent 6 months as a visiting scholar at the Human Sensing Lab, with Fernando De la Torre,
between 2011 and 2012. He was given a Fundación Pedro Barrié Scholarship to do a 5 month
research stay at the Computer Vision Lab (CVL), University of Nottingham, with Michel Valstar, between 2013 and 2014.
Georgios (Yorgos) Tzimiropoulos received the
M.Sc. and Ph.D. degrees in Signal Processing
and Computer Vision from Imperial College London, U.K. He is Assistant Professor with the
School of Computer Science at the University of
Nottingham, U.K. Prior to this, he was Assistant
Professor with the University of Lincoln, U.K. and
Senior Researcher in the iBUG group, Department of Computing, Imperial College London.
He is currently Associate Editor of the Image and
Vision Computing Journal. His main research
interests are in the areas of face and object recognition, alignment and
tracking, and facial expression analysis.
Brais Martinez (M’10) is a Research Fellow with
the Computer Vision Laboratory at the University of Nottingham, UK. He has previously obtained his PhD in Computer Science from the
Universitat Autonoma de Barcelona, and held
a postdoctoral research position at the Imperial
College London. He has worked on the topics
of Computer Vision and Pattern Recognition,
mainly focusing on model-free tracking and on
face analysis problems such as face alignment,
face detection and facial expression recognition,
publishing his findings on these topics on a variety of authoritative
venues such as CVPR, ICCV, ECCV or TPAMI.
Fernando De la Torre is an Associate Research
Professor in the Robotics Institute at Carnegie
Mellon University. He received his B.Sc. degree
in Telecommunications, as well as his M.Sc. and
Ph. D degrees in Electronic Engineering from
La Salle School of Engineering at Ramon Llull
University, Barcelona, Spain in 1994, 1996, and
2002, respectively. His research interests are
in the fields of Computer Vision and Machine
Learning. Currently, he is directing the Component Analysis Laboratory (http://ca.cs.cmu.edu)
and the Human Sensing Laboratory (http://humansensing.cs.cmu.edu)
at Carnegie Mellon University. He has over 150 publications in referred
journals and conferences and is Associate Editor at IEEE TPAMI. He has
organized and co-organized several workshops and has given tutorials
at international conferences on component analysis.
Michel Valstar (M’09) is an Associate Professor
in the Computer Vision and Mixed Reality Labs
at the University of Nottingham. He received his
masters degree in Electrical Engineering at Delft
University of Technology in 2005 and his PhD in
computer science with the intelligent Behaviour
Understanding Group (iBUG) at Imperial College
London in 2008. His main interest is in automatic
recognition of human behaviour. In 2011 he was
the main organiser of the first facial expression
recognition challenge, FERA 2011. In 2007 he
won the BCS British Machine Intelligence Prize for part of his PhD
work. He has published technical papers at authoritative conferences
including CVPR, ICCV and SMC-B and his work has received popular
press coverage in New Scientist and on BBC Radio.
| 1cs.CV
|
Learning Depth-Three Neural Networks in Polynomial Time
Surbhi Goel∗ and Adam Klivans†
arXiv:1709.06010v3 [cs.DS] 24 Oct 2017
Department of Computer Science, University of Texas at Austin
{surbhi,klivans}@cs.utexas.edu
October 25, 2017
Abstract
We give a polynomial-time algorithm for learning neural networks with one hidden layer of
sigmoids feeding into any smooth, monotone activation function (e.g., sigmoid or ReLU). We
make no assumptions on the structure of the network, and the algorithm succeeds with respect
to any distribution on the unit ball in n dimensions (hidden weight vectors also have unit norm).
This is the first assumption-free, provably efficient algorithm for learning neural networks with
more than one hidden layer.
Our algorithm– Alphatron– is a simple, iterative update rule that combines isotonic regression
with kernel methods. It outputs a hypothesis that yields efficient oracle access to interpretable
features. It also suggests a new approach to Boolean function learning via smooth relaxations of
hard thresholds, sidestepping traditional hardness results from computational learning theory.
Along these lines, we give improved results for a number of longstanding problems related to Boolean concept learning, unifying a variety of different techniques. For example,
we give the first polynomial-time algorithm for learning intersections of halfspaces with a margin (distribution-free) and the first generalization of DNF learning to the setting of probabilistic
concepts (queries; uniform distribution). Finally, we give the first provably correct algorithms
for common schemes in multiple-instance learning.
∗
†
Supported by University of Texas at Austin Graduate School Summer 2017 Fellowship.
Supported by NSF Algorithmic Foundations Award AF-1717896.
1
1
Introduction
Giving provably efficient algorithms for learning neural networks is a longstanding and fundamental challenge in the theory of machine learning. Despite the remarkable achievements obtained in
practice from applying tools for learning neural networks, surprisingly little is known from a theoretical perspective. In fact, much theoretical work has led to negative results showing that– from a
worst-case perspective– even learning the simplest architectures seems computationally intractable
[LSSS14, SVWX17]. For example, there are known hardness results for agnostically learning a single
halfspace (learning a halfspace with adversarial noise) [KS09b, Dan16].
As such, much work has focused on finding algorithms that succeed after making various restrictive assumptions on both the network’s architecture and the underlying marginal distribution.
Recent work gives evidence that for gradient-based algorithms these types of assumptions are actually necessary [Sha16]. In this paper, we focus on understanding the frontier of efficient neural
network learning: what is the most expressive class of neural networks that can be learned, provably,
in polynomial-time without taking any additional assumptions?
1.1
Our Results
We give a simple, iterative algorithm that efficiently learns neural networks with one layer of sigmoids
feeding into any smooth, monotone activation function (for example, Sigmoid or ReLU). Both the
first hidden layer of sigmoids and the output activation function have corresponding hidden weight
vectors. We assume nothing about these weight vectors other than the standard normalization that
they have 2-norm equal to one. The algorithm succeeds with respect to any distribution on the unit
ball in n dimensions. This is the first provably efficient, assumption-free result for learning neural
networks with more than one hidden layer.
Our algorithm, which we call Alphatron, combines the expressive power of kernel methods with an
additive update rule inspired by work from isotonic regression. Alphatron also outputs a hypothesis
that gives efficient oracle access to interpretable features. That is, if the output activation function
is u, Alphatron constructs a hypothesis of the form u(f (x)) where f is an implicit encoding of
products of features from the instance space, and f yields an efficient algorithm for random access
to the coefficients of these products.
More specifically, we obtain the following new supervised learning results by choosing an appropriate kernel function in conjunction with Alphatron:
• Let c(x1 , . . . , xn ) be any feedforward neural network with one hidden layer of sigmoids of size
k feeding into any activation function u that is monotone and L-Lipschitz. Given independent
draws (x, y) from Sn−1 × [0, 1] with E[y|x] = c(x), we obtain an efficiently computable hypothesis u(f (x)) such that E[(c(x) − u(f (x)))2 ] ≤ ǫ with running time and sample complexity
poly(n, k, 1/ǫ, L) (the algorithm succeeds with high probability).
• We obtain the first efficient PAC algorithm for learning intersections of polynomially many
halfspaces (with a margin) with respect to any distribution on Sn−1 (prior work due to [KS08]
gave a quasipolynomial-time algorithm). We show that this is a special case of a more general
class of Boolean learning problems where the goal is to learn Lipschitz-bounded combinations
of Boolean functions in the probabilistic concept model due to Kearns and Schapire [KS94].
In this framework, we can learn smooth combinations of halfspaces (with a margin) whose
sample complexity is independent of the number of halfspaces.
2
• We give the first generalization of well-known results for PAC learning DNF formulas with
respect to the uniform distribution (given query access to the unknown DNF) to the setting
of probabilistic concepts. Concretely, we give a query algorithm– KMtron– that learns any
random variable whose conditional mean is a smooth, monotone combination of functions of
bounded L1 -norm with respect to the uniform distribution on {0, 1}n1 . It is easy to see this
captures the function class of polynomial-size DNF formulas.
• We give the first provably efficient algorithms for nontrivial schemes in multiple instance
learning (MIL). Consider an MIL scheme where a learner is given a set or bag of instances
x1 , . . . , xt , and the learner is told only some function of their labels, namely u(c(x1 ), . . . , c(xt ))
for some unknown concept c and combining function u. We give the first provably efficient
algorithms for correctly labeling future bags even if the instances within each bag are not
identically distributed. Our algorithms hold if the underlying concept c is sigmoidal or a
halfspace with a margin. If the combining function averages label values (a common case), we
obtain bounds that are independent of the bag size.
Almost all of our results holds in the probabilistic concept model of learning due to Kearns
and Schapire [KS94] and only require that the conditional mean of the label y given instance x is
approximated by bounded-norm elements from a Reproducing Kernel Hilbert Space (RKHS). We
learn specifically with respect to square-loss, though this will imply polynomial-time learnability for
most commonly studied loss functions.
1.2
Relationship to Traditional Boolean Function Learning
PAC learning simple Boolean concept classes has proved challenging. For example, the best known
distribution-free algorithm for learning an intersection of just two halfspaces runs in exponential
time in the dimension. For learning intersections of polynomially many halfspaces, there are known
hardness results based on cryptographic assumptions [KS09b] or the hardness of constraint satisfaction problems [Dan16]. A source of angst in machine learning theory is that these hardness results
do not square with recent practical successes for learning expressive function classes.
A key conceptual aspect of this work is to shift from the PAC model to the probabilistic concept
model. We hope to revive the probabilistic concept model as a fertile area for constructing supervised
learning algorithms, as it can handle both Boolean and real-valued concepts. Additionally, this
model has the following interesting benefit that we feel has been overlooked: it allows for Boolean
learning problems where an output hypothesis can answer “don’t know” by giving the value 1/2. As
mentioned above, hardness results from computational learning theory indicate that simple Boolean
function classes such as intersections of halfspaces can encode pseudorandom outputs. For these
classes, PAC learning is hopeless. On the other hand, in the probabilistic concept model, we measure
error with respect to square-loss. In this model, for subsets of inputs encoding pseudorandom labels,
a hypothesis may simply output 1/2, which is essentially the optimal strategy.
The models we study in this paper still capture Boolean learning problems. More specifically,
let (x, y) be a random draw from a distribution where y ∈ {0, 1} with E[y|x] = c(x). If c always
outputs 0 or 1, then we are in the typical PAC scenario. On the other hand, c may be a realvalued function. Our approach is to consider Boolean learning problems where the conditional
1
Since we only put a requirement on the conditional mean, this model naturally inherits strong noise-tolerant
properties that previous work does not handle.
3
mean function is computed by a real-valued neural network. These networks can be viewed as
relaxations of Boolean function classes.
For example, one natural relaxation of an AND of k Boolean inputs would be a piecewise-linear
combining function u that is 0 for all inputs in [0, . . . , k − 1], equal to 1 on input k, and a line
that interpolates between u(k − 1) and u(k). Additionally, we could relax any halfspace sign(w · x)
defined on the unit sphere to σ(w · x) where σ(z) = 1+e1−z , a sigmoid. Although PAC learning
an intersection of halfspaces seems out of reach, we can give fully polynomial-time algorithms for
learning sums of polynomially many sigmoids feeding into u as a probabilistic concept.
1.3
Our Approach
The high-level approach is to use algorithms for isotonic regression to learn monotone combinations of functions approximated by elements of a suitable RKHS. Our starting point is the Isotron
algorithm, due to Kalai and Sastry [KS09a], and a refinement due to Kakade, Kalai, Kanade and
Shamir [KKKS11] called the GLMtron. These algorithms efficiently learn any generalized linear
model (GLM): distributions on instance-label pairs (x, y) where the conditional mean of y given
x is equal to u(w · x) for some (known) smooth, non-decreasing function u and unknown weight
vector w. Their algorithms are simple and use an iterative update rule to minimize square-loss, a
non-convex optimization problem in this setting. Both of their papers remark that their algorithms
can be kernelized, but no concrete applications are given.
Around the same time, Shalev-Shwartz, Shamir, and Sridharan [SSSS11] used kernel methods
and general solvers for convex programs to give algorithms for learning a halfspace under a distributional assumption corresponding to a margin in the non-realizable setting (agnostic learning). Their
kernel was composed by Zhang et al. [ZLJ16] to obtain results for learning sparse neural networks
with certain smooth activations, and Goel et al. [GKKT16] used a similar approach in conjunction
with general tools from approximation theory to obtain learning results for a large class of nonlinear
activations including ReLU and Sigmoid.
1.4
Our Algorithm
We combine the two above approaches into an algorithm called Alphatron that inherits the best
properties of both: it is a simple, iterative update rule that does not require regularization, and it
learns broad classes of networks whose first layer can be approximated via an appropriate feature
expansion into an RKHS. It is crucial that we work in the probabilistic concept model. Even
learning a single ReLU in the non-realizable or agnostic setting seems computationally intractable
[GKKT16].
One technical challenge is handling the approximation error induced from embedding into an
RKHS. In some sense, we must learn a noisy GLM. For this, we use a learning rate and a slack
variable to account for noise and follow the outline of the analysis of GLMtron (or Isotron). The
resulting algorithm is similar to performing gradient descent on the support vectors of a target
element in an RKHS. Our convergence bounds depend on the resulting choice of kernel, learning rate,
and quality of RKHS embedding. We can then leverage several results from approximation theory
and obtain general theorems for two different notions of RKHS approximation: 1) the function class
can be uniformly approximated by low-norm elements of a suitable RKHS or 2) the function class
is separable (similar to the notion of a margin) by low-norm elements of an RKHS.
4
For generalizing uniform-distribution DNF learning algorithms to the probabilistic concept setting, we re-interpret the KM algorithm for finding large Fourier coefficients [KM93] as a query
algorithm that gives accurate estimates of sparse, high-dimensional gradients. For the case of
square-loss with respect to the uniform distribution on the hypercube, we can combine these estimates with a projection operator to learn smooth, monotone combinations of L1 -bounded functions
(it is easy to see that DNF formulas fall into this class).
For Multiple Instance Learning (MIL), we observe that the problem formulation is similar to
learning neural networks with two hidden layers. We consider two different types of MIL: deterministic (the Boolean label is a deterministic function of the instance labels) and probabilistic (the
Boolean label is a random variable whose mean is a function of the instance labels). We make
use of some further kernel tricks, notably the mean-map kernel, to obtain a compositional feature
map for taking averages of sets of instances. This allows us to prove efficient run-time and sample
complexity bounds that are, in some cases, independent of the bag size.
1.5
Related Work
The literature on provably efficient algorithms for learning neural networks is extensive. In this
work we focus on common nonlinear activation functions: sigmoid, ReLU, or threshold. For linear
activations, neural networks compute an overall function that is linear and can be learned efficiently
using any polynomial-time algorithm for solving linear regression. Livni et al. [LSSS14] observed
that neural networks of constant depth with constant degree polynomial activations are equivalent
to linear functions in a higher dimensional space (polynomials of degree d are equivalent to linear
functions over nd monomials). It is known, however, that any polynomial that computes or even
ǫ-approximates a single ReLU requires degree Ω(1/ǫ) [GKKT16]. Thus, linear methods alone do
not suffice for obtaining our results.
The vast majority of work on learning neural networks takes strong assumptions on either the
underlying marginal distribution, the structure of the network, or both. Works that fall into these
categories include [KOS04, KM13, JSA15, SA14, ZPS17, ZLJ16, DFS16, ZSJ+ 17, GK17]. In terms
of assumption-free learning results, Goel et al. [GKKT16] used kernel methods to give an efficient,
agnostic learning algorithm for sums of sigmoids (i.e., one hidden layer of sigmoids) with respect to
any distribution on the unit ball.
Another line of work related to learning neural networks focuses on when local minima found by
gradient descent are actually close to global minima. In order to give polynomial-time guarantees
for finding a global minimum, these works require assumptions on the underlying marginal or the
structure of the network (or both) [CHM+ 15, Kaw16, BG17, SC16, LY17, Sol17]2 . All of the
problems we consider in this paper are non-convex optimization problems, as it is known that a
single sigmoid with respect to square-loss has exponentially many bad local minima [AHW96].
For classical generalization and VC dimension bounds for learning neural networks we refer the
reader to Anthony and Bartlett [AB99] (the networks we consider in this paper over the unit ball
have VC dimension Ω(nk) where k is the number of hidden units in the first layer).
2
In their setting, even the case of linear activations are interesting as the goal is explicitly recovering the parameters
of the hidden units’ weight vectors.
5
1.6
Organization
In the preliminaries we define the learning models we use and review the core tools we need from
kernel methods and approximation theory. We then present our main algorithm, Alphatron, and
give a proof of its correctness. We then combine Alphatron with various RKHS embeddings to
obtain our most general learning results. Using these general results, we subsequently describe how
to obtain all of our applications, including our main results for learning depth-three neural networks.
2
Preliminaries
Notation. Vectors are denoted by bold-face and || · || denotes the standard 2-norm of the vector.
We denote the space of inputs by X and the space of outputs by Y. In our paper, X is usually the
unit sphere/ball and Y is [0, 1] or {0, 1}. Standard scalar (dot) products are denoted by a · b for
vectors a, b ∈ Rn , while inner products in a Reproducing Kernel Hilbert Space (RKHS) are denoted
by ha, bi for elements a, b in the RKHS. We denote the standard composition of functions f1 and
f2 by f1 ◦ f2 .
2.0.1
Learning Models
We consider two learning models in our paper, the standard Probably Approximately Correct (PAC)
learning model and a relaxation of the standard model, the Probabilistic Concept (p-concept) learning model. For completeness, we define the two models and refer the reader to [Val84, KS90] for a
detailed explanation.
Definition 1 (PAC Learning [Val84]). We say that a concept class C ⊆ {0, 1}X is Probably Approximately Correct (PAC) learnable, if there exists an algorithm A such that for every c ∈ C,δ, ǫ > 0
and D over X , if A is given access to examples drawn from D and labeled according to c, A outputs
a hypothesis h : X → {0, 1}, such that with probability at least 1 − δ,
P rx∼D [h(x) 6= c(x)] ≤ ǫ.
(1)
Furthermore, we say that C is efficiently PAC learnable to error ǫ if A can output an h satisfying
the above with running time and sample complexity polynomial in n, 1/ǫ, and 1/δ.
Definition 2 (p-concept Learning [KS90]). We say that a concept class C ⊆ Y X is Probabilistic
Concept (p-concept) learnable, if there exists an algorithm A such that for every δ, ǫ > 0, c ∈ C and
distribution D over X × Y with E[y|x] = c(x) we have that A, given access to examples drawn from
D, outputs a hypothesis h : X → Y, such that with probability at least 1 − δ,
E(x,y)∼D [(h(x) − c(x))2 ] ≤ ǫ.
(2)
Furthermore, we say that C is efficiently p-concept learnable to error ǫ if A can output an h satisfying
the above with running time and sample complexity polynomial in n, 1/ǫ, and 1/δ.
Here we focus on square loss for p-concept since an efficient algorithm for square-loss implies
efficient algorithms of various other standard losses.
6
2.0.2
Generalization Bounds
The following standard generalization bound based on Rademacher complexity is useful for our
analysis. For a background on Rademacher complexity, we refer the readers to [BM02].
Theorem 1 ([BM02]). Let D be a distribution over X × Y and let L : Y ′ × Y (where Y ⊆ Y ′ ⊆ R)
be a b-bounded loss function that is L-Lipschitz in its first argument.
Let F ⊆ (Y ′ )X and for
1 Pm
b
any f ∈ F, let L(f ; D) := E(x,y)∼D [L(f (x), y)] and L(f ; S) := m i=1 L(f (xi ), yi ), where S =
((x1 , y1 ), . . . , (xm , ym )) ∼ D m . Then for any δ > 0, with probability at least 1 − δ (over the random
sample draw for S), simultaneously for all f ∈ F, the following is true:
r
b ; S)| ≤ 4 · L · Rm (F) + 2 · b · log(1/δ)
|L(f ; D) − L(f
2m
where Rm (F) is the Rademacher complexity of the function class F.
For a linear concept class, the Rademacher complexity can be bounded as follows.
Theorem 2 ([KST09]). Let X be a subset of a Hilbert space equipped with inner product h·, ·i such
that for each x ∈ X , hx, xi ≤ X 2 , and let W = {x 7→ hx, wi | hw, wi ≤ W 2 } be a class of linear
functions. Then it holds that
r
1
Rm (W) ≤ X · W ·
.
m
The following result is useful for bounding the Rademacher complexity of a smooth function of
a concept class.
Theorem 3 ([BM02, LT91]). Let φ : R → R be Lφ -Lipschitz and suppose that φ(0) = 0. Let
Y ⊆ R, and for a function f ∈ Y X . Finally, for F ⊆ Y X , let φ ◦ F = {φ ◦ f : f ∈ F}. It holds that
Rm (φ ◦ F) ≤ 2 · Lφ · Rm (F).
2.1
Kernel Methods
We assume the reader has a basic working knowledge of kernel methods (for a good resource on
kernel methods in machine learning we refer the reader to [SS02]). We denote a kernel function
by K(x, x′ ) = hψ(x), ψ(x′ )i where ψ is the associated feature map and H is the corresponding
reproducing kernel Hilbert space (RKHS).
Here we define two kernels and a few of their properties that we will use for our analysis. First,
we define a variant of the polynomial kernel, the multinomial kernel due to Goel et al. [GKKT16]:
Definition 3 (Multinomial Kernel [GKKT16]). Define ψd : Rn → RNd , where Nd = 1+ n + · · ·+ nd ,
indexed by tuples (k1 , . . . , kj ) ∈ [n]j for each j ∈ {0, 1, . . . , d}, where the entry of ψd (x) corresponding
to tuple (k1 , . . . , kj ) equals xk1 · · · xkj . (When j = 0 we have an empty tuple and the corresponding
entry is 1.) Define kernel MKd as follows:
d
X
(x · x′ )j .
MKd (x, x ) = hψd (x), ψd (x )i =
′
′
j=0
Also define HMKd to be the corresponding RKHS.
7
It is easy to see that the multinomial kernel is efficiently computable. A multivariate polynomial
p of degree d can be represented as an element v ∈ HMKd . Also, every v ∈ HMKd can be interpreted
as a multivariate polynomial of degree d such that
X
p(x) = hv, ψd (x)i =
β(i1 , . . . , in )xi11 · · · xinn .
(i1 ,...,in )∈{0,...,d}n
ii +···+in ≤d
where coefficient β(i1 , . . . , in ) is as follows,
β(i1 , . . . , in ) =
X
v(k1 , . . . , kj ).
k1 ,...,kj ∈[n]j
j∈{0,...,d}
Here, v(·) is used to index the corresponding entry in v.
The following lemma is due to [GKKT16], following an argument of Shalev-Shwartz et al.
[SSSS11]:
P
P
Lemma 1. Let p(t) = di=0 βi ti be a given univariate polynomial with di=1 βi2 ≤ B 2 . For w such
that ||w|| ≤ 1, the polynomial p(w · x) equals hpw , ψ(x)i for some pw ∈ HM Kd with ||pw || ≤ B.
Remark. Observe that we can normalize the multinomial feature map such that
≤
P ∀x ∈ X , MK
Pdd (x, x)
j
1 for bounded space X . More formally, maxx∈X MKd (x, x) = maxx∈X dj=0 ||x||j ≤
X
j=0
where X = maxx∈X ||x||, hence we can normalize using this value. Subsequently, in the above,
||pw || will need to be multiplied by the same value. For X = Sn−1 , the scaling factor is d + 1
[GKKT16]. Throughout the paper, we will assume the kernel to be normalized as discussed.
For our results on Multiple Instance Learning, we make use of the following known kernel defined
over sets of vectors:
Definition 4 (Mean Map Kernel [SGSS07]). Let X ∗ denote the Kleene closure of X .
The mean map kernel Kmean : X ∗ × X ∗ → R of kernel K : X × X → R with feature vector
ψ : X → Z is defined as,
X
1
K(s, t).
Kmean (S, T ) =
|S||T |
s∈S,t∈T
The feature map ψmean : X ∗ → Z corresponding to this kernel is given by
ψmean (S) =
1 X
ψ(s).
|S|
s∈S
Also define Hmean to be the corresponding RKHS.
Fact. If ∀x, x′ ∈ X , K(x, x′ ) ≤ M then ∀S, S ′ ∈ X ∗ , Kmean (S, S ′ ) ≤ M .
2.2
Approximation Theory
We will make use of a variety of tools from approximation theory to obtain specific embeddings of
function classes into a RKHS. The following lemma for approximating the Boolean sign function
was given by [Dan15]:
8
Lemma 2. Let a, γ, τ > 0. There exists a polynomial p of degree O
• For x ∈ [−a, a], |p(x)| < 1 + τ .
1
γ
· log
1
τ
such that
• For x ∈ [−a, a]\[−γ · a, γ · a], |p(x) − sign(x)| < τ .
The above lemma assumes sign takes on values {±1}, but a simple linear transformation also
works for {0, 1}.
A halfspace is a Boolean function h : Rn → {0, 1} defined by a vector w ∈ Rn and threshold θ,
given input x ∈ Rn , h(x) = sign(w · x + θ) where sign(a) is 0 if a < 0 else 1. Given X ⊆ Rn and
halfspace h over Rn , h is said to have margin ρ with respect to X if ρ = min{||z − y|| : z ∈ X , y ∈
Rn , h(z) 6= h(y)}/ max{||z|| : z ∈ X }. Let hwi (x) for i ∈ [t] with ||wi ||2 ≤ 1 be the t halfspaces
and fAND = AND(hw1 , . . . , hwt (x)) equals the intersection. The following lemma due to Klivans
and Servedio [KS08] gives a construction of a polynomial whose sign always equals an intersection
of halfspaces with a margin. We give the proof as it is useful for a subsequent lemma.
p
Lemma 3. For ρ > 0, there exists a polynomial PAND of degree O( 1/ρ log t) such that, if each
of the t halfspaces has margin ρ on X ⊆ Sn−1 , then ∀x ∈ X , PAND (x) ≥ 1/2 if fAND (x) = 1 and
PAND (x) ≤ −1/2 if fAND (x) = 0. Constructively, PAND is as follows,
t
PAND (x) = t +
1 X
p(wi · x)
−
2
i=1
where p(a) = Tr (1 − a)⌈log 2t⌉
kind.
p
for r = ⌈ 1/ρ⌉ and Tr is the rth Chebyshev polynomial of the first
Proof. We have ∀i, ρ ≤ |wi · x| ≤ 1. From the properties of Chebyshev polynomials, we know that
|Tr (1 − a)| ≤ 1 for a ∈ [0, 1] and Tr (1 − a) ≥ 2 for a ≤ ρ. Hence,
• If fAND (x) = 1, for each i, ρ ≤ wi ·x ≤ 1 hence |p(wi ·x)| ≤ 1, implying PAND (x) ≥ t+1/2−t =
1/2.
• If fAND (x) = 0, then there exists i such that −1 ≤ wi · x ≤ ρ, thus p(wi · x) ≥ 2⌈log 2t⌉ ≥ 2t.
Also observe that ∀i, p(wi · x) ≥ −1. This implies PAND (x) ≤ t + 1/2 − 2t + t − 1 ≤ −1/2.
We extend the above lemma to give a threshold function for OR3 of a fixed halfspace with
margin over a set of vectors. Let hw (x) = sign(w · x) be a halfspace with ||w||2 ≤ 1 and fOR (S) =
OR(hw (s1 ), . . . , hw (st )) represent the conjunction over any set S = {s1 , . . . , st } of vectors. Let
X ⊆ Sn−1 over which the halfspace has margin.
Lemma 4. For ρ, N > 0, there exists a function POR such that, if the halfspace h has margin ρ
on X ⊆ Sn−1 , then for all sets S ⊆ X ∗ such that |S| ≤ N , POR (S) ≥ 1/2 if fOR (S) = 1 and
POR (S) ≤ −1/2 if fOR (S) = 0. Constructively, POR is as follows,
1
N X
POR (S) = − − N +
p(w · s)
2
|S|
s∈S
p
where p(a) = Tr (1 + a)⌈log 2N ⌉ for r = ⌈ 1/ρ⌉ and Tr is the rth Chebyshev polynomial of the first
kind.
3
A similar construction would hold for AND, with PAND (S) =
9
1
2
+N −
N
|S|
P
s∈S
p(w · s).
Proof. Similar to the previous proof, we have ∀x ∈ X , ρ ≤ |w · x| ≤ 1. From the properties of
Chebyshev polynomials, we know that |Tr (1 + a)| ≤ 1 for a ∈ [−1, 0] and Tr (1 + a) ≥ 2 for a ≥ ρ.
Hence,
• If fOR (S) = 0, for each s ∈ S, −1 ≤ w · s ≤ −ρ hence |p(w · s)| ≤ 1, implying POR (S) ≤
−1/2 − N + N = −1/2.
• If fOR (S) = 1, then there exists s ∈ S such that ρ ≤ w · s ≤ 1, thus p(w · s) ≥ 2⌈log 2N ⌉ ≥ 2N .
Also observe that ∀s ∈ S, p(w · s) ≥ −1. This implies POR (S) ≥ −1/2 − N + 2N − N + 1 ≥ 1/2
since N ≥ |S|.
Finally we state the following lemmas that bound the sum of squares of coefficients of a univariate
polynomial:
P
Lemma 5 ([She12]). Let p(t) = di=0 βi ti be a univariate polynomial of degree d. Let M be such
d
X
βi2 ≤ (d + 1) · (4e)2d · M 2 .
that max |p(t)| ≤ M . Then
t∈[−1,1]
i=0
Lemma 6. [Fact 3 [KS08]] Let p(t) =
|βi | ≤ M for all i ∈ [d]. For any r
dr
X
i=0
i
i=0 βi t be a univariate
∈ Z+ consider pr (t) =
ηi2 ≤ (M d)2r .
Proof. We have pr (t) =
bounded above by
P
degree d such that
of
polynomial
r
Pd
Pdr
i
i
=
i=0 βi t
i=0 ηi t then,
i1 +···+ir . It follows that
i1 ,··· ,ir ∈[d] βi1 · · · βir t
X
i1 ,··· ,ir ∈[d]
3
Pd
2
|βi1 · · · βir | ≤ M 2r
X
i1 ,··· ,ir ∈[d]
2
P
d
i
i=0 βi t
r
=
Pdr
i=0 ηi t
i
is
1 = (M d)2r .
The Alphatron Algorithm
Here we present our main algorithm Alphatron (Algorithm 1) and a proof of its correctness. In the
next section we will
our most general learning results.
P use tthis algorithm to obtain
t (x) = u(hvt , ψ(x)i). Let ε(h) = E
2
Define vt = m
α
ψ(x
)
implying
h
x,y [(h(x) − E[y|x]) ]
i
i=1 i
2
and err(h) = Ex,y [(h(x) − y) ]. It is easy to see that ε(h) = err(h) − err(E[y|x]). Let εb, err
c be the
empirical versions of the same.
The following theorem generalizes Theorem 1 of [KKKS11] to the bounded noise setting in a
high dimensional feature space. We follow the same outline, and their theorem can be recovered by
setting ψ(x) = x and ξ as the zero function.
10
Algorithm 1: Alphatron
n
4
Input : data h(xi , yi im
i=1 ∈ R × [0, 1], non-decreasing L-Lipschitz function u : R → [0, 1],
kernel function K corresponding to feature map ψ, learning rate λ > 0, number of
n
iterations T , held-out data of size N haj , bj iN
j=1 ∈ R × [0, 1]
1
m
1 α := 0 ∈ R
2 for t = 1, . . . , T do
P
t
3
ht (x) := u( m
i=1 αi K(x, xi )) for i = 1, 2, . . . , m do
t+1
λ
(yi − ht (xi ))
4
αi := αti + m
5
end
6 end
P
2
t
Output: hr where r = arg mint∈{1,...,T } N
j=1 (h (aj ) − bj )
Theorem 4. Let K be a kernel function corresponding to feature map ψ such that ∀x ∈ X , ||ψ(x)|| ≤
1. Consider samples (xi , yi )m
i=1 drawn iid from distribution D on X × [0, 1] such that E[y|x] =
u(hv, ψ(x)i + ξ(x)) where u : R → [0, 1] is a known L-Lipschitz non-decreasing function, ξ : Rn →
[−ǫ0 , ǫ0 ] for ǫ0 ∈ (0,p
1) and ||v|| ≤ B. Then for δ ∈ (0, 1), with probability 1 − δ, Alphatron with
λ = 1/L, T = CBL m/ log(1/δ) and N = C ′ m log(T /δ)) for large enough constants C, C ′ > 0
outputs a hypothesis h such that,
!
r
log(1/δ)
.
ε(h) ≤ O Lǫ0 + BL
m
1 Pm
1 Pm
t
t
Proof. Let ∆ = m
i=1 (yi −u(hv, ψ(xi )i)+ξ(xi ))ψ(x i ) and ∆ = m
i=1 (yi −u(hv , ψ(xi )i))ψ(x i ).
We will first prove the following lemma and subsequently use it to prove the theorem.
Lemma 7. At iteration t in Alphatron, suppose ||vt − v|| ≤ B for B > 1, then if ||∆|| ≤ η < 1,
then
2
t
2
t
2
t+1
2
− λ εb(h ) − 2ǫ0 − 2Bη − λη − 2λη .
||v − v|| − ||v
− v|| ≥ λ
L
Proof. Expanding the left hand side of the equation above, we have
||vt − v||2 − ||vt+1 − v||2
t
(3)
t
2
t 2
= 2λhv − v , ∆ i − λ ||∆ ||
(4)
= 2λhv − vt , ∆t − ∆i + 2λhv − vt , ∆i − λ2 ||∆t ||2
(5)
m
2λ X
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))hv − vt , ψ(xi )i − 2λB||∆|| − λ2 ||∆t ||2
≥
m
i=1
(6)
=
2λ
m
m
X
i=1
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))(hv, ψ(xi )i + ξ(xi ) − hvt , ψ(xi )i)
4
We present the algorithm and subsequent results for non-decreasing function u. Non-increasing functions can
λ
also be handled by negating the update term, i.e., αt+1
:= αti − m
(yi − ht (xi )).
i
11
m
2λ X
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))ξ(xi ) − 2λB||∆|| − λ2 ||∆t ||2
m
−
≥
2λ
Lm
≥
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))2
i=1
m
X
2λ
m
−
(7)
i=1
m
X
i=1
|u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i)||ξ(xi )| − 2λB||∆|| − λ2 ||∆t ||2
(8)
2λ
εb(ht ) − 2λǫ0 − 2λBη − λ2 ||∆t ||2
L
(9)
Here (4) follows from substituting the expression of vt+1 , (6) follows from bounding ||vt − v|| ≤ B
and, (8) follows from u being monotone and L-Lipschitz, that is, (u(a) − u(b))(a − b) ≥ L1 (u(a) −
u(b))2 . (9) follows from observing that the first term equals εb(ht ), the second term is bounded in
norm by ǫ0 since the range of u is [0, 1] and using the assumption ||∆|| ≤ η.
We now bound ||∆t || as follows.
m
t 2
||∆ || =
1 X
(yi − u(hv, ψ(xi )i + ξ(xi )) + u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))ψ(xi )
m
2
(10)
i=1
m
1 X
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))ψ(xi )
m
2
≤ ||∆|| +
1
m
+ 2||∆||
2
i=1
m
X
i=1
(u(hv, ψ(xi )i + ξ(xi )) − u(hvt , ψ(xi )i))ψ(x i )
(11)
≤ η 2 + εb(ht ) + 2η
(12)
Here (13) follows by expanding the square and (12) follows by applying Jensen’s inequality to show
2
1 Pm
1 Pm
that for all a, b ∈ Rm and vectors vi for i ∈ {1, · · · , m}, m
≤ m
i=1 (ai − bi )vi
i=1 (ai −
bi )2 ||vi ||2 and subsequently using the fact that ||ψ(xi )|| ≤ 1. Combining (9) and (12) gives us the
result.
By definition we have that (yi − u(hv, ψ(xi )i + ξ(xi )))ψ(xi ) are zero mean iid random variables with norm bounded by 1. Using Hoeffding’s inequality (and the fact that that the xi ’s are
independent draws), with probability 1 − δ we have
p
1 X
1
(yi − u(hv, ψ(xi )i + ξ(xi )))ψ(xi ) ≤ √
1 + 2 log(1/δ) .
m
m
m
||∆|| =
i=1
Now using the previous lemma with λ = 1/L and η =
t
2
||v − v|| − ||v
t+1
1
− v|| ≥
L
2
√1
m
p
1 + 2 log(1/δ) , we have
η 2 2η
εb(ht )
− 2ǫ0 − 2Bη −
−
L
L
L
.
Thus, for each iteration t of Alphatron, one of the following two cases needs to be satisfied,
12
Case 1: ||vt − v||2 − ||vt+1 − v||2 ≥
Bη
L
q
Case 2: εb(ht ) ≤ 3LBη + 2Lǫ0 + η 2 + 2η = O Lǫ0 + BL log(1/δ)
(assuming that B > η and
m
L > 1)
Let t be the first iteration where Case 2 holds. We need to show that such an iteration exists.
Assume the contradictory, that is, Case 2 fails for each iteration. Since ||v0 − v||2 ≤ B 2 , however,
BL
in at most BL
η iterations Case 1 will be violated and Case 2 will have to be true. If η ≤ T then t
exists such that
!
r
log(1/δ)
.
εb(ht ) ≤ O Lǫ0 + BL
m
We need to bound ε(h) in terms of εb(h). Define F = {x → u(hz, ψ(x)i) : ||z|| ≤ 2B}, and
p Z =
{x → f (x)−u(hv, ψ(x)i+ξ(x)) : f ∈ F}. Using Theorem 2 and 3 we have Rm (F) = O(BL 1/m).
By definition of Rademacher complexity, we have
!#
"
m
2 X
σi z(xi )
Rm (Z) = Exi ,σi sup
m
z∈Z
i=1
"
!#
m
2 X
= Exi ,σi sup
σi (f (xi ) − u(hv, ψ(xi )i + ξ(xi )))
m
f ∈F
i=1
"
!#
"
#
m
m
X
2
2 X
= Exi ,σi sup
σi f (xi )
− E xi
Eσi [σi ]u(hv, ψ(xi )i + ξ(xi ))
m
m
f ∈F
i=1
i=1
= Rm (F)
Here, σi ∈ {±1} are iid Rademacher variables hence ∀ i, E[σi ] = 0 and xi are drawn iid from D.
Recall that h(x) = u(hvT , ψ(x)i) is an element of U as ||vT − v||2 ≤ B 2 (case 1 is satisfied in
iteration t−1) and ||v|| ≤ B. A direct application of Theorem 1 on Z with loss function L(a, ·) = a2 ,
gives us the following bound on ε(ht ) with probability 1 − δ,
!
!
r
r
r
1
log(1/δ)
log(1/δ)
t
= O Lǫ0 + BL
.
+
ε(h ) ≤ εb(h) + O BL
m
m
m
The last step is to show that we can indeed find a hypothesis satisfying the above guarantee.
Since for all h, ε(h) is up to constants equal to err(h) we can do so by choosing the hypothesis
with the minimum err using a fresh sample set of size O(log(T /δ)/ǫ2 ) ≤ N . This holds as given
the sample size, by Chernoff bound using the fact that εb(ht ) is bounded in [0, 1], each ht for t ≤ T
will have empirical error within ǫ of the true error with probability 1 − δ/T and hence all will
simultaneously satisfy this with probability 1 − δ,
!
r
log(1/δ)
.
ε(h) ≤ O ǫ + Lǫ0 + BL
m
√
Setting ǫ = 1/ m will give us the required bound.
Alphatron runs in time poly(n, m, log(1/δ), tK ) where tK is the time required to compute the
kernel function K.
13
4
Some General Theorems Involving Alphatron
In this section we use Alphatron to give our most general learnability results for the p-concept
model and PAC learning setting. We then state several applications in the next section. Here we
show that if a function can be uniformly approximated by an element of an appropriate RKHS
then it is p-concept learnable. Similarly, if a function is separable by an element in an appropriate
RKHS then it is PAC learnable. We assume that the kernel function is efficiently computable, that
is, computable in polynomial time in the input dimension. Formally, we define approximation and
separation as follows:
Definition 5 ((ǫ, B)-approximation). Let f be a function mapping domain X to R. Let K be a
kernel function with corresponding RKHS H and feature vector ψ. We say f is (ǫ, B)-approximated
by K if for all x ∈ X , |f (x) − hv, ψ(x)i| ≤ ǫ for some v ∈ H such that ||v|| ≤ B.
Definition 6 (B-separation). Let f be a boolean function mapping domain X to {0, 1}. Let K be
a kernel function with corresponding RKHS H and feature vector ψ. We say f is B-separated by K
if for all x ∈ X , if f (x) = 1 then hv, ψ(x)i ≥ 1/2 and if f (x) = 0 then hv, ψ(x)i ≤ −1/2 for some
v ∈ K such that ||v|| ≤ B.
Combining Alphatron and the approximation (separation) guarantees, we have the following
general learning results:
Theorem 5. Consider distribution D on X × [0, 1] such that E[y|x] = u(f (x)) where u is a known
L-Lipschitz non-decreasing function and f is (ǫ, B)-approximated by some kernel function K such
that K(x, x′ ) ≤ 1. Then for δ ∈ (0, 1), there exists an algorithm that draws m iid samples from D
2
· log(1/δ)
and outputs a hypothesis h such that with probability 1 − δ, ε(h) ≤ O(Lǫ) for m ≥ BL
ǫ
in time poly(n, B, L, 1/ǫ, log(1/δ)) where n is the dimension of X .
Proof. Let H be the RKHS corresponding to K and ψ be the feature vector. Since f is (ǫ, B)approximated by kernel function K, we have ∀ x, |f (x) − hv, ψ(x)i| ≤ ǫ for ||v|| ≤ B. This implies
that f (x) = hv, ψ(x)i + ξ(x) for some function ξ : X → [−ǫ, ǫ]. Thus E[y|x] = u(f (x)) =
u (hv, ψ(x)i + ξ(x)). Applying Theorem 4, we have that Alphatron outputs a hypothesis h such
that
!!
r
log(1/δ)
ε(h) ≤ CL ǫ + B
m
p
for some constants C > 0. Also Alphatron requires at most O(BL m/ log(1/δ)) iterations. Setting
2
m = BL
· log(1/δ) gives us the required result.
ǫ
Theorem 6. Consider a sample of size m > 0 drawn from distribution D over X and labeled
by Boolean function f . Assume f is B-separated by some kernel function K such that K(x, x′ ) ≤
1. Then there exists an algorithm that outputs a hypothesis h such that with probability 1 − δ,
2
P rx∼D [h(x) 6= f (x)] ≤ ǫ for m ≥ C Bǫ · log(1/δ) for sufficiently large constant C > 0 in time
poly(n, B, 1/ǫ, log(1/δ)) where n is the dimension of X .
Proof. Let v be the vector that B-separates f and P (x) = hv, ψ(x)i. Consider u as follows:
a ≤ −1/2
0
u(a) = a + 1/2 −1/2 < a < 1/2
1
a ≥ 1/2
14
Note that u is monotone and 1-Lipschitz. Observe that,
• If f (x) = 1 then u(P (x)) = 1 since P (x) ≥ 1/2.
• If f (x) = 0 then u(P (x)) = 0 since P (x) ≤ −1/2.
From above, we can see that the samples drawn from the distribution satisfy E[y|x] = f (x) =
2
u(P (x)). Thus we can apply Theorem 4 with ǫ0 = 0, L = 1 and m = C Bǫ log(1/δ) (for
sufficiently large constant C > 0) to obtain output hypothesis h′ with ε(h′ ) ≤ ǫ (with probability
1 − δ).
Recall that h′ may be real-valued as Alphatron learns with square loss. Let us define h(x) to
equal 1 if h′ (x) ≥ 1/2 and 0 otherwise. We will show that Ex [h(x) 6= f (x)] ≤ O(ǫ). Using Markov’s
inequality, we have P rx [(h′ (x) − u(P (x)))2 ≥ 1/4] ≤ 4ǫ. For x, suppose (h′ (x) − u(P (x)))2 <
1/4 =⇒ |h′ (x) − u(P (x))| < 1/2. Since u(P (x)) = f (x) ∈ {0, 1} then clearly h(x) = f (x). Thus,
P rx [h(x) 6= f (x)] ≤ 4ǫ. Scaling ǫ appropriately, we get the required result.
5
Main Applications
The general framework of the previous section can be used to give new learning results for well
studied problems. In this section we give polynomial time learnability results for depth-three neural
networks with sigmoidal activations in the p-concept model. We follow this by showing how to obtain
the first polynomial-time algorithms PAC learning a polynomial number of intersections/majorities
of halfspaces with a margin.
5.1
Learning Depth-3 Neural Networks
Following standard convention (see for example [SS16]), we define a neural network with one hidden
layer (depth-2) with k units as follows:
N2 : x →
k
X
i=1
bi σ(ai · x)
for x ∈ Rn , ai ∈ Sn−1 for i ∈ {1, · · · , k}, b ∈ Sk−1 . We subsequently define a neural network with
two hidden layers (depth-3) with one unit in hidden layer 2 and k units in hidden layer 1 as:
!
k
X
bi σ(ai · x)
N3 : x → σ ′ (N1 (x)) = σ ′
i=1
for x ∈ Rn , ai ∈ Sn−1 for i ∈ {1, · · · , k}, b ∈ Sk−1 and σ, σ ′ : R → R.
[GKKT16] showed that activation functions sigmoid: σsig (a) = 1+e1−a and ReLU: σrelu (a) =
max(0, a) can be (ǫ, B)-approximated by the multinomial kernel for B dependent on ǫ, more formally
they showed the following:
Lemma 8 (Approximating a Single Hidden Unit). We have,
1. Sigmoid: For all a ∈ Sn−1 there exists a corresponding v ∈ HMKd for d = O(log(1/ǫ)), such
that
∀ x ∈ Sn−1 , |σsig (a · x) − hv, ψ(x)i| ≤ ǫ.
Further, ||v|| ≤ (1/ǫ)O(1) . This implies that σsig is (ǫ, (1/ǫ)O(1) )-approximated by MKd .
15
2. ReLU: For all a ∈ Sn−1 there exists a corresponding v ∈ HMKd for d = O(1/ǫ), such that
∀ x ∈ Sn−1 , |σrelu (a · x) − hv, ψ(x)i| ≤ ǫ.
Further, ||v|| ≤ 2O(1/ǫ) . This implies that σrelu is (ǫ, 2O(1/ǫ) )-approximated by MKd .
The following lemma extends the approximation guarantees to linear combinations of function
classes.
P
Lemma 9. If for all i ∈ [k], fi is (ǫ, B)-approximated in kernel K then ki=1 ai fi (x) for a ∈ Rk
such that ||a||1 ≤ W is (ǫW, W B)-approximated in kernel K.
Proof. We have
each i ∈ [k], ∀x ∈ X , |fi (x)−hvi , ψ(x)i| ≤ ǫ for some vi ∈ H such that ||vi || ≤ B.
Pfor
k
Consider v = i=1 ai vi . We have ∀x ∈ X ,
k
X
i=1
ai fi (x) − hv, ψ(x)i =
Also ||v|| =
Pk
i=1 ai vi
≤
k
X
i=1
Pk
ai (fi (x) − hvi , ψ(x)i) ≤
i=1 |ai |||vi ||
k
X
i=1
|ai ||fi (x) − hvi , ψ(x)i| ≤ ǫ||a||1 = ǫW.
≤ W B. Thus v satisfies the required approximation.
The following theorem is our main result for learning classes of depth-three neural networks in
polynomial time:
n−1 × [0, 1] such that
Theorem 7. Consider samples (xi , yi )m
i=1 drawn iid from distribution D on S
E[y|x] = N3 (x) with σ ′ : R → [0, 1] is a known L-Lipschitz non-decreasing function and σ = σsig
is the sigmoid function. There exists an algorithm that outputs a hypothesis h such that, with
probability 1 − δ,
Ex,y∼D (h(x) − N3 (x))2 ≤ ǫ
O(1)
for m = kL
· log(1/δ). The algorithm runs in time polynomial in m and n.
ǫ
√ √
Proof. Combining Lemmas 8 and 9 we have that N2 for activation function σsig is (ǫ0 k, ( k/ǫ0 )C )approximated by some kernel MKd with d = O(log(1/ǫ0 )) and sufficiently large constant C > 0.
Thus by Theorem 5, we have that there exists an algorithm that outputs a hypothesis h such that,
with probability 1 − δ,
√ !C r
√
k
log(1/δ)
·
Ex,y∼D (h(x) − N3 (x))2 ≤ C ′ L ǫ0 k +
ǫ0
m
4 log(1/δ)
ǫ
2kCL 2C
·
for some constants C ′ > 0. Setting ǫ0 = 2√kC
gives us the required
and
m
=
′L
ǫ
ǫ2
result (the claimed bounds on running time also follow directly from Theorem 5).
We also obtain results for networks of ReLUs, but the dependence on the number of hidden
units, ǫ, and L are exponential (the algorithm still runs in polynomial-time in the dimension):
16
n−1 × [0, 1] such that
Theorem 8. Consider samples (xi , yi )m
i=1 drawn iid from distribution D on S
E[y|x] = N3 (x) with σ ′ : R → [0, 1] is a known L-Lipschitz non-decreasing function and σ = σrelu is
the ReLU function. There exists an algorithm that outputs a hypothesis h such that with probability
1 − δ,
Ex,y∼D (h(x) − N3 (x))2 ≤ ǫ
for m = 2O(kL/ǫ) · log(1/δ). The algorithm runs in time polynomial in m and n.
Although our algorithm does not recover the parameters of the network, it still outputs a hypothesis with interpretable features. More specifically, our learning algorithm outputs the hidden layer
as a multivariate polynomial. Given
x1 , · · · , xm , the hypothesis output by P
our algorithm
Pminputs
∗
∗
Alphatron is of the form h(x) = u( i=1 αi MKd (x, xi )) = u(hv, ψd (x)i) where v = m
i=1 αi ψd (xi )
and d is dependent on required approximation. As seen in the preliminaries, hv, ψd (x)i can be
expressed as a polynomial and the coefficients can be computed as follows,
∗
αi
β(i1 , . . . , in ) =
i=1
m
X
X
k1 ,...,kj ∈[n]j
j∈{0,...,d}
M (k1 ,...,kj )=(i1 ,...,in )
X
m
α∗i C (i1 , . . . , in ) (xi )i11 · · · (xi )inn .
(xi )k1 · · · (xi )kj =
i=1
Here, we follow the notation from [GKKT16]; M maps ordered tuple (k1 , . . . , kj ) ∈ [n]j for j ∈ [d]
to tuple (i1 , . . . , in ) ∈ {0, . . . , d}n such that xk1 · · · xkj = xi11 · · · xinn and C maps ordered tuple
(i1 , . . . , in ) ∈ {0, . . . , d}n to the number of distinct orderings of the ij ’s for j ∈ {0, . . . , n}. The
function C can be computed from the multinomial theorem (cf. [Wik16]). Thus, the coefficients
of the polynomial can be efficiently indexed. Informally, each coefficient can be interpreted as the
correlation between the target function and the product of features appearing in the coefficient’s
monomial.
5.2
Learning Smooth Functions of Halfspaces with a Margin
In this section we consider the problem of learning a smooth combining function u of k halfspaces
with a margin ρ. We assume that all examples lie on the unit ball Sn−1 and that for each weight
vector w, ||w|| = 1. For simplicity we also assume each halfspace is origin-centered, i.e. θ = 0
(though our techniques easily handle the case of nonzero θ).
n−1 × [0, 1] such that
Theorem 9.PConsider samples
(xi , yi )m
i=1 drawn iid from distribution D on S
t
n
E[y|x] = u
i=1 ai hi (x) where u : R → [0, 1] is a L-Lipschitz non-decreasing function, hi are
origin-centered halfspaces with margin ρ on X and ||a||1 = A. There exists an algorithm that outputs
a hypothesis h such that with probability 1 − δ,
!!2
t
X
≤ǫ
Ex,y∼D h(x) − u
ai hi (x)
i=1
for m =
LA O(1/ρ)
ǫ
log(1/δ). The algorithm runs in time polynomial in m and n.
17
Proof. We use Lemma 2 to show the existence of polynomial p of degree d = O 1ρ · log ǫ10
such
that for a ∈ [−1, 1], |p(a)| < 1 + ǫ0 and for a ∈ [−1, 1]\[−ρ, ρ], |p(a) − sign(a)| < ǫ0 .
Since for each i, ρ ≤ |wi · x| ≤ 1, we have |p(wi · x) − sign(wi · x)| ≤ ǫ0 such that p is bounded
in [−1, 1] by 1 + ǫ0 . From Lemma 5 and 1, we have that for each i, p(wi · x) = hvi , ψd (x)i such that
O(1/ρ)
where ψd is the feature vector corresponding to the multinomial kernel of degree
||vi || = ǫ10
O(1/ρ)
Pt
d. Using Lemma 9, we have that i=1 ai hi (x) is ǫ0 A, A ǫ10
-approximated by MKd .
Subsequently, applying Theorem 5, we get that there exists an algorithm that outputs a hypothesis h such that with probability 1 − δ,
!
C ′ /ρ r
log(1/δ)
1
·
ε(h) ≤ CLA ǫ0 +
ǫ0
m
2C ′ /ρ 4 log(1/δ)
ǫ
for some constants C, C ′ > 0. Setting ǫ0 = 2CLA
to gives us the
and m = 2CLA
·
ǫ
ǫ2
required result.
P
Remark. If E[y|x] = u 1t ti=1 hi (x) , that is, a function of the fraction of true halfspaces, then
the run-time is independent of the number of halfspaces t. This holds since A = 1 in this case.
5.3
PAC Learning Intersection/Majority of Halfspaces with a Margin
Consider t-halfspaces {h1 , . . . , ht }. An intersection of these t-halfspaces
is given by fAND (x) =
Pt
t
∧i=1 hi (x). A majority of t-halfspaces is given by fmaj (x) = sign
i=1 hi (x) − t/2 .
The following lemmas show that there exists kernels that separate intersection/majority of halfspaces.
Lemma 10. Intersection of√ t-halfspaces on Sn−1 is B-separated by a kernel K such that ∀x ∈
H, K(x, x) ≤ 1 with B = tO( 1/ρ) .
Proof. Consider PAND given by Lemma 3 for given ρ. We know that ∀x ∈ Sn−1 , PAND (x) ≥ 1/2
if fAND (x) = 1 and PAND (x) ≤ −1/2 if fAND (x) = 0. We know that each coefficient of Tr (a) is
bounded by 2r . It is easy to see that each coefficient of Tr (1 − a) is bounded by 2O(r) . Using Lemma
6, the sum of squares of the coefficients of p(a) are bounded by 2O(r⌈log 2t⌉) = tO(r) . Now using
Lemma 1, we know that p(wi · x) = hvi , ψd (x)i for the feature vector ψd of the multinomial kernel
with d = r⌈log 2t⌉. such that ||vi || ≤ tO(r) . Thus, PAND (x) = hv, ψd (x)i (adjusting the constant
term corresponding to the 1 entry in the feature vector) such that ||v|| ≤ tO(r) . We can assume that
||ψ(x)|| ≤ 1 by scaling the weight vector appropriately as before. Thus, fAND is tO(r) -separated by
kernel MKd .
Lemma 11. Majority of t-halfspaces on Sn−1 is B-separated by a kernel K such that ∀x ∈ H, K(x, x) ≤
1 with B = tO(1/ρ) .
Proof. We use Lemma 2 to show the existence of polynomial p of degree d = O ρ1 · log t such that
for a ∈ [−1, 1], |p(a)| < 1 + 1/2t and for a ∈ [−1, 1] [−ρ, ρ], |p(a) − sign(a)| < 1/2t. Consider
Pmaj (x) =
t
X
i=1
p(wi · x) −
18
t
2
Since for each hP
i , ρ ≤ |wi · x| ≤ 1, we have |p(wi · x) − sign(wi · x)| ≤ 1/2t. Thus if fmaj (x) = 1
then Pmaj (x) ≥ ti=1 hi (x) − t/2 − 1/2 ≥ 1/2. Similarly if fmaj (x) = 0 then Pmaj (x) ≤ −1/2.
From Lemma 5 and 1, we have that for each i, p(wi · x) = hvi , ψ(x)i such that ||vi || = tO(1/ρ)
where ψ is the feature vector corresponding to the multinomial kernel of degree d. Thus, Pmaj (x) =
hv, ψ(x)i (adjusting the constant term corresponding to the 1 entry in the feature vector) such that
||v|| = tO(1/ρ) . We can assume that ||ψ(x)| ≤ 1 by scaling the weight vector appropriately as before.
Thus, fmaj is tO(1/ρ) -approximated by kernel K.
Applying Theorem 6 to the above lemmas, we obtain the following results for PAC learning:
Corollary 1. There exists√an algorithm that PAC learns any intersection of t-halfspaces with margin
ρ > 0 on Sn−1 in time tO( 1/ρ) · poly(n, 1/ǫ, log(1/δ)).
Thisq
result improves the previous best bound by [KS08] that had running time roughly n ·
O 1 log t
ρ
log t
. For constant ρ, our algorithm is the first that has running time polynomial in
ρ
the number of halfspaces.
Corollary 2. There exists√an algorithm that PAC learns any majority of t-halfspaces with margin
ρ > 0 on Sn−1 in time tO( 1/ρ) · poly(n, 1/ǫ, log(1/δ)).
6
The KMtron Algorithm
In this section, we propose an algorithm, KMtron, which combines isotonic regression with the
KM algorithm [KM93] for finding large Fourier coefficients of a function (given query access to the
function). In some sense, the KM algorithm takes the place of the “kernel trick” used by Alphatron
to provide an estimate for the update step in isotonic regression. Viewed this way, the KM algorithm
can be re-interpreted as a query-algorithm for giving estimates of the gradient of square-loss with
respect to the uniform distribution on Boolean inputs.
The main application of KMtron is a generalization of celebrated results for PAC learning DNF
formulas [Jac97] to the setting of probabilistic concepts. That is, we can efficiently learn any conditional mean that is a smooth, monotone combination of L1 -bounded functions.
Notation. We follow the notation of [GKK08]. For any function P : {−1, 1}n → R, we denote
the Fourier coefficients by Pb(S) for all S ⊆ [n]. The support of P , i.e., the number of non-zero
Fourier coefficients, is denoted by supp(P ). The norms of the coefficient vectors are defined as
P
1/p
p
b
Lp (P ) =
for p ≥ 1 and L∞ (P ) = maxS |Pb (S)|. Similarly, the norm of the
S |P (S)|
P
function P are defined as ||P ||p = Ex∈{−1,1}n [ S |P (x)|p ]1/p for p ≥ 1. Also, the inner product
P · Q = Ex∈{−1,1}n [P (x)Q(x)].
KM Algorithm. The KM algorithm learns sparse approximations to boolean functions given query
access to the underlying function. The following lemmas about the KM algorithm are important to
our analysis.
Lemma 12 ([KM93]). Given an oracle for P : {−1, 1}n → R, KM(P, θ) returns Q : {−1, 1}n → R
with |supp(Q)| ≤ O(L2 (P )2 θ −2 ) and L∞ (P − Q) ≤ θ. The running time is poly(n, θ −1 , L2 (P )).
19
ǫ2
Lemma 13 ([KM93]). If P has L1 (P ) ≤ k, then KM P, 2k
returns Q s.t. ||P − Q||2 ≤ ǫ.
Projection Operator. The projection operator projK (P ) for P : {−1, 1}n → R maps P to the
closest Q in convex set K = {Q : {−1, 1}n → R | L1 (Q) ≤ k}, i.e., projK (P ) = arg minQ∈K ||Q−P ||2 .
[GKK08] show that projK is simple and easy to compute for sparse polynomials. We use the following
lemmas by [GKK08] about the projection operator in our analysis.
Lemma 14 ([GKK08]). Let P, P ′ be such that L∞ (P − P ′ ) ≤ ǫ. Then L∞ (projK (P ) − projK (P ′ )) ≤
2ǫ.
Lemma
15 ([GKK08]). Let P, P ′ be such that L∞ (P − P ′ ) ≤ ǫ. Then ||projK (P ) − projK (P ′ )||2 ≤
√
2 ǫk.
KMtron. The algorithm KMtron is as follows:
Algorithm 2: KMtron
Input : Function u : R → [0, 1] non-decreasing and L-Lipschitz, query access to u ◦ P for
some function P : {−1, 1}n → R, learning rate λ ∈ (0, 1], number of iterations T ,
error parameter θ
1 P0 = 0
2 for t = 1, . . . , T do
3
Pt′ := Pt−1 + λKM(u ◦ P − u ◦ Pt−1 , θ)
4
Pt = KM(projK (Pt′ ), θ)
5 end
Output: Return u ◦ Pt where Pt is the best over t = 1, · · · , T
To efficiently run KMtron, we require efficient query access to u ◦ P − u ◦ Pt−1 . Since Pt−1 is
stored as a sparse polynomial, and we are given query access for u ◦ P , we can efficiently compute
u(P (x)) − u(Pt−1 (x)) for any x. We can extend the algorithm to handle distribution queries (pconcept), i.e., for any x of our choosing we obtain a sample of y where E[y|x] = u(P (x)). [GKK08]
(c.f. Appendix A.1) observed that using distribution queries instead of function queries to the
conditional mean is equivalent as long as the number of queries is polynomial.
The following theorem proves the correctness of KMtron.
Theorem 10. For any non-decreasing L-Lipschitz u : R → [0, 1] and function P : {−1, 1}n → R
′ 4
2 2
ǫ
such that L1 (P ) ≤ k, given query access to u ◦ P , KMtron run with λ = 2L
, T = 2kǫ2L and θ ≤ LC4 kǫ 3
for sufficiently small constant C ′ > 0 and outputs P ∗ such that Ex∈{−1,1}n [(u(P (x)) − u(P ∗ (x)))2 ] ≤
ǫ. The runtime of KMtron is poly(n, k, L, 1/ǫ).
Proof. Let ε(h) = Ex∈{−1,1}n [(u(P (x)) − u(h(x)))2 ] = ||u ◦ P − u ◦ h||22 . Similar to Alphatron, we
will show that with each iteration t we will move closer to the true solution as long as ε(Pt ) is large.
Lemma 16. For a suitable choice of θ, ||Pt − P ||22 − ||Pt+1 − P ||22 ≥
2λ
L (ε(Pt ) −
Lλ).
Proof. Let us define the following polynomials for t ≤ T , Q′t = Pt−1 + λ(u ◦ P − u ◦ Pt−1 ) and Qt =
projK (Q′t ). For all t ≤ T ,
Pt′ − Q′t = (Pt−1 + λKM(u ◦ P − u ◦ Pt−1 , θ)) − (Pt−1 + λ(u ◦ P − u ◦ Pt−1 ))
= λ(KM(u ◦ P − u ◦ Pt−1 , θ) − (u ◦ P − u ◦ Pt−1 )).
20
From Lemma 12, L∞ (KM(u◦P −u◦Pt−1 , θ)−(u◦P −u◦Pt−1 )) ≤ θ implying L∞ (Pt′ −Q′t ) ≤ λθ ≤ θ
since λ ≤ 1.
√
2 θk. Since Pt = KM(projK (Pt′ ), θ) and
Using Lemma 15, we have ||projK (Pt′ ) − projK (Q′t )||2 ≤ √
L1 (projK (Pt′ )) ≤ k, using Lemma 13, ||Pt − projK (Pt′ )||2 ≤ 2θk. Using Triangle inequality, we get,
√
||Pt − Qt ||2 ≤ ||Pt − projK (Pt′ )||2 + ||projK (Pt′ ) − projK (Q′t )||2 < 4 θk.
Observe that ||Qt − P ||2 = L2 (Qt − P ) ≤ L1 (Qt − P ) ≤ L1 (Qt ) + L1 (P ) ≤ 2k. Combining these
two observations, we have
√
√
||Pt − P ||22 ≤ (||Pt − Qt ||2 + ||Qt − P ||2 )2 ≤ ||Qt − P ||2 + 16k θk + 16θk ≤ ||Qt − P ||2 + Ck θk
for large enough constant C > 0. Therefore,
√
||Pt − P ||22 − ||Pt+1 − P ||22 ≥ ||Pt − P ||22 − ||Qt+1 − P ||2 − Ck θk
√
≥ ||Pt − P ||22 − ||Q′t+1 − P ||2 − Ck θk
(13)
√
(14)
= ||Pt − P ||22 − ||Pt − P + λ(u ◦ P − u ◦ Pt )||22 − Ck θk
2
= −2λ(u ◦ P − u ◦ Pt ) · (Pt − P ) − λ ||u ◦ P − u ◦
√
2λ
≥
· ε(Pt ) − λ2 − Ck θk.
L
Pt ||22
(15)
√
− Ck θk (16)
Here, (13) follows from the triangle inequality and (5), (14) follows from projecting to a convex set
reducing the distance to points in the convex set, (16)√follows from u being monotone, L-Lipschitz
with output bounded in [0, 1]. Setting θ such that Ck θk ≤ λ2 gives the required result.
As long as ε(Pt ) ≥ 2Lλ, we have ||Pt − P ||22 − ||Pt+1 − P ||22 ≥ 2λ2 . Since ||P0 − P ||22 = ||P ||22 ≤
k2
2
2
L1 (P )2 ≤ k2 , after T = 2λ
2 , there must be some r ≤ T such that ||Pr − P ||2 ≥ 2λ does not hold,
ǫ
at this iteration, ε(Pt ) ≤ 2Lλ = ǫ for λ = 2L . The last step of choosing the best Pt would give us
the required hypothesis (similar to Alphatron). Observe that each iteration of KMtron runs in time
poly(n, k, L, 1/ǫ) (Lemma 12) and KMtron is run for poly(k, L, 1/ǫ) iterations giving us the required
runtime.
Corollary 3. Let PP
i be such
that L1 (Pi ) ≤ k for i ∈ [s]. If we have query access to y for all x such
that E[y|x] = u 1s si=1 Pi for non-decreasing L-Lipschitz u, then using the above, we can learn
the conditional mean function in time poly(n, k, L, 1/ǫ).
Observe that the complexity
bounds are independent of the number of terms. This follows
1 Ps
from the fact that L1 s i=1 Pi ≤ k. This leads to the following new learning result for DNF
formulas: fix a DNF f and let frac(f (x)) denote the fraction of terms of f satisfied by x. Fix monotone, L-Lipschitz function u. For uniformly chosen input x, label y is equal to 1 with probability
u(frac(f (x))). Then in time polynomial in n, 1/ǫ, and L, KMtron outputs a hypothesis h such that
E[(h(x) − u(frac(f (x))))2 ] ≤ ǫ (recall L1 (AND) = 1). Note that the running time has no dependence
on the number of terms.
As an easy corollary, we also obtain a simple (no Boosting required) polynomial time queryalgorithm for learning DNFs under the uniform distribution5 :
5
Feldman [Fel12] was the first to obtain a query-algorithm for PAC learning DNF formulas with respect to the
uniform distribution that did not require a Boosting algorithm.
21
Corollary 4. Let f be a DNF formula from {−1, 1}n → {0, 1} with s terms. Then f is PAC
learnable under the uniform distribution using membership queries in time poly(n, s, 1/ǫ).
P
Proof. Let {Ti }si=1 be the ANDs corresponding to each term of the DNF. Let T = si=1 Ti . By
definition of f , if f (x) = 1 then T (x) ≥ 1 and if f (x) = 0 then T (x) = 0. Observe that L1 (T ) ≤
P
s
i=1 L1 (Ti ) ≤ s using the well known fact that AND has L1 bounded by 1.
Consider the following u,
0 a ≤ 0
u(a) = a 0 < a < 1
1 a≥1
Observe that u is 1-Lipschitz. It is easy to see that f (x) = u(T (x)) on {−1, 1}n . Hence, given query
access to f is the same as query access to u ◦ T over {−1, 1}n .
Since L1 (T ) is bounded, we can apply Theorem 10 for the given u. We get that in time
poly(n, s, 1/ǫ), KMtron outputs a polynomial P such that Ex∈{−1,1}n [(u(T (x)) − u(P (x)))2 ] ≤ ǫ.
Following the proof of Theorem 6, we can move from square-loss to zero-one loss and show that
Prx∈{−1,1}n [f (x) 6= sign(u(P (x)))] ≤ 4ǫ. Scaling ǫ appropriately, we obtain the required result.
7
Multiple Instance Learning
Multiple Instance Learning (MIL) is a generalization of supervised classification in which a label is
assigned to a bag, that is, a set of instances, instead of an individual instance [DLLP97]. The bag
label is induced by the labels of the instances in it. The goal we focus on in this work is to label
future bags of instances correctly, with high probability. We refer the reader to [Amo13, HVB+ 16]
for an in-depth study of MIL. In this section we apply the previously developed ideas to MIL and
give the first provable learning results for concrete schemes that do not rely on unproven assumptions.
Comparison to Previous Work. Under the standard MI assumption, various results are known
in the PAC learning setting. Blum and Kalai [BK98] showed a simple reduction from PAC learning
MIL to PAC learning with one-sided noise under the assumption that the instances in each bag were
drawn independently from a distribution. Sabato and Tishby [ST12] removed the independence assumption and gave sample complexity bounds for learning future bags. All the above results require
the existence of an algorithm for PAC learning with one-sided noise, which is itself a challenging
problem and not known to exist for even simple concept classes.
In this work, we do not assume instances within each bag are independently distributed, and we
do not require the existence of PAC learning algorithms for one-sided noise. Instead, we give efficient
algorithms for labeling future bags when the class labeling instances is an unknown halfspace with
a margin or an unknown depth-two neural network. We succeed with respect to general monotone,
smooth combining functions.
Notation. Let us denote the space of instances as X and the space of bags as B ⊆ X ∗ . Let N be
an upper bound on the size of the bags, that is, N = maxβ∈B |β|. Let the instance labeling function
be c : X → R and the bag labeling function be fbag . We assume a distribution D over the bags and
allow the instances within the bag to be dependent on each other. We consider two variants of the
relationship between the instance and bag labeling functions and corresponding learning models.
22
7.1
Deterministic MIL
In the deterministic case we assume that there is a deterministic map from instance labels to bag
labels. This is the standard model that has been studied in literature.
Definition 7 (Deterministic MI Assumption). Given combining function u : {0, 1}∗ → {0, 1},
for bag β = {x1 , . . . , xr }, fbag (β) is a fixed function such that fbag (β) = u(c(β)) where c(β) =
(c(x1 ), · · · , c(xr )) and c is the instance labeling function.
Definition 8 (Deterministic MIL). The concept class C is (ǫ, δ)-Deterministic MIL for u with
sample complexity M and running time T if under the deterministic MIL assumption for u, there
exists an algorithm A such that for all c ∈ C as the instance labeling function and any distribution
D on B, A draws at most M iid bags and runs in time T , to output a bag-labeling hypothesis h such
that with probability 1 − δ,
P rβ∼D [h(β) 6= fbag (β)] ≤ ǫ.
The following theorem follows directly from Theorem 6 setting the input space as B.
2
Theorem 11. The concept class C is (ǫ, δ)-Deterministic MIL for u with sample complexity C( BL
ǫ ) ·
log(1/δ) for sufficiently large constant C > 0 and running time poly(n, B, L, 1/ǫ, log(1/δ)) if for all
c ∈ C, u ◦ c is B-separated by some kernel K.
Thus, we need to show that OR of halfspaces with a margin are B-separated for an appropriate
B by some kernel function K. We have the following lemma:
Lemma 17. The class of OR of halfspaces with margin ρ on bags of size ≤ N√ with elements in
Sn−1 is B-separated by a kernel K such that ∀β ∈ B, K(β, β) ≤ 1 with B = N O( 1/ρ) .
Proof. Let w correspond to the weight vector of the halfspace that generates the instance labeling
function. Since OR is the combining function, we consider POR given by Lemma 4 for given ρ and
bag size (sets of size) ≤ N which separates the required class.
1
N X
POR (β) = − − N +
p(w · x)
2
|β|
x∈β
p
where p(a) = Tr (1 + a) for r = O( 1/ρ). From the proof of Lemma 10, we know that the sum of
squares of the coefficients of p(a) are bounded by 2O(r⌈log 2N ⌉) = N O(r) . Now using Lemma 1, we
know that ∀x ∈ β, p(w · x) = hv, ψd (x)i for the feature vector
ψd of the multinomialPkernel MKd
1
N P
O(r)
for d = r⌈log 2N ⌉. We also have ||v|| ≤ N
. Thus, |β| x∈β p(w · x) = hN v, |β|
x∈β ψ(s)i =
6
hN v, ψmean (β)i where ψmean is the feature vector of the mean map kernel of MKd . The constant
term 1/2 can be incorporated by adding an additional entry to the mean map feature vector ψmean
corresponding to constant 1, that is, consider new feature vector ψ ′ (S) = [ψmean (S), 1]. Observe
that the corresponding kernel K′ (S, T ) = Kmean (S, T ) + 1. Note that ||ψ ′ (S)||2 = ||ψmean (S)||2 + 1 =
O(r⌈log 2N ⌉). We can normalize this to norm 1 by scaling the weight vector appropriately as before.
Thus, OR ◦ c is N O(r) -separated by kernel K′ .
6
The mean map kernel is essential for our analysis as it handles varying bag sizes requiring only an upper bound
on the size of the bags. An alternative would be to concatenate the feature vectors of each instance in a bag to form
a new kernel but this would only work if all bags had the same fixed size.
23
This in turn implies deterministic MIL for the concept class of halfspaces with a constant margin
and OR as the combining function. A similar proof can be given for AND. Note that for the proof
we do not require independence among the instances but just among the bags as we can embed each
possible bag in B into an RKHS.
Corollary 5. The concept class of halfspaces with constant margin is (ǫ, δ)-Deterministic MIL for
OR/AND and bag size ≤ N with sample complexity and running time poly(n, N, 1/ǫ, log(1/δ)).
Similarly, following Lemma 11, we can extend the above result to majority.
Corollary 6. The concept class of halfspaces with constant margin is (ǫ, δ)-Deterministic MIL for
majority and bag size ≤ N with sample complexity and running time poly(n, N, 1/ǫ, log(1/δ)).
7.2
Probabilistic MIL
We generalize the deterministic model to allow the labeling function to induce a probability distribution over the labels. This assumption seems more intuitive and less restrictive than the deterministic
case as it allows for noise in the labels.
Definition 9 (Probabilistic MI Assumption). Given combining
function u: R → [0, 1], for bag β,
P
1
fbag (β) is a random variable such that P r[fbag (β) = 1] = u |β| · x∈β c(x) where c is the instance
labeling function.
Definition 10 (Probabilistic MIL). The concept class C is (ǫ, δ)-Probabilistic MIL for u with sample
complexity M and running time T if under the probabilistic MI assumption for u, there exists an
algorithm A such that for all c ∈ C as the instance labeling function and any distribution D on B,
A draws at most M iid bags and runs in time at most T to return a bag-labeling hypothesis h such
that with probability 1 − δ,
2
X
1
·
c(x) ≤ ǫ.
Eβ∼D h(β) − u
|β|
x∈β
The following lemma along with Theorem 5 give us learnability results in the Probabilistic MIL
setting.
Lemma 18. Let c be the instance labeling function mapping X to R such that c is (ǫ, B)-approximated
P
1
· x∈β c(x) is
by kernel K and feature vector ψ. Then the function f : B → R given by f (β) = |β|
(ǫ, B)-approximated by the mean map kernel of K.
Proof. We have that ∀x ∈ X , |c(x) − hv, ψ(x)i| ≤ ǫ for v such that ||v|| ≤ B. Let Kmean be the mean
map kernel of K and ψmean be the corresponding vector. We will show that v (ǫ, B)-approximates
f in Kmean . This follows from the following,
|f (β) − hv, ψmean (β)i| =
1 X
1 X
·
·
c(x) −
hv, ψ(x)i
|β|
|β|
x∈β
x∈β
1 X
·
|c(x) − hv, ψ(x)i| ≤ ǫ.
≤
|β|
x∈β
24
Theorem 12. The concept class C is (ǫ, δ)-Probabilistic MIL for monotone L-Lipschitz u with
2
sample complexity ( BL
ǫ ) · log(1/δ) and running time poly(n, B, L, 1/ǫ, log(1/δ)) if all c ∈ C are
(ǫ/CL, B)-approximated by some kernel K for large enough constant C > 0.
Proof. Consider c ∈ C that is (ǫ/CL, B)-approximated by some kernel K for large P
enough con1
· x∈β c(x) is
stant C > 0 (to be chosen later). Using Lemma 18 we know that f (β) = |β|
(ǫ/CL, B)-approximated by the mean map kernel Kmean of K. Applying Theorem 5, we get that
with probability 1 − δ,
Eβ∼D h(β) − u
1
·
|β|
X
x∈β
2
C′
ǫ.
c(x) ≤
C
for sufficiently large constant C ′ > 0. Choosing C ≥ C ′ gives the result.
Combining Theorem 12 with learnability Theorems 7 and 9 we can show the following polynomial
time Probabilistic MIL results.
Corollary 7. For any monotone L-Lipschitz function u, the concept class of sigmoids over Sn−1
are (ǫ, δ)-Probabilistic MIL with sample complexity and running time poly(n, L, 1/ǫ, log(1/δ)).
Corollary 8. For any monotone L-Lipschitz function u, the concept class of halfspaces with a
constant margin over Sn−1 are (ǫ, δ)-Probabilistic MIL with sample complexity and running time
poly(n, L, 1/ǫ, log(1/δ)).
8
Conclusions and Future Work
We have given the first assumption-free, polynomial-time algorithm for learning a class of nonlinear neural networks with more than one hidden layer. We only require that the conditional mean
function is computed by such a network, and therefore our results also capture Boolean learning
problems. In contrast, PAC learning corresponding Boolean function classes seems intractable.
Understanding the broadest class of conditional mean functions that admit efficient learning
algorithms is a natural open problem. For example, can we learn two hidden layers with multiple
hidden units in the second layer?
We remark that while our algorithms are noise-tolerant in the sense that we make assumptions
only on the conditional mean function, it is unlikely that our results can be extended to the fully
non-realizable (agnostic) setting, as there are hardness results for agnostically learning even a single
ReLU in fully-polynomial time [GKKT16].
Finally, it is easy to see that depth-3 neural networks with one hidden layer of ReLUs feeding
into a sigmoid can approximate (with negligible error), polynomial-size DNF formulas. Thus, it is
unlikely we can obtain polynomial-time algorithms for such architectures [DSS16].
References
[AB99]
Martin Anthony and Peter L. Bartlett. Neural Network Learning: Theoretical Foundations. Cambridge University Press, Cambridge, England, 1999.
25
[AHW96]
Peter Auer, Mark Herbster, and Manfred K. Warmuth. Exponentially many local minima for single neurons. In Advances in Neural Information Processing Systems, volume 8,
pages 316–322. The MIT Press, 1996.
[Amo13]
Jaume Amores. Multiple instance classification: Review, taxonomy and comparative
study. Artificial Intelligence, 201:81–105, 2013.
[BG17]
Alon Brutzkus and Amir Globerson. Globally optimal gradient descent for a convnet
with gaussian inputs. CoRR, abs/1702.07966, 2017.
[BK98]
Avrim Blum and Adam Kalai. A note on learning from multiple-instance examples.
Machine Learning, 30(1):23–29, 1998.
[BM02]
Peter L. Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk
bounds and structural results. Journal of Machine Learning Research, 3:463–482, 2002.
[CHM+ 15] Anna Choromanska, Mikael Henaff, Michaël Mathieu, Gérard Ben Arous, and Yann
LeCun. The loss surfaces of multilayer networks. In AISTATS, volume 38 of JMLR
Workshop and Conference Proceedings. JMLR.org, 2015.
[Dan15]
Amit Daniely. A ptas for agnostically learning halfspaces. In Conference on Learning
Theory, pages 484–502, 2015.
[Dan16]
Amit Daniely. Complexity theoretic limitations on learning halfspaces. In STOC, pages
105–117. ACM, 2016.
[DFS16]
Amit Daniely, Roy Frostig, and Yoram Singer. Toward deeper understanding of neural
networks: The power of initialization and a dual view on expressivity. In NIPS, pages
2253–2261, 2016.
[DLLP97] Thomas G Dietterich, Richard H Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence, 89(1):31–71,
1997.
[DSS16]
Amit Daniely and Shai Shalev-Shwartz. Complexity theoretic limitations on learning
dnf’s. In Proceedings of the 29th Conference on Learning Theory, COLT 2016, New York,
USA, June 23-26, 2016, volume 49 of JMLR Workshop and Conference Proceedings,
pages 815–830. JMLR.org, 2016.
[Fel12]
Vitaly Feldman. Learning dnf expressions from fourier spectrum. In Shie Mannor,
Nathan Srebro, and Robert C. Williamson, editors, COLT 2012 - The 25th Annual
Conference on Learning Theory, June 25-27, 2012, Edinburgh, Scotland, volume 23 of
JMLR Proceedings, pages 17.1–17.19. JMLR.org, 2012.
[GK17]
Surbhi Goel and Adam Klivans. Eigenvalue decay implies polynomial-time learnability
of neural networks. In NIPS, 2017.
[GKK08]
Parikshit Gopalan, Adam Tauman Kalai, and Adam R Klivans. Agnostically learning
decision trees. In Proceedings of the fortieth annual ACM symposium on Theory of
computing, pages 527–536. ACM, 2008.
26
[GKKT16] Surbhi Goel, Varun Kanade, Adam Klivans, and Justin Thaler. Reliably learning the
relu in polynomial time. arXiv preprint arXiv:1611.10258, 2016.
[HVB+ 16] Francisco Herrera, Sebastián Ventura, Rafael Bello, Chris Cornelis, Amelia Zafra, Dánel
Sánchez-Tarragó, and Sarah Vluymans. Multiple instance learning. In Multiple Instance
Learning, pages 17–33. Springer, 2016.
[Jac97]
Jeffrey C. Jackson. An efficient membership-query algorithm for learning dnf with respect to the uniform distribution. J. Comput. Syst. Sci, 55(3):414–440, 1997.
[JSA15]
Majid Janzamin, Hanie Sedghi, and Anima Anandkumar. Beating the perils of nonconvexity: Guaranteed training of neural networks using tensor methods. arXiv preprint
arXiv:1506.08473, 2015.
[Kaw16]
Kenji Kawaguchi. Deep learning without poor local minima. In Daniel D. Lee, Masashi
Sugiyama, Ulrike V. Luxburg, Isabelle Guyon, and Roman Garnett, editors, NIPS, pages
586–594, 2016.
[KKKS11] Sham M. Kakade, Adam Kalai, Varun Kanade, and Ohad Shamir. Efficient learning
of generalized linear and single index models with isotonic regression. In NIPS, pages
927–935, 2011.
[KM93]
Eyal Kushilevitz and Yishay Mansour. Learning decision trees using the fourier spectrum. SIAM Journal on Computing, 22(6):1331–1348, 1993.
[KM13]
Adam R. Klivans and Raghu Meka. Moment-matching polynomials. Electronic Colloquium on Computational Complexity (ECCC), 20:8, 2013.
[KOS04]
A. Klivans, R. O’Donnell, and R. Servedio. Learning intersections and thresholds of
halfspaces. JCSS: Journal of Computer and System Sciences, 68, 2004.
[KS90]
Michael J Kearns and Robert E Schapire. Efficient distribution-free learning of probabilistic concepts. In Foundations of Computer Science, 1990. Proceedings., 31st Annual
Symposium on, pages 382–391. IEEE, 1990.
[KS94]
Michael J. Kearns and Robert E. Schapire. Efficient distribution-free learning of probabilistic concepts. J. of Comput. Syst. Sci., 48(3):464–497, 1994.
[KS08]
Adam R Klivans and Rocco A Servedio. Learning intersections of halfspaces with a
margin. Journal of Computer and System Sciences, 74(1):35–48, 2008.
[KS09a]
Adam Kalai and Ravi Sastry. The isotron algorithm: High-dimensional isotonic regression. In COLT, 2009.
[KS09b]
Adam R. Klivans and Alexander A. Sherstov. Cryptographic hardness for learning
intersections of halfspaces. J. Comput. Syst. Sci, 75(1):2–12, 2009.
[KST09]
Sham M Kakade, Karthik Sridharan, and Ambuj Tewari. On the complexity of linear
prediction: Risk bounds, margin bounds, and regularization. In Advances in neural
information processing systems, pages 793–800, 2009.
27
[LSSS14]
Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of
training neural networks. In Advances in Neural Information Processing Systems, pages
855–863, 2014.
[LT91]
Michel Ledoux and Michel Talagrand. Probability in Banach Spaces: Isoperimetry and
Processes. Springer, 1991.
[LY17]
Yuanzhi Li and Yang Yuan. Convergence analysis of two-layer neural networks with
relu activation. In NIPS 2017, 2017.
[SA14]
Hanie Sedghi and Anima Anandkumar. Provable methods for training neural networks
with sparse connectivity. arXiv preprint arXiv:1412.2693, 2014.
[SC16]
Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error
guarantees for multilayer neural networks. CoRR, abs/1605.08361, 2016.
[SGSS07]
Alex Smola, Arthur Gretton, Le Song, and Bernhard Schölkopf. A hilbert space embedding for distributions. In International Conference on Algorithmic Learning Theory,
pages 13–31. Springer, 2007.
[Sha16]
Ohad Shamir. Distribution-specific hardness of learning neural networks. arXiv preprint
arXiv:1609.01037, 2016.
[She12]
Alexander A Sherstov. Making polynomials robust to noise. In Proceedings of the fortyfourth annual ACM symposium on Theory of computing, pages 747–758. ACM, 2012.
[Sol17]
Mahdi Soltanolkotabi. Learning relus via gradient descent. In NIPS, 2017.
[SS02]
Bernhard Schölkopf and Alexander J Smola. Learning with kernels: support vector
machines, regularization, optimization, and beyond. MIT press, 2002.
[SS16]
Itay Safran and Ohad Shamir. Depth separation in relu networks for approximating
smooth non-linear functions. CoRR, abs/1610.09887, 2016.
[SSSS11]
Shai Shalev-Shwartz, Ohad Shamir, and Karthik Sridharan. Learning kernel-based halfspaces with the 0-1 loss. SIAM J. Comput., 40(6):1623–1646, 2011.
[ST12]
Sivan Sabato and Naftali Tishby. Multi-instance learning with any hypothesis class.
Journal of Machine Learning Research, 13(Oct):2999–3039, 2012.
[SVWX17] Le Song, Santosh Vempala, John Wilmes, and Bo Xie. On the complexity of learning
neural networks. arXiv preprint arXiv:1707.04615, 2017.
[Val84]
Leslie G Valiant. A theory of the learnable. Communications of the ACM, 27(11):1134–
1142, 1984.
[Wik16]
Wikipedia. Multinomial theorem — Wikipedia, the free encyclopedia, 2016. URL:
https://en.wikipedia.org/wiki/Multinomial_theorem.
[ZLJ16]
Yuchen Zhang, Jason Lee, and Michael Jordan. ℓ1 networks are improperly learnable in
polynomial-time. In ICML, 2016.
28
[ZPS17]
Qiuyi Zhang, Rina Panigrahy, and Sushant Sachdeva. Electron-proton dynamics in deep
learning. CoRR, abs/1702.00458, 2017.
[ZSJ+ 17]
Kai Zhong, Zhao Song, Prateek Jain, Peter L. Bartlett, and Inderjit S. Dhillon. Recovery
guarantees for one-hidden-layer neural networks. In ICML, volume 70, pages 4140–4149.
JMLR.org, 2017.
29
| 8cs.DS
|
SCIENCE CHINA
Information Sciences
. RESEARCH PAPER .
arXiv:1703.09499v2 [cs.CV] 16 Nov 2017
Locality preserving projection on SPD matrix Lie
group: algorithm and analysis
Yangyang LI1,2* & Ruqian LU1
1Academy
of Mathematics and Systems Science Key Lab of MADIS
Chinese Academy of Sciences, Beijing 100190, China;
2University of Chinese Academy of Sciences, Beijing 100049, China
Abstract Symmetric positive definite (SPD) matrices used as feature descriptors in image recognition are
usually high dimensional. Traditional manifold learning is only applicable for reducing the dimension of
high-dimensional vector-form data. For high-dimensional SPD matrices, directly using manifold learning
algorithms to reduce the dimension of matrix-form data is impossible. The SPD matrix must first be
transformed into a long vector, and then the dimension of this vector must be reduced. However, this
approach breaks the spatial structure of the SPD matrix space. To overcome this limitation, we propose a
new dimension reduction algorithm on SPD matrix space to transform high-dimensional SPD matrices into
low-dimensional SPD matrices. Our work is based on the fact that the set of all SPD matrices with the
same size has a Lie group structure, and we aim to transform the manifold learning to the SPD matrix Lie
group. We use the basic idea of the manifold learning algorithm called locality preserving projection (LPP)
to construct the corresponding Laplacian matrix on the SPD matrix Lie group. Thus, we call our approach
Lie-LPP to emphasize its Lie group character. We present a detailed algorithm analysis and show through
experiments that Lie-LPP achieves effective results on human action recognition and human face recognition.
Keywords Manifold learning, SPD matrix Lie group, Locally preserving projection, Laplace operator,
Log-Euclidean metric
Citation
Yangyang LI, Ruqian LU. Locality Preserving Projection on SPD Matrix Lie Group. Sci China Inf
Sci, for review
1
Introduction
Image recognition, including dynamic action recognition and static image recognition, is a popular research subject in the field of machine vision and pattern recognition [1] [2] [3] [14]. This technique has a
wide range of applications in many fields, such as intelligent video retrieval and perceived interaction. One
key step of image recognition is to construct a high-quality image feature descriptor, which determines
the accuracy rate of recognition. The original image feature descriptor is a pixel matrix, which is usually
transformed into a high-dimensional row feature vector because image recognition on the pixel matrix
space is difficult. Given the high dimension of the feature vector, manifold learning algorithms are applied
to implement image recognition [1] [9] [15] [16]. However, the vector-form feature descriptor breaks the
geometric structure of the pixel matrix space and is highly sensitive to various factors such as illumination
intensity, background, and object location. To avoid these disadvantages, tensor space dimensionality
reduction based on locality preserving projection (LPP) [5] was proposed in [6]; this approach is linear
* Corresponding author (email: [email protected])
Yangyang LI, et al.
Sci China Inf Sci 2
and deals with the image pixel matrix directly. F. Porikli et al. [12] presented a new feature descriptor
by computing a feature covariance matrix within any size region in an image, which preserves the local
geometric structure of the pixel matrix (see Appendix A in the supplementary file). Covariance matrix,
which is also called symmetric positive definite (SPD) matrix descriptor, has mainly been used in static
image recognition [10] [12] [14] [11]. For human action recognition, A. Sanin et al. [4] proposed a new
method based on spatial-temporal covariance descriptors. H. Tabia et al. [13] applied SPD matrices as
descriptors of 3D skeleton location by constructing covariance matrices on 3D joint locations.
A set of SPD matrices with the same size forms a Riemannian manifold. This SPD Riemannian
D
manifold has a group structure that forms an SPD matrix Lie group S+
. The group operation in [7] is
shown as
.
D
S1 , S2 ∈ S+
, S1 ⊙ S2 = exp (log (S1 ) + log (S2 )) .
We proposed that the dimensions of SPD matrices calculated from image regions are especially high in
general. As a result of the matrix-form data of the matrix Lie group, directly reducing the dimension of
high-dimensional matrix Lie group is difficult without breaking the structure of the matrix form.
Harandi et al. [18] proposed to learn a kernel function to first map the SPD matrices into a higherdimensional Euclidean space and then use LPP [5] to reduce their dimension. However, this method would
distort the geometric and algebraic structure of the SPD manifold and would lose a considerable amount of
important structure information. To overcome this limitation, Harandi et al. [19] suggested mapping the
original SPD manifold to a Grassmann manifold and then solving the dimension reduction problem on the
latter. However, this method has a high time cost. A similar study was conducted by Huang et al. [7],
who proposed transforming SPD matrices to their corresponding tangent space and learning a linear
dimensionality reduction map on that tangent space. However, this algorithm needs several parameters,
which are sensitive factors that influence the algorithm. Overall, all three methods [7] [18] [19] require a
linear space and rely on nonlinear dimensionality reduction mappings.
According to the definition of covariance matrix in [8] [12], each covariance matrix can be represented
by the product of a set of feature vectors. Covariance matrix summarizes the linear bivariate relationships
among a set of variables. Thus, we can solve the dimensionality reduction problem directly on the SPD
matrix Lie group. We extend the idea of LPP [5] to the dimensionality reduction learning on the SPD
matrix Lie group in a new approach called Lie-LPP.
The main contributions of our work can be summarized as follows:
• The LPP algorithm in [5] is extended to Lie-LPP and applied to the SPD matrix Lie group. A
bilinear dimensionality reduction mapping on the SPD matrix Lie group is obtained, which preserves the
intrinsic geometric and algebraic structure of the SPD matrix Lie group.
• To overcome the limitation of other methods regarding dimensionality reduction of the SPD matrix
Lie group, our method solves the dimensionality reduction problem of the SPD matrix Lie group directly
without mapping to other spaces and without needing numerous sensitive parameters, thereby resulting
in a simple and straightforward approach.
• The graph Laplacian matrix is constructed on the SPD matrix Lie group to reflect the intrinsic
geometric and algebraic structure of the original SPD matrix Lie group.
• A detailed algorithm analysis in theory is given to analyze the difference and relationship between
Lie-LPP and LPP. The main conclusions are shown in Theorems 4.1 and 4.2 and in Proposition 4.1.
This paper is a further extension of the algorithm and analysis in our previous paper [24]. In [24], we
simply presented a simple description of our algorithm. In this paper, we present the detailed algorithm
analysis in theory, as well as full experiments on human action and face databases.
Yangyang LI, et al.
2
2.1
Sci China Inf Sci 3
Background
Covariance Matrix
Suppose a set of n feature vectors from an image region are expressed as the following matrix F , where
the dimensionality of each feature vector is d:
F = (f1 , f2 , . . . , fn ) ,
fi = (fi1 , fi2 , . . . , fid ) is the ith feature vector. The corresponding covariance matrix C with respect to
these feature vectors is defined as [8]
n
C=
1
1X
fi fiT .
FFT =
n
n i=1
(1)
In this definition, we assume the expectation of feature vectors is zero and each term fi fiT in the summation is the outer product of the feature vector fi . Obviously, covariance matrix is a semi-positive
definite matrix. In this paper, we consider only the positive covariance matrix called the SPD matrix.
When the feature vectors are adjacent, the corresponding covariance matrices are also adjacent. Thus,
the covariance matrices preserve the local geometric structure of the corresponding feature vectors. The
detailed proof is shown in Appendix A.
Covariance matrices (SPD matrices) have several advantages as the feature descriptors of images.
First, covariance matrices can fuse all the features of images. Second, they provide a way of filtering
noisy information such as illumination and the location of the object in the image. In addition, the size
of the covariance matrix is dependent on the dimensionality of feature vectors other than the size of the
image region. Thus, we can construct the covariance matrices with the same size from different regions.
2.2
Geometric structure of SPD Matrix Lie group
In machine learning, we usually have to learn an effective metric for comparing data points. In particular,
in the image recognition step, a metric is required to measure the distance between two different image
feature descriptors. In this paper, we use the SPD matrices as feature descriptors of images. Thus, the
corresponding Riemannian metric needs to be constructed on the SPD matrix Lie group to compute the
intrinsic geodesic distances between any two SPD matrices.
D
, where every point
The SPD matrix Lie group that we consider in this paper is represented by S+
D
D
S1 ∈ S+ is a D × D size matrix. The tangent space of S+ at the identity is Sym(D), a bilinear space
d
of symmetric matrices. The learned lower-dimensional SPD matrix Lie group is represented by S+
. The
family of all scalar products on all tangent spaces of SPD matrix Lie group is known as the Riemannian
D
metric. The geodesic distance dG (S1 , S2 ) between any two points S1 , S2 ∈ S+
can be computed under
this Riemannian metric. In this paper, we choose the Log-Euclidean metric (LEM) from [23] as the
Riemannian metric of SPD matrix Lie group. The detailed definition of LEM is shown in Definition 2.1.
D
Definition 2.1. (Log-Euclidean Metric) [23] The Riemannian metric at a point S1 ∈ S+
is a
D
scalar product defined in the tangent space TS1 S+ :
hT1 , T2 i = hDlog.T1 , Dlog.T2 i,
(2)
D
where T1 , T2 ∈ TS1 S+
, Dlog.T1 indicates the directional derivative of the matrix logarithm along T1 [7].
LEM is a bi-invariant metric defined on the SPD matrix Lie group. The corresponding theoretical
conclusion is shown in [23].
Definition 2.2. (Bi-invariant Metric) [23] Any bi-invariant metric on the Lie group of SPD matrices
is also called a LEM because it corresponds to a Euclidean metric (EM) in the logarithm domain.
The logarithm domain is the tangent space of the SPD matrix Lie group.
Corollary 2.3. (Flat Riemannian Manifold) [23] Endowed with a bi-invariant metric, the space
of SPD matrices is a flat Riemannian space; its sectional curvature is null everywhere.
Yangyang LI, et al.
Sci China Inf Sci 4
D
Thus, under LEM, S+
is a flat manifold and locally isometric to the tangent space Sym(D). In a local
neighborhood, the mapping between the SPD matrix Lie group and the corresponding tangent space is
represented by the exponential map, and the inverse map is the logarithm shown in Eqs. 3 and 4.
expS1 (T1 ) = exp (log (S1 ) + DS1 log · T1 ) ,
(3)
logS1 (S2 ) = Dlog(S1 ) exp · (log (S2 ) − log (S1 )) ,
(4)
D
D
a tangent vector. expS1 (T1 ) is
where the exponential map is defined at point S1 ∈ S+
, T1 ∈ TS1 S+
D
the exponential representation of T1 on S+ , and the corresponding logarithmic representation of S2 at
D
is logS1 (S2 ).
TS1 S+
The geodesic distance between S1 , S2 under LEM is defined as follows [7]:
dG (S1 , S2 ) = hlogS1 (S2 ) , logS1 (S2 )i =k log (S1 ) − log (S2 ) k2F ,
(5)
where k · kF represents the Frobenius norm of the matrix and logS1 (S2 ) is the logarithm of S2 at S1 .
Under LEM, the SPD Matrix Lie group is a complete manifold. Thus, every two points on the SPD
matrix Lie group are linked by the shortest geodesic line.
3
Algorithm
In this section, we first analyze the LPP algorithm. Then, we give the construction process of graph
Laplacian matrix on the SPD matrix Lie group. Finally, we describe the process of our proposed dimensionality reduction algorithm Lie-LPP.
3.1
Locally Preserving Projection (LPP)
LPP aims to learn a linear dimensionality reduction map to reduce the dimension of high-dimensional
vector-form data points, which can be seen as a linear approximation of LEP [17]. This linear dimensionality reduction map optimally preserves the local neighborhood geometric structure of a dataset by
building a graph Laplacian matrix on the dataset. The graph Laplacian matrix is a discrete approximation of a differential Laplace operator that arises from the manifold. Let X = [x1 , x2 , · · · , xN ] ⊂ RD is the
input data set distributed on a d-dimensional manifold M, which is embedded into RD . a is the linear dimensionality reduction map. The learned lower-dimensional data set is Y = aT X = [y1 , y2 , · · · , yN ] ⊂ Rd .
The algorithm of LPP is stated as follows:
• Constructing the adjacency graph: Denote G a graph with N nodes. If xi and xj are ”close”,
then a connection exists between nodes i and j. The ”closenes” between two nodes is measured by the
K-nearest neighbor method.
• Choosing the weights: Here, the authors denote matrix W as the corresponding weight matrix,
where W is a sparse symmetric matrix with size N × N . The element Wij is defined as follows:
kxi −xj k2
t
, if nodes i and j are connected,
Wij = e−
Wij = 0, if nodes i and j are not connected.
• Eigenmaps: The following generalized eigenvector problem is solved to obtain the corresponding
dimension reduction map a:
XLX T a = λXDX T a,
(6)
P
where D is a diagonal matrix, Dii = j Wji . L = D − W is the graph Laplacian matrix.
Thus, the critical step of LPP is to construct the graph Laplacian matrix L on data points. The global
lower dimensional representations are learned by solving the corresponding generalized eigenfunction.
Yangyang LI, et al.
3.2
Sci China Inf Sci 5
Laplace operator on SPD matrix Lie group
The Laplace operator is a significant operator defined on Riemannian manifold [28]. It measures the
intrinsic structure of manifold such as the curvature of manifold, the similarities among different points
on Riemannian manifold. The Laplacian matrix on a graph is a discrete analog of the Laplace operator
that we are familiar with in functional analysis [27]. The critical step of the LPP algorithm is to construct
a graph Laplacian matrix to represent the intrinsic local geometric structure of data points, which is
applied only on vector-form data points. For Lie-LPP, we aim to uncover the intrinsic structure of SPD
matrices, which are essentially different from vector-form data points in a spatial structure. The Laplace
operator is defined based on the Riemannian metric. For vector-form data points, the Laplacian matrix
is constructed based on the EM in each local patch. For SPD matrices, we need to use LEM [22] [23] to
construct the corresponding Laplacian matrix.
Several difficulties are encountered for the learning of the Laplacian matrix on SPD matrices. One
critical difficulty is the discrete representations of the first- and second-order derivatives on SPD matrix
space. Another difficulty is the need to find the representation of the Laplace operator on the SPD matrix
Lie group, which is different from the structure on vector space. To construct a Laplacian matrix on SPD
matrices, we need to solve these difficulties during the learning process. In addition, the core of Lie-LPP is
to construct an accurate Laplacian matrix on the SPD matrix Lie group before dimensionality reduction.
The construction process of a Laplacian matrix on SPD matrices and the descriptions for solving these
difficulties are given in detail in a separate section.
To better understand the construction of the Laplacian matrix, we first present an intuitive example
of a set of one dimensional nodes. Consider a graph G with n nodes. Every node i is adjacent to two
nodes i − 1 and i + 1. If we assign value vi to node i, then the Laplacian is represented as (vi+1 − vi ) −
(vi − vi−1 ). Thus, vi+1 − vi is the discrete analog for a first-order derivative defined over the real number
line. −2 vi − 12 vi−1 − 12 vi+1 is the discrete approximation of the second-order derivative. For higher
dimensions the ’normalized’ graph Laplacian ∆nm about function f is defined as:
n
(∆nm f ) (i) = f (i) −
1 X
Wij f (j) ,
deg (i) j=1
(7)
Pn
where the degree function deg is defined as deg (i) = j=1 Wji . Wji is the heat kernel weight defined in
the same way as in the second algorithmic procedure of LPP [5].
The abovementioned Laplacian matrix is defined on vector-form data points only. In the following, we
construct the ’normalized’ graph Laplacian matrix on the SPD matrix Lie group endowed with LEM. Sup−−−−−−−−−−→
D
pose a parameterized SPD matrix Lie group Σ defined as Σ : Rd → S+
, we call the vector Σ (x) Σ (x + u)
D
as the standard first-order derivative on S+
[21]:
1
1
−−−−−−−−−−→
Σ (x) Σ (x + u) = Σ (x) 2 (logΣ (x + u) − logΣ (x)) Σ (x) 2 .
(8)
The Laplace-Beltrami operator ∆ about function Σ is defined as:
∆Σ =
∆i Σ
d
X
∆i Σ,
i=1
= ∂i2 Σ
− 2 (∂i Σ) Σ
(9)
(−1)
(∂i Σ) .
To approximate the graph Laplacian matrix on SPD matrices, we first need to approximate the firstand second-order derivatives on SPD matrix Lie group. Under the approximation of the second-order
D
derivative, the corresponding approximation of the Laplace-Beltrami operator on S+
[21] is:
−−−−−−−−−−→ −−−−−−−−−−→
∆u Σ = ∂u2 Σ − 2 (∂u Σ) Σ(−1) (∂u Σ) = Σ (x) Σ (x + u) + Σ (x) Σ (x − u) + O kuk4 .
(10)
To compute the complete Laplacian matrix of the SPD matrix Lie group of Eq. 9, we only have to
compute the Laplace operator along d orthonormal directions.
Yangyang LI, et al.
Sci China Inf Sci 6
For a discrete dataset, suppose Σ1 , Σ2 , · · · , ΣN are a set of SPD matrices generated from Σ. The
?normal’ graph Laplacian matrix on this data set is
n
X
1
1
fij log (Σj ) Σ 2 ,
W
(11)
(∆nm Σi ) = Σi2 log (Σi ) −
i
j=1
klog(Σ
)−log(Σ )k2
i
j
F
fij = e−
fij = 0. The corresponding graph
t
where W
, if i and j are connected, else W
e
e
f
f
fij defined
Laplacian matrix on a set of SPD matrices is L = D − W , where W is a symmetric matrix, W
P
e is a diagonal matrix, D
e ii =
f e
as above and D
j Wji . L is a discrete representation of the Laplace operator
on the SPD matrix Lie group.
3.3
Lie-LPP Algorithm
D
On the basis of the definition of graph Laplacian matrix on S+
and the algorithmic procedures of LPP,
we show our Lie-LPP algorithm as follows:
3.3.1 Lie-LPP Algorithm
D
For the SPD matrix Lie group S+
, the SPD matrix logarithms in the tangent space are also symmetric
matrices. The bilinear mapping between tangent spaces is defined as follows:
f (log (S1 )) = AT log (S1 ) A.
(12)
The corresponding mapping between SPD matrix Lie groups is
g (S1 ) = exp ◦ f (log (S1 )) = exp AT log (S1 ) A ,
(13)
D
where S1 ∈ S+
, A is a linear map matrix, f is the corresponding map defined on Lie algebras, and g is
the derived map defined on SPD matrix Lie groups.
g (S1 ) is still an SPD matrix; this idea is easily proven. In this paper, we attempt to learn a transforD
, and d × d is
mation matrix A where A ∈ RD×d is a full column rank matrix. D × D is the size of S+
d
the mapped size of S+ , D >> d. The linear map matrix A is proven to preserve the algebraic structure
D
d
of S+
. To obtain a discriminative SPD matrix Lie group S+
, A should also inherit and preserve the
D
geometric structure of S+ . According to the idea of LPP, the key step of the Lie-LPP algorithm is to
D
D
construct the Laplacian matrix L on S+
, which reflects the local geometrical structure of S+
. Under
D
D
LEM, S+ is locally isometric to the tangent space of S+ . Thus, the geodesic distance between two SPD
matrices is equal to the Euclidean distance between the corresponding points on tangent space.
D
Suppose the input data points are S1 , S2 , . . . , SN ∈ S+
, and the output sample points are Y1 , Y2 , . . . , YN ∈
d
S+ , where N is the number of sampled points.
The algorithm steps of Lie-LPP are as follows:
• The first step is to divide the input SPD matrices into a set of local patches. We use the K-nearest
method to find the K-nearest neighborhoods Ui of every point Si , where the distance metric between
D
two points is defined as their geodesic distance on the SPD matrix Lie group S+
using Eq. 5.
• The second step is to construct a weight matrix W on each local patch Ui to represent the local
intrinsic geometric structure of Ui .
klog(S
)−log(S )k2
i
j
F
fij = e−
t
, if Sj ∈ Ui ,
W
f
Wij = 0, else if Sj ∈
/ Ui .
fij is based on the construction of a Laplacian matrix on input SPD
The definition of weight value W
matrices from Eq. 11.
Yangyang LI, et al.
Sci China Inf Sci 7
• The third step is to compute the eigenvalues and the corresponding eigenvectors for the generalized
eigenfunction problem
e
e
S T LSA
= λS T DSA,
(14)
where S = [log (S1 ) , log (S2 ) , · · · , log (SN )]T is a partitioned matrix.
3.3.2 Optimal Embedding
The optimal dimension reduction map A is obtained by minimizing the following energy function:
1X
fij ,
dG (Yi , Yj ) W
2 i,j
(15)
fij is the corresponding weight.
where dG (Yi , Yj ) is the geodesic distance between Yi and Yj , while W
According to the definition of geodesic distance and LEM on the SPD matrix Lie group in Eq. 5, the
energy function Eq. 15 can be transformed into the following equation:
1X
fij ,
k logYi − logYj k2F W
2 i,j
(16)
According to Eq. 12, the optimization function Eq. 16 is represented as follows:
1X
fij
k logYi − logYj k2F W
2 i,j
1X
fij
k AT logSi A − AT logSj A k2F W
=
2 i,j
e −W
f SP
= tr P T S T D
e
= tr P T S T LSP
,
(17)
f and D
e are two sparse N ×N block matrices, W
f = [W
fij ID×D ], D
e = diag D
e 11 ID×D , · · · , D
e N N ID×D .
where W
e = D
e −W
f are all semi-SPD matrices. Thus, P T S T D
e −W
f SP is also a
P = AAT ∈ RD×D and L
e
semi-SPD matrix, and its eigenvalues are all non-negative. We have tr P T S T LSP
> 0. To compute
e .
the minimum value of Eq. 17, we just need to compute the minimum eigenvalues of matrix P T S T LSP
To avoid obtaining a singular solution, we impose a constraint as follows:
e
P T S T DSP
= I.
(18)
Then, the corresponding minimization problem turns to
e ),
min tr(P T S T LSP
e
s.t. P T S T DSP
= I.
We use the Lagrange multiplier method to solve the minimization problem
e
e
L (A, λ) = tr P T S T LSP
− λ tr P T S T DSP
−I
T
T
e
e
= tr AAT LSAA
− λ AAT S T DSAA
−I .
(19)
(20)
As L’s derivative to A, we obtain the following:
∂L (A, λ)
T
T
e
e
,
= 4tr AT S T LSAA
− λAT S T DSAA
∂A
(21)
Yangyang LI, et al.
Sci China Inf Sci 8
T
e . Eq. 21
e are both semi-SPD matrices, tr S T LSAA
e
= tr AAT S T LS
where P = AAT and S T LS
can be derived as
∂L (A, λ)
e − λAT AAT S T DS
e
e
e
= 4tr AT AAT S T LS
= 4tr S T LSA
− λS T DSA
AT A .
(22)
A
A is a D × d full rank matrix; thus, AT A is a d × d SPD matrix. To obtain
minimize the following generalized eigenfunction problem:
∂L(A,λ)
∂A
= 0, we need to
e
e
S T LSA
= λS T DSA.
(23)
We obtain the bottom smallest d eigenvalues 0 6 λ1 6 λ2 6 . . . 6 λd , and the corresponding eigenvectors a1 , a2 , . . . , ad ∈ RD×1 . A = [a1 , a2 , . . . , ad ] ∈ RD×d is the learned linear dimensionality reduction
D
d
map matrix. The corresponding dimension reduction map g between S+
and S+
is
Yi = g (Si ) = exp AT log (Si ) A .
(24)
d
The lower-dimensional SPD matrix Lie group S+
preserves the local geometric and algebraic structure
D
D
of S+ , which is maintained by the Laplacian matrix L on S+
. Good similarity between two points
corresponds to a large weight between them. In addition, through the construction of a graph Laplacian
matrix on the SPD matrices, the reduction map is learned based on global data points. The Laplacian
matrix can be viewed as an alignment matrix that aligns a set of local patch structures to obtain global
lower-dimensional representations by solving a generalized eigenfunction. Unlike other methods [7] [18]
[19], our method uncovers the intrinsic structure of SPD matrices by constructing this discrete Laplacian
matrix without the help of other spaces.
4
Algorithm Analysis
In this section, we mainly analyze the relationships between the proposed Lie-LPP and LPP [5] in theory.
In the first subsection, we present dimension reduction error comparisons between these two algorithms.
In the second subsection, we analyze the similarity relation between them.
4.1
Comparison with LPP
We analyze the reconstruction errors during dimension reduction of Lie-LPP and LPP from two aspects.
First, we analyze the local weight matrix construction. Second, we analyze the global alignment matrix
and the null space of the alignment matrix. The dimension reduction losses of the two algorithms are
determined by the corresponding graph Laplacian matrices defined on data points. In this section, we
mainly compare Lie-LPP and LPP in theory to analyze the improvements of our algorithm.
First, we show the relationship between two graph Laplacian matrices, which are defined on vector-form
data points and SPD matrix-form data points.
The local weight matrix W of LPP is defined in the second step of the LPP algorithm:
Wij = e−
kxi −xj k2
t
,
where xi and xj are in the same neighborhood.
f of Lie-LPP is defined in the second step of the Lie-LPP algorithm
The local weight matrix W
fij = e−
W
klog(Si )−log(Sj )k2
F
t
,
where Si and Sj are also in the same neighborhood.
The distance kxi − xj k between xi and xj is computed based on EM. However, the distance k log (Si ) −
log (Sj ) kF between Si and Sj is computed based on LEM. EM is not the real Riemannian metric of
the embedded manifold M defined in the LPP algorithm. We mentioned in Subsection 2.2 that LEM
Yangyang LI, et al.
Sci China Inf Sci 9
is the intrinsic Riemannian metric defined on the SPD matrix Lie group. Thus, the distance kxi − xj k
under EM is not the intrinsic geodesic distance distM (xi , xj ) on M and obviously distM (xi , xj ) >
kxi − xj k. Under LEM, the intrinsic geometric structure of SPD matrix Lie group can be determined,
and k log (Si ) − log (Sj ) kF is the real geodesic distance between Si and Sj . To our knowledge, xi and Si
are two different feature descriptors of the same image in computer vision. In the Appendix, we prove
that the SPD matrix descriptors preserve the geometric structures of vector-form feature descriptors.
Thus, we have
k log (Si ) − log (Sj ) k2F > kxi − xj k2 .
Under this analysis, we have
fij .
Wij > W
(25)
The corresponding graph Laplacian matrices defined on {xi } and {Si }, i = 1, 2, · · · , N are represented
e respectively. On the basis of the above analysis, we present our first comparison conclusion
as L and L,
in Theorem 4.1.
Theorem 4.1. If the datasets {xi } and {Si }, (i = 1, 2, · · · , N ) are two different feature descriptors
e that is,
of the same N images, then we have L L,
e ,
λi (L) > λi L
for all i = 1, 2, · · · , N .
f and W , we obtain Eq. 25. Then D − D
e is a
Proof: According to the definition of weight matrices W
e=D
e −W
f and L = D − W ; thus, L − L
e = D−D
e− W −W
f
positive diagonal matrix. We know that L
e is a semi-positive definite symmetric matrix. We have
is also a Laplacian matrix. L − L
e > 0,
λi (L − L)
e
L L,
for all i = 1, 2, · · · , N .
e if L − L
e 0, then we write L L.
e If L L,
e then
For two symmetric matrices L, L,
e ,
λi (L) > λi L
for every i.
For a special situation, if the original sub-manifold M in LPP is highly curved and the Riemannian
e
curvature isnot
zero everywhere, then the eigenvalues of L are strictly greater than those of L, that is,
e for every i under this situation. Therefore, we have proven this theorem.
λi (L) > λi L
e in Theorem 4.1, we need to analyze the dimension
After analyzing the relationship between L and L
reduction errors of Lie-LPP and LPP. In the dimension reduction step, the two algorithms need to
minimize the following generalized eigenvalue function:
E=
1X
2
(yi − yj ) Wij = Y T LY,
2 i,j
where {y1 , y2 , · · · , yN } are lower-dimensional representations of {xi } or {Si }, for i = 1, 2, · · · , N .
The dimension reduction errors of Lie-LPP and LPP are measured by the smallest eigenvalues of graph
Laplacian matrices. Suppose the dimension reduction error under LPP is represented as E and the error
e On the basis of Theorem 4.1, we present our second comparison
under Lie-LPP is represented as E.
conclusion in Theorem 4.2.
e under Lie-LPP is less than the dimension reduction
Theorem 4.2. The dimension reduction error E
error E under LPP, that is, we have
e F 6 kEkF .
kEk
Yangyang LI, et al.
Sci China Inf Sci 10
Proof: According to the algorithm procedures of the Laplacian eigenmap, the dimension reduction errors
E of LPP and Lie-LPP are mainly determined by the smallest eigenvalues of the graph Laplacian matrix.
Thus, the norm of general reconstruction error E is measured as
kEkF =
d
X
λi ,
i=1
where d is the intrinsic dimension of lower-dimensional representations.
From Theorem 4.1, we can deduce that for the same image database, the graph Laplacian matrix
e
L constructed
on the SPD matrices is lower than that on the vector-form descriptor. Thus, we have
e , for all i = 1, 2, · · · , N . Then, on the basis of the definition of the dimension reduction
λi (L) > λi L
error norm of Lie-LPP and LPP, we have
e F 6 kEkF .
kEk
Under the same special situation in Theorem 4.1, if the embedded manifold M in LPP is highly curved,
e
then the reconstruction error E is strictly greater than E.
e F < kEkF .
kEk
The key reason for this situation is that in LPP, the authors used EM to determine the local intrinsic
geometric structure of M, which is not the real local Riemannian metric of M.
4.2
Connection to LPP
We also present a theoretical analysis of the similarity relation between Lie-LPP and LPP aside from
reconstruction error comparisons between these two algorithms. Through the analysis, we can see that
under the following special situation, Lie-LPP is equivalent to LPP by defining a new weight matrix.
Suppose the vector form descriptor of an object is represented as xi a row vector. The corresponding
SPD matrix descriptor of this object is shown as Si = xTi xi . SPD matrix Lie group is a flat Riemannian
manifold, where the local neighborhood of the Lie group is locally isometric to the corresponding tangent
space. Thus, the local tangent space can be approximately represented by the local neighborhood of the
SPD matrix Lie group. Under this special situation and theoretical analysis, Lie-LPP can be transformed
into LPP with a special weight matrix. The theoretical analysis is stated in Proposition 4.1.
Proposition 4.1. The vector form descriptor of an object be xi and the corresponding SPD matrix
descriptor be xTi xi . Under this special situation, Lie-LPP can be transformed into LPP by defining a
new weight matrix.
Proof: First, we give the following representation of the generalized eigenvalue function of Lie-LPP:
e
e
S T LSA
= λS T DSA,
f SA = (1 − λ) S T DSA,
e
ST W
e = D
e −W
f is the corresponding graph Laplacian matrix defined on a set of SPD matrices S
where L
T
represented as S = [xT1 x1 , xT2 x2 , · · · , xTN xN ], and S T is the transpose matrix of S.
By rewriting S T as a matrix representation, we obtain the following representation:
0
..
. ,
0 · · · xN
x1 · · ·
T
T
T
T .. . .
S = [x1 , x2 , · · · , xN ] .
.
Yangyang LI, et al.
Sci China Inf Sci 11
Under this representation, we obtain
x1
Tf
T
T
T ..
S W SA = [x1 , x2 , · · · , xN ] .
0
Define
xT1 · · · 0
x1 · · · 0
. .
. .
.
f . . . ..
. . . .. W
WV =
.
.
.
0 · · · xTN
0 · · · xN
0
xT1 · · ·
.. f .. . .
.
. W .
· · · xN
0 ···
···
..
.
Suppose X =
x1
x2
..
.
x1
0
x2
..
A.
.
...
xTN
xN
S W SA =
xT1 · · · 0
x1 · · · 0
. .
. .
.
.
e
, DV = . . . . D
. .. . . ..
.
0 · · · xTN
0 · · · xN
f SA as follows:
Under this new weight matrix WV , we rewrite S T W
Tf
[xT1 , xT2 , · · ·
, xTN ]WV
x1
x2
..
.
xN
.
A.
Tf
, S W SA = X T WV XA.
xN
Thus, under this new weight matrix WV , Lie-LPP is equivalent to LPP. The above analysis shows that
the Lie-LPP algorithm can be transformed into LPP by defining a new weight matrix WV if the feature
descriptors of vector form and SPD matrix form are {xi } and {xTi xi }, i = 1, 2, · · · , N , respectively.
5
Experiments
In this section, we first report the results we obtained after running Lie-LPP on two human action
databases, namely, Motion Capture HDM05 [25] and CMU Motion Graph. We compare our algorithm
with traditional manifold learning algorithms through two experiments. In the second part, we test our
algorithms on a static face database e.g., extended Yale Face Database B (YFB DB). We compare LEML
algorithm [7] and SPD-ML algorithm [19] with our algorithm, after which we present the experimental
comparison between Lie-LPP and LPP [5].
5.1
Human Action Recognition
In this subsection, we test Lie-LPP on two human action databases. Each action segment trajectory
can be seen as a curve traversing a manifold. Action recognition involves classifying the different action
curves. In the recognition step, we use the nearest neighborhood framework to classify human action
sequences. After feature extraction, the embedded covariance feature descriptors form an SPD matrix,
which belongs to a low-dimensional SPD matrix Lie group.
Yangyang LI, et al.
Sci China Inf Sci 12
Table 1 Time comparison, Accuracy rates of HDM05 database under different reduced dimensions and different Riemannian metrics, i.e., EM and LEM.
Methods
Dimension d × d
time
Accuracy Rate
Lie-LPP-EM
10 × 10
16.1961s
0.5905
Lie-LPP-EM
15 × 15
21.0645s
0.6095
Lie-LPP-EM
20 × 20
30.2843s
0.6095
Lie-LPP-LEM
10 × 10
275.156s
0.7619
Lie-LPP-LEM
15 × 15
822.289s
0.7857
Lie-LPP-LEM
20 × 20
1003.13s
0.7905
Lie-LPP-EM
93 × 93
438.777s
0.5857
Lie-LPP-LEM
93 × 93
21133.9s
0.6000
5.1.1 Motion Capture HDM05 Database
The HDM05 database [25] contains more than 70 motion classes in 10 to 50 realizations executed by
various actors. The actions are performed by five subjects whose names are ’bd’,’bk’, ’dg’, ’mm’, and ’tr’.
For each subject, we choose the following 14 actions: ’clap above head’, ’deposit floor’, ’elbow to knee’,
’grab high’, ’hop on both legs’, ’jog’, ’kick forward’, ’lie down on the floor’, ’rotate both arms backward’,
’sit down on chair’, ’sneak’, ’stand up’, and ’throw basketball’. We choose three motion fragments per
subject per action. Thus, the total number of motion fragments in this experiment is 210. The dataset
provides the 3D locations of 31 joints over time acquired at the speed of 120 frames per second. In our
experiment, we describe each action observed over T frames [3] by its joint covariance descriptor, which
is an SPD matrix of size 93 × 93. T is a parameter. In this experiment, we choose T = 20.
If we directly perform the classification in the original 93 × 93 dimensional SPD matrix Lie group, the
experimental time cost would be as high as up to 21133.9s. In practice, each joint action is controlled by
only a few features, which can be much less than 93. Therefore, dimension reduction is necessary before
recognition. We perform eight groups of experiments. Each experiment is divided into two parts. The first
part involves using the use Lie-LPP algorithm to reduce the dimension of SPD matrices, while the second
part involves using the leave-one-out method to compute the recognition rate. We measure the similarity
between two SPD matrices under two Riemannian metrics, namely, EM and LEM, for comparison. The
final comparison results are shown in Table 1. The purpose of this experiment is to analyze the time
complexity of Lie-LPP. Thus, we present comparisons of our method under different Riemannian metrics
and different reduced dimensions. Under the same reduced dimensions d × d, the accuracy rates under
LEM are higher than those under EM overall, but the time cost under LEM is higher than that under
EM because of the especially high time cost of the log operation on the SPD matrix. Under the same
Riemannian metric, the time cost increases along with the growth of the reduced dimensions. If we do
not reduce the dimensionality of SPD matrices before recognition, then the recognition accuracies under
both Riemannian metrics are 0.5857 and 0.6000, respectively. However, the recognition rates under the
low-dimensional SPD matrix Lie group with LEM are 0.7619, 0.7857, and 0.7905 for dimensions reduced
10 × 10, 15 × 15, and 20 × 20, respectively (Table 1); these rates are much higher than they would be
without dimension reduction. In conclusion, reducing the dimension of the original SPD matrices is
necessary before recognition. Our method reduces the time cost of the experiment and improves the
recognition rate. In addition, on the basis of the LEM, the recognition rates are relatively higher. Thus,
the intrinsic structure of SPD matrices is more accurately determined based on LEM.
5.1.2 CMU Motion Graph Database
We consider four different action classes, namely, walking, running, jumping, and swing. Each class
contains 10 sequences or a total of 40 sequences for our experiment. A total of 31 joints are marked in
the skeletal joint sequences. Only the root joint is represented by a 6D vector; the other 30 joints are
Yangyang LI, et al.
Sci China Inf Sci 13
Table 2 Classification performance of CMU Motion Graph database together with the comparison results for Lie-LPP
and traditional manifold learning algorithms.
Methods
CMU
Dimension d
Lie-LPP
0.9750
3×3
LPP [5]
0.8500
9
LEP [17]
0.9000
9
PCA [20]
0.2500
9
represented by 3D rotation angle vectors. Each action frame is represented by a 96D vector. The SPD
matrix, which is a 96 × 96 matrix, is constructed by computing the covariance of T frames subwindows,
where we choose T = 20. To guarantee connectivity between subwindows, we take a T2 frames that
overlap between adjacent subwindows, as mentioned in [3].
In this experiment, we use leave-one-out method to compute the recognition rate. Every time, we
choose one sequence of each class as the test set and the remaining sequences as the training set. The
comparison results in Table 2 show the recognition accuracies for different action classes, where the
recognition accuracy of our method is 0.975, which is higher than that of the other three methods.
Notably, the SPD matrix descriptor under joint locations in our method obtains a surprising result.
5.2
Human Face Recognition
In [24], we tested Lie-LPP on human face databases, namely, Yale Face DB and YTC DB. In this
subsection, we further test our algorithm on another static human face database, namely, YFB DB. Only
two other algorithms, namely, LEML [7] and SPD-ML [19], are similar to our method in that they reduce
the dimensionality of SPD matrix manifold. In this subsection, we report the comparison results for
our method with LEML, SPD-ML, the manifold learning algorithm LPP [5], and linear dimensionality
reduction PCA [20].
5.2.1 Extended Yale Face Database B
YFB DB [26] contains 2414 single-light-source images of 38 individuals each seen under approximately
64 near-frontal images and under different illuminations per individual. For every subject in a particular
pose, an image with ambient illumination was captured. The face region in each image is resized to
32 × 32. We use the raw intensity feature to construct the corresponding SPD matrix, in which we
follow [7] to construct the SPD matrix for each image. For this dataset, because the image size is 32 × 32,
the size of the corresponding SPD matrix is 1024 × 1024.
5.2.2 Recognition
In the recognition step, SPD-ML [19] used the affine-invariant metric and the Stein divergence metric
to measure the similarities among SPD matrices. LEML [7] used LEM, under which the SPD matrix
Lie group is locally isometric to Lie algebra. LPP and PCA used EM to measure the similarities among
vector-form data points. For Lie-LPP, we also use LEM to compute the geodesic distances between two
SPD matrices. Unlike the other two algorithms, our algorithm aims to construct a Laplace?Beltrami
operator on the SPD matrix Lie group and then learn a more discriminable Lie group, which preserves
the geometric and algebraic structure of the original one.
For YFB DB, we choose 60 images per subject and run the experiment four times by using each algorithm. In each experiment, we randomly choose p (p = 20, 30, 40, 50) images per subject as the training
dataset and the remaining 60 − p images as the test dataset respectively. For PCA and LPP, we choose
the dimension of low dimensional space with d(d + 1)/2 after dimension reduction, where d = 10. For
Lie-LPP, LEM, and SPD-ML, we choose the low size of the SPD matrix, that is, d × d. Under these
choices, the low dimension of PCA and LPP is equal to the low dimension of Lie-LPP. The recognition
accuracy results are reported in Table 3. We choose the same classification methods with LEML and
Yangyang LI, et al.
Sci China Inf Sci 14
Table 3 Classification performance of YFB DB together with the comparison results for Lie-LPP and traditional manifold
learning algorithms PCA, LPP, LEML, SPD-ML-Stain, and SPD-ML-Airm.
YFB DB
YFD-trn20/tst40
YFD-trn30/tst30
YFD-trn40/tst20
YFD-trn50/tst10
PCA [20]
46.4 ± 1.8
50.2 ± 2.2
61.8 ± 1.7
63.7 ± 1.7
LPP [5]
48.2 ± 1.5
57.2 ± 2.4
62.8 ± 1.8
69.4 ± 1.4
LEML [7]
46.5 ± 1.6
53.6 ± 2.1
57.3 ± 1.7
72.9 ± 1.6
SPD-Stain [19]
44.6 ± 1.9
52.4 ± 2.1
56.7 ± 1.3
69.4 ± 1.7
SPD-Airm [19]
45.2 ± 1.8
53.1 ± 1.7
58.3 ± 1.6
68.5 ± 1.9
Lie-LPP
50.8 ± 2.4
66.3 ± 1.2
70.2 ± 1.7
73.9 ± 2.1
SPD-ML in the recognition step. Table 3 shows that the recognition results of Lie-LPP outperform PCA
and LPP in any case. Comparisons among Lie-LPP, LEM, and SPD-ML show that Lie-LPP is slightly
higher than LEML only when we choose p = 50 images per subject. The accuracy recognition rates
of Lie-LPP especially outperform the accuracies of LEML and SPD-ML when we choose p = 20, 30, 40.
These results show that the effect of dimensionality reduction for SPD matrices by Lie-LPP is better
than the effects of LEML and SPD-ML. In addition, LEML and SPD-ML need several parameters when
performing their algorithms. Our algorithm only needs to analyze the Laplacian operator of the SPD
matrix Lie group and solves the dimensionality reduction problem directly on the SPD matrix Lie group.
6
Conclusions and Future directions
In summary, the following are the main conclusions of this paper:
1. We construct a Laplace?Beltrami operator on the SPD matrix Lie group and give the corresponding
discrete Laplacian matrix.
2. We extend the manifold learning algorithm LPP to Lie-LPP and used it on the SPD Matrix Lie
group. We have shown that Lie-LPP can be successfully applied to human action recognition and
human face recognition.
3. Our analysis of the geometric and algebraic structure of the SPD matrix Lie group shows that the
SPD matrix Lie group is a complete and compact manifold. The sectional curvature on every point
is zero; thus, the SPD matrix Lie group is a flat manifold and locally isometric to Lie algebra;
4. However, this is not a simple application of the idea in paper [5]. We developed a new algorithm
called Lie-LPP, which is a substantial extension of the LPP algorithm;
5. We analyze the relationships between Lie-LPP and LPP in theory and obtain three theoretical
conclusions;
6. We conducted practical experiments, the results of which show that Lie-LPP outperforms the
existing ones significantly.
In the future, we will further improve this algorithm. We will attempt to add a time dimension to
enhance human action recognition, introduce manifold learning algorithms to other types of Lie groups,
and introduce new manifold learning algorithms on higher-dimensional tensor space.
Acknowledgments
This work is supported by the National Key Research and Development Program of China under grant no.
2016YFB1000902; the National Natural Science Foundation of China project Nos. 61232015, 61472412,
Yangyang LI, et al.
Sci China Inf Sci 15
and 61621003; the Beijing Science and Technology Project on Machine Learning-based Stomatology; and
the Tsinghua-Tencent-AMSS-Joint Project on WWW Knowledge Structure and its Application.
Conflict of interest
The authors declare that they have no conflict of interest.
References
1 Ma A J, Yuen P C, Zou W W W, et al. Supervised spatio-temporal neighborhood topology learning for action
recognition. IEEE Trans. Circuits and Systems for Video Technology, 2013, 23(8): 1447-1460
2 Hussein M E, Torki M, Gowayyed M A, et al. Human action recognition using a temporal hierarchy of covariance
descriptors on 3D joint locations. In: Proceedings of the Twenty-Third International Joint Conference on Artificial
Intelligence, Beijing, 2013. 2466-2472
3 Vemulapalli R, Arrate F, Chellappa R. Human action recognition by representing 3D skeletons as points in a Lie group.
In: Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, 2014. 82
4 Sanin A, Sanderson C, T.Harandi M, et al. Spatio-temporal covariance descriptors for action and gesture recognition.
In: Proceedings of the 2013 IEEE Workshop on Applications of Computer Vision, Washington, 2013. 103-110
5 He X F, Niyogi P. Locality preserving projections. In: Proceedings of the sixteenth Annual Conference on Neural
Information Processing Systems, Chicago, 2003
6 He X F, Cai D, Niyogi P. Tensor subspace analysis. In: Proceedings of the eighteenth Annual Conference on Neural
Information Processing Systems, British Columbia, 2005. 499-506
7 Huang Z W, Wang R P, Shan S G, et al. Log-Euclidean metric learning on symmetric positive definite manifold with
application to image set classification. In: Proceedings of the 32th International Conference on Machine Learning,
Lille, 2015. 720-729
8 Kwatra V, Han M. Fast covariance computation and dimensionality reduction for sub-window features in image. In:
Proceedings of the 11th European conference on Computer vision: Part II, Heraklion, 2010. 156-169
9 Wang L, Suter D. Learning and matching of dynamic shape manifolds for human action recognition. IEEE transactions
on image processing, 2007, 16(6):1646-1661
10 Tuzel O, Porikli F, Meer P. Pedestrian detection via classification on Riemannian manifolds. IEEE Transaction on
PAMI, 2008, 30(10):1713-1727
11 Porikli F, Tuzel O, Meer P. Covariance tracking using model update based on Lie algebra. In: Proceedings of the 2006
IEEE Conference on Computer Vision and Pattern Recognition, Washington, 2006. 94
12 Porikli F, Tuzel O. Fast Construction of covariance matrices for arbitrary size image windows. In: Proceedings of the
International Conference on Image Processing, Atlanta, 2006. 1581-1584
13 Tabia H, Laga H, Picard D, et al. Covariance descriptors for 3D shape matching and retrieval. In: Proceedings of the
2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, 2014. 533
14 Tuzel O, Porikli F, Meer P. Region covariance: a fast descriptor for detection and classification. In: Proceedings of
the 7th European conference on Computer vision: Part II, Graz, 2006. 589-600
15 Roweis S, Saul L. Nonlinear dimensionality reduction by locally linear embedding. Science, 2000, 290(5500):2323-2326
16 B.Tenenbaum J, Silva V, Langford J. A global geometric framework for nonlinear dimensionality reduction. Science,
2000, 290(5500):2319-2323
17 Belkin M, Niyogi P. Laplacian eigenmaps and spectral techniques for embedding and clustering. In: Proceedings of
the International Conference Advances in Neural Information Processing Systems, 2001. 585-591
18 Harandi M, Sanderson C, Wiliem A, Lovell B. Kernel analysis over Riemannian manifolds for visual recognition of
actions, pedestrians and textures. In: Proceedings of the IEEE Workshop on the Applications of Computer Vision,
Breckenridge, 2012. 6163005
19 Harandi M, Salzmann M, Hartley R. From manifold to manifold: geometry-aware dimensionality reduction for SPD
matrices. In: Proceedings of the 15th European conference on Computer vision, Zurich, 2014. 17-32
20 Smith L. A tutorial on principal components analysis. February 26, 2002
21 Pennec X, Fillard P, Ayache N. A Riemannian framework for tensor computing. IJCV, 2006, 66(1):41-66
22 Li X, Hu W, Zhang Z, et al. Visual tracking via incremental log-Euclidean Riemannian subspace learning. In:
Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition, Anchorage, 2008. 4587516
23 Arsigny V, Fillard P, Pennec X, et al. Geometric means in a novel vector space structure on symmetric positive-definite
matrices. SIAM J.Matrix Analysis and Applications, 2007, 29(1):328-347
24 Li Y Y. Locally preserving projection on symmetric positive definite matrix Lie group. In: Proceedings of the International Conference on Image Processing, Beijing, 2017
25 Muller M, Roder T, Clausen M, et al. Documentation: Mocap database HDM05. Tech. Rep. CG-2007-2, Universitat
Bonn, 2007
26 Yale Univ. Face database. http://cvc.yale.edu/projects/yalefaces/yalefaces/ html
27 Belkin M, Niyogi P. Towards a theoretical foundation for Laplacian-based manifold methods. Journal of Computer
and System Sciences, 2005, 74(8):1289-1308
28 Willmore T. Riemannian Geometry. Oxford University Press, Oxford, 1997
| 1cs.CV
|
Distributed Bilevel Energy Allocation Mechanism with Grid
Constraints and Hidden User Information
M.Nazif Faqiry, IEEE Member, Sanjoy Das
Published in : IEEE Transactions on Smart Grid
DOI: 10.1109/TSG.2017.2779826
Publication Date: 2017
Citation (APA)
M.Nazif Faqiry, S. Das. (2017).Distributed Bilevel Energy Allocation Mechanism with Grid Constraints and Hidden
User Information. IEEE Transactions on Smart Grid, Issue (99) 10.1109/TSG.2017.2779826
© © 2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in
any current or future media, including reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this
work in other works.
Distributed Bilevel Energy Allocation
Mechanism with Grid Constraints and Hidden
User Information
M.Nazif Faqiry, IEEE Member, Sanjoy Das
Abstract— A novel distributed energy allocation mechanism for
distribution system operator (DSO) market through a bilevel
iterative auction is proposed. With the locational marginal price
known at the substation node, the DSO runs an upper level auction
with aggregators as intermediate agents competing for energy.
This DSO level auction takes into account physical grid constraints
such as line flows, transformer capacities and node voltage limits.
This auction mechanism is a straightforward implementation of
projected gradient descent on the social welfare (SW) of all home
level agents. Aggregators, which serve home level agents - both
buyers and sellers, implement lower level auctions in parallel,
through proportional allocation of power, until market
equilibrium is established. This is accomplished without asking
for the agents’ utility functions and generation capacities that are
considered private information. The overall bilevel auction is
shown to be globally efficient and weakly budget balanced.
Index Terms—distribution system; agents; trading; auction;
bid; social welfare
NOMENCLATURE
𝒩
𝑁
𝒜
𝐴
𝛿
𝑃0 , 𝑄0
𝑉0
𝑘, 𝑙
𝒟(𝑘)
𝓊(𝑘)
𝒰(𝑘)
𝑟𝑘 , 𝑥𝑘
𝑝𝑘
𝑞𝑘
𝑃𝑘
𝑄𝑘
Δ𝑉𝑘
𝑉𝑘
𝑆𝑘̅
𝜃𝑘
𝑐𝑘
𝒩𝐵𝑘
𝒩𝑆𝑘
𝑁𝑆𝑘
𝑁𝐵𝑘
𝑖, 𝑗
Set of nodes, excluding root
Cardinality |𝒩| of 𝒩
Set of nodes with aggregators
Cardinality of 𝒜
Maximum allowable per unit (pu) voltage deviation
Active, Reactive power from root node in pu
Voltage at root in pu
The 𝑘 th , 𝑙 th nodes, 𝑘, 𝑙 ∈ 𝒩
Set of downstream nodes of 𝑘 (immediate & separated)
Index of immediate upstream node of node 𝑘
Index of all upstream nodes of 𝑘 to root, 𝑘 ∈ 𝒰(𝑘)
Resistance and reactance of line (𝓊(𝑘), 𝑘) in pu
Real power injected into 𝑘 in pu, 𝑘 ∉ 𝒜 ⇒ 𝑝𝑘 = 0
Reactive power injected into 𝑘 in pu, 𝑘 ∉ 𝒜 ⇒ 𝑞𝑘 = 0
Active power flowing through line (𝓊(𝑘), 𝑘) in pu
Reactive power flowing through line (𝓊(𝑘), 𝑘) in pu
Voltage drop through line (𝓊(𝑘), 𝑘) in pu
Voltage at node 𝑘 in pu
MVA limit of line going into node 𝑘 ( line 𝑘) in pu
Fraction of 𝑝𝑘 as 𝑞𝑘 at node 𝑘 when 𝑘 ∈ 𝒜
Per unit price of energy of node 𝑘 in (¢/pu), 𝑘 ∈ 𝒜
Set of buyers at node 𝑘, 𝑘 ∈ 𝒜
Set of sellers at node 𝑘, 𝑘 ∈ 𝒜
Cardinality of 𝒩𝑆𝑘
Cardinality of 𝒩𝐵𝑘
The 𝑖 th buyer and 𝑗 th seller
This work was partially supported by the National Science
Foundation-CPS under Grant CNS-1544705.
𝑔𝑗𝑘
𝑑𝑖𝑘
𝑠𝑗𝑘
𝑏𝑖𝑘
𝑢𝑖𝑘 , 𝑣𝑗𝑘
𝑐0
Θ𝑘
Ω
Max generation of 𝑗 th seller in pu at node 𝑘
Demand in pu delivered to the 𝑖 th buyer at node 𝑘
Supply in pu by to the 𝑗 th seller at node 𝑘
Total bid money for demand 𝑑𝑖𝑘 by the 𝑖 th buyer
Utility functions of buyer 𝑖 and seller 𝑗 at node 𝑘
Price function at root (substation) node in ¢/pu
Social welfare function of aggregator 𝑘
Global (DSO level) social welfare function
I. INTRODUCTION
proliferation of renewable energy resources (RES) at the
THE
distribution level is reshaping the market structure of
distribution system operators (DSO). The electricity sector has
devolved from a highly regulated system operated by vertically
integrated utilities to a relatively decentralized system based
more fully on market valuation and allocation mechanisms [1].
A RES owner with a surplus of energy is anticipated to
participate in such mechanisms more strategically while
seeking profit [2], [3]. Specifically, it sells energy with the
objective of maximizing its payoff, i.e. the sum of the monetary
gain from supplying an amount of energy and utility it gains
from retaining any surplus energy that is not traded. In a similar
manner, the payoff of an energy consumer, i.e. a buyer, is the
difference between its utility gained from consuming energy
and the cost of procuring that energy.
DSOs on the other hand, are expected to leverage the
available local resources in order to capture additional value by
optimizing the system for least cost operation while
maintaining the physical system operation constraints [4]. One
of the key challenges for efficient energy distribution
mechanisms is its design to motivate active participation of
customers [5]. Without their participation, the benefits of smart
grid is not fully realized [6]. Therefore, suitable mechanisms for
it to operate effectively within its physical constraints, while
incentivizing customer participation, are needed.
The existing literature on the energy grid focuses mostly on
pricing based on the distribution locational marginal price
(DLMP) that is defined as the marginal cost of serving the next
increment of the electrical demand at a distribution node.
DLMP is typically determined in a centralized fashion by the
Lagrange multiplier of the distribution node energy balance
constraint [7]-[12]. It has been shown in [13] that introducing
price responsive customers causes distribution line congestion.
In [9], DLMP based pricing is used as the means of dynamic
pricing tariffs to alleviate distribution system congestion.
IEEE TRANSACTIONS ON SMART GRID
Similar to [8], [9], the recent work in [10] proposes two
benchmark pricing methodologies, namely DLMP and iterative
DLMP (iDLMP), for a congestion free energy management by
buildings providing flexible demand. Aggregators in this model
have contracts with flexible buildings to decide their reserve
and energy schedule by interacting with the DSO in a cost
optimal manner in order to avoid any congestion in the
distribution grid. Standard dual decomposition [14] is used to
determine iDLMP at the DSO level to alleviate the need for data
transfer among the DSO and aggregators. This study does not
simulate market conditions. Its lower level agents, i.e.
buildings, do not have to place bids in order to obtain energy,
which is supplied as per contractual obligations with the
aggregator.
In this paper, a novel bilevel distribution auction mechanism
is proposed that converges to a socially optimal solution, while
maintaining physical grid constraints. The lower level auction,
referred to as the aggregator level auction (ALA), is conducted
by the local aggregators assigned to each distribution node
among downstream consumers and prosumers.
The upper level auction, referred to as the DSO level auction
(DLA), is implemented iteratively by the DSO among
aggregators competing for the share of energy that the DSO
receives from the wholesale market. The goal of the DSO’s
auction is draw a suitable amount of energy from the wholesale
market and allocate it among competing aggregators in such a
manner that maximizes the global SW, while maintaining
physical grid constraints such as voltage, line, and transformer
limits.
There are several aspects of the proposed auction mechanism
that deviate from earlier research. Firstly, consistent with
economic theory, equilibrium is established by means of a
market-driven mechanism, where both the volume of energy
traded as well as the equilibrium price, are determined entirely
through the ALA bidding process. Thus, determining the
energy available to each aggregator is shifted to the upper level
DLA. Therefore, the underlying optimization algorithm is
parallelized through primal (instead of dual) decomposition.
Next, the proposed bilevel mechanism is efficient – it
maximizes the social welfare (SW) of the grid, which is the sum
total of the utilities of the consumer and prosumer agents. This
is different from revenue-oriented (or optimal) auctions, which
try to maximize the revenue accrued by the DSO or aggregators
[15]-[17]. The underlying optimization formulation in [10]
which minimizes the total cost of procuring energy, is another
example of an optimal mechanism. Efficient mechanisms that
have been proposed so far for the energy grid, are indivisible
(i.e. discrete) resource auctions, designed for load scheduling
[18], [19], whereas this research treats energy as a divisible (i.e.
continuous) resource.
Information privacy is a significant advantage of the
proposed mechanism, where the agents reveal neither their
utility functions nor their individual PV-generated energy to the
aggregators. This is possible as at market equilibrium, SW is
automatically maximized separately within each aggregator,
without the need for any specific optimization algorithm.
Lastly, asymmetric bidding [3] is another feature of the
proposed mechanism. Buying agents declare monetary amounts
to the aggregator and are allocated energy. On the other hand,
selling agents receive unit costs of energy from the aggregator
3
Substation
Node 0
1
k
k+1
N
P0+iQ0
P1+iQ1
Pk+iQk
p1+iq1
sj
Pk+1+iQk+1
pk+iqk
pk+1+iqk+1
pn+iqn
di
Aggregator k
Fig. 1. Schematic diagram of a unidirectional single branch radial distribution
system with N nodes excluding the root node.
and submit to the latter, the quantity of energy they are willing
to sell. Although it is possible for an agent to switch roles at any
stage of the mechanism, in doing so, it would have to change its
bidding strategy accordingly. For simplicity, in the simulations
reported here (Section III), it is assumed that agents declare
their intention to act as buyers or sellers a priori. Bidding
asymmetry is the means through which proportional fairness
can be incorporated in a double-sided market-driven auction
mechanism [3].
The buying agents are allocated power directly in proportion
to the monetary bids that they place, whereas the sellers are
reimbursed monetary amounts that are proportional to the
energy sold. Although such proportionally fair auctions possess
certain desirable properties [40], these mechanisms are
susceptible to price anticipation, where each agent bids in a
manner to influence the equilibrium market price in its favor.
Unfortunately, when all agents become price anticipators, the
overall social welfare decreases. A novel approach to avoid
price anticipation, through virtual bidding, has been addressed
in details in [3].
II. FRAMEWORK
Consider the radial distribution network in Fig. 1, with 𝑁
nodes excluding the root. The proposed bilevel auction
mechanism is implemented in two levels among aggregators
assigned to primary distribution nodes through DLA, and
among strategic consumers and prosumers, i.e. agents residing
on a lateral feeder, by the aggregators through ALA. In ALA,
each agent’s objective is to maximize its own profit by
participating in its ALA. Each aggregator’s objective is to
maximize its agents’ SW, without access to their private utility
functions and generation capacities, which are shown in Fig. 2
with dotted lines and boxes. The aggregators make this possible
by using double-sided proportional allocation [3] and
participate in DLA by competing with other aggregators in
order to get their optimal demand/supply share of the real
power 𝑝𝑘 . The DSO’s objective is to implement DLA
iteratively until convergence, so that maximum global SW is
attained.
As shown in Fig. 2, at each iteration of DLA, aggregator 𝑘
receives real power 𝑝𝑘 , implements ALA and submits its per
unit price 𝑐𝑘 to trade 𝑝𝑘 . The prices 𝑐𝑘 are market equilibrium
price of multiple parallel ALA algorithms. Upon receiving
IEEE TRANSACTIONS ON SMART GRID
4
updated 𝑐𝑘 from each aggregator 𝑘 ∈ 𝒜, the DSO executes the
next iteration of DLA to find the new supply 𝑝𝑘 . This procedure
continues until convergence is reached. From the flow of
information in Fig. 2, it is clear that the buying and selling
agents do not reveal their utilities to the aggregator. Similarly,
the agents’ generations are not disclosed either.
Buyer
Agent 1
Supply
𝑐𝑘
𝑑𝑖
Price
Distribution
System Operator
𝑝𝑘
Allocation
Buyer
Agent i
𝑐𝑘
Price
Aggregator k
Bid
Supply
𝑠𝑗
𝑏𝑖
Social Welfare
Utility
Function of
Buyer 1
Seller
Agent j
Seller
Agent 1
Utility
Function of
Seller 1
Fig. 2. Schematic showing flow of bidding information among agents in the
bilevel mechanism.
A. Distribution System Constraints
The DSO needs to ensure that grid constraints are not violated.
It is assumed in this research that the distribution system is
balanced and all quantities can be represented per phase at each
seller or buyer node. The simplified version of DistFlow
equations [20], [21] that has been extensively used in literature
[22]-[24] is adopted here. These equations are as shown below,
𝑃𝑘 = 𝑝𝑘 + ∑ 𝑝𝑙 ,
(1)
𝑙∈𝒟(𝑘)
𝑄𝑘 = 𝑞𝑘 + ∑ 𝑞𝑙 ,
(2)
𝑙∈𝒟(𝑘)
𝑉𝑘 = 𝑉0 − ∑ 𝛥𝑉𝑙 ,
(3)
𝑙∈𝒰(𝑘)
where Δ𝑉𝑙 =
𝑟𝑙 𝑃𝑙 +𝑥𝑙 𝑄𝑙
𝑉0
. 𝑉0 is the root node’s pu voltage and Δ𝑉𝑙
is the voltage drop at line segment entering node 𝑘. 𝒟(𝑘) is the
set of immediate and separated downstream nodes of 𝑘 and
𝒰(𝑘) is the set of upstream nodes (including 𝑘) to the root node.
For instance, in Fig. 3, 𝒟(17) = {18, 19, 20, 21} , 𝒰(27) =
{1,2,27}. The following system architecture matrices are
defined,
𝑘 = aggregator 𝑙
[𝐀]𝑘𝑙 = {1
0 otherwise,
𝑙 ∈ 𝒟(𝑘) or k = 𝑙
[𝐃]𝑘𝑙 = {1
0 otherwise,
[𝐔]𝑘𝑙 = {1 𝑙 ∈ 𝒰(k)
0 otherwise.
In the above, 𝐀 is an 𝑁 × 𝐴 matrix and 𝐃, 𝐔 are 𝑁 × 𝑁 matrices
associated with the spatial topology of the radial distribution
network. The matrix 𝐀 is the node-aggregator matrix that has
an entry of unity (‘1’) at every column on the row, i.e. node,
with which it is associated. The matrices 𝐃 and 𝐔 correspond
to the descendant and ancestor nodes. Every row (node) of 𝐃
and 𝐔 has an entry of unity where the corresponding column
(node) is its descendant or ancestor, and a zero elsewhere. The
voltage drop in line 𝑘 is Δ𝑉𝑘 = 𝑟𝑘 𝑃𝑘 + 𝑥𝑘 𝑄𝑘 , so that,
(4)
Δ𝑽 = 𝑉0−1 (𝐫 ∘ 𝑷 + 𝐱 ∘ 𝑸)
where ∘ is the elementwise (Hadamard) product and Δ𝑽 is the
𝑁 × 1 vector whose 𝑘 th entry is Δ𝑉𝑘 . This notation has been
followed throughout the remainder of this paper – bolded
variables are vectors of their italicized scalar counterparts, with
the entry position specified in the latters’ subscripts.
Using the matrices defined above, (1) – (3) can be written as
follow,
(5)
𝑷 = 𝐃𝐀𝐩,
(6)
𝑸 = 𝐃𝐀𝐪,
(7)
𝑽 = 𝑉0 𝟏N − 𝐔Δ𝑽.
Note that it is assumed that homes are furnished with smart
meters, and in case of sellers, also inverters [25], [26]. They are
also capable of communicating their reactive power
supply/demand to/from the aggregators at the end of each ALA.
In order to enforce the physical grid constraints, the DSO
requires the reactive powers 𝑞𝑘 of the aggregators in addition to
their 𝑝𝑘 s, with the latter directly available through the bidding
process in ALA. The ratio between 𝑞𝑘 and 𝑝𝑘 is denoted as 𝜃𝑘 ,
so that 𝑞𝑘 = 𝜃𝑘 𝑝𝑘 . Each aggregator 𝑘 can obtain 𝑞𝑘 as the sum
of the reactive power requirement obtained directly from its
agents in 𝒩𝐵𝑘 ∪ 𝒩𝑆𝑘 at the termination of each round of ALA.
Alternately, 𝑞𝑘 can be estimated from historical data. In this
research, it is assumed that the aggregator has access to the
numerical value of 𝜃𝑘 , that is communicated to the DSO along
with 𝑐𝑘 . With 𝛉 = [𝜃𝑘 ]𝑘∈𝒜 and 𝐪 = 𝛉 ∘ 𝐩 as elementwise
product of 𝛉 and 𝐩, using (4), the expressions in (5) – (7) can
be rewritten as,
(8)
𝑷 = 𝐃𝐀𝐩,
𝑸 = 𝐃𝐀(𝛉 ∘ 𝐩),
𝑽 = 𝑉0 𝟏N −
(9)
1
𝐔(𝐫 ∘ 𝐃𝐀𝐩 + 𝐱 ∘ 𝐃𝐀(𝛉 ∘ 𝐩)). (10)
𝑉0
In the simplified DistFlow equations above, the real/reactive
branch flows and node voltages are entirely functions of real
power injection 𝐩, substation per unit voltage 𝑉0 and
distribution grid topology. The DSO can implement DLA to
determine 𝐩, using (8) – (10) to set up grid constraints, which
are explained below.
A.1 Transformer Capacity: Since the distribution
transformer(s) at the substation node has limited capacity, the
total apparent power that the DSO can draw from the wholesale
market is bounded as,
(11)
𝐩T 𝐙𝟎 𝐩 ≤ 𝑆02 .
Here, the matrix 𝐙𝟎 is given by,
𝐙𝟎 = 𝟏𝐴 𝟏𝐴T + 𝛉𝛉T .
(12)
A.2 Line Limits: The total apparent power at line 𝑘 cannot
exceed its MVA limit 𝑆𝑘̅ so that 𝑃𝑘2 + 𝑄𝑘2 ≤ 𝑆𝑘̅ 2. Defining the
matrix 𝐄𝑘 as a square matrix with the only non-zero entry of
unity appearing at the 𝑘th row and the 𝑘th column, it is seen that,
IEEE TRANSACTIONS ON SMART GRID
5
𝐏 T 𝐄𝑘 𝐏 + 𝐐T 𝐄𝑘 𝐐 ≤ 𝑆𝑘̅ 2 .
Whence, using (8) and (9), the following line limit constraint is
obtained,
(13)
𝐩T 𝐙𝑘 𝐩 ≤ 𝑆𝑘̅ 2 ,
∀𝑘 ∈ 𝒩.
Each matrix 𝐙𝑘 above is given by,
𝐙𝑘 = 𝐀T 𝐃T 𝐄𝑘 𝐃𝐀 + 𝑑𝑖𝑎𝑔𝛉𝐀T 𝐃T 𝐄𝑘 𝐃𝐀𝑑𝑖𝑎𝑔𝛉.
(14)
A.3 Node Voltage Limits: The voltage at node 𝑘 is given by
(10), and expressed solely in terms of 𝐩 and other grid
parameters. The total voltage deviation at node 𝑘 must not
exceed a numerical value of 𝛿 (typically 0.05), yielding the
following constraint,
(15)
𝟏𝑁 − 𝛅 ≤ 𝑽 ≤ 𝟏𝑁 + 𝛅.
Whence using (10) in (15), and upon further simplification the
following bounds on the power vector 𝐩 are obtained,
(16)
𝐥 ≤ 𝐌𝐩 ≤ 𝐥.
𝐥 = (V0 − 1)𝟏𝑁 + 𝛅.
The matrix 𝐌 that shows the sensitivity of voltage deviation to
power injection 𝐩 is equal to,
(17)
𝐌 = 𝐌𝑃 + 𝐌𝑄 𝑑𝑖𝑎𝑔𝛉,
where, 𝐌𝑃 = 𝑉0−1 𝐔𝐫 ∘ 𝐃𝐀, 𝐌𝑄 = 𝑉0−1 𝐔𝐱 ∘ 𝐃𝐀.
A.4 Budget Balance: As in [27], [13], [10], the market price at
the substation node is coupled with the demand drawn by the
grid from the wholesale market and is modeled as,
𝑐0 = 𝑐0𝑏 + 𝛽0 ∑ 𝑝𝑘 .
𝐩 = [𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 ]
𝐵
𝑆
(22)
,
𝑘∈𝒜
(23)
𝐬𝑘 ≤ 𝐠𝑘 .
In this constrained optimization problem formulation, Ω is the
global SW, obtained by aggregating that of each aggregator Θ𝑘 .
The equality constraint in (22) restricts the quantity pk allocated
to each aggregator k as the difference between the total energy
demanded by the buyers (𝟏T𝑁𝑘 𝐝𝑘 ) and that supplied by the
𝐵
sellers (𝟏T𝑁𝑘 𝐬 𝑘 ). The inequality constraint in (23) is present so
𝑆
that the power supplied 𝑠𝑗𝑘 , by any seller 𝑗 stays below its
generation 𝑔𝑗𝑘 .
Each aggregator level SW Θ𝑘 in (21) is given by,
Θ𝑘 (𝐝𝑘 , 𝐬 𝑘 |p𝑘 ) = 𝟏T𝑁𝑘 𝐮𝑘 + 𝟏T𝑁𝑘 𝐯 𝑘 ,
𝐵
𝑆
(24)
where 𝐮𝑘 and 𝐯 𝑘 are the hidden utility functions of buyers and
sellers in aggregator 𝑘.
In the above expression, the lower and upper bounds are,
𝐥 = (V0 − 1)𝟏𝑁 − 𝛅,
Subject to:
𝐩 ∈ 𝔊,
(18)
Under the assumption that the utility functions 𝒖𝑘 and 𝒗𝑘 are
concave, the SW optimization problem is convex. Note that in
the DSO SW problem, 𝐝𝑘 and 𝐬 𝑘 are local variable vectors
pertaining to aggregator 𝑘’s buyers and sellers demand and
supply whereas 𝐩 is the global variable vector of injections to
all aggregators. Due to lack of access to the utility functions
𝑢𝑘 , 𝑣 𝑘 , and for scalability [28], [29], the DSO implements a
distributed algorithm by decomposing its original problem into
a master and multiple sub-problems that are solved in a
distributed fashion. The master problem is solved by DLA and
the sub-problems are solved by ALAs locally.
𝑘
Here, c0b is the base demand price and β0 is the elasticity
coefficient that can be obtained using statistics of historical data
of locational marginal price as explained in [27]. The subscript
‘0’ appears in all variables to indicate association with the root
node. Note that to model c0, the DSO can add a reasonable fixed
amount to c0b to account for any grid usage tariff. The
underlying mechanism remains weakly budget balanced despite
this increment. In other words, 𝐜 T 𝐩 − 𝑐0 𝟏𝐴T 𝐩 ≥ 0. Replacing c0
from (18) yields the following DSO budget balance constraint,
(19)
𝑐0𝑏 𝟏𝐴T 𝐩 − 𝐜 T 𝐩 + 𝛽0 𝐩T 𝟏𝐴 𝟏𝐴T 𝐩 ≤ 0.
B. DSO Level Auction
B.1 Social Welfare Problem: The feasible set 𝔊 consists of all
power vectors 𝐩 that meet constraints in (11), (13), (16) and
(19) pertaining to substation transformer MVA limit, lines
MVA limit, nodes voltage limits, and DSO budget balance, i.e.
∀𝑘 ∈ 𝒩 ∪ {0}: 𝐩T 𝐙𝑘 𝐩 ≤ 𝑆𝑘̅ 2 ,
(20)
}.
∀𝑘 ∈ 𝒩: 𝐥 ≤ 𝐌𝐩 ≤ 𝐥,
T
T
T
− 𝐜 𝐩 + 𝛽0 𝐩 𝟏𝐴 𝟏𝐴 𝐩 ≤ 0
The goal of DSO is to solve the SW problem as stated below.
Maximize w.r.t. [𝐝𝑘 ]𝑘∈𝒜 , [𝐬 𝑘 ]𝑘∈𝒜 , 𝐩:
𝔊 = {𝐩|
𝑐0𝑏 𝟏𝐴T 𝐩
Ω([𝐝𝑘 ]𝑘∈𝒜 , [𝐬 𝑘 ]𝑘∈𝒜 , 𝐩)
= ∑ Θ𝑘
𝑘∈𝒜
(𝐝𝑘
𝑘
, 𝐬 , 𝐩) .
(21)
B.2 Distributed Implementation: At the lower level, each
aggregator 𝑘, implements the following sub-problem in parallel
by means of ALA:
Maximize w.r.t. 𝐝𝑘 , 𝐬 𝑘 ∶ Θ𝑘 (𝐝𝑘 , 𝐬 𝑘 , 𝐩).
Subject to (22) and (23).
At the upper level, each iteration of DLA realizes a projected
gradient descent step of the decomposed SW problem with 𝐩 as
the global variable. The DSO sends to each aggregator 𝑘 the
energy 𝑝𝑘 and receives the gradient direction 𝛌 = ∇𝐩 Ω from it.
It will be shown in Proposition 7 that 𝛌 = 𝐜 the vector of prices
towards which ALA converges. Next, the vector 𝐩 is
incremented by an amount 𝜖𝛌 to 𝐩′′ , where 𝜖 is the gradient step
size. The vector 𝐩′′ is then projected onto the feasible region 𝔊,
after the constraint parameters are updated according to (14)
and (17). Parameter updates are required because the reactive
power changes according to the numerical value of the
projection 𝐩′ , which in turn causes the fraction 𝛉 to change.
The DLA algorithm terminates when any aggregator 𝑘
returns a flag 𝒶𝑘 = 𝐹 defined in the following, indicating that
the constraints in (22) and (23) were not satisfied,
𝒶k = {
𝑇 if 𝑝𝑘 = 𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 ,
𝐵
𝑆
𝐹 if 𝑝𝑘 ≠ 𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 .
𝐵
𝑆
(25)
IEEE TRANSACTIONS ON SMART GRID
6
Algorithm DLA:
lim
∂𝑐𝑘
𝑘
𝑠0 →∞ ∂𝑠𝑗
𝐩 ← initial
Repeat:
Send to aggregators 𝑘 ∈ 𝒜: 𝑝𝑘
Receive from aggregators 𝑘 ∈ 𝒜: 𝜆𝑘 = 𝑐𝑘 , 𝒶𝑘
If ⋀𝑘 𝒶𝑘 ≠ 𝑇 then exit loop
𝐩′′ ← 𝐩 + 𝜖𝛌
Compute constraint parameters: 𝛉, 𝐌, 𝐙𝑘
𝐩′ ← argmin‖𝐩 − 𝐩′′ ‖
= lim
1
T
𝑠0 →∞ 𝑝𝑘 +𝑠0 +𝟏 𝑘 𝐬𝑘
(1 −
𝑐𝑘0 𝑠0 +𝟏T 𝑘 𝐛𝑘 −𝟏T 𝑘 𝐬𝑘
𝑁𝑆
𝑁𝐵
𝑁𝑆
𝑝𝑘 +𝑠0 +𝟏T 𝑘 𝐬𝑘
𝑁𝑆
)
∂𝑐𝑘0
∂𝑠𝑗𝑘
= 0,
so that (27) is valid in the limiting scenario. In other words, the
presence of the virtual bidder that place very high bids allows
the auction to behave as one with a large number of agents.
Under these circumstances, the participating agents bid as price
takers.
Proposition-2: ALA is strongly budget and energy balanced.
𝐩∈𝔊
Proof: It follows from virtual bidding (26) that 𝑐𝑘 𝑝𝑘 +
𝑐𝑘 𝟏T𝑁𝑘 𝐬 𝑘 = 𝟏T𝑁𝑘 𝐛𝑘 . The RHS is the total monetary amount that
Until convergence
Output: 𝐩, 𝐜, [𝐝𝑘 ]𝑘∈𝒜 , [𝐬 𝑘 ]𝑘∈𝒜
𝑆
C. Aggregator Level Auction
C.1 Virtual Bidding: Price anticipation is an undesirable effect
that occurs in proportional auctions with relatively few bidders
[3], [30]. In such scenarios, the bidders are aware of the
∂𝑐
sensitivity of the equilibrium price, i.e. that 𝑘𝑘 ≠ 0 for a buyer,
and
∂𝑏𝑖
∂𝑐𝑘
∂𝑠𝑗𝑘
≠ 0 for a seller. As the bidders place bids to maximize
their individual payoffs, price anticipation leads to inefficiency.
In [3], virtual bidding is shown to approach price-taking
conditions for isolated microgrids by lowering the market
powers of the bidders. Virtual bidding involves the presence of
a virtual agent, which acts as both a seller and a buyer. Unlike
other agents, the virtual bidder is merely an algorithmic entity
that is incorporated within the aggregator, ergo has access to 𝐛𝑘
and 𝐬 𝑘 . By incorporating the virtual bidder, whose buying and
selling bids can be as large as possible, the aggregator enables
the price approach that of a market with potentially infinite
number of agents. The larger the number of bidder agents, the
less the effect of price anticipation [3]. Before addressing the
implementation of ALA, the following propositions are
established.
Proposition–1: Due to virtual bidding, ALA can be arbitrarily
close to price taking mechanism. In other words, the following
expressions hold,
𝑐𝑘 =
𝟏TN𝑘 𝐛𝑘
B
𝑝𝑘 + 𝟏T𝑁𝑘 𝐬 𝑘
(26)
,
𝑆
𝜕𝑐𝑘
𝜕𝑏𝑖𝑘
= 0,
𝜕𝑐𝑘
𝜕𝑠𝑗𝑘
𝐵
the aggregator receives from the buyers. The LHS is the sum of
the payment that the aggregator makes to the DSO and the
reimbursement amount given to the sellers. This establishes
strong budget balance. Energy balance is established under the
proportional allocation auction paradigm [31], where the
energy to each buyer to be proportional to its bid, i.e. 𝐝𝑘 =
1 𝑘
𝐛 . Energy balance immediately follows from (26). Note that
𝑐𝑘
this satisfies the constraint in in (22).
C.2 Distributed Implementation: ALA receives 𝑝𝑘 from the
DSO, and initializes the price 𝑐𝑘 (see Fig. 2). In each step, it
sends 𝑐𝑘 to the sellers and receives 𝐬 𝑘 . Using proportional
allocation, it determines 𝐝𝑘 which is communicated to the
buyers. The buyers return their bids 𝐛𝑘 . The price 𝑐𝑘 is
determined as a two-step procedure using virtual bidding.
Algorithm ALA(𝑘):
Receive from DSO: 𝑝𝑘
Initialize: 𝑐𝑘
Repeat:
Send to sellers: 𝑐𝑘
Receive from sellers: 𝐬 𝑘
1
𝐝𝑘 ← 𝐛𝑘
𝑐𝑘
Send to buyers: 𝐝𝑘
Receive from buyers: 𝐛𝑘
𝟏T𝑁𝑘 𝐛𝑘
𝐵
0
𝑐𝑘 ←
𝑝𝑘 + 𝟏T𝑁𝑘 𝐬 𝑘
𝑆
(27)
= 0.
𝑐𝑘 ←
𝑐𝑘0 𝑠0 + 𝟏T𝑁𝑘 𝐛𝑘
𝐵
𝑝𝑘 + 𝑠0 + 𝟏T𝑁𝑘 𝐬 𝑘
𝑆
Proof: The virtual bidder bids a large amount of energy 𝑠0 ,
which it buys for a total amount 𝑐𝑘0 𝑠0 . Here 𝑐𝑘0 =
𝟏T 𝑘 𝐛𝑘
𝑁𝐵
𝑝𝑘 +𝟏T 𝑘 𝐬𝑘
is the
𝑁𝑆
desired price under price taking. The actual price, 𝑐𝑘 =
𝑐𝑘0 𝑠0 +𝟏T 𝑘 𝐛𝑘
𝑁𝐵
𝑝𝑘 +𝑠0 +𝟏T 𝑘 𝐬𝑘
, is computed by the aggregator. It can be readily
𝑁𝑆
𝑠0 →∞
Likewise
lim
∂𝑐𝑘
𝑘
𝑠0 →∞ ∂𝑏𝑖
𝑐𝑘0 𝑠0 +𝟏T 𝑘 𝐛𝑘
𝑁𝐵
T
𝑠0 →∞ 𝑝𝑘 +𝑠0 +𝟏 𝑘 𝐬𝑘
𝑁𝑆
established that lim 𝑐𝑘 = lim
= lim
1
T
𝑠0 →∞ 𝑝𝑘 +𝑠0 +𝟏 𝑘 𝐬𝑘
𝑁𝑆
, justifying (26).
(1 +
∂𝑐𝑘0
∂𝑏𝑖𝑘
)=0
Until equilibrium
Send to DSO: 𝑐𝑘
and
D. Equilibrium Analysis
D.1 Bidding Strategies: Buyers and sellers’ bidding strategies
are established by means of the following propositions.
Proposition-3: The bid 𝑏𝑖𝑘 placed by each buyer 𝑖 is such that
its marginal utility equals its per unit price,
𝜕𝑢𝑖𝑘
𝜕𝑑𝑖𝑘
= 𝑐𝑘 .
(28)
IEEE TRANSACTIONS ON SMART GRID
7
Proof: The buyer’s payoff 𝑢𝑖𝑘 (𝑑𝑖𝑘 ) − 𝑏𝑖𝑘 , is maximized when its
𝜕
derivative is zero, i.e.
𝜕𝑏𝑖𝑘
𝑢𝑖𝑘 (𝑑𝑖𝑘 ) − 1 =
𝜕𝑢𝑖𝑘 1
𝜕𝑑𝑖𝑘 𝑐𝑘
(1 −
1 𝜕𝑐𝑘
𝑐𝑘 𝜕𝑏𝑖𝑘
)−
1 = 0. Under virtual bidding, (27) holds so that (28) is satisfied.
Proposition-4: The bid 𝑠𝑗𝑘 placed by each seller 𝑗 is such that if
the seller does not bid its entire generation (𝑠𝑗𝑘 < 𝑔𝑗𝑘 ), its
marginal utility, with 𝛾𝑗𝑘 ≥ 0 being a positive scalar quantity,
equals its per unit price,
𝜕𝑣𝑗𝑘
𝜕𝑠𝑗𝑘
{
= 𝛾𝑗𝑘 − 𝑐𝑘 ,
𝛾𝑗𝑘 = 0,
𝑠𝑗𝑘 < 𝑔𝑗𝑘 ,
𝛾𝑗𝑘 > 0,
𝑠𝑗𝑘 = 𝑔𝑗𝑘 .
(29)
and (29),
Proof: In fact, 𝛾𝑗𝑘 is a dual variable as shall be seen here. The
seller’s payoff 𝑣𝑗𝑘 (𝑔𝑗𝑘 − 𝑠𝑗𝑘 ) + 𝑐𝑘 𝑠𝑗𝑘 is maximized under the
constraint 𝑠𝑗𝑘 ≤ 𝑔𝑗𝑘 . The Lagrangian for this problem
is 𝐿𝑗 (𝑠𝑗𝑘 ) = 𝑣𝑗𝑘 (𝑔𝑗𝑘 − 𝑠𝑗𝑘 ) + 𝑐𝑘 𝑠𝑗𝑘 − 𝛾𝑗𝑘 (𝑠𝑗𝑘 − 𝑔𝑗𝑘 ).
At
stationarity,
𝜕
𝐿
𝜕𝑠𝑗𝑘 𝑗
=
𝜕
𝑣 𝑘 (𝑔𝑗𝑘
𝜕𝑠𝑗𝑘 𝑗
− 𝑠𝑗𝑘 ) + 𝑐𝑘 + 𝑠𝑗𝑘
𝜕𝑐𝑘
𝜕𝑠𝑗𝑘
− 𝛾𝑗𝑘 =
0. Under virtual bidding, (27) holds so that (29) is satisfied.
D.2 Aggregator Equilibrium: During each iteration of DLA, an
aggregator establishes equilibrium conditions to return price 𝑐𝑘 .
Proposition-5: The equilibrium of the aggregator 𝑘’s auction
maximizes the social welfare Θ𝑘 (𝐝𝑘 , 𝐬 𝑘 , 𝐩) as defined in Eqn.
(24) with respect to 𝐝𝑘 , 𝐬 𝑘 under the energy balance constraint,
𝑝𝑘 = 𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 and with no seller selling more energy
𝐵
𝑆
than its generated capacity, 𝐬 𝑘 ≤ 𝐠 𝑘 .
Proof: The statement above defines a constrained optimization
problem with the following Lagrangian ℒ𝑘 (𝐝𝑘 , 𝐬 𝑘 ),
ℒ𝑘 (𝐝𝑘 , 𝐬 𝑘 ) = Θ𝑘 (𝐝𝑘 , 𝐬 𝑘 ) − 𝛄T𝑘 (𝐬 𝑘 − 𝐠 𝑘 )
−λ𝑘 (𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 − 𝑝𝑘 ) .
𝐵
Proof: The matrix 𝐙0 in (12) can be recognized to be symmetric
positive semidefinite (SPSD). The matrix 𝐄𝑘 in (14) is also an
SPSD matrix with a single non-zero eigenvalue of unity. The
first and second terms in the expression for 𝐙𝑘 in (14) being
similar to 𝐄𝑘 under the transform matrices 𝐃𝐀 and 𝐃𝐀𝑑𝑖𝑎𝑔𝛉,
have the same eigenvalues and are therefore SPSD. Each matrix
𝐙𝑘 (𝑘 ∈ 𝒩) being the sum of two such SPSD matrices is also
SPSD. Hence, suppose 𝐜 were constant, the region 𝔊 would be
convex in 𝐩 since it would contain only convex quadratic and
linear inequality constraints.
However, 𝐜 is dependent on 𝐩 as the equilibrium price 𝑐𝑘 of
any aggregator 𝑘 is determined by the power 𝑝𝑘 and the market
𝜕Θ
mechanism. From (21) and (31), 𝑘 = 𝑐𝑘 whereas from (28)
𝑆
(30)
The constrained optimum satisfies ALA’s energy balance
condition. Stationarity at the optimum implies that ∇𝐝𝑘 ℒ𝑘 = 0
and ∇𝐬𝑘 ℒ𝑘 = 0. From (24), it is that ∇𝐝𝑘 𝐮𝑘 = 𝜆𝑘 𝟏𝑁𝑘
𝐵
and ∇𝐬𝑘 𝐯 𝑘 = 𝛄𝑘 − 𝜆𝑘 𝟏𝑁𝑘 , so that the optimum of Θ𝑘 coincides
𝑆
with the auction equilibrium when 𝜆𝑘 = 𝑐𝑘 in (28) and (29) and
with 𝛄𝑘 being the vector of entries 𝛾𝑗𝑘 in (29).
Proposition-6: The price vector 𝐜 returned by ALA is also the
gradient of the overall SW function given by (21), i.e.
(31)
∇𝐩 Ω([𝐝𝑘 ]𝑘∈𝒜 , [𝐬 𝑘 ]𝑘∈𝒜 , 𝐩) = 𝐜.
Proof: From (28) and (29), it is seen that 𝜆𝑘 = 𝑐𝑘 . It follows
𝜕
that
Θ𝑘 (𝐝𝑘 , 𝐬 𝑘 ) = 𝑐𝑘 . The expression in (31) directly
𝜕𝑝𝑘
follows from (21).
D.3 Global Equilibrium:
Proposition–7: Under certain assumptions, the bilevel auction
mechanism converges to the unique global maximum of the
social welfare.
𝜕𝑢𝑖𝑘
𝜕𝑑𝑖𝑘
=−
𝜕𝑣𝑗𝑘
𝜕𝑠𝑗𝑘
𝜕𝑝𝑘
= 𝑐𝑘 . Hence
𝜕𝑢𝑖𝑘
𝜕𝑝𝑘
=
𝜕𝑣𝑗𝑘
𝜕𝑝𝑘
= 𝑐𝑘 . We assume
that the utilities are strictly concave and increasing, i.e. that the
marginal utility of the energy consumed is always positive no
matter how small. Under these circumstances, neglecting cubic
and higher order terms, the SW can be approximated as, Θ𝑘 =
𝑐𝑘 𝑝𝑘 + ℎ𝑘 𝑝𝑘2 , where ℎ𝑘 < 0. In fact, some recent research have
modeled utilities as quadratic functions [32], [33], [18].Thus
Ω ≈ 𝐜 T 𝐩 + 𝐩T 𝐇𝐩, where 𝐇 is a negative definite diagonal
matrix thereby rendering Ω as strictly concave in 𝐩. As the
feasible region is strictly convex, termination of DLA due to a
constraint violation can take place only when the constrained
optimum has been reached with the violating constraint being
active. DLA is a straightforward implementation of projected
gradient ascent of the SW problem, whose convergence to the
global minimum has been well studied [29], [34], [35].
Proposition–8: The bilevel auction is weakly budget balanced.
Proof: From Proposition-2, ALA is strongly budget balanced,
so that 𝑐𝑘 𝟏T𝑁𝑘 𝐝𝑘 − 𝑐𝑘 𝟏T𝑁𝑘 𝐬 𝑘 = 𝑐𝑘 𝑝𝑘 . (19) explicitly imposes
𝐵
𝑆
the weak budget balance constraint at the DSO level so that
𝐜 T [𝟏T𝑁𝑘 𝐝𝑘 − 𝟏T𝑁𝑘 𝐬 𝑘 ]
≥ 𝑐0b 𝟏𝐴T 𝐩 + 𝛽0 𝐩T 𝟏𝐴 𝟏𝐴T 𝐩. The left
𝐵
𝑆
𝑘∈𝒜
side of this inequality is the total revenue collected from the
sum total of all agents and the right side is the total payment
made to the wholesale market. Since the latter can never exceed
the money collected, the bilevel auction is weakly budget
balanced.
III. SIMULATION RESULTS
Simulation results corroborates the theory presented in section
II. A modified IEEE 37 node system as shown in Fig. 3 has been
used to simulate the proposed bilevel energy allocation
mechanism. A total of 483 agents, 303 buyers and 180 sellers,
were generated and assigned to different nodes with
aggregators. As summarized in Table I, seventeen aggregators
(labeled A𝑘) with different numbers of buyer and seller agents
each with their own sets of parameters were created and
assigned to nodes that have load. Fig. 3. shows a schematic of
the IEEE 37 node system, where the nodes with aggregators are
enclosed within shaded circles.
Due to their wide use in the literature for quantifying user
satisfaction with diminishing returns, the utility functions of the
agents were assumed to follow concave logarithmic curves
[36]-[38] according to the equations 𝑢𝑖 (𝑑𝑖 ) = 𝑥𝑖 log(𝑦𝑖 𝑑𝑖 + 1)
IEEE TRANSACTIONS ON SMART GRID
8
and 𝑣𝑗 (𝑔𝑗 − 𝑠𝑗 ) = 𝑥𝑗 log(𝑦𝑗 (𝑔𝑗 − 𝑠𝑗 ) + 1).The
quantities 𝑥𝑖 , 𝑦𝑖 , 𝑥𝑗 and 𝑦𝑗 in these equations were generated
randomly, separately for each agent. They were adjusted so that
agents’ marginal utilities are scaled to reasonable per unit
prices. The generation 𝑔𝑗 for sellers were also drawn at random,
uniformly in the interval [0.1, 0.5] pu based on a system base
value of 100 kVA.
TABLE I
AGGREGATOR ASSIGNMENT WITH THEIR RESPECTIVE NUMBER
OF BUYER AND SELLER AGENTS TO NODES
1 2 3 4 5 6 7 8 9 10 11
𝐀𝒌
Node 1 8 12 13 17 18 22 23 25 26 27
𝑵𝒌𝑩 14 11 25 27 20 8 21 27 22 21 5
13 7 5 21 22 10 15 6 4 22 3
𝑵𝒌𝑺
12
29
13
9
0
35
5
27
2
24
26
A11
4
16
35
26
6
17
36
27
9
36
25
A10
15
33
10
5
A16
A1
1
14
31
19
19
A17
SUBSTATION BUS
A9
13
30
7
4
base price c0b and demand sensitivity coefficient β0 were
increased significantly to 800¢/pu and 40¢/pu2 . In Scenarios
II and III, equal base prices but different sensitivity coefficients
were selected to simulate two normal scenarios. In Scenario IV,
a base price of 200¢/pu with zero sensitivity coefficient was
selected so that the price is inelastic to demand to observe
energy trade and allocation under cheap supply from the
wholesale market. Furthermore, the transformer capacity was
increased to 40 pu to draw high amount of energy from the
wholesale market and observe activation of physical grid
constraints in DLA.
All case studies were simulated using MATLAB. Fig. 4
shows the outcome of the bilevel mechanism for each scenario.
Here bars show the energy allocation 𝑝𝑘 to each node that has
aggregator and dotted lines show its corresponding price. The
energy allocation 𝑝𝑘 to each aggregator 𝑘 (shown in Fig. 3 as
3
34
28
31 A14
A12 29
32
7
A13
6
30
8 A2
10
11
A3
9
23
33
A15
22
A7
A8
12
16
18
13
14
A4
15
17
A5
19
21
A6
20
Fig. 3. Modified IEEE 37 node system with aggregators indexed. 𝐴𝑘 is used as
an abbreviation for aggregator 𝑘.
Four different scenarios were created to examine the impact
that the price and substation node’s capacity have on DLA’s
allocation of energy to each aggregator and ALA allocations to
local buyer and seller agents along with the resulting global
SW. These four scenarios were created with differing wholesale
market price and substation transformer capacity parameters,
according to (18) and (11) as follow.
I) c0b = 800¢/pu (high), β0 = 40¢/pu2 (elastic), S0 = 25 pu
II) c0b = 200¢/pu (low), β0 = 30¢/pu2 (elastic), S0 = 25 pu
III) c0b = 200¢/pu (low), β0 = 10¢/pu2 (elastic), S0 = 25 pu
IV) c0b = 200¢/pu (low), β0 = 0¢/pu2 (inelastic), S0 = 40 pu
In Scenario I, in order to draw minimum energy from the
substation bus and observe energy trade among aggregators, the
Fig. 4. DLA outcome for aggregators’ share of energy (in pu) and its associated
price (in ¢/𝑝𝑢)
an A𝑘) and the equilibrium price 𝑐𝑘 are the solutions of DLA
and ALA at equilibrium when the global SW given by (21) has
been reached its constrained maximum. Furthermore, the
market equilibrium price 𝑐𝑘 are different for each aggregator,
illustrating that DLA is price heterogeneous.
In Scenario I, due to significantly high wholesale price, the
aggregators are willing to trade among themselves at different
IEEE TRANSACTIONS ON SMART GRID
prices ranging from 469 ¢/pu to 592 ¢/pu. Aggregators with
more buyers, i.e. A3 on node 12 (𝑁𝐵3 = 25, 𝑁𝑆3 = 5 ), buy more.
Several aggregators, especially those with more sellers, sell
energy. For example, A5 on node 17 supplies a high amount of
energy as it has more sellers (𝑁𝑆5 = 22), compared to others.
The DSO draws only 1.72 pu at c0 of 869 ¢/pu from the
wholesale market at which it becomes strongly budget
balanced, i.e. the constraint in (19) is active and the DSO makes
zero profit.
In Scenario II, the price c0 decreases to 525 ¢/pu. As a result,
a smaller number of aggregators sell energy to the DSO. The
ALA equilibrium prices ck in this scenario range from 387 ¢/pu
to 524 ¢/pu. The DSO draws 10.83 pu from the wholesale
market making zero profit. Aggregators import more energy
compared to Scenario I. For instance, A4 on node 13 that was
supplying energy as a seller in Scenario I imports energy as a
buyer in Scenario II.
In Scenario III, the wholesale market price drops down to
419 ¢/pu at which point, all aggregators switch to being buyers.
At this point, the DSO draws 21.87 pu from the wholesale
market while still not attaining any budget surplus.
In Scenario IV, the wholesale market price at 200 ¢/pu is
inelastic and inexpensive. All aggregators import energy at
different equilibrium prices ranging from 233 ¢/pu to 448
¢/pu. Aggregators with more buyers and with steeper utilities,
such as those in nodes 12, 25, and 35, import high amounts. The
DSO in this Scenario draws 30.14 pu and makes 4992 cents in
the role of an arbitrager. It is worth emphasizing that the
proposed mechanism is an SW maximizing bilevel auction in
which the goal is to maximize the overall SW. The DSO’s
budget is merely set as a constraint. As pointed earlier, a profit
seeking DSO can incorporate its surcharge in the pricing model
in (18).
Note that the volume of energy traded at each aggregator
node depends internally on the numbers of buyer and seller
agents (𝑁𝑆𝑘 and 𝑁𝐵𝑘 ), generation capacities 𝑔𝑗𝑘 , and marginal
utilities (𝑢′𝑘𝑖 and 𝑣′𝑗𝑘 ) as well as all external factors such as the
wholesale price 𝑐0 and other aggregators’ market conditions. In
general, aggregators with a surplus of energy that settle down
to lower equilibrium prices 𝑐𝑘 tend to supply more 𝑝𝑘 to the rest
of the network, whereas those that converge to higher 𝑐𝑘 supply
less. Similarly, aggregators with deficit energy and higher
equilibrium prices 𝑐𝑘 , are assigned more 𝑝𝑘 .
Convergence to the global optimum of the SW (sum of
utilities of 483 agents) under each scenario is shown in Fig. 5.
As can be seen, in all cases, DLA converges to within 1% of the
optimum within as little as 10 iterations. The maximum
allowable iterations of ALA to converge to equilibrium is 100
per DLA iteration in all cases, although it was routinely
observed that equilibrium could be established well before this
limit.
In Fig. 5, it can be observed that as the injection 𝟏𝐴T 𝐩 from
the wholesale market increases, so does the SW. This is because
the availability of more energy to the buyers results in their
acquiring less from sellers. Consequently, selling agents
consume a larger portion of their own generated energy.
In Fig. 6, the node voltages 𝑉𝑘 are shown, sorted according
to the aggregator index 𝑘, separately for each scenario. The
upper and lower bounds on the pu voltages appear as horizontal
9
Fig. 6. Node pu voltages within given bounds.
Fig. 7. Real, reactive, and apparent power flows in line 𝑘 (line entering node
𝑘) within line MVA limits.
dotted lines. In all cases, it can be seen that the voltages remain
within these bounds, indicating that the voltage constraint in
(15) is met. Increasingly higher voltage drops can be seen from
IEEE TRANSACTIONS ON SMART GRID
10
In Scenario IV, A6 imports 0.958 pu at a lower equilibrium
price of 273 ¢/pu. Buyers’ demand increase and only sellers
with marginal utilities 𝑣𝑗′6 equal to 𝑐6 trade nonzero supply 𝑠𝑗6 .
Notice that sellers 3, 6, 7, and 8 whose marginal utilities are
higher than 𝑐6 , i.e. 𝑣𝑗′6 > 𝑐6 , are assigned zeros supply 𝑠𝑗6 ,
allowing them to consume all their generation 𝑔𝑗6 instead of
selling so that their utilities are maximized. Notice that the
marginal utilities of all agents (both buyers and sellers) that are
trading becomes uniform and equal to the market equilibrium
price 𝑐6 for both scenarios.
Fig. 5. Optimum SW under each scenario. In scenarios other than III,
some aggregator 𝑘 returns 𝒶𝑘 = 𝐹 and DLA terminates.
Scenarios I through IV. This is a direct outcome of increasing
supply from the wholesale market.
Line power flows are shown in Fig. 7. The dotted lines
pertain to the line MVA limits (𝑆𝑘̅ ). Adjacent vertical bars
correspond to real (𝑃𝑘 ), reactive (𝑄𝑘 ), and apparent (𝑆𝑘 ) power
flows of each line segment, which are again ordered according
to the indices of their corresponding nodes. Higher energy
supplied from the wholesale market results in more power flows
(top – bottom). Moreover, in Scenario I (top), the power flows
in several lines are in the reverse direction, as it appear below
the x-axis. Conversely, in Scenario IV (bottom), all lines that
have downstream aggregators can be seen to be delivering
power to them. Additionally, the powers flowing through three
lines (indexed 2, 8, 27) are at their limits (13). The adjacent set
of lines labeled 12, 13 and 17 as per their node indices (see Fig.
3) which provide power directly to aggregators A3, A4, and A5
are taken up as examples to further highlight some
observations.
In Scenario I, due to significantly high wholesale price 𝑐0 ,
with high number of sellers and low price, A4 and A5 on nodes
13 and 17 supply to A3 on node 12 that has 25 buyers and only
5 sellers (see Fig. 4 and Fig. 7). In Scenario II, however, as 𝑐0
is decreased, A4 on node 13 starts to import (see Fig. 4) while
A5 keeps exporting to feed A3 as well as A4. This forces power
flow in lines indexed 17 and 13 to become negative regardless
of the fact that A4 imports. Notice also that A3 imports from
other upstream aggregators in addition to A5 resulting to a
positive power flow in line 12. In Scenarios III and IV, due to
further decrease in the wholesale price 𝑐0 , all aggregators
import power which result to positive power flow in all lines.
So far, we discussed results pertaining to DLA, distribution
grid constraints, and power flows. In order to illustrate the
results of ALAs, auction outcome for A6 (𝑁𝐵6 = 8, 𝑁𝑆6 = 10 )
under Scenarios I and IV are summarized in Table II. In
Scenario I, A6 exports 0.834 pu to the grid which is obtained as
the sum of supplies (∑𝑗 𝑠𝑗6 ) of its sellers minus sum of demands
of its buyers (∑𝑖 𝑑𝑖6 ). The price 𝑐6 stabilizes at 533 ¢/pu at
which only those buyers and sellers’ whose marginal utilities
𝑢𝑖′6 and 𝑣𝑗′6 becomes equal to this price can trade. Furthermore,
because of high wholesale price 𝑐0 , all sellers are willing to
trade and declare nonzero 𝑠𝑗6 . For example, seller 1 generates
𝑔16 = 0.434 pu and sells 𝑠16 = 0.325 pu at 𝑣𝑗′6 = 𝑐0 = 553.
TABLE II
ALA OUTCOME FOR AGGREGATOR 6 (A6) UNDER SCENARIOS I
AND IV
Aggregator 6 on node 18 (𝒩𝐵6 = 5, 𝒩𝑆6 = 20)
𝑝6 (pu)
I
IV
-0.834
0.958
𝑑𝑖6 (pu)
I
0.115
0.105
0.118
0.079
0.131
0.126
0.081
0.123
IV
0.234
0.214
0.240
0.160
0.267
0.257
0.165
0.251
𝑔𝑗6 (pu)
I
0.434
0.437
0.110
0.232
0.270
0.153
0.249
0.185
0.274
0.432
IV
0.434
0.437
0.110
0.232
0.270
0.153
0.249
0.185
0.274
0.432
𝑐6 (¢/pu)
I
IV
553
273
Buyer Agents
𝑐𝑖6 (¢/pu)
I
IV
553
273
553
273
553
273
553
273
553
273
553
273
553
273
553
273
Seller Agents
𝑠𝑗6 (pu)
I
0.325
0.351
0.017
0.151
0.141
0.018
0.127
0.086
0.182
0.315
IV
0.211
0.262
0
0.067
0.007
0
0
0
0.087
0.194
𝑉6 (pu)
I
IV
1.032
0.972
𝑢𝑖′6 (¢/pu)
I
IV
553
273
553
273
553
273
553
273
553
273
553
273
553
273
553
273
𝑣𝑗′6 (¢/pu)
I
553
553
553
553
553
553
553
553
553
553
IV
273
273
467
273
273
490
272
295
273
273
IV. CONCLUSION
A globally efficient energy allocation mechanism that does
not violate any grid constraints has been proposed here. At the
DSO level, the auction is price heterogeneous among
aggregators, while the lower level auction is price uniform
among home agents. Instead of determining the energy price as
in other recently proposed mechanisms [8]-[10] the upper level
agent (DSO) allocates power to the aggregators, that in turn
conduct their separate ALA to establish market equilibrium
conditions locally within their agents.
One of the advantages of this new approach is that the
aggregators can conduct their own auctions in islanded
conditions. During islanding, the aggregator would assume zero
power allocated from the DSO. Its ALA would converge to an
appropriate unit price entirely via market equilibrium, allowing
locally energy trade among its agents to proceed.
The proposed mechanism is able to maximize SW requiring
access to neither the agents’ utility functions nor their power
generations. ALA is able to do so by inferring the derivatives
IEEE TRANSACTIONS ON SMART GRID
of the agents’ utility functions from bidding information (see
Proposition 5). This is another advantage of the proposed
bilevel mechanism.
As ALA relies on market equilibrium prices, it is an iterative
mechanism, requiring several iterations to establish
equilibrium. As a tradeoff, DLA requires relatively few
iterations to converge to the SW maximizing power allocation.
This is because it uses primal problem decomposition instead
of the dual decomposition approach. An algorithm to maximize
SW that would use dual decomposition, would need gradient
information, whereas unlike in primal decomposition, the
derivatives of the agents’ utilities with respect to allocated
energy would not be readily available to the DSO. The
alternative is for a dual decomposition optimization scheme to
apply subgradient descent, which is known to converge very
slowly (see [28] for details).
There are certain limitations of the research described below
that can be addressed effectively in future research. The
simplified DistFlow equations used here neglect power loss.
The underlying physical grid model can be extended to consider
line losses by using a more elaborate set of power flow
equations [2].
The bilevel auction assumes a radial distribution system. This
is because radial networks are very common, and prevalent
within the United States. In order to adapt the proposed bilevel
auction mechanism to other networks, only the manner in which
the physical grid constraints are imposed (using DistFlow
equations) would need to be replaced appropriately.
The present approach assumes a balanced network. The
mechanism could be extended for use in 3-phase unbalanced
networks by implementing additional grid constraints
specifying the maximum allowable power imbalance, along
with some algorithmic modifications.
Future research can also investigate how to extend the current
model to multiple time slots. Aggregators that do not include
any energy storage or conventional generators can implement
an auction during each time slot independently of the others.
For aggregators that do so, temporal constraints would need to
be taken into account.
11
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
D.-Y. Heo and L. S. Tesfatsion, “Facilitating appropriate
compensation of electric energy and reserve through standardized
contracts with swing,” Cit. Inf. J. Energy Mark., vol. 8, no. 4, pp. 93–
21, 2015.
Y. Wang, W. Saad, Z. Han, H. V Poor, and T. Basar, “A GameTheoretic Approach to Energy Trading in the Smart Grid,” Smart
Grid, IEEE Trans., vol. 5, no. 3, pp. 1439–1450, 2014.
M. N. Faqiry and S. Das, “Double-Sided Energy Auction in
Microgrid: Equilibrium under Price Anticipation,” IEEE Access, vol.
4, pp. 3794–3805, 2016.
RMI, “Detailed Overview of Services Technical Appendix a :
Detailed Overview of Services,” 2015.
W. Tushar, J. A. Zhang, D. B. Smith, H. V. Poor, and S. Thiébaux,
“Prioritizing consumers in smart grid: A game theoretic approach,”
IEEE Trans. Smart Grid, vol. 5, no. 3, pp. 1429–1438, 2014.
W. H. E. Liu, K. Liu, and D. Pearson, “Consumer-centric smart grid,”
IEEE PES Innov. Smart Grid Technol. Conf. Eur. ISGT Eur., 2011.
N. O’Connell, Q. Wu, J. Østergaard, A. H. Nielsen, S. T. Cha, and Y.
Ding, “Day-ahead tariffs for the alleviation of distribution grid
congestion from electric vehicles,” Electr. Power Syst. Res., vol. 92,
pp. 106–114, 2012.
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Z. Liu, Q. Wu, S. Oren, S. Huang, R. Li, and L. Cheng, “Distribution
Locational Marginal Pricing for Optimal Electric Vehicle Charging
through Chance Constrained Mixed-Integer Programming,” IEEE
Trans. Smart Grid, vol. 3053, no. c, pp. 1–1, 2016.
S. Huang, Q. Wu, S. S. Oren, R. Li, and Z. Liu, “Distribution
Locational Marginal Pricing Through Quadratic Programming for
Congestion Management in Distribution Networks,” IEEE Trans.
Power Syst., vol. 30, no. 4, pp. 2170–2178, 2015.
S. Hanif, T. Massier, H. B. Gooi, T. Hamacher, and T. Reindl, “Cost
Optimal Integration of Flexible Buildings in Congested Distribution
Grids,” IEEE Trans. Power Syst., vol. PP, no. 99, pp. 1–14, 2016.
M. N. Faqiry, A. K. Zarabie, F. Nassery, H. Wu, and S. Das, “A Day
Ahead Market Energy Auction for Distribution System Operation,”
in Proceedings of IEEE Conference on Electro-Information
Technology, pp. 1–6, 2017.
S. Parhizi, A. Khodaei, and S. Bahramirad, “Distribution Market
Clearing and Settelement,” IEEE PES Gen. Meet., pp. 0–4, 2016.
R. A. Verzijlbergh, L. J. De Vries, and Z. Lukszo, “Renewable
Energy Sources and Responsive Demand. Do We Need Congestion
Management in the Distribution Grid?,” Power Syst. IEEE Trans.,
vol. 29, no. 5, pp. 2119–2128, 2014.
S. Boyd, L. Xiao, A. Mutapcic, and J. Mattingley, “Notes on
Decomposition Methods,” 2008.
L. Gkatzikis, I. Koutsopoulos, and T. Salonidis, “The role of
aggregators in smart grid demand response markets,” IEEE J. Sel.
Areas Commun., vol. 31, no. 7, pp. 1247–1257, 2013.
M. R. Sarker, M. A. Ortega-Vazquez, and D. S. Kirschen, “Optimal
Coordination and Scheduling of Demand Response via Monetary
Incentives,” IEEE Trans. Smart Grid, vol. 6, no. 3, pp. 1341–1352,
2015.
I. Momber, S. Member, and S. Wogrin, “Retail Pricing : A Bilevel
Program for PEV Aggregator Decisions Using Indirect Load
Control,” vol. 31, no. 1, pp. 464–473, 2016.
P. Samadi, H. Mohsenian-Rad, R. Schober, and V. W. S. Wong,
“Advanced demand side management for the future smart grid using
mechanism design,” IEEE Trans. Smart Grid, vol. 3, no. 3, pp. 1170–
1180, 2012.
A. C. Chapman and G. Verbiˇ, “An Iterative On-Line Auction
Mechanism for Aggregated Demand-Side Participation,” vol. 8, no.
1, pp. 158–168, 2017.
F. F. Wu, “Network reconfiguration in distribution systems for loss
reduction and load balancing,” vol. 4, no. 2, pp. 1401–1407, 1989.
M. E. Baran and F. F. Wu, “Optimal sizing of capacitors placed on a
radial distribution system.,” IEEE Trans. Power Deliv., vol. 4, no. 1,
pp. 735–743, 1989.
H. Yeh, S. Member, D. F. Gayme, and S. H. Low, “Adaptive VAR
Control for Distribution Circuits With Photovoltaic Generators,” vol.
27, no. 3, pp. 1656–1663, 2012.
Z. Wang, B. Chen, J. Wang, and J. Kim, “Decentralized Energy
Management System for Networked Microgrids in Grid-Connected
and Islanded Modes,” IEEE Trans. Smart Grid, vol. 7, no. 2, pp.
1097–1105, 2016.
K. Turitsyn, P. Šulc, S. Backhaus, and M. Chertkov, “Distributed
control of reactive power flow in a radial distribution circuit with high
photovoltaic penetration,” in IEEE PES General Meeting, PES 2010,
2010.
A. R. Malekpour and A. Pahwa, “A Dynamic Operational Scheme for
Residential PV Smart Inverters,” IEEE Trans. Smart Grid, pp. 1–10,
2016.
A. Maknouninejad, N. Kutkut, I. Batarseh, and Z. Qu, “Analysis and
control of PV inverters operating in VAR mode at night,” in IEEE
PES Innovative Smart Grid Technologies Conference Europe, ISGT
Europe, 2011.
T. K. Kristoffersen, K. Capion, and P. Meibom, “Optimal charging
of electric drive vehicles in a market environment,” Appl. Energy,
vol. 88, no. 5, pp. 1940–1948, 2011.
D. P. Palomar and M. Chiang, “A tutorial on decomposition methods
for network utility maximization,” IEEE J. Sel. Areas Commun., vol.
24, no. 8, pp. 1439–1451, 2006.
S. Boyd, L. Xiao, A. Mutapcic, and J. Mattingley, “Notes on
IEEE TRANSACTIONS ON SMART GRID
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
Decomposition Methods,” 2008.
R. Johari and J. Tsitsiklis, “Efficiency Loss in Network Resource
Allocation Game,” Math. Oper. Res., vol. 29, no. 3, pp. 407–435,
2004.
F. Kelly, “Charging and rate control for elastic traffic,” Eur. Trans.
Telecommun., vol. 8, no. 1, pp. 33–37, 1997.
P. Samadi, A.-H. Mohsenian-Rad, R. Schober, V. W. S. Wong, and
J. Jatskevich, “Optimal Real-Time Pricing Algorithm Based on
Utility Maximization for Smart Grid,” 2010 First IEEE Int. Conf.
Smart Grid Commun., pp. 415–420, 2010.
J. Ma, J. Deng, L. Song, and Z. Han, “Incentive Mechanism for
Demand Side Management in Smart Grid Using Auction,” IEEE
Trans. Smart Grid, vol. 5, no. 3, pp. 1379–1388, 2014.
D. Bertsekas, A. Nedi, and A. Ozdaglar, “Convex analysis and
optimization,” 2003.
D. P. Bertsekas, “Nonlinear programming,” Proceedings of the
Second Berkeley Symposium on Mathematical Statistics and
Probability, J. Neyman , University of California Press, Berkeley,.
pp. 481–492, 1999.
J. Lee, J. Guo, J. K. Choi, and M. Zukerman, “Distributed Energy
Trading in Microgrids: A Game Theoretic Model and Its Equilibrium
Analysis,” IEEE Trans. Ind. Electron., vol. 62, no. 6, pp. 3524-3533
1–1, 2015.
T. Alpcan and T. Başar, “A globallly stable adaptive congestion
control scheme for internet-style networks with delay,” IEEE/ACM
Trans. Netw., vol. 13, no. 6, pp. 1261–1274, 2005.
D. Niyato and Z. Han, “Dynamics of multiple-seller and multiplebuyer spectrum trading in cognitive radio networks: A game-theoretic
modeling approach,” IEEE Trans. Mob. Comput., vol. 8, no. 8, pp.
1009–1022, 2009.
J. A. Taylor, Convex Optimization of Power Systems. Cambridge
University Press, 2015.
W. Saad, Z. Han, M. Debbah, A. Hjorungnes, T. Basar, “Coalitional
game theory for communication networks,” IEEE Signal Processing
Magazine, vol. 26, no. 5, pp.77-97, 2009.
12
| 3cs.SY
|
arXiv:1602.02603v2 [math.GT] 12 Feb 2016
QUASI-ISOMETRIC RIGIDITY
OF PIECEWISE GEOMETRIC MANIFOLDS
R. FRIGERIO
Abstract. Two groups are virtually isomorphic if they can be obtained
one from the other via a finite number of steps, where each step consists in taking a finite extension or a finite index subgroup (or viceversa). Virtually isomorphic groups are always quasi-isometric, and a
group Γ is quasi-isometrically rigid if every group quasi-isometric to Γ
is virtually isomorphic to Γ. In this survey we describe quasi-isometric
rigidity results for fundamental groups of manifolds which can be decomposed into geometric pieces. After stating by now classical results
on lattices in semisimple Lie groups, we focus on the class of fundamental groups of 3-manifolds, and describe the behaviour of quasi-isometries
with respect to the Milnor-Kneser prime decomposition (following Papasoglu and Whyte) and with respect to the JSJ decomposition (following
Kapovich and Leeb). We also discuss quasi-isometric rigidity results for
fundamental groups of higher dimensional graph manifolds, that were
recently defined by Lafont, Sisto and the author. Our main tools are
the study of geometric group actions and quasi-actions on Riemannian
manifolds and on trees of spaces, via the analysis of the induced actions
on asymptotic cones.
1. Introduction
The study of finitely generated groups up to quasi-isometry is a broad and
very active research area which links group theory to geometric topology.
As stated, the task of classifying groups up to quasi-isometry is certainly too
ambitious. Nevertheless, when restricting to classes of groups coming from
specific algebraic or geometric contexts, several astonishing results have been
obtained in the last three decades. In this survey we will go through some
of them, paying a particular attention to the case of fundamental groups of
manifolds obtained by gluing “geometric” pieces.
Once a symmetric finite set S of generators for a group Γ has been fixed,
one can construct a graph having as vertices the elements of Γ, in such a
way that two vertices are connected by an edge whenever the corresponding
elements of the group are obtained one from the other by right multiplication
by a generator. The resulting graph CS (Γ) is the Cayley graph of Γ (with
2000 Mathematics Subject Classification. 20F65, 20F67, 20F69, 57M07, 57M50, 57M60,
20E06, 20E08, 53C23.
Key words and phrases. Prime decomposition, JSJ decomposition, relatively hyperbolic
group, asymptotic cone, Bass-Serre tree, non-positive curvature, hyperbolic manifolds.
1
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
2
respect to S). Its small-scale structure depends on the choice of the set of
generators S, but the quasi-isometry class of CS (Γ) is actually independent
of S, so the quasi-isometry class of the group Γ is well-defined (see Section 2
for the definition of quasi-isometry and for further details). It is natural
to ask which algebraic and/or geometric properties of the group Γ may be
encoded by the quasi-isometry class of its Cayley graphs. Another way to
phrase the same question is the following: if Γ, Γ′ are quasi-isometric groups,
what sort of properties must Γ share with Γ′ ? Surprisingly enough, it turns
out that the (apparently quite loose) relation of being quasi-isometric implies
many other relations which may look much finer at first glance (see the
paragraph after Definition 1.1 and Subsection 1.1 for a list of many results
in this spirit).
In this paper, we will mainly focus on the particular case of fundamental
groups of manifolds which decompose into specific geometric pieces. For
the sake of simplicity, henceforth we confine ourselves to the case of orientable manifolds. In dimension 3 the picture is particularly well understood. The now proved Thurston’s geometrization conjecture implies that
every closed (i.e. compact without boundary) 3-manifold M can be canonically decomposed into locally homogeneous Riemannian manifolds. Namely,
Milnor–Kneser prime decomposition Theorem [64] implies that M can first
be cut along spheres into summands M1 , . . . , Mk such that each Mi (after
filling the resulting boundary spheres with disks) either is homeomorphic to
S 2 × S 1 or is irreducible, meaning that every 2-sphere in it bounds a 3-ball.
Then, every irreducible manifold admits a canonical decomposition along
a finite family of disjoint tori, named JSJ decomposition after Jaco–Shalen
and Johansson [48, 49, 50, 59], such that every component obtained by cutting along the tori is either Seifert fibered or hyperbolic. A quite vague
but natural question is then the following: can quasi-isometries detect the
canonical decomposition of 3-manifolds? And what can one say in higher
dimensions? The main aim of this paper is to survey the work done by many
mathematicians in order to provide answers to these questions.
Before going on, let us first describe some phenomenon that quasi-isometries
cannot detect. It readily follows from the definition that every finite group
is quasi-isometric to the trivial group, but even more is true. Following [19],
we say that two groups Γ1 , Γ2 are virtually isomorphic if there exist finite
index subgroups Hi < Γi and finite normal subgroups Fi ⊳ Hi , i = 1, 2, such
that H1 /F1 is isomorphic to H2 /F2 . It is an exercise to check that virtual
isomorphism is indeed an equivalence relation. In fact, it is the smallest
equivalence relation for which any group is equivalent to any of its finite
index subgroups, and any group is equivalent to any of its finite extensions,
where we recall that Γ is a finite extension of Γ′ if it fits into an exact
sequence of the form
1 −→ F −→ Γ −→ Γ′ −→ 1 ,
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
3
where F is finite. It is not difficult to show that virtually isomorphic groups
are quasi-isometric (see Remark 2.2). On the other hand, being virtually isomorphic is a very strict condition, so there is no reason why quasi-isometric
groups should be virtually isomorphic in general. The following definition
singles out those situations where quasi-isometry implies virtual isomorphism.
Definition 1.1. A group Γ is quasi-isometrically rigid (or QI-rigid for short)
if any group quasi-isometric to Γ is in fact virtually isomorphic to Γ. A
collection of groups C is quasi-isometrically rigid (or QI-rigid for short) if
any group quasi-isometric to a group in C is virtually isomorphic to (possibly
another) group in C.
We are not giving here the most complete list of groups or of classes of
groups that are (or that are not) quasi-isometrically rigid (the interested
reader is addressed e.g. to [51] or to [19, Chapter 23]). However, it is maybe
worth mentioning at least some results that can help the reader to put the
subject of this survey in a more general context. Some important properties of groups are immediately checked to be preserved by quasi-isometries
(this is the case, for example, for amenability). Moreover, many algebraic
properties are preserved by quasi-isometries. For example, free groups are
QI-rigid [89, 21] (and this fact is closely related to our discussion of the invariance of the prime decomposition under quasi-isometries in Appendix A),
as well as nilpotent groups [40] and abelian groups [72]. On the contrary,
the class of solvable groups is not QI-rigid [24]. Also algorithmic properties of groups are often visible to quasi-isometries: for example, the class
of finitely presented groups with solvable word problem is QI-rigid (in fact,
the class of finitely presented groups is QI-rigid, and the growth type of the
Dehn function of a finitely presented group is a quasi-isometry invariant –
see e.g. the nice survey by Bridson [7]).
1.1. QI-rigidity of lattices in semisimple Lie groups. In this survey
we are mainly interested in quasi-isometric rigidity of fundamental groups
of manifolds with specific geometric properties. At least in the case of locally symmetric spaces (which are in some sense the most regular manifolds,
hence the most natural spaces to first deal with) the subject is now completely understood, thanks to the contribution of several mathematicians.
The general strategy that lead to the complete classification up to quasiisometry of lattices in semisimple Lie groups applies also to the case we
are interested in, so we briefly discuss it here. First of all, the fundamental Milnor-Svarc Lemma (see Proposition 2.1) asserts that the fundamental
group of any closed Riemannian manifold is quasi-isometric to the universal
covering of the manifold. Therefore, uniform lattices in the same semisimple Lie group G are quasi-isometric, since they are all quasi-isometric to the
symmetric space X associated to G. Moreover, the classification of groups
quasi-isometric to such lattices coincides with the classification of groups
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
4
quasi-isometric to X. Now, whenever a group Γ is quasi-isometric to a geodesic space X, it is possible to construct a geometric quasi-action of Γ on X
by quasi-isometries (see Proposition 2.4). In order to show quasi-isometric
rigidity of uniform lattices in G, one would like to turn this quasi-action into
a genuine action by isometries. Having shown this, one is then provided with
a homomorphism of Γ into G, and it is routine to show that such a homomorphism has a finite kernel and a discrete image. In the case when X has
higher rank, QI-rigidity then holds because every quasi-isometry stays at
bounded distance from an isometry [57, 28]. The same is true also for every
quaternionic hyperbolic space of dimension at least two [73], while in the case
of real or complex hyperbolic spaces there exist plenty of quasi-isometries
which are not at finite distance from any isometry. In the real hyperbolic
case, as already Gromov pointed out [41], Sullivan’s and Tukia’s results on
uniformly quasi-conformal groups [93, 90] imply that, in dimension strictly
bigger than 2, every group quasi-action by uniform quasi-isometries is just
the perturbation of a genuine isometric action. This concludes the proof of
QI-rigidity for the class of uniform lattices in G = Isom(Hn ), n ≥ 3 (see
also [11]). The complex hyperbolic case is settled thanks to an analogous
argument due to Chow [13]. Finally, QI-rigidity also holds for surface groups
(i.e. for uniform lattices in G = Isom(H2 )) thanks to a substantially different
proof obtained from the combination of results by Tukia [94], Gabai [38] and
Casson–Jungreis [12]. Summarizing, we have the following:
Theorem 1.2. Let G be a semisimple Lie group. Then the class of irreducible uniform lattices in G is QI-rigid.
Note however that, since every uniform lattice in G is quasi-isometric to
the associated symmetric space X, a single uniform lattice is usually very
far from being QI-rigid. Surprisingly enough, much more can be said when
dealing with non-uniform lattices. In fact, a non-uniform lattice Γ in G is
quasi-isometric to the complement in the corresponding symmetric space X
of an equivariant collection of horoballs. Such a space is much more rigid
than X itself, and this fact can be exploited to show the following:
Theorem 1.3 ([81, 82, 28, 27]). Let G be a semisimple Lie group distinct
from SL(2, R). Then every non-uniform lattice in G is QI-rigid.
The fact that patterns inside X (such as the previously mentioned collection of horoballs) can be of use in proving rigidity of maps will clearly
emerge in the discussion of quasi-isometric rigidity of fundamental groups
of manifolds with distinguished submanifolds (such as the tori of the JSJdecomposition of irreducible 3-manifolds, or the tori separating the pieces
of higher dimensional graph manifolds introduced in [36]).
For a detailed survey on QI-rigidity of lattices in semisimple Lie groups
we refer the reader to [31].
1.2. Quasi-isometric invariance of the prime decomposition. Let us
recall that every manifold will be assumed to be orientable throughout the
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
5
paper. A 3-manifold M is prime if it does not admit any non-trivial decomposition as a connected sum (a connected sum is trivial if one of its
summands is the 3-sphere S 3 ), and it is irreducible if every 2-sphere embedded in M bounds an embedded 3-ball. Every irreducible manifold is prime,
and the only prime manifold which is not irreducible is S 2 × S 1 . As mentioned above, every closed 3-manifold M admits a canonical decomposition
M = M1 # . . . #Mk into a finite number of prime summands. Of course we
have π1 (M ) = π1 (M1 ) ∗ . . . ∗ π1 (Mk ), and on the other hand the (proved)
Kneser conjecture asserts that every splitting of the fundamental group of a
closed 3-manifold as a free product is induced by a splitting of the manifold
as a connected sum (see e.g. [46, Theorem 7.1]). As a consequence, the study
of the quasi-isometry invariance of the prime decomposition boils down to
the study of the behaviour of quasi-isometries with respect to free decompositions of groups, a task that has been completely addressed by Papasoglu
and Whyte in [76].
It is not true that the quasi-isometry type of the fundamental group recongnizes whether a manifold is prime, since π1 (S 2 ×S 1 ) = Z and π1 (P3 (R)#P3 (R)) =
Z ⋊ Z2 are obviously virtually isomorphic (however, Theorem 1.4 implies
that irreducibility is detected by the quasi-isometry class of the fundamental group). Moreover, if Γ1 , Γ2 contain at least three elements and F is
virtually cyclic, then Γ1 ∗ Γ2 ∗ F is quasi-isometric to Γ1 ∗ Γ2 [76], so quasiisometries cannot see in general all the pieces of the prime decomposition of
a manifold. However, the following result shows that quasi-invariance of the
prime decomposition with respect to quasi-isometries holds as strongly as
the just mentioned examples allow. For the sake of convenience, let us say
that a prime manifold is big if its fundamental group is not virtually cyclic
(in particular, it is infinite).
Theorem 1.4. Let M, M ′ be closed orientable 3-manifolds. Then π1 (M )
is quasi-isometric to π1 (M ′ ) if and only if one of the following mutually
exclusive conditions holds:
(1) Both M and M ′ are prime with finite fundamental group.
(2) Both M and M ′ are irreducible with infinite quasi-isometric fundamental groups.
(3) M and M ′ belong to the set {S 2 × S 1 , P3 (R)#P3 (R)}.
(4) both M and M ′ are not prime and distinct from P3 (R)#P3 (R); moreover, if N is a big piece in the prime decomposition of M , then there
exists a big piece N ′ in the prime decomposition of M ′ such that
π1 (N ) is quasi-isometric to π1 (N ′ ), and viceversa.
In the spirit of Definition 1.1, one may also ask what can be said about
groups which are quasi-isometric to the fundamental group of a closed 3manifold. The following holds:
Theorem 1.5. Let M be a closed non-prime 3-manifold distinct from P3 (R)#P3 (R),
and let Γ be a finitely generated group. Then Γ is quasi-isometric to π1 (M ) if
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
6
and only if Γ has infinitely many ends and splits as the fundamental group of
a graph of groups with finite edge groups such that the set of quasi-isometry
types of one-ended vertex groups coincides with the set of quasi-isometry
types of big summands in the prime decomposition of M .
Theorems 1.4 and 1.5 readily descend from [76]. Nevertheless, since apparently they are not available in an explicit form elsewhere, we will provide
a quick proof of them in the appendix. By Theorems 1.4 and 1.5, the problem of classifying fundamental groups of 3-manifolds up to quasi-isometry
is now reduced to the study of irreducible ones.
1.3. QI-rigidity of the eight 3-dimensional geometries. As already
mentioned, the now proved Thurston’s geometrization conjecture states that
every closed irreducible 3-manifold decomposes along a finite family of disjoint π1 -injective tori (and Klein bottles) into a finite collection of manifolds
each of which admits a complete finite-volume locally homogeneous Riemannian metric (henceforth, such a manifold will be said to be geometric).
Strictly speaking, this decomposition (which will be called geometric form
now on) is not exactly the same as the classical JSJ decomposition of the
manifold, the only differences being the following ones:
(1) If M is a closed manifold locally isometric to Sol (see below), then
the geometric decomposition of M is trivial. On the other hand, M
either is a torus bundle over S 1 , or it is obtained by gluing along their
toric boundaries two twisted orientable I-bundles over the Klein bottle. In the first case, the JSJ decomposition of M is obtained by
cutting along a fiber of the bundle, in the second one it is obtained
by cutting along the torus that lies bewteen the I-bundles.
(2) If M is not as in (1), then each torus of the JSJ decomposition
which bounds a twisted orientable I-bundle over a Klein bottle is
replaced in the geometric decomposition by the core Klein bottle of
the bundle.
In fact, neither (the internal part of) the twisted orientable I-bundle over the
Klein bottle nor S 1 × S 1 × (0, 1) supports a locally homogeneous Riemannian metric of finite volume. The geometric decomposition is still canonical
(i.e. preserved by homeomorphisms) and cuts the manifolds into geometric
pieces in the sense described above.
Up to equivalence, there exist exactly eight homogeneous Riemannian
3-manifolds which admit a finite-volume quotient: the constant curvature
spaces S 3 , R3 , H3 , the product spaces S 2 × R, H2 × R, and the 3-dimensional
^
g2 = SL
Lie groups SL
2 (R), Sol, N il [92, 83] (the equivalence that needs
to be taken into account is not quite the relation of being homothetic, because a single Lie group may admit several non-homothetic left-invariant
metrics, and we don’t want to consider the same Lie group endowed with
two non-homotetic metrics as two distinct geometries). We will call these
spaces “geometries” or “models”. We will say that a manifold is geometric
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
7
if it admits a complete finite-volume Riemannian metric locally isometric to
one of the eight models just introduced (with the exception of P3 (R)#P3 (R)
and of S 2 × S 1 , every geometric manifold is automatically irreducible). It
is well-known that, with the exception of H2 × R that is quasi-isometric
g2 , the eight 3-dimensional geometries are pairwise not quasi-isometric
to SL
(see Proposition 3.10). Together with Milnor–Svarc Lemma, this information is sufficient to classify irreducible 3-manifolds with trivial geometric
decomposition, up to quasi-isometries of their fundamental groups:
Proposition 1.6. Let M1 , M2 be closed irreducible 3-manifolds, and suppose
that Mi admits a locally homogeneous Riemannian metric locally isometric
to Xi , where Xi is one of the eight geometries described above. Then π1 (M1 )
is quasi-isometric to π1 (M2 ) if and only if either X1 = X2 or X1 , X2 ∈
g2 , H2 × R}.
{SL
Together with the fact that a closed 3-manifold cannot support both a
g2 , Propostructure locally modeled on H2 ×R and a structure modeled on SL
sition 1.6 implies tha the geometry of a single closed irreducible manifold
with trivial geometric decomposition is uniquely determined by its fundamental group.
We have already seen that the class of fundamental groups of closed hyperbolic 3-manifolds is QI-rigid. The following result extends quasi-isometric
rigidity to the remaining seven 3-dimensional geometries.
Theorem 1.7. Let X be any of the eight 3-dimensional geometries, and let
Γ be a group quasi-isometric to the fundamental group of a closed manifold
^
modeled on X. If X 6= H2 × R and X 6= SL
2 (R), then Γ is virtually
isomorphic to the fundamental group of a closed manifold modeled on X.
^
If X = H2 × R or X = SL
2 (R), then Γ is virtually isomorphic to the
^
fundamental group of a closed manifold modeled on H2 × R or on SL
2 (R).
Proof. The cases of S 3 and of S 2 × R are obvious due to the quasi-isometric
^
rigidity of cyclic groups. The case of H2 × R and SL
2 (R) is due to Rieffel [78] (see also [58]), while quasi-isometric rigidity of abelian groups follows
from [40, 72]. The case of Sol is settled in [29]. Finally, quasi-isometric
rigidity of N il may be proved as follows. Let Γ be quasi-isometric to N il.
Quasi-isometric rigidity of nilpotent groups [40] implies that Γ is virtually
nilpotent. Since every finitely generated nilpotent group is linear, by Selberg
Lemma we may suppose, up to virtual isomorphism, that Γ is nilpotent and
torsion-free. The Malcev closure G of Γ is a simply connected nilpotent Lie
group in which Γ embeds as a uniform lattice [77]. Being quasi-isometric,
the groups Γ, G and N il must have equivalent growth functions. Now results of Guivarc’h [45] imply that the growth of N il is polynomial of degree
4, and that the only simply connected nilpotent Lie groups with polynomial
growth of degree 4 are N il and R4 . In order to conclude it is now sufficient
to observe that the case G = R4 cannot hold, since otherwise the Heisenberg
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
8
group (which is a uniform lattice in N il) would be quasi-isometric to Z4 ,
whence virtually abelian.
1.4. Quasi-isometric invariance of the JSJ decomposition. Let M
be a closed irreducible 3-manifold. We are now interested in the case when
the decomposition of M is not trivial. In this case we say that the manifold is non-geometric (because indeed it does not support any complete
finite-volume locally homogeneous metric), and all the pieces resulting from
cutting along the tori and the Klein bottles of the geometric decomposition are the internal parts of compact irreducible 3-manifolds bounded by
tori. As it is customary in the literature, we will often confuse the concrete
pieces of the decomposition (which are open) with their natural compactifications obtained by adding some boundary tori. The possibilities for the
geometry of these pieces are much more restricted than in the closed case:
the only geometries that admit non-compact finite-volume quotients (whose
g2 and H2 × R. Moreover, if
compactification is) bounded by tori are H3 , SL
the internal part of a compact irreducible 3-manifold bounded by tori can
g2 , then it can be modeled also on H2 × R, and viceversa
be modeled on SL
(while as in the closed case, complete finite-volume hyperbolic manifolds
cannot support a metric locally isometric to H2 × R). Pieces modeled on
H2 × R admit a foliation by circles, and are called Seifert fibered (or simply
Seifert). They are finitely covered (as foliated manifolds) by manifolds of
the form Σ×S 1 (endowed with the obvious foliation by circles), where Σ is a
punctured surface of finite type with negative Euler characteristic. The first
question that comes to mind is then whether quasi-isometries of the fundamental group can detect the presence or the absence of hyperbolic pieces
and/or of Seifert fibered pieces. This question was answered in the positive
by Kapovich and Leeb in [54]. Before stating their result, let us recall that,
if N ⊆ M is a piece of the geometric decomposition of M , then the inclusion
N ֒→ M induces an injective map on fundamental groups. Therefore, π1 (N )
may be identified with (the conjugacy class of) a subgroup of π1 (M ), and
this fact will be always tacitly understood in the sequel. Kapovich and Leeb
first proved that quasi-isometries recognize the presence of Seifert fibered
pieces [52], and then they improved their result as follows:
Theorem 1.8 ([54]). Let M, M ′ be closed irreducible 3-manifolds, let f : π1 (M ) →
π1 (M ′ ) be a quasi-isometry, and let N ⊆ M be a piece of the geometric decomposition of M . Then there exists a piece N ′ of the geometric decomposition of M ′ such that f (π1 (N )) stays at bounded Hausdorff distance from (a
conjugate of ) π1 (N ′ ). Moreover, f |π1(N ) is at finite distance from a quasiisometry between π1 (N ) and π1 (N ′ ).
In particular, the set of the quasi-isometry types of the fundamental
groups of the pieces of an irreducible manifold is a quasi-isometry invariant of the fundamental group of the manifold. Since Seifert manifolds with
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
9
non-empty boundary and cusped hyperbolic 3-manifolds cannot have quasiisometric fundamental groups (see Corollary 3.7), it readily follows that the
presence of a hyperbolic and/or a Seifert piece is a quasi-isometry invariant
of the fundamental group (another proof of the fact that quasi-isometries
detect the presence of a hyperbolic piece was given by Gersten [39], who
showed that M contains a hyperbolic piece if and only if the divergence of
π1 (M ) is exponential).
Regarding quasi-isometric rigidity, in the same paper Kapovich and Leeb
proved the following:
Theorem 1.9 ([54]). The class of fundamental groups of closed non-geometric
irreducible 3-manifolds is QI-rigid.
Theorems 1.8 and 1.9 will be the main object of this survey. Following
Kapovich and Leeb, we will make an extensive use of asymptotic cones, a
very useful quasi-isometry invariant of geodesic spaces introduced by Gromov and studied by a number of mathematicians in the last decades. Nowadays, asymptotic cones of groups themselves are the subject of an independent research field. We will introduce them and briefly describe their
properties in Section 3.
The following result by Behrstock and Neumann shows that Theorem 1.8
may be strengthened in the case of graph manifolds, i.e. of irreducible 3manifolds whose geometric decomposition is not trivial and does not contain
any hyperbolic piece.
Theorem 1.10 ([4]). Let M, M ′ be closed non-geometric graph manifolds.
Then π1 (M ) is quasi-isometric to π1 (M ′ ).
Another big progress towards the classification of quasi-isometry types of
non-geometric irreducible 3-manifolds has been done in a subsequent paper
by Behrstock and Neumann [5], where the case when at least one of the
hyperbolic pieces of the decomposition is not arithmetic is completely addressed. Since their results cannot be properly stated without introducing
a bit of terminology, and since our attention in this paper will be primarily
concentrated on Theorems 1.8 and 1.9 above, we address the reader to [5]
for the precise statements.
Before jumping into the realm of higher dimensional manifolds, let us
stress the fact that all the theorems stated in this subsection make an essential use of the geometry supported by the pieces in which an irreducible
manifold decomposes. While in the case of the prime decomposition of
3-manifolds the fact that quasi-isometries recognize summands can be deduced from more general facts regarding the quasi-isometry classification of
free products, no general theorem on amalgamated products is known which
can be exploited to show that quasi-isometries also detect the pieces of the
geometric decomposition of irreducible 3-manifolds. During the last 25 years
the notion of JSJ decomposition has been extended from the case of (fundamental groups of) irreducible 3-manifolds to the context of finitely presented
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
10
groups. The JSJ decomposition was first defined by Sela for torsion-free
word-hyperbolic groups [86]. After Bowditch developed the theory for general word-hyperbolic groups [6], other notions of JSJ decomposition were
introduced by Rips and Sela [79], Dunwoody and Sageev [23], Scott and
Swarup [84], and Fujiwara and Papasoglu [37] in the general context of
finitely presented groups. However, uniqueness of such decompositions up
to isomorphism is already delicate (see e.g. Forester [33] and Guirardel and
Levitt [44]), and no quasi-isometry invariance result may be directly applied
to the case we are interested in. Moreover, in many versions of the theory the outcome of the decomposition is a graph of groups with two-ended
(i.e. infinite virtually cyclic) subgroups as edge groups, rather than rank2 free abelian edge groups like in the classical case. We refer the reader
e.g. to [74, 75, 65, 66] for some approaches to quasi-isometry invariance of
decompositions of groups as amalgamated products.
1.5. A glimpse to the higher-dimensional case. In dimension greater
than three generic manifolds do not admit decompositions into pieces with
controlled geometry. Therefore, in order to study to what extent quasiisometries capture the geometry of the pieces of a manifold obtained by
gluing “geometric pieces” one first need to define the appropriate class of
objects to work with. For example, Nguyen Phan defined in [67, 68] classes
of manifolds obtained by gluing non-positively curved finite-volume locally
symmetric spaces with cusps removed. In the cited papers, Nguyen Phan
proved (smooth) rigidity results for manifolds obtained this way. In the context of non-positively curved manifolds, Leeb and Scott defined a canonical
decomposition along embedded flat manifolds which is meant to generalize
to higher dimensions the JSJ decomposition of irreducible 3-manifolds [61].
A different class of manifolds, called higher dimensional graph manifolds,
was defined by Lafont, Sisto and the author in [36] as follows.
Definition 1.11. A compact smooth n-manifold M , n ≥ 3, is a higher
dimensional graph manifold (or HDG manifold for short) if it can be constructed in the following way:
(1) For every i = 1, . . . , r, take a complete finite-volume non-compact
hyperbolic ni -manifold Ni with toric cusps, where 3 ≤ ni ≤ n.
(2) Denote by N i the manifold obtained by “truncating the cusps” of
Ni , i.e. by removing from Ni a horospherical neighbourhood of each
cusp.
(3) Take the product Vi = N i × T n−ni , where T k = (S 1 )k is the kdimensional torus.
(4) Fix a pairing of some boundary components of the Vi ’s and glue the
paired boundary components via diffeomorphisms, so as to obtain a
connected manifold of dimension n.
Observe that ∂M is either empty or consists of tori. The submanifolds
V1 , . . . , Vr are the pieces of M . The manifold N i is the base of Vi , while
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
11
every subset of the form {∗} × T n−ni ⊆ Vi is a fiber of Vi . The boundary
tori which are identified together are the internal walls of M (so any two
distinct pieces in M are separated by a collection of internal walls), while
the components of ∂M are the boundary walls of M .
Therefore, HDG manifolds can be decomposed into pieces, each of which
supports a finite-volume product metric locally modeled on some Hk × Rn−k
(k ≥ 3).
Remark 1.12. In the original definition of HDG manifold the gluing diffeomorphisms between the paired tori are required to be affine with respect to
the canonical affine structure that is defined on the boundary components
of the pieces. Without this restriction, no smooth rigidity could hold for
HDG manifolds. Nevertheless, the quasi-isometric rigidity results we are
investigating here are not affected by allowing generic diffeomorphisms as
gluing maps.
Probably the easiest (and somewhat less interesting) examples of HDG
manifolds are the so-called purely hyperbolic HDG manifolds, i.e. those HDG
manifolds whose pieces are all just truncated hyperbolic manifolds. Such
manifolds enjoy additional nice properties, that are of great help in understanding their geometry: for example, they support non-positively curved
Riemannian metrics (at least when the gluings are affine), and their fundamental groups are relatively hyperbolic [36]. Examples of purely hyperbolic
HDG manifolds include both the classical “double” of a finite volume hyperbolic manifold with toric cusps, as well as twisted doubles of such manifolds
(in the sense of Aravinda and Farrell [2]).
A number of rigidity results (like smooth rigidity à la Mostow within the
class of HDG manifolds obtained via affine gluings and topological rigidity à
la Farrell–Jones within the class of aspherical manifolds) are proved in [36]
for generic HDG manifolds. In order to get quasi-isometric rigidity, however,
a further assumption is needed, which ensures that the (fundamental group
of) each piece is undistorted in the fundamental group of the manifold. Let
M be a HDG graph manifold, and V + , V − a pair of adjacent (not necessarily
distinct) pieces of M . We say that the two pieces have transverse fibers along
the common internal wall T provided that, under the gluing diffeomorphism
ψ : T + → T − of the paired boundary tori corresponding to T , the image
of the fiber subgroup of π1 (T + ) under ψ∗ intersects the fiber subgroup of
π1 (T − ) only in {0}.
Definition 1.13. A HDG manifold is irreducible if every pair of adjacent
pieces has transverse fibers along every common internal wall.
In the case of 1-dimensional fibers (and when restricting only to affine
gluings between the pieces), a HDG graph manifold is irreducible if and only
if the S 1 -bundle structure on each piece cannot be extended to the union
of adjacent pieces, a phenomemon which always occurs in 3-dimensional
graph manifolds built up by products of hyperbolic surfaces times S 1 . This
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
12
suggests that irreducible HDG manifolds should provide a closer analogue
to 3-dimensional graph manifolds than generic HDG manifolds.
Every purely hyperbolic HDG manifold is irreducible. However, the class
of irreducible HDG manifolds is much richer than the class of purely hyperbolic ones: for example, in each dimension n ≥ 4, there exist infinitely many
irreducible HDG n-manifolds which do not support any locally CAT(0) metric, in contrast with the fact that every purely hyperbolic HDG manifold
(obtained via affine gluings) supports a non-positively curved Riemannian
metric.
In this paper we will be mainly concerned with the behaviour of the fundamental groups of HDG manifolds with respect to quasi-isometries. Using
the technology of asymptotic cones (together with other tools, of course), in
Section 6 we will sketch the proof of the following results, which are taken
from [36]. The following result provides a higher-dimensional analogue of
Theorem 1.8:
Theorem 1.14. Let M1 , M2 be a pair of irreducible HDG manifolds, and
Γi = π1 (Mi ) their respective fundamental groups. Let Λ1 ≤ Γ1 be a subgroup
conjugate to the fundamental group of a piece in M1 , and ϕ : : Γ1 → Γ2
be a quasi-isometry. Then, the set ϕ(Λ1 ) is within finite Hausdorff distance
from a conjugate of Λ2 ≤ Γ2 , where Λ2 is the fundamental group of a piece
in M2 .
After one knows that fundamental groups of the pieces are essentially
mapped to fundamental group of the pieces under quasi-isometries, the next
goal is to understand the behavior of groups quasi-isometric to the fundamental group of a piece.
Theorem 1.15. Let N be a complete finite-volume hyperbolic m-manifold,
m ≥ 3, and let Γ be a finitely generated group quasi-isometric to π1 (N )×Zd ,
d ≥ 0. Then there exist a finite-index subgroup Γ′ of Γ, a finite-sheeted
covering N ′ of N , a group ∆ and a finite group F such that the following
short exact sequences hold:
1
1
/ Zd
/F
j
/∆
/ Γ′
/∆
/ π1 (N ′ )
/ 1,
/ 1.
Moreover, j(Zd ) is contained in the center of Γ′ . In other words, Γ′ is a
central extension by Zd of a finite extension of π1 (N ′ ).
In the case of purely hyperbolic pieces, Theorem 1.15 reduces to Schwartz’s
results on the quasi-isometric rigidity of non-uniform lattices in the group
of isometries of real hyperbolic space [81]. The analogous result in the case
where N is closed hyperbolic has been proved by Kleiner and Leeb [58].
Once Theorems 1.14 and 1.15 are established, it is not difficult to put the
pieces together to obtain the following:
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
13
Theorem 1.16. Let M be an irreducible HDG n-manifold obtained by gluing the pieces Vi = N i × T di , i = 1, . . . , k. Let Γ be a group quasi-isometric
to π1 (M ). Then either Γ itself or a subgroup of Γ of index two is isomorphic to the fundamental group of a graph of groups satisfying the following
conditions:
• every edge group contains Zn−1 as a subgroup of finite index;
• for every vertex group Γv there exist i ∈ {1, . . . , k}, a finite-sheeted
covering N ′ of Ni and a finite-index subgroup Γ′v of Γv that fits into
the exact sequences
1
1
/ Zdi
/F
j
/∆
/ Γ′
v
/∆
/ π1 (N ′ )
/ 1,
/ 1,
where F is a finite group, and j(Zdi ) is contained in the center of
Γ′v .
1.6. Plan of the paper. In Section 2 we recall the definition of quasiisometries, we briefly discuss the fundamental Milnor-Svarc Lemma and we
state a well-known characterization of quasi-isometry among groups based
on the notion of quasi-action. In Section 3 we introduce asymptotic cones,
and we discuss an asymptotic characterization of (relatively) hyperbolic
groups due to Drutu and Sapir that will prove useful in the subsequent
sections. Being the fundamental groups of manifolds which decompose into
pieces, the groups we are interested in are quasi-isometric to trees of spaces,
that are introduced in Section 4, where their asymptotic cones are also carefully analyzed. We come back from homeomorphisms between asymptotic
cones to quasi-isometries of the original spaces in Section 5, where we give
the proofs of Theorems 1.8 and 1.14. QI-rigidity of the classes of fundamental groups of irreducible non-geometric 3-manifolds and of irreducible
HDG manifolds is discussed in Section 6. In Section 7 we collect some open
questions. Finally, the appendix is devoted to the description of the behaviour of quasi-isometries with respect to free decompositions of groups as
exhaustively illustrated in [76], and to the proofs of Theorems 1.4 and 1.5.
Acknowledgments
Many thanks go to Elia Fioravanti for several interesting conversations on
Kapovich and Leeb’s proof of the invariance of the JSJ decomposition under
quasi-isometries. The author is indebted to Yves de Cornulier for suggesting
a quick proof of the quasi-isometric rigidity of N il (see Proposition 1.7).
2. Quasi-isometries and quasi-actions
Let (X, d), (Y, d′ ) be metric spaces and k ≥ 1, c ≥ 0 be real numbers.
A (not necessarily continuous) map f : X → Y is a (k, c)-quasi-isometric
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
14
embedding if for every p, q ∈ X the following inequalities hold:
d(p, q)
− c ≤ d′ (f (p), f (q)) ≤ k · d(p, q) + c.
k
Moreover, a (k, c)-quasi-isometric embedding f is a (k, c)-quasi-isometry
if there exists a (k, c)-quasi-isometric embedding g : Y → X such that
d′ (f (g(y)), y) ≤ c, d(g(f (x)), x) ≤ c for every x ∈ X, y ∈ Y . Such a
map g is called a quasi-inverse of f . It is easily seen that a (k, c)-quasiisometric embedding f : X → Y is a (k′ , c′ )-quasi-isometry for some k′ ≥ 1,
c′ ≥ 0 if and only if its image is r-dense for some r ≥ 0, i.e. if every point in
Y is at distance at most r from some point in f (X) (and in this case k′ , c′
only depend on k, c, r).
In the introduction we recalled the definition of Cayley graph of a group
with respect to a finite set of generators, and it is immediate to check that
different finite sets of generators for the same group define quasi-isometric
Cayley graphs, so that every finitely generated group is endowed with a
metric which is well-defined up to quasi-isometry. For later reference we
recall that left translations induce a well-defined isometric action of every
group on any of its Cayley graphs.
The following fundamental result relates the quasi-isometry type of a
group to the quasi-isometry type of a metric space on which the group acts
geometrically. A metric space X is geodesic if every two points in it can be
joined by a geodesic, i.e. an isometric embedding of a closed interval. An
isometric action Γ × X → X of a group Γ on a metric space X is proper if
for every bounded subset K ⊆ X the set {g ∈ Γ | g · K ∩ K 6= ∅} is finite, and
cobounded if there exists a bounded subset Y ⊆ X such that Γ · Y = X (or,
equivalently, if one or equivalently every orbit of Γ in X is r-dense for some
r > 0, which may depend on the orbit). An isometric action is geometric if
it is proper and cobounded.
Theorem 2.1 (Milnor-Svarc Lemma). Suppose Γ acts geometrically on a
geodesic space X. Then Γ is finitely generated and quasi-isometric to X, a
quasi-isometry being given by the map
ψ : Γ → X,
ψ(γ) = γ(x0 ),
where x0 ∈ X is any basepoint.
A proof of this result can be found e.g. in [8, Chapter I.8.19]. As a
corollary, if M is a compact Riemannian manifold with Riemannian universal
f, then the fundamental group of M is quasi-isometric to M
f.
covering M
Another interesting consequence of Milnor-Svarc Lemma is the fact that
virtual isomorphism implies quasi-isometry:
Remark 2.2. Suppose that the groups Γ, Γ′ are such that there exists a
short exact sequence
1
/F
/Γ
ϕ
/ Γ′′
/1,
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
15
where Γ′′ is a finite index subgroup of Γ′ , and let X be any Cayley graph
of Γ′ . Then every γ ∈ Γ isometrically acts on X via the left translation
by ϕ(γ). This action is cocompact because Γ′′ has finite index in Γ′ , and
proper because F is finite. By Milnor-Svarc Lemma, this implies that Γ is
quasi-isometric to Γ′ .
Thanks to Milnor-Svarc Lemma, two groups acting geometrically on the
same proper geodesic space are quasi-isometric. The converse implication
does not hold: namely, there are examples of quasi-isometric groups that do
not share any geometric model, i.e. for which there exist no geodesic space
on which both groups can act geometrically. In fact, it is proved in [65] that,
if p, q are distinct odd primes, then the groups Γ = Zp ∗ Zq and Γ′ = Zp ∗ Zp
are quasi-isometric (since they are virtually isomorphic to the free group on
two generators) but do not share any geometric model. Things get easier if
one weakens the notion of action into the one of quasi-action:
Definition 2.3. Suppose (X, d) is a geodesic metric space, let QI(X) be
the set of quasi-isometries of X into itself, and let Γ be a group. For k ≥ 1,
a k-quasi-action of Γ on X is a map h : Γ → QI(X) such that the following
conditions hold:
(1) h(γ) is a (k, k)-quasi-isometry with k-dense image for every γ ∈ Γ;
(2) d(h(1)(x), x) ≤ k for every x ∈ X;
(3) the composition h(γ1 ) ◦ h(γ2 ) is at distance bounded by k from the
quasi-isometry h(γ1 γ2 ), i.e.
d h(γ1 γ2 )(x), h(γ1 )(h(γ2 )(x)) ≤ k for every x ∈ X, γ1 , γ2 ∈ Γ.
A k-quasi-action h as above is cobounded if one orbit of Γ in X is k′ -dense
for some k′ > 0 (or, equivalently, if every orbit of Γ in X is k′ -dense, for a
maybe bigger k′ ), and proper if for every bounded subset K ⊆ X the set
{g ∈ Γ | g ·K ∩K 6= ∅} is finite. A quasi-action is geometric if it is cobounded
and proper.
Throughout the whole paper, by an abuse of notation, when h is a quasiaction as above we do not distinguish between γ and h(γ).
We have already observed that the property that the group Γ be quasiisometric to X is not sufficient to guarantee that Γ acts geometrically on X.
On the contrary, we can ask for geometric quasi -actions:
Proposition 2.4. Let ϕ : Γ → X be a quasi-isometry between a group and
a geodesic metric space with quasi-inverse ψ : X → Γ. Then the formula
h(γ)(x) = ϕ(γ · ψ(x))
for every x ∈ X
defines a geometric quasi-action h(γ) : X → X.
Proof. The fact that h is indeed a quasi-action readily follows from the fact
that left translations are isometries of Γ. Properness and coboundedness of
h are easily checked.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
16
The usual proof of Milnor-Svarc’s Lemma may easily be adapted to the
context of quasi-actions to yield the following:
Lemma 2.5 ([36], Lemma 1.4). Let X be a geodesic space with basepoint
x0 , and let Γ be a group. Let h : Γ → QI(X) be a geometric quasi-action
of Γ on X. Then Γ is finitely generated and the map ϕ : Γ → X defined by
ϕ(γ) = γ(x0 ) is a quasi-isometry.
Corollary 2.6. Let Γ, Γ′ be finitely generated groups. Then Γ is quasiisometric to Γ′ if and only if there exists a geodesic metric space X such
that both Γ and Γ′ geometrically quasi-act on X.
Proof. The “if” implication follows from Lemma 2.5. On the other hand,
if Γ is quasi-isometric to Γ′ and X is a fixed Cayley graph for Γ′ , then Γ
geometrically quasi-acts on X by Proposition 2.4, and G′ geometrically acts
on X ′ by left translations.
3. Asymptotic cones
Roughly speaking, the asymptotic cone of a metric space gives a picture
of the metric space as “seen from infinitely far away”. It was introduced by
Gromov in [40], and formally defined in [95]. Being uninfluenced by finite
errors, the asymptotic cone is a quasi-isometry invariant, and turns discrete
objects into continuous spaces. Citing Gromov, “This space [the Cayley
graph of a group Γ] may appear boring and uneventful to a geometer’s eye
since it is discrete and the traditional local (e.g. topological and infinitesimal)
machinery does not run in Γ. To regain the geometric perspective one has
to change his/her position and move the observation point far away from Γ.
Then the metric in Γ seen from the distance d becomes the original distance
divided by d and for d → ∞ the points in Γ coalesce into a connected
continuous solid unity which occupies the visual horizon without any gaps
or holes and fills our geometer’s hearth with joy” [43].
Gromov himself provided a characterization of word hyperbolicity in terms
of asymptotic cones [42, 43] (see also [16] and [30]). Also relative hyperbolic groups admit a neat (and very useful) characterization via asymptotic
cones [20]. It should maybe worth mentioning that, while having nice metric
properties (e.g., asymptotic cones of Cayley graphs of groups are complete,
homogeneous and geodesic), asymptotic cones are quite wild from the topological point of view. They are often not locally compact (for example, a
group is virtually nilpotent if and only if all its asymptotic cones are locally compact [40, 16] if and only if one of its asymptotic cone is locally
compact [80, 47, 88]), and their homotopy type can be quite complicated:
Gromov himself [43] conjectured that the fundamental group of any asymptotic cone of any group either is simply connected, or has an uncountable
fundamental group. While this conjecture eventually turned out to be false
in general [71], usually non-simply connected asymptotic cones of groups are
rather complicated: Erschler and Osin showed that every countable group
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
17
is a subgroup of the fundamental group of an asymptotic cone of a finitely
generated group [26] (this result was then sharpened in [20], where it is
shown that, for every countable group G, there exists an asymptotic cone of
a finitely generated group whose fundamental group is the free product of
uncountably many copies of C). Moreover, Gromov’s conjectural dichotomy
was proved to hold in several cases (see e.g. [56, 14] for results in this direction).
Let us recall the definition of asymptotic cone of a space. A filter on N
is a set ω ⊆ P(N) satisfying the following conditions:
(1) ∅ ∈
/ ω;
(2) A, B ∈ ω =⇒ A ∩ B ∈ ω;
(3) A ∈ ω, B ⊇ A =⇒ B ∈ ω.
For example, the set of complements of finite subsets of N is a filter on N,
known as the Fréchet filter on N.
A filter ω is a ultrafilter if for every A ⊆ N we have either A ∈ ω or
Ac ∈ ω, where Ac := N \ A. For example, fixing an element a ∈ N, we
can take the associated principal ultrafilter to consist of all subsets of RbbN
which contain a. An ultrafilter is non-principal if it does not contain any
finite subset of N.
It is readily seen that a filter is an ultrafilter if and only if it is maximal
with respect to inclusion. Moreover, an easy application of Zorn’s Lemma
shows that any filter is contained in a maximal one. Thus, non-principal
ultrafilters exist (just take any maximal filter containing the Fréchet filter).
From this point on, let us fix a non-principal ultrafilter ω on N. As usual,
we say that a statement Pi depending on i ∈ N holds ω-a.e. if the set of
indices such that Pi holds belongs to ω. If X is a topological space, and
(xi ) ⊆ X is a sequence in X, we say that ω-lim xi = x∞ if xi ∈ U ω-a.e.
for every neighbourhood U of x∞ . When X is Hausdorff, an ω-limit of a
sequence, if it exists, is unique. Moreover, any sequence in any compact
space admits an ω-limit. For example, any sequence (ai ) in [0, +∞] admits
a unique ω-limit.
Now let (Xi , xi , di ), i ∈ N, be a sequence of pointed metric spaces. Let C
be the set of sequences (yi ), yi ∈ Xi , such that ω-lim di (xi , yi ) < +∞, and
consider the following equivalence relation on C:
(yi ) ∼ (zi )
⇐⇒
ω- lim di (yi , zi ) = 0.
We set ω-lim(Xi , xi , di ) = C/∼ , and we endow
ω-lim(Xi , xi , di ) with the welldefined distance given by dω [(yi )], [(zi )] = ω-lim di (yi , zi ). The pointed
metric space (ω-lim(Xi , xi , di ), dω ) is called the ω-limit of the pointed metric
spaces Xi .
Let (X, d) be a metric space, (xi ) ⊆ X a sequence of base-points, and
(ri ) ⊂ R+ a sequence of rescaling factors diverging to infinity. We introduce
the notation (Xω ((xi ), (ri )), dω ) := ω-lim(Xi , xi , d/ri ).
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
18
Definition 3.1. The metric space Xω (xi ), (ri ) , dω is the asymptotic
cone of X with respect to the ultrafilter ω, the basepoints (xi ) and the rescal-
ing factors (ri ). For conciseness, we will occasionally just write Xω (xi ), (ri )
for the asymptotic cone, the distance being implicitly understood to be dω ,
or even Xω when basepoints and rescaling factors are fixed.
If ω is fixed and (ai ) ⊆ R is any sequence, we say that (ai ) is o(ri )
(resp. O(ri )) if ω-lim ai /ri = 0 (resp. ω-lim |ai |/ri < ∞). Let (xi ) ⊆ X,
(ri ) ⊆ R be fixed sequences of basepoints and rescaling factors, and set
Xω = (Xω ((xi ), (ri )), dω ). Sequences of subsets in X give rise to subsets of
Xω : if for every i ∈ N we are given a subset ∅ =
6 Ai ⊆ X, we set
ω- lim Ai = {[(pi )] ∈ Xω | pi ∈ Ai for every i ∈ N}.
It is easily seen that for any choice of the Ai ’s, the set ω-lim Ai is closed in
Xω . Moreover, ω-lim Ai 6= ∅ if and only if the sequence (d(xi , Ai )) is O(ri ).
It is usually quite difficult to describe asymptotic cones of spaces. Nevertheless, in some cases the situation is clear: for example, if the metric of
X is homogeneous and scale-invariant (meaning that (X, d) is isometric to
(X, d/r) for every r > 0) then every asymptotic cone of X is isometric to X
(this is the case e.g. for X = Rn ). We will see below that also asymptotic
cones of hyperbolic spaces are easily understood.
3.1. Asymptotic cones and quasi-isometries. Quasi-isometries are just
maps that are bi-Lipschitz up to a finite error. Therefore, it is not surprising
that they asymptotically define bi-Lipschitz homeomorphisms. In fact, it is
an easy exercise to check that, once a non-principal ultrafilter ω is fixed,
if (Xi , xi , di ) (Yi , yi , d′i ), i ∈ N are pointed metric spaces, (ri ) ⊂ R is a
sequence of rescaling factors, and fi : Xi → Yi are (ki , ci )-quasi-isometries
such that k = ω-lim ki < ∞, ci = o(ri ) and d′i (fi (yi ), xi ) = O(ri ), then the
formula [(pi )] 7→ [fi (pi )] provides a well-defined k-bi-Lipschitz embedding
fω : ω-lim(Xi , xi , di /ri ) → ω − lim(Yi , yi , d′i /ri ).
As a corollary, quasi-homogeneous spaces (i.e. metric spaces whose isometry group admits r-dense orbits for some r > 0) have homogeneous asymptotic cones, whose isometry type does not depend on the choice of basepoints.
Moreover, quasi-isometric metric spaces have bi-Lipschitz homeomorphic asymptotic cones. In particular, quasi-isometric groups have homogeneous
bi-Lipschitz homeomorphic asymptotic cones.
Actually, the last sentence of the previous paragraph should be stated
more precisely as follows: once a non-principal ultrafilter and a sequence of
rescaling factors are fixed, any quasi-isometry between two groups induces a
bi-Lipschitz homeomorphism between their asymptotic cones. Here and in
what follows we will not focus on the dependence of asymptotic cones on the
choice of ultrafilters and rescaling factors. In fact, in many applications (and
in all the arguments described in this paper) the role played by such choices
is very limited (but see the proof of Theorem 5.1). Let us just mention
here that, answering to a question of Gromov [42], Thomas and Velickovic
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
19
exhibited a finitely generated group with two non-homeomorphic asymptotic
cones [91] (the first finitely presented example of such a phenomenon is due
to Ol’shanskii and Sapir [70]).
3.2. Asymptotic cones of (relatively) hyperbolic groups. Recall that
a geodesic metric space X is δ-hyperbolic, where δ is a non-negative constant, if every side of every geodesic triangle in X is contained in the δneighbourhood of the union of the other two sides (in this case one says
that the triangle is δ-fine). A space is hyperbolic if it is δ-hyperbolic for
some δ ≥ 0. A fundamental (and not completely trivial) fact is that being
hyperbolic is a quasi-isometry invariant, so it makes sense to speak of hyperbolic groups. A real tree is a 0-hyperbolic geodesic metric space. Simplicial
trees are real trees, but many wilder examples of real trees can be easily
constructed (and naturally arise as asymptotic cones of groups!).
Every pinched negatively curved simply connected Riemannian manifold
is hyperbolic, so by Milnor-Svarc Lemma the fundamental group of every
compact negatively curved Riemannian manifold is hyperbolic. In fact, such
groups were probably the motivating examples that lead Gromov to the
definition of hyperbolicity. Of course, by dividing the metric by a factor
rn we can turn a δ-hyperbolic space into a δ/rn -hyperbolic space, and this
implies that any asymptotic cone of a hyperbolic space is a real tree (after
one proves that geodesics in Xω are ω-limits of geodesics in X: this is false
for generic geodesic spaces, and true if X is hyperbolic). More precisely, if
Γ is a non-virtually cyclic hyperbolic group, then every asymptotic cone of
Γ is a homogeneous real tree each point of which is a branching point with
uncountably many branches. In particular, all the asymptotic cones of all
non-virtually cyclic hyperbolic groups are pairwise isometric [25].
As we mentioned above, being asymptotically 0-hyperbolic indeed characterizes hyperbolic spaces: if every asymptotic cone of X is a real tree, then
X is hyperbolic [42, 43, 16, 30]. This implies, for example, the non-trivial
fact that if every triangle of diameter D in X is δ(D)-fine, where δ(D) grows
sublinearly with respect to D, then X is hyperbolic (so δ is bounded). It is
maybe worth mentioning that indeed one needs all the asymptotic cones of
X to be real trees in order to get hyperbolicity of X: groups having at least
one asymptotic cone which is a real tree are known as lacunary hyperbolic,
and were studied in [70]; note however that a result of Gromov [42] ensures
that a finitely presented lacunary hyperbolic group is indeed hyperbolic.
Just as hyperbolic groups generalize fundamental groups of negatively
curved compact manifolds, the notion of relatively hyperbolic group was introduced by Gromov to extend the class of fundamental groups of pinched
negatively curved complete manifolds of finite volume (see e.g. the fundamental paper of Farb [32]). Here we will define relative hyperbolic groups
by means of a later asymptotic characterization which is due to Druţu and
Sapir [20]. If X is a set, then we denote by |X| the cardinality of X.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
20
Definition 3.2. A geodesic metric space X is said to be tree-graded with
respect to a collection of closed subsets {Pi }i∈I , called pieces, if
S
(1) Pi = X,
(2) |Pi ∩ Pj | ≤ 1 if i 6= j,
(3) any simple geodesic triangle in X is contained in a single piece.
In fact, it is proved in [20, Lemma 2.15] that if A is a path-connected
subset of a tree-graded space Y and A has no cut-points, then A is contained
in a piece. In order to avoid trivialities, we will always assume that no piece
is equal to the whole tree-graded space.
Definition 3.3. A geodesic metric space X is asymptotically tree-graded
with respect to a collection of subsets A = {Hi }i∈I if the following conditions
hold:
(1) for each choice of basepoints (xi ) ⊆ X and rescaling factors (ri ), the
associated asymptotic cone Xω = Xω ((xi ), (ri )) is tree-graded with
respect to the collection of subsets P = {ω-lim Hi(n) | Hi(n) ∈ A},
and
(2) if ω-lim Hi(n) = ω-lim Hj(n) , where i(n), j(n) ∈ I, then i(n) = j(n)
ω-a.e.
We are now ready to give the definition of relatively hyperbolic group.
Definition 3.4. Let Γ be a group and let {H1 , . . . , Hk } be a finite collection of proper infinite subgroups of Γ. Then Γ is hyperbolic relative to
{H1 , . . . , Hk } if it is asymptotically tree-graded with respect to the family
of all left cosets of the Hi ’s. If this is the case, the Hi ’s are called peripheral
subgroups of Γ.
In some sense, the asymptotic cone of a relatively hyperbolic group is
obtained from a real tree by blowing up points into pieces where the nontree-like behaviour is concentrated.
Definition 3.4 implies that a group admitting an asymptotic cone without cut-points cannot be relatively hyperbolic (such a group is said to be
unconstricted) [20]. In fact, the following stronger result holds:
Proposition 3.5 ([20, 3]). If Γ is unconstricted, then every (k, c)-quasiisometric embedding of Γ into a relatively hyperbolic group lies in a Dneighbourhood of a coset of a peripheral subgroup, where D only depends
on k, c (and on the geometry of the relatively hyperbolic group).
Recall that the geometric pieces of the manifolds we are interested in
(i.e. irreducible 3-manifolds and HDG manifolds) each consist of the product of a cusped hyperbolic manifold N with a torus (where the cusped
manifold may be just a surface in the case of pieces modeled on H2 × R,
and the torus may be reduced to a circle or to a point). The fundamental
groups of cusped hyperbolic manifolds are hyperbolic relative to the cusp
subgroups (see e.g. [32] for a much more general result dealing with variable
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
21
negative curvature). As usual, in order to better understand the geometry
of the group one would like to make use of Milnor-Svarc Lemma, and to this
aim we need to replace our cusped n-manifold N with the one obtained by
truncating the cusps, which will be denoted by N . The universal covering
of N is the complement in hyperbolic space of an equivariant collection of
disjoint horoballs whose union is just the preimage of the cusps under the
universal covering map Hn → N . Such a space B is usually called a neutered
space, and will be always endowed with the path metric induced by its Riemannian structure. Its boundary is given by an equivariant collection of
disjoint horospheres, each of which is totally geodesic (in a metric sense)
and flat. Now the quasi-isometry provided by Milnor-Svarc Lemma sends
each coset of a peripheral subgroups of π1 (N ) into a component of ∂B (this
is true as stated if N has one cusp and the basepoint of B is chosen on ∂B,
and true only up to finite errors otherwise). Together with the invariance
of asymptotic cones with respect to quasi-isometries, we may conclude that
B is asymptotically tree–graded with respect to the family of the connected
components of ∂B.
We recall that an n-flat is a totally geodesic embedded subspace isometric
to Rn , and that whenever X, Y are metric spaces we understand that X×Y is
endowed with the induced ℓ2 -metric (which coincides with the usual product
metric in the case of Riemannian manifolds). Since the asymptotic cone of a
product is the product of the asymptotic cones of the factors, the following
result summarizes the previous discussion:
Proposition 3.6. Let M = N × (S 1 )k , where N is a complete finite volume
hyperbolic n-manifold with cusps removed and (S 1 )k is a flat k-dimensional
torus, let X be the metric universal covering of M and let Xω be any asymptotic cone of X. Then Xω is isometric to Y × Rk , where Y is a tree-graded
space and every piece of Y is an (n − 1)-flat. In particular, if dim N = 2,
then Xω is isometric to the product of a real tree with Rk .
The last sentence of the previous proposition can be explained in two
ways. If dim N = 2, then π1 (N ) is free, whence hyperbolic. On the other
hand, π1 (N ) is also hyperbolic relative to the peripheral subgroups, which
are isomorphic to Z, so its asymptotic cone is tree-graded with pieces isometric to lines; and it is easy to see that a tree-graded space with pieces
that are real trees is itself a real tree.
Corollary 3.7. Let M1 , M2 be a Seifert 3-manifold with non-empty boundary and a non-compact complete finite-volume hyperbolic 3-manifold, respectively. Then π1 (M1 ) is not quasi-isometric to π1 (M2 ).
Proof. By Proposition 3.6, it is sufficient to observe that the product of a
real tree with the real line does not contain cut-points, while any tree-graded
space does.
Of course free abelian groups are unconstricted, so Proposition 3.5 (together with Milnor-Svarc Lemma) readily implies the following:
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
22
Proposition 3.8. Let B ⊆ Hn be an n-dimensional neutered space as above.
Then the image of any (k, c)-quasi-isometric embedding of Rn−1 into B lies
in the D-neighbourhood of a component of ∂B, where D only depends on
(k, c) (and on the geometry of B).
Proposition 3.8 was first proved by Schwartz in [81], where it provided
one of the key steps in the proof of the quasi-isometric rigidity of nonuniform lattices in the isometry group of real hyperbolic n-space, n ≥ 3. In
fact, Proposition 3.8 implies that any quasi-isometry f : B → B ′ between
neutered spaces must coarsely send ∂B into ∂B ′ . As a consequence, the
map f can be extended to a quasi-isometry f of the whole of Hn . An additional argument exploiting the fact that f sends an equivariant family of
horospheres into another equivariant family of horospheres and involving a
fine analysis of the trace of f on ∂Hn allows to conclude that f is uniformly
close to an isometry g. Then it is not difficult to show that g almost conjugates the isometry group of B into the isometry group of B ′ , so that these
isometry groups (which are virtually isomorphic to the fundamental groups
of any compact quotient of B, B ′ , respectively) are virtually isomorphic (see
the discussion in Subsection 6.1). This argument should somewhat clarify
the statement we made in the introduction, according to which the existence
of patterns with a distinguished geometric behaviour in a space X usually
imposes some rigidity on geometric maps of X into itself. In the case just
described the pattern is given by a family of flats, whose intrinsic geometry
already make them detectable in the hyperbolic context where they lie. In
other situations the single objects of the pattern do not enjoy peculiar intrinsic features. This is the case, for example, for the boundary components
of the universal covering of a compact hyperbolic n-manifold with geodesic
boundary. Such boundary components are themselves hyperbolic hyperplanes in Hn (in particular, their asymptotic cones are not unconstricted),
so in order to prove that they are recognized by quasi-isometries one cannot
rely on general results on relatively hyperbolic groups. Nevertheless, the
general strategy described by Schwartz still applies to get quasi-isometric
rigidity [34], once one replaces Proposition 3.5 with an argument making
use of a suitable notion of coarse separation. Such a notion was also at the
hearth of Schwartz’s original argument for Proposition 3.8.
Schwartz’s strategy to prove QI-rigidity of non-uniform lattices may be
in fact pursued also in the more general context of relatively hyperbolic
groups. In fact, building on Proposition 3.5 it is possible to show that, if Γ is
hyperbolic relative to {H1 , . . . , Hk }, where each Hi is unconstricted, and Γ′ is
quasi-isometric to Γ, then also Γ′ is hyperbolic relative to a a finite collection
of subgroups each of which is quasi-isometric to one of the Hi ’s [20, 3]. In
fact, the hypothesis on the peripheral subgroups of Γ being unconstricted
may be weakened into the request that each Hi be not relatively hyperbolic
with respect to any finite collection of proper subgroups. However, this is not
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
23
sufficient to conclude that the class of relatively hyperbolic groups is quasiisometrically rigid, since there exist relatively hyperbolic groups that have no
list of peripheral subgroups composed uniquely of non-relatively hyperbolic
groups (indeed, the inaccessible group constructed by Dunwoody [22] is also
an example of such a relatively hyperbolic group, see [3]; see the appendix for
a brief discussion of (in)accessibility of groups). Nevertheless, it eventually
turns out that the whole class of relatively hyperbolic groups is QI-rigid:
Theorem 3.9 ([18]). The Γ be a group hyperbolic relative to a family of
subgroups {H1 , . . . , Hk }, and let Γ′ be a group quasi-isometric to Γ. Then
Γ′ is hyperbolic relative to a family {H1′ , . . . , Hs′ } of subgroups, where each
Hi′ can be quasi-isometrically embedded in Hj for some j = j(i).
Before concluding the section, let us come back to what asymptotic cones
can and cannot distinguish. We have already seen that the asymptotic cones
of hyperbolic groups are all quasi-isometric to each other, so they cannot be
exploited to distinguish non-quasi-isometric hyperbolic groups. However,
asymptotic cones can tell apart the quasi-isometry classes of the eight 3dimensional geometries:
g2 ,
Proposition 3.10. With the exception of the case of H2 × R and SL
the eight 3-dimensional geometries can be distinguished by looking at their
asymptotic cones. In particular, with the exception of the case of H2 × R
g2 , distinct geometries are not quasi-isometric.
and SL
Proof. Every asymptotic cone of the sphere is a point. The Euclidean 3space and N il are the unique geometries such that each of their asymptotic
cones is unbounded and locally compact. Moreover, both these spaces admit
a nontrivial 1-parameter group of non-isometric similarities, so they are isometric to their asymptotic cones. However, it is not diffult to show that R3 is
not bi-Lipschitz equivalent to N il (in fact, they are not even quasi-isometric,
since they have different growth rates). Every asymptotic cone of S 2 × R,
of H3 and of H2 × R is isometric to the real line, to the homogeneous real
tree T with uncountably many branches at every point, and to the product
T × R, respectively. In particular, they are all simply connected and pairwise non-bi-Lipschitz equivalent. Finally, Sol is the unique 3-dimensional
geometry admitting non-simply connected asymptotic cones [9].
4. Trees of spaces and their asymptotic cones
Let now M be either an irreducible non-geometric 3-manifold, or an irreducible n-dimensional HDG manifold, n ≥ 3. For the sake of simplicity, in
the case when M is not a HDG manifold, henceforth we will assume that M
is closed. It follows by the very definitions that M decomposes into pieces
of the form N × T k , where N is a truncated hyperbolic manifold and T k
is a k-dimensional torus. Moreover, dim N ≥ 3 unless dim M = 3, while k
may be any non-negative integer. We put on M a Riemannian metric (in
the 3-dimensional case, we will carefully choose one later), and we denote by
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
24
X the Riemannian universal covering of M . The decomposition of M into
pieces lifts to a decomposition of X into chambers. Since the fundamental
group of each piece injects into π1 (M ), every chamber is the universal covering of a geometric piece of M . Chambers are adjacent along walls, where
a wall is the lift to X of a torus (or Klein bottle) of the decomposition of
M . This decomposition of X can be encoded by a graph T whose vertices
(resp. edges) correspond to chambers (resp. walls), and the edge corresponding to the wall W joins the vertices corresponding to the chambers adjacent
to W . Moreover, each chamber is foliated by fibers, which are just the lifts to
X of the fibers of the pieces of M (where fibers of a purely hyperbolic piece
are understood to be points). Since X is simply connected, the graph T is
a (simplicial, non-locally finite) tree. In fact, the decomposition of M into
pieces realizes π1 (M ) as the fundamental group of a graph of groups having
the fundamental groups of the pieces as vertex groups and the fundamental
pieces of the splitting tori (or Klein bottles) as edge groups. The tree T is
exactly the Bass-Serre tree associated to this graph of groups (see [87, 85] for
the definition and the topological interpretation of the fundamental group
of a graph of groups).
By Milnor-Svarc Lemma, Theorems 1.8 and 1.14 readily follow from the
following:
Theorem 4.1. Let M, M ′ be either irreducible non-geometric graph manifolds or irreducible HDG manifolds, and denote by X, X ′ the universal coverings of M, M ′ respectively. Let also f : X → X ′ be a quasi-isometry.
Then for every chamber C of X there exists a chamber C ′ of X ′ such that
f (C) lies within finite Hausdorff distance from C ′ (as a consequence, C is
quasi-isometric to C ′ ). Moreover, f preserves the structures of X, X ′ as
trees of spaces, i.e. it induces an isomorphism between the trees encoding
the decomposition of X, X ′ into chambers and walls.
We would like to study the coarse geometry of X starting from what we
know about the coarse geometry of its chambers. This strategy can be more
easily pursued provided that chambers are quasi-isometrically embedded in
X. We first observe that, if the manifold M carries a Riemannian metric of
non-positive curvature, then by the Flat Torus Theorem (see e.g. [8]) the decomposition of M into pieces can be realized geometrically by cutting along
totally geodesic embedded flat tori (and Klein bottles in the 3-dimensional
case). This readily implies that walls and chambers (endowed with their
intrinsic path metric) are isometrically embedded in X. Now it turns out
that non-geometric irreducible 3-manifolds “generically” admit metrics of
non-positive curvature: namely, this is true for every manifold containing at
least one hyperbolic piece [60], while it can fail for graph manifolds (Buyalo
and Svetlov [10] have a complete criterion for deciding whether or not a
3-dimensional graph manifold supports a non-positively curved Riemannian
metric). Nevertheless, it is proved in [55] that for every non-geometric irreducible 3-manifold M , there exists a non-positively curved non-geometric
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
25
irreducible 3-manifold M ′ such that the universal covers X of M and X ′
of M ′ are bi-Lipschitz homeomorphic by a homeomorphism which preserves
their structures as trees of spaces (in particular, walls and chambers are
quasi-isometrically embedded in X). In fact, thanks to this result it is not
restrictive (with respect to our purposes) to consider only non-positively
curved non-geometric irreducible 3-manifolds.
In the higher dimensional case it is not possible to require non-positive
curvature. Indeed, it is easy to construct HDG manifolds whose fundamental group cannot be quasi-isometric to the fundamental group of any
non-positively curved closed Riemannian manifold [36, Remark 2.23]. At
least from the coarse geometric point of view, things get a bit better when
restricting to irreducible manifolds:
Theorem 4.2 (Theorem 0.16 in [36]). Suppose that M is an irreducible
HDG manifold with universal covering X. Then walls and chambers are
quasi-isometrically embedded in X.
Nevertheless, it is possible to construct irreducible HDG manifolds which
do not support any non-positively curved metric [36, Theorem 0.20]. It is
still an open question whether the fundamental group of every irreducible
HDG manifold is quasi-isometric to the fundamental group of a non-positively
curved HDG manifold.
4.1. Irreducible 3-manifolds versus HDG manifolds. If C is any chamber of the tree of spaces X introduced above, then the walls of C are r-dense
in C. Using this fact, it is easy to realize that the key step in the proof of
Theorem 4.1 consists in showing that quasi-isometries must preserve walls
(and send walls that do not separate each other to walls that do not separate each other). In the case of irreducible HDG n-manifolds, using that
the codimension of the fibers is big enough one can show that walls are the
unique (n − 1)-dimensional quasi-isometrically embedded copies of Rn−1 in
X (see Theorem 5.1), and this provides the key step towards Theorem 4.1
in that case. On the contrary, chambers corresponding to Seifert pieces of
irreducible 3-manifolds contain a lot of 2-dimensional flats coming from the
lifts to X of the product of closed simple geodesics in the 2-dimensional base
with the S 1 -fiber. As a consequence, some more work is needed to provide
a quasi-isometric characterization of walls. On the positive side, in the 3dimensional case one can use the peculiar features of non-positive curvature,
that are not available in higher dimensions.
4.2. Asymptotic cones of trees of spaces. Let M be either a 3-dimensional
irreducible non-geometric 3-manifold or an irreducible HDG manifold. For
the sake of simplicity, in the first case we also assume that M is closed,
and endowed with a non-positively curved metric (see the previous section).
We have seen that the universal covering of M decomposes as a tree of
spaces into chambers separated by walls. Moreover, chambers and walls are
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
26
quasi-isometrically embedded in X (and even isometrically embedded when
dim M = 3).
Let ω be a fixed non-principal ultrafilter on N, let (xi ) ⊆ X, (ri ) ⊆
R be fixed sequences of basepoints and rescaling factors, and set Xω =
(Xω , (xi ), (ri )).
Definition 4.3. An ω-chamber (resp. ω-wall, ω-fiber) in Xω is a subset Yω ⊆
Xω of the form Yω = ω- lim Y i , where each Y i ⊆ X is a chamber (resp. a
wall, a fiber). An ω-wall Wω = ω- lim Wi is a boundary (resp. internal)
ω-wall if Wi is a boundary (resp. internal) wall ω-a.e.
The decomposition of a tree-graded space into its chambers induces a
decomposition of Xω into its ω-chambers. Indeed, since a constant k exists such that each point of X has distance at most k from some wall,
every point of Xω lies in some ω-wall. Recall that, in a tree-graded space,
subspaces homeomorphic to Euclidean spaces of dimension bigger than one
(which do not have cut-points) are contained in pieces. We would like to
prove that a similar phenomenon occurs in our context, and this is indeed
the case when dealing with HDG manifolds. In fact, in that case ω-walls
can be characterized as the only subspaces of Xω which are bi-Lipschitz
homeomorphic to Rn−1 , where n = dim M (see Proposition 4.11). As a
consequence, every bi-Lipschitz homeomorphism of Xω preserves the decomposition of Xω into ω-walls. Together with an argument which allows
us to recover quasi-isometries of the original spaces from bi-Lipschitz homeomorphisms of asymptotic cones, this will imply Theorem 4.1 in the case
of HDG manifolds. As anticipated above, extra work is needed in the case
of 3-manifolds, where the presence of Seifert pieces implies the existence of
2-dimensional flats in Xω that are not ω-walls.
We begin by collecting some facts that will prove useful for the proof of
Theorem 4.1. Henceforth we denote by n ≥ 3 the dimension of M (and
of its unviersal covering X). Being a quasi-isometric embedding, the inclusion of any chamber in X induces a bi-Lipschitz embedding of the asymptotic cone of the chamber into Xω . Therefore, from Proposition 3.6 we
deduce that for any ω-chamber Cω there exists a bi-Lipschitz homeomorphism ϕ : Cω → Y × Rl , where Y is a tree-graded space whose pieces are
bi-Lipschitz homeomorphic to Rn−l−1 , such that the following conditions
hold:
(1) For every p ∈ Y , the subset ϕ−1 ({p} × Rl ) is an ω-fiber of Xω .
(2) For every piece P of Y , the set ϕ−1 (P × Rl ) is an ω-wall of Xω .
In particular, every ω-wall of X is bi-Lipschitz homeomorphic to Rn−1 , and
every ω-fiber of X is bi-Lipschitz homeomorphic to Rh for some h ≤ n − 3.
A fiber of Cω is an ω-fiber of Xω of the form described in item (1) above.
A wall of Cω is an ω-wall of Xω of the form described in item (2) above. If
Wω is a wall of Cω , then we also say that Cω is adjacent to Wω .
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
27
Definition 4.4. Let Wω = ω- lim Wi be an ω-wall. A side S(Wω ) of Wω is
a subset S(Wω ) ⊆ Xω which is defined as follows. For every i, let Ωi be a
connected component of X \ Wi . Then
S(Wω ) = (ω- lim Ωi ) \ Wω .
An internal wall has exactly two sides S(Wω ), S ′ (Wω ) and is adjacent
to exactly two ω-chambers Cω , Cω′ . Up to reordering them, we also have
S(Wω ) ∩ Cω′ = S ′ (Wω ) ∩ Cω = ∅, Cω \ Wω ⊆ S(Wω ) and Cω′ \ Wω ⊆ S ′ (Wω ).
Moreover, any Lipschitz path joining points contained in distinct sides of
Wω must pass through Wω .
Definition 4.5. A subset A ⊆ Xω is essentially separated by an internal
ω-wall Wω ⊆ Xω if A intersects both sides of Wω .
Proposition 4.6. Let A ⊆ Xω be a subset which is not essentially separated
by any ω-wall. Then A is contained in an ω-chamber.
Proof. In the 3-dimensional case, this is Lemma 3.4 in Kapovich and Leeb’s
paper [54]. Their proof applies verbatim also to the case of irreducible HDG
manifolds (thus providing a positive answer to the problem posed at page
122 in [36]).
As already mentioned in the introduction, a key ingredient for the analysis
of the geometry of Xω (or of X) is the understanding of which subspaces
separate (or coarsely separate) some relevant subsets of Xω (or of X). Let
S(Wω ) be a side of Wω , and let Cω be the unique ω-chamber of Xω which
intersects S(Wω ) and is adjacent to Wω . A fiber of Wω associated to S(Wω )
is a fiber of Cω that is contained in Wω . The following observation follows
from the fact that the gluings defining our manifold M are transverse, and
it is crucial to our purposes:
Lemma 4.7. Let S + (Wω ) and S − (Wω ) be the sides of the internal ω-wall
Wω , and let Fω+ , Fω− be fibers of Wω associated respectively to S + (Wω ),
S − (Wω ). Then |Fω+ ∩ Fω− | ≤ 1.
If P, P ′ are distinct pieces of a tree-graded space Y , then there exist p ∈ P ,
∈ P ′ such that, for any continuous path γ : [0, 1] → Y with γ(0) ∈ P
and γ(1) ∈ P ′ , we have p, p′ ∈ Im γ (see e.g. [36, Lemma 8.8]). Since ωchambers are (bi-Lipschitz homeomorphic to) products of tree-graded spaces
with Euclidean factors, this immediately implies that, if Wω and Wω′ are
distinct ω-walls of the ω-chamber Cω , then there exists an ω-fiber Fω ⊆ Wω
of Cω such that every continuous path in Cω joining a point in Wω with
a point in Wω′ has to pass through Fω . With some work it is possible to
extend this result to pairs of ω-walls which are not contained in the same
ω-chamber:
p′
Lemma 4.8 (Lemma 8.24 in [36]). Let Wω , Wω′ be distinct ω-walls, and let
S(Wω ) be the side of Wω containing Wω′ \ Wω . Then there exists an ω-fiber
Fω of Wω such that
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
28
(1) Fω is associated to S(Wω ), and
(2) every Lipschitz path joining a point in Wω′ with a point in Wω passes
through Fω .
As every point in Xω is contained in an ω-wall, Lemma 4.8 implies the
following:
Corollary 4.9. Let Wω be an ω-wall, let p ∈ Xω \Wω , and let S(Wω ) be the
side of Wω containing p. Then there exists an ω-fiber Fω of Wω associated
to S(Wω ) such that every Lipschitz path joining p with Wω passes through
Fω .
4.3. A characterization of bi-Lipschitz flats in higher dimension.
Throughout this subsection we assume that n = dim M ≥ 4. As already
mentioned in Subsection 4.1, in this case the fact that fibers have higher
codimension allows us to provide an easy characterization of ω-walls.
A bi-Lipschitz m-flat in Xω is the image of a bi-Lipschitz embedding
f : Rm → Xω . This section is aimed at proving that ω-walls are the only biLipschitz (n−1)-flats in Xω . We say that a metric space is L.-p.-connected if
any two points in it may be joined by a Lipschitz path. The following lemma
provides a fundamental step towards the desired characterization of ω-walls,
so we give a complete proof of it. It breaks down in the 3-dimensional case.
Lemma 4.10. Let A ⊆ Xω be a bi-Lipschitz (n − 1)-flat. Then for every
ω-fiber Fω the set A \ Fω is L.-p.-connected.
Proof. Let f : Rn−1 → Cω be a bi-Lipschitz embedding such that f (Rn−1 ) =
A, and let l ≤ n−3 be such that Fω is bi-Lipschitz homeomorphic to Rl . The
set f −1 (Fω ) is a closed subset of Rn−1 which is bi-Lipschitz homeomorphic to
a subset of Rl . But it is known that the complements of two homeomorphic
closed subsets of Rn−1 have the same singular homology (see e.g. [15]), so
Rn−1 \ f −1 (Fω ) is path-connected. It is immediate to check that any two
points in a connected open subset of Rn−1 are joined by a piecewise linear
path, so Rn−1 \ f −1 (Fω ) is L.-p.-connected. The conclusion follows from the
fact that f takes Lipschitz paths into Lipschitz paths.
Let now A ⊆ Xω be a bi-Lipschitz (n − 1)-flat. We first observe that
A is not essentially separated by any ω-wall of Xω . In fact, if this were
not the case, then there would exist an ω-wall Wω and points p, q ∈ A on
opposite sides of Wω . Then the fiber Fω of Wω such that every path joining p
with Wω passes through Fω (see Corollary 4.9) would disconnect A, against
Lemma 4.10. By Proposition 4.6 we can then suppose that A is contained
in an ω-chamber Cω . Recall that Cω is homeomorphic to a product Y × Rl ,
where Y is a tree-graded space and l ≤ n − 3. Lemma 4.10 implies that the
projection of A onto Y does not have cut-points, so it is contained in a piece
of Y . This is equivalent to say that A is contained in an ω-wall of Cω , so it
is actually equal to such an ω-wall by invariance of domain. We have thus
proved the following:
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
29
Proposition 4.11. Let A be a bi-Lipschitz (n − 1)-flat in Xω . Then A is
an ω-wall.
Corollary 4.12. Let M, M ′ be irreducible HDG manifolds with universal
coverings X, X ′ , and let f : X → X ′ be a quasi-isometry. Then the map
fω : Xω → Xω′ induced by f on the asymptotic cones takes any ω-wall of Xω
onto an ω-wall of Xω′ .
4.4. The 3-dimensional case. We would like to extend Corollary 4.12 to
the 3-dimensional case. Let M, M ′ be closed irreducible non-geometric 3manifolds with universal coverings X, X ′ . We can suppose that M, M ′ are
non-positively curved. We will say that an ω-chamber of the asymptotic
cone Xω (or Xω′ ) is hyperbolic (resp. Seifert) if it is the ω-limit of chambers
covering a hyperbolic (resp. Seifert) piece of M (or M ′ ). Observe that every
ω-chamber is either Seifert or hyperbolic. Being bi-Lipschitz homeomorphic
to the product of a real tree with the real line, every Seifert ω-chamber
contains many flats that are not ω-walls, so Proposition 4.11 cannot hold
in this case. Therefore, in order to obtain Corollary 4.12 some additional
arguments are needed, that we briefly outline here. The reader is addressed
to [54] for complete proofs.
We first collect some facts about (bi-Lipschitz) flats in Xω .
(1) Any bi-Lipschitz flat contained in an ω-chamber is a flat (this easily
follows from the explicit description of the geometry of the chambers,
together with the basic properties of tree-graded spaces and the fact
that, thanks to non-positive curvature, ω-chambers are isometrically
embedded in Xω ).
(2) Any flat Zω ⊆ Xω must be contained in a single ω-chamber. In
fact, otherwise Zω is essentially separated by an ω-wall Wω . The
arguments described in the previous subsection may be exploited to
show that no ω-chamber adjacent to Wω can be hyperbolic, i.e. the
two ω-chambers adjacent to Wω must both be Seifert. Moreover,
Zω ∩ Wω must contain two transversal fibers of Wω , and this implies
that indeed Zω = Wω , a contradiction.
(3) A bi-Lipschitz flat B ⊆ Xω is an ω-wall which is not adjacent to
any Seifert ω-chamber if and only of its intersection with any other
bi-Lipschitz flat in Xω contains at most one point. In fact, if the
latter condition is true, then B cannot be essentially separated by
any ω-wall, so it is contained in an ω-chamber that cannot be Seifert
(because otherwise B would intersect many other flats in more than
one point). Conversely, if B is an ω-wall not adjacent to any Seifert
ω-chamber and B ′ is a bi-Lipschitz flat intersecting F , then the
closure of any component of B ′ \ B intersects B in one point, which
of course cannot disconnect B. Therefore B ′ \ B is connected and
B ′ ∩ B is one point.
(4) If T is a real tree which branches at every point, then the image Ω of
any bi-Lipschitz embedding f : T × R → Xω is contained in a Seifert
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
30
ω-chamber. In fact, if Wω is an ω-wall such that S ± (Wω ) ∩ Ω = Ω±
are both non-empty, then the boundaries l± of Ω± are transversal
fibers in Wω . However, their inverse image under f separate T ×
R, so they should be parallel lines in T × R, and this contradicts
the fact that f is bi-Lipschitz. The conclusion now follows from
Proposition 4.6.
These facts imply the following:
Proposition 4.13. Let f : Xω → Xω′ be a bi-Lipschitz homeomorphism.
Then:
(i) f maps flats to flats;
(ii) f maps each ω-wall that is not adjacent to a Seifert ω-chamber into
an ω-wall of the same kind;
(iii) f maps any Seifert ω-chamber of Xω into a Seifert ω-chamber of
Xω′ .
Proof. By fact (2) above, any flat in Xω is contained in an ω-chamber.
Thus assertion (i) follows from fact (3) for ω-walls adjacent to at least one
hyperbolic ω-chamber and from fact (4) for flats contained in Seifert ωchambers. Assertions (ii) and (iii) follow from facts (3) and (4), respectively.
We are now ready to prove the following result, which shows that also in
the 3-dimensional case ω-walls and ω-chambers are preserved by bi-Lipschitz
homeomorphisms:
Corollary 4.14. Let f : Xω → Xω′ be a bi-Lipschitz homeomorphism. Then:
(1) f maps any ω-chamber of Xω into an ω-chamber of the same type;
(2) f maps ω-walls into ω-walls.
Proof. Since every ω-wall may be expressed as the intersection of adjacent ωchambers, (2) readily follows from (1). In order to show (1), since we already
know that Seifert ω-chambers are sent to Seifert ω-chambers (see Proposition 4.13–(iii)), it is sufficient to prove that any hyperbolic ω-chamber
Cω ⊆ Xω is sent to an ω-chamber of Xω′ (that cannot be Seifert by Proposition 4.13–(iii) applied to f −1 ). Suppose by contradiction that f (Cω ) is
essentially separated by the ω-wall Wω′ ⊆ Xω′ . We know that Cω is the
union of its ω-walls, that every ω-wall lies in an ω-chamber, and that f takes
flats to flats (see Proposition 4.13–(i)), so there exist flats Z + , Z − ∈ f (Cω )
such that Z ± ∩ S ± (Wω′ ) 6= ∅. Now the bi-Lipschitz flats f −1 (Z ± ) lie in the
hyperbolic ω-chamber Cω , so they are flats, and are essentially separated by
the flat f −1 (Wω′ ) ⊆ Xω . This provides the desired contradiction, because
two flats in the same hyperbolic ω-chamber cannot be essentially separated
by any other flat.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
31
5. Proof of Theorem 4.1
Let now M, M ′ both be either closed irreducible non-geometric 3-manifolds,
or irreducible n-dimensional HDG manifolds, n ≥ 3. We denote by X, X ′
the Riemannian universal coverings of M, M ′ , respectively. In the case when
dim M = dim M ′ = 3, we can (and we do) suppose that X, X ′ are endowed
with a non-positively curved equivariant metric. We also fix a quasi-isometry
f : X → X ′ . We have already observed that Theorems 1.8 and 1.14 reduce
to Theorem 4.1, which asserts that:
(1) for every chamber C of X there exists a chamber C ′ of X ′ such that
f (C) lies within finite Hausdorff distance from C ′ ;
(2) f preserves the structures of X, X ′ as tree of spaces.
Our study of bi-Lipschitz homeomorphisms between the asymptotic cones
Xω and Xω′ was mainly aimed at showing that such homeomorphisms must
preserve ω-walls (see Corollaries 4.12 and 4.14). We first show how this
fact may be used to show that f must coarsely preserve walls. The proof
of the following result illustrates a very general strategy to get back from
homeomorphisms between asymptotic cones to quasi-isometries between the
original spaces.
Theorem 5.1. There exists β ≥ 0 such that, if W is a wall of X, then f (W )
lies at Hausdorff distance bounded by β from a wall W ′ ⊆ X ′ . Moreover, f
stays at distance bounded by β from a quasi-isometry between W and W ′ .
Proof. It is well-known that a quasi-isometric embedding between spaces
both quasi-isometric to Rn−1 is itself a quasi-isometry (for example, because
otherwise by taking asymptotic cones with suitably chosen sequences of
basepoints and of rescaling factors, one could construct a non-surjective
bi-Lipschitz embedding of Rn−1 into itself, which cannot exist). Therefore,
it is sufficient to show that there exists a wall W ′ ⊆ X ′ such that f (W ) ⊆
Nβ (W ′ ), where Nβ denotes the β-neighbourhood in X ′ .
Suppose by contradiction that for each m ∈ N and wall W ′ ⊆ X ′ we have
f (W ) * Nm (W ′ ). Fix a point p ∈ W . The quasi-isometry f induces a biLipschitz homeomorphism fω from the asymptotic cone Xω = Xω ((p), (m))
to the asymptotic cone Xω′ (f (p), (m)). By Corollaries 4.12 and 4.14, if Wω
is the ω-limit of the constant sequence of subsets all equal to W , then there
′ such that f (W ) = W ′ . By hypothesis, for
is an ω−wall Wω′ = ω-lim Wm
ω
ω
ω
′ ) ≥ m. Set r = d(p , p).
each m there is a point pm ∈ W with d(f (pm ), Wm
m
m
By choosing pm as close to p as possible, we may assume that no point q
′ ) ≥ m, so
such that d(p, q) ≤ rm − 1 satisfies d(f (q), Wm
(1)
′
d(f (q), Wm
)≤m+k+c
for every q ∈ W s.t. d(p, q) ≤ rm .
Notice that ω-lim rm /m = ∞, for otherwise [(pm )] should belong to Wω ,
[f (pm )] should belong to Xω′ ((f (p)), (m)), and, since fω (Wω ) = Wω′ , we
′ ) = o(m).
would have d(f (pm ), Wm
Let us now change basepoints, and consider instead the pair of asymptotic
cones W ω = Wω ((pm ), (m)) and Xω′ ((f (pm )), (m)). The quasi-isometry f
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
32
induces a bi-Lipschitz embedding f between these asymptotic cones (note
that f 6= f , simply because due to the change of basepoints, f and f are
defined on different spaces with values in different spaces!). Let Am = {q ∈
W | d(q, p) ≤ rm } and Aω = ω-lim Am ⊆ W ω . Since ω-lim rm /m = ∞,
it is easy to see that Aω is bi-Lipschitz homeomorphic to a half-space in
W ω . Moreover, by (1) each point in f (Aω ) is at a distance at most 1 from
′
′
W ω = ω-lim Wi′ ⊆ Xω′ ((f (pm )), (m)) (as before, observe that the sets W ω
and Wω′ live in different spaces). Again by Corollaries 4.12 and 4.14, we
have that f (Aω ) ⊆ f (W ω ) = Wω′′ for some ω−wall Wω′′ . Moreover, since
′
′
[(f (pm ))] ∈ Wω′′ \ W ω , we have W ω 6= Wω′′ .
′
By Lemma 4.8 there exists a fiber Fω ⊆ W ω ∩ Wω′′ such that every path
′
joining a point in Wω′′ with a point in W ω has to pass through Fω . Now,
′
if a ∈ f (Aω ) we have d(a, W ω ) ≤ 1, so there exists a geodesic of length at
′
most one joining a ∈ Wω′′ with some point in W ω . Such a geodesic must pass
through Fω , so every point of f (Aω ) must be at a distance at most 1 from Fω .
If h : f (Aω ) → Fω is such that d(b, h(b)) ≤ 1 for every b ∈ f (Aω ), then h is
a (1, 2)-quasi-isometric embedding. Therefore the map g = h ◦ f : Aω → Fω
is a quasi-isometric embedding. But this is not possible, since if n − 1 > l
there are no quasi-isometric embeddings from a half space in Rn−1 to Rl
(as, taking asymptotic cones, such an embedding would provide an injective
continuous function from an open set in Rn−1 to Rl ). This completes the
proof of the theorem.
If W ⊆ X is a wall, then we denote by f# (W ) the wall of X ′ which
lies at finite Hausdorff distance from f (W ) (this wall exists by the previous
theorem, and it is obviously unique because distinct walls in X ′ are at infinite
Hausdorff distance from each other). It is not difficult to show that, if
W1 , W2 , W3 are walls of X such that W3 does not separate W1 from W2 ,
then we can connect W1 and W2 by a curve supported outside Nr (W3 ),
where r > 0 is any chosen constant. If r is sufficiently large, then the image
of such a curve via f may be replaced by a continuous path that connects
f# (W1 ) and f# (W2 ) without intersecting f# (W3 ). This readily implies that
f maps walls that are adjacent to the same chamber (close to) walls that
are adjacent to the same chamber. Together with the fact that the walls
of a chamber are r-dense in the chamber itself for some r > 0, this implies
in turn that f must coarsely preserve chambers. Moreover, f must also
preserve the structures of X and of X ′ as trees of spaces. This concludes
the proof of Theorem 4.1, whence of Theorems 1.8 and 1.14.
6. Quasi-isometric rigidity
Theorem 4.1 allows us to understand quite well groups that are quasiisometric to fundamental groups of irreducible non-geometric 3-manifolds or
of irreducible HDG manifolds. We begin by studying groups quasi-isometric
to fundamental groups of the pieces. As usual, the case when Seifert pieces
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
33
are allowed requires more care, so we first focus on the higher dimensional
case.
6.1. QI-rigidity of higher dimensional pieces. Let N be a complete
finite-volume hyperbolic m-manifold, m ≥ 3, and let Γ be a finitely generated group quasi-isometric to π1 (N ) × Zd , d ≥ 0. Without loss of generality,
we may assume that the cusps of N are toric.
By definition, the group Γ is quasi-isometric to π1 (M ), where M = N ×
d
T is an (obviously irreducible) HDG manifold of dimension n = m + d
consisting of a single piece. The universal covering X of M is isometric to
the Riemannian product B × Rd , where B is a neutered space. The walls of
X coincide with the boundary components of X. As discussed in Section 2,
a quasi-isometry between Γ and π1 (M ) induces a geometric quasi-action of
Γ on X that will be fixed from now on. As usual, we will identify every
element γ ∈ Γ with the corresponding quasi-isometry of X defined by the
quasi-action.
We want to prove that every quasi-isometry γ : X → X, γ ∈ Γ can be
coarsely projected on B to obtain a quasi-isometry of B. Recall first that
Theorem 5.1 implies that every wall of X is taken by γ close to another wall.
Since every fiber of X may be expressed as the coarse intersection of two
walls, this readily implies that γ must coarsely preserve also fibers. This
fact can be exploited to define a quasi-action of Γ on B as follows: for every
γ ∈ Γ, we define a map ψ(γ) : B → B by setting ψ(γ)(b) = πB (γ((b, 0)))
for every b ∈ B, where πB : X ∼
= B × Rd → B is the natural projection. It
is not difficult that ψ is indeed a cobounded quasi-action. On the contrary,
the fact that the quasi-action ψ is proper is a bit more delicate.
We first observe that, from the way the action of Γ on B was defined, every
γ ∈ Γ coarsely permutes the components of ∂B. Recall that m = n − d is
the dimension of the neutered space B, and let G be the isometry group
of (B, dB ). Every element of G is the restriction to B of an isometry of
the whole hyperbolic space Rm containing B. We will denote by Comm(G)
the commensurator of G in Isom(Rm ), i.e. the group of those elements h ∈
Isom(Rm ) such that the intersection G ∩ (hGh−1 ) has finite index both in
G and in hGh−1 .
The following rigidity result is an important step in Schwartz’s proof of
QI-rigidity of non-uniform lattices in G:
Proposition 6.1 (Lemma 6.1 in [81]). For every γ ∈ Γ there exists a
unique isometry θ(γ) ∈ Isom(Rm ) whose restriction to B stays at finite
distance from ψ(γ). Moreover, for every γ ∈ Γ the isometry θ(γ) belongs to
Comm(G), and the resulting map θ : Γ → Comm(G) is a group homomorphism.
In order to conclude our study of Γ we now need to understand the structure of the kernel and of the image of θ. We set Λ = θ(Γ) < Isom(Rm ) the
image of the homomorphism θ, and we show that Λ is commensurable with
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
34
π1 (N ). It is a result of Margulis that a non-uniform lattice in Isom(Rm ) is
arithmetic if and only if it has infinite index in its commensurator (see [96]).
As a result, things would be quite a bit easier if N were assumed to be nonarithmetic. To deal with the general case, one needs to observe the following
facts:
(1) Since elements of Λ are uniformly close to quasi-isometries of B,
each of them must send every horosphere in ∂B into a horosphere
O′ which is parallel and uniformly close to a horosphere O′′ ⊆ B;
(2) using (1), one can slightly change the “heights” of the horospheres in
b which is left invariant
∂B in order to define a new neutered space B
by the action of Λ;
b is discrete, one then gets that Λ is
(3) since the isometry group of B
b the action of Λ on Hm has a finite
discrete; being cobounded on B,
covolume, so Λ is a non-uniform lattice;
b and B respectively, the groups Λ and
(4) being quasi-isometric to B
π1 (N ) are quasi-isometric to each other, so one may use again Schwartz’s
results to conclude that Λ is commensurable with π1 (N ).
The study of ker θ is easier. In fact, it is not difficult to show that the
quasi-action of ker θ on X may be slightly perturbed to define a geometric
quasi-action of ker θ on one fiber of X. As a consequence, ker θ is finitely
generated, quasi-isometric to Zd and quasi-isometrically embedded in Γ.
Since groups quasi-isometric to Zd are virtually isomorphic to Zd , we have
thus shown that Γ is isomorphic to the extension of a non-uniform lattice
commensurable with π1 (N ) by a group virtually isomorphic to Zd . Since
abelian undistorted normal subgroups are always virtually central (see [36,
Proposition 9.10]), this concludes the proof of Theorem 1.15.
6.2. QI-rigidity of 3-dimensional pieces. Something more can be said
in the 3-dimensional case. Namely, if Γ is quasi-isometric to the fundamental
group of a hyperbolic piece, then Schwartz’s results imply that Γ is a finite
extension of a non-uniform lattice in Isom(H3 ), i.e. Γ fits into a short exact
sequence
/Γ
/ Γ′
/1,
/F
1
where F is finite and Γ′ is the fundamental group of a finite-volume hyperbolic 3-orbifold with flat cusps. Moreover, (quasi-)stabilizers in Γ of boundary flats of X are sent to peripheral subgroups of Γ′ , and the subgroup F
can be characterized as the maximal finite normal subgroup of Γ′ , and is
also the unique maximal finite normal subgroup of every (quasi-)stabilizer
of boundary flats of X.
In the Seifert case we have X = B × R, where B may be chosen to
be the complement in H2 of an equivariant collection of disjoint open halfplanes (horoballs are replaced by half-planes because the base surface is now
geometrized as a surface with geodesic boundary rather than with cusps).
Schwart’s results are no longer available in dimension 2, but using the fact
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
35
that all the (quasi-)isometries involved (quasi)-preserve the boundary of B
it is possible to slighlty modify the strategy described above to show that Γ
fits into a short exact sequence
1
/K
/Γ
/ Γ′
/1,
where K has a unique maximal finite normal subgroup F , the group K/F
is isomorphic either to Z or to the infinite dihedral group, and Γ′ is the fundamental group of a compact hyperbolic 2-orbifold with geodesic boundary.
Again the (quasi-)stabilizer of any boundary component of X is sent to a
peripheral subgroup of π1 (O), and F is also the unique maximal normal
subgroup of the (quasi-)stabilizers of the boundary components of X.
6.3. Quasi-isometric rigidity: the final step. Let now M be either an
irreducible non-geometric 3-manifold or an irreducible HDG manifold with
universal covering X, and take a group Γ quasi-isometric to π1 (M ). By
Proposition 2.4 we have a geometric quasi-action of Γ on X. By Theorem 4.1,
this quasi-action induces an action by automorphisms on the simplicial tree
T which encodes the structure of X as a tree of spaces. Now a fundamental
result from Bass-Serre theory says that any group acting on a simplicial
tree without inversions is isomorphic to the fundamental group of a graph
of groups whose vertex groups coincide with (conjugates of) the stabilizers
of vertices, and edge groups coincide with (conjugates of) the stabilzers of
edges (recall that G acts on T without inversions if no element of G switches
the endpoints of an edge of T ).
Now the action of Γ on T might include some inversions, but it is easy to
construct a subgroup Γ0 of Γ of index at most two that acts on T without
inversions. Moreover, it readily follows from the construction that vertex
groups of Γ0 are quasi-isometric to stabilizers of chambers, while edge groups
are quasi-isometric to stabilizers of walls. This already concludes the proof
of quasi-isometric rigidity in the higher dimensional case (i.e. Theorem 1.16).
In the 3-dimensional case a stronger result holds, thanks to the extra
information we described above. In fact, after replacing Γ with Γ0 (which is
obviously virtually isomorphic to Γ), we can make use of the fact that the
unique maximal finite normal subgroups of all vertex and edge stabilizers
coincide, and therefore coincide with the kernel F of the action of Γ on T .
The vertex stabilizers for the action of Γ′ = Γ/F on T are fundamental
groups of hyperbolic or Seifert 3-orbifolds with boundary. One can then
glue these orbifolds together (according to the combinatorics described by
the graph T /Γ) to get a 3-dimensional orbifold with fundamental group Γ′ .
This 3-orbifold is finitely covered by a manifold [63], and this concludes the
proof of Theorem 1.9.
7. Open questions
As already stated in the introduction, in order to conclude the classification of 3-manifold groups up to quasi-isometry only the case of non-geometric
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
36
irreducible manifolds with at least one arithmetic hyperbolic piece still has
to be understood (see [5]). Therefore, in this section we pose some questions
about the quasi-isometric rigidity of higher dimensional piecewise geometric
manifolds. Many of the following problems are taken from [36, Chapter 12].
The following question in addressed in [35]:
Problem 7.1. In [67, 68] Nguyen Phan defined the class of cusp decomposable manifolds, each element of which decomposes into cusped locally
symmetric pieces. Is is true than any quasi-isometry between the fundamental groups of two cusp decomposable manifolds induces a quasi-isometry
between the fundamental groups of their pieces? To what extent does quasiisometric rigidity hold for fundamental groups of cusp decomposable manifolds?
Problem 7.2. In [61], Leeb and Scott defined a canonical decomposition
for non-positively curved closed Riemannian manifolds, which provides a
generalization to higher dimensions of the JSJ decomposition of irreducible
3-manifolds. Is it true than every quasi-isometry between the fundamental
groups of two closed non-positively curved Riemannian manifolds induces a
quasi-isometry between the fundamental groups of their pieces?
Specializing to the class of HDG manifold groups, Theorem 1.16 describes
a necessary condition for deciding whether the fundamental groups of two
irrfeducible HDG manifolds M1 and M2 are quasi-isometric to each other:
the two HDG manifolds Mi must essentially be built up from the same collection of pieces (up to commensurability), with the same patterns of gluings
(see [36, Theorem 10.7] for a precise statement). The only distinguishing
feature between M1 and M2 would then be in the actual gluing maps used
to attach pieces together. We are thus lead to the following questions:
Problem 7.3. Take pieces V1 and V2 each having exactly one boundary
component, and let M1 , M2 be a pair of irreducible HDG manifolds obtained
by gluing V1 with V2 . Must the fundamental groups of M1 and M2 be quasiisometric?
Problem 7.4. Is there a pair of irreducible HDG manifolds with quasiisometric fundamental groups, with the property that one of them supports a
locally CAT(0) metric, but the other one cannot support any locally CAT(0)
metric?
Problem 7.5. Is it true that the fundamental group of every irreducible
HDG manifold is quasi-isometric to the fundamental group of a non-positively
curved HDG manifold?
Problem 7.6. Is it true that the fundamental group of every irreducible
HDG manifold is semihyperbolic in the sense of Alonso and Bridson [1]?
A positive answer to Problem 7.3 would imply positive answers both to
Problem 7.4 and to Problem 7.5, and a positive answer to Problem 7.5 would
imply in turn a positive answer to Problem 7.6.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
37
Concerning Problem 7.4, Nicol’s thesis [69] exhibits in each dimension
≥ 4 infinitely many pairs of non-irreducible HDG manifolds with quasiisometric fundamental groups, with the property that each pair consists of
one manifold that supports a locally CAT(0) metric and one manifold that
cannot support any locally CAT(0) metric.
Notice that in the proof of Theorem 1.16 each vertex stabilizer is studied
separately. It might be possible to obtain additional information by studying the interaction between vertex stabilizers of adjacent vertices, just as
Kapovich and Leeb did in the 3-dimensional case [54]:
Problem 7.7. Is it possible to strenghten the conclusion of Theorem 1.16?
Appendix A. Quasi-isometric invariance of the prime
decomposition
One of the most influential results in geometric group theory is Stallings’
Theorem, which asserts that a group has more than one end if and only if
it splits non-trivially as a free product or an HNN-extension with amalgamation over a finite subgroup (recall that the number of ends of any proper
geodesic metric space is a quasi-isometry invariant (see e.g. [8, Proposition
8.29]), so it makes sense to speak of the number of ends of a group). Before
going into the details of the proofs of Theorems 1.4 and 1.5, let us recall
some terminology and the main results from [76]. We say that a graph of
groups is terminal if every edge group is finite and every vertex group of ∆
cannot be expressed as a non-trivial free product or HNN-extension amalgamated along finite subgroups. By Stallings’ Theorem, this is equivalent
to say that every vertex group has less than two ends. We also say that
a terminal graph of groups is a terminal decomposition of its fundamental
group. It is a striking fact that not all finitely generated groups admit a
terminal decomposition: in other words, there may be cases when one can
go on splitting a finitely generated group along finite subgroups an infinite
number of times [22]. Groups admitting a realization as the fundamental
group of a terminal graph of groups are called accessible. In the torsionfree case, free products amalgamated along finite subgroups are just free
products, so accessibility is guaranteed by Grushko Theorem, which asserts
that the minimal number of generators of a free product is the sum of the
minimal number of generators of the factors. Moreover, it is a deep result
of Dunwoody [21] that every finitely presented group is also accessible, so
fundamental groups of closed 3-manifolds are accessible (this may also be
easily deduced from the existence of a prime decomposition, together with
Grushko Theorem and the fact that aspherical manifolds have torsion free
fundamental groups).
For any group Γ, let us denote by e(Γ) the number of ends of Γ. Theorem 0.4 in [76] states the following:
Theorem A.1 ([76]). Let Γ be an accessible group with terminal decomposition G. A group Γ′ is quasi-isometric to Γ if and only if the following
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
38
holds: Γ′ is also accessible, e(Γ) = e(Γ′ ), and any terminal decomposition of
Γ′ has the same set of quasi-isometry types of one-ended factors as G.
Let us now apply this result in the context of 3-manifolds. We first prove
the following:
Lemma A.2. Let M be a closed 3-manifold with fundamental group Γ.
Then:
(1) M is irreducible with finite fundamental group if and only if e(Γ) = 0.
(2) M is irreducible with infinite fundamental group if and only if e(Γ) =
1.
(3) M ∈ {S 2 × S 1 , P3 (R)#P3 (R)} if and only if e(Γ) = 2.
(4) M is not prime and distinct from P3 (R)#P3 (R) if and only if e(G) =
∞.
Proof. Since being irreducible with finite fundamental group, being irreducible with infinite fundamental group, belonging to {S 2 ×S 1 , P3 (R)#P3 (R)},
and being not prime but distinct from P3 (R)#P3 (R) are mutually exclusive conditions, it is sufficient to prove the “if” implications. Let M =
M1 # . . . #Mk be the prime decomposition of M , and set Γi = π1 (Mi ) so
that Γ = Γ1 ∗ . . . ∗ Γk . By the Poincaré conjecture, we have Γi 6= {1} for
every i = 1, . . . , k.
If e(Γ) = 0, then Γ is finite, so k = 1 and Γ 6= π1 (S 2 × S 1 ). Therefore, M
has a finite fundamental group, and it is prime and distinct from S 2 × S 1 ,
whence irreducible.
Let now e(Γ) = 1, and suppose by contradiction that M is not irreducible.
Since e(π1 (S 2 × S 1 )) 6= 2 we have M 6= S 2 × S 1 , so M is not prime and
Γ splits as a non-trivial free product Γ = Γ1 ∗ . . . ∗ Γk , k ≥ 2. Now it is
well-known that such a product has two ends if k = 2 and Γ1 = Γ2 = Z2 and
infinitely may ends otherwise (see e.g. [8, Theorem 8.32]). This contradicts
the fact that e(Γ) = 1, and shows that M is irreducible (the fact that Γ in
infinite obviously follows from e(Γ) = 1).
Let now e(Γ) = 2. We can argue as above to deduce that either M is
prime, or k = 2 and Γ1 = Γ2 = Z2 . In the second case, by the Poincaré
conjecture both M1 and M2 are doubly covered by S 3 , and this implies that
M1 = M2 = P3 (R), since any fixed-point-free involution of the 3-sphere is
conjugated to the antipodal map [62]. Therefore, we may suppose that M is
prime, and since S 2 × S 1 is the only prime manifold which is not irreducible,
we are left to show that M is not irreducible. Suppose by contradiction that
M is irreducible. Since groups with two ends are virtually infinite cyclic
f of M with
(see e.g. [8, Theorem 8.32]), we can choose a finite covering M
f is still irreducible, and it
infinite cyclic fundamental group. The manifold M
is well-known that irreducible 3-manifolds with infinite fundamental groups
f) should
are aspherical. Therefore, the cohomological dimension of Z = π1 (M
be equal to three, a contradiction. This concludes the analysis of the case
e(Γ) = 2.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
39
Finally suppose that e(Γ) = ∞. Since π1 (P3 (R)#P3 (R)) has two ends,
we need to show that M is not prime. Since π1 (S 2 × S 1 ) has two ends
and finite groups have 0 ends, we may suppose by contradiction that M
is irreducible with infinite fundamental group. We have already observed
that this implies that M is aspherical, so Γ is torsion free. By Stallings’
Theorem, since e(Γ) > 1 the group Γ must split as an HNN-extension or
an amalgamated product over a finite subgroup. Being Γ torsion free, this
implies that Γ actually splits as a non-trivial free product, so by the Kneser
conjecture M cannot be prime.
Remark A.3. Thanks to the previous lemma, a prime manifold M is big
if and only if its fundamental group has one end.
Let now M be a closed 3-manifold with prime decomposition M = M1 # . . . #Mk .
The graph of groups G corresponding to this decomposition is not quite
terminal according to the definition above, because every summand homeomorphic to S 2 × S 1 gives rise to a vertex group which may be expressed
as the unique HNN-extension of the trivial group. However, if in G one
replaces every vertex labelled by Z with a vertex labelled by {1} and a loop
based at it, then the new graph of groups G0 still has π1 (M ) as fundamental
group, and is indeed terminal. We call G0 the terminal graph of groups
corresponding to the prime decomposition of M . By Remark A.3 one-ended
vertex groups of G0 are precisely the fundamental groups of the big prime
summands of M .
We are now ready to prove Theorem 1.4. Let M = M1 # . . . #Mk and
M ′ = M1′ # . . . #Mk′ ′ be the prime decompositions of M, M ′ respectively,
and let G0 , G0′ be the corresponding terminal graphs of groups. As usual,
we set Γ = π1 (M ), Γi = π1 (Mi ), Γ′ = π1 (M ′ ), Γ′i = π1 (Mi′ ). We first
show that the conditions on M, M ′ described in points (1), (2), (3), (4)
are sufficient to ensure that Γ is quasi-isometric to Γ′ . This is obvious
if M, M ′ are both prime with finite fundamental groups, or if M, M ′ are
irreducible with infinite quasi-isometric fundamental groups, or if M, M ′ ∈
{S 2 × S 1 , P3 (R)#P3 (R)}, so we may suppose that both M and M ′ are not
prime and distinct from P3 (R)#P3 (R). In this case, Lemma A.2 ensures
that e(Γ) = e(Γ′ ) = ∞. Moreover, big summands in the decompositions of
M, M ′ are exactly the one-ended vertex groups respectively of G0 , G0′ , so the
groups Γ and Γ′ are quasi-isometric by Theorem A.1.
Let us now suppose that Γ is quasi-isometric to Γ′ . Of course we have
e(Γ) = e(Γ′ ). By Lemma A.2 if e(Γ) = e(Γ′ ) < ∞ we are done, so we may
suppose that both M and M ′ are not prime and distinct from P3 (R)#P3 (R).
The fact that the set of quasi-isometry types of fundamental groups of big
summands of M coincides with the set of quasi-isometry types of fundamental groups of big summands of M ′ is now a consequence of Theorem A.1 and
Remark A.3. This concludes the proof of Theorem 1.4.
The proof of Theorem 1.5 is very similar. Let M be a non-prime manifold
which is distinct from P3 (R)#P3 (R). By Lemma A.2 we have e(π1 (M )) =
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
40
∞, and we know that one-ended vertex groups of the terminal decomposition of π1 (M ) induced by the prime decomposition of M correspond to
fundamental groups of big summands of M . Therefore, Theorem A.1 implies that a group Γ is quasi-isometric to π1 (M ) if and only if e(Γ) = ∞ and
the set of quasi-isometry classes of one-ended vertex groups in a terminal decomposition of Γ is equal to the set of quasi-isometry classes of fundamental
groups of big summands of M . This finishes the proof of Theorem 1.5.
References
[1] J. M. Alonso, M. R. Bridson, Semihyperbolic groups Proc. Lond. Math. Soc. 70
(1995), 56-114.
[2] C. S. Aravinda, F. T. Farrell, Twisted doubles and non-positive curvature, Bull. Lond.
Math. Soc. 41 (2009), 1053–1059.
[3] J. Behrstock, C. Druţu, L. Mosher, Thick metric spaces, relative hyperbolicity, and
quasi-isometric rigidity, Math. Ann. 344 (2009), no. 3, 543–595.
[4] J. A. Behrstock, W. D. Neumann, Quasi-isometric classification of graph manifold
groups, Duke Math. J. 141 (2008), no. 2, 217–240.
[5] J. A. Behrstock, W. D. Neumann, Quasi-isometric classification of non-geometric
3-manifold groups, J. Reine Angew. Math. 669 (2012), 101–120.
[6] B. Bowditch, Cut points and canonical splittings of hyperbolic groups, Acta Math.
180 (1998), no. 2, 145–186.
[7] M. R. Bridson, The geometry of the word problem, Invitations to geometry and topology, 29–91, Oxf. Grad. Texts Math., 7, Oxford Univ. Press, Oxford, 2002.
[8] M. R. Bridson, A. Haefliger, Metric spaces of non-positive curvature, Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
319. Springer-Verlag, Berlin, 1999. xxii+643 pp.
[9] J. Burillo, Dimension and fundamental groups of asymptotic cones, J. London Math.
Soc. (2) 59 (1999), no. 2, 557–572.
[10] S. V. Buyalo, P. Svetlov, Topological and geometric properties of graph-manifolds, St.
Petersburg Math. J. 16 (2005), 297-340.
[11] J. W. Cannon, D. Cooper, A characterization of cocompact hyperbolic and finitevolume hyperbolic groups in dimension three, Trans. Amer. Math. Soc. 330 (1992),
no. 1, 419–431.
[12] A. Casson, D. Jungreis, Convergence groups and Seifert fibered 3-manifolds, Invent.
Math. 118 (1994), no. 3, 441–456.
[13] R. Chow, Groups quasi-isometric to complex hyperbolic space, Trans. Amer. Math.
Soc. 348 (1996), no. 5, 1757–1769.
[14] G. R. Conner, C. Kent, Local topological properties of asymptotic cones of groups,
Algebr. Geom. Topol. 14 (2014), no. 3, 1413–1439.
[15] A. Dold, A simple proof of the Jordan-Alexander complement theorem, Amer. Math.
Monthly 100 (1993), 856-857.
[16] C. Druţu, Quasi-isometry invariants and asymptotic cones, Int. J. Alg. Comp. 12
(2002), 99-135.
[17] C. Druţu, Cônes asymptotiques et invariants de quasi-isométrie pour des espaces
métriques hyperboliques, Ann. Inst. Fourier 51 (2001), 81–97.
[18] C. Druţu, Relatively hyperbolic groups: geometry and quasi-isometric invariance,
Comment. Math. Helv. 84 (2009), no. 3, 503–546.
[19] C. Druţu, M. Kapovich, Lectures on geometric group theory, available at
https://www.math.ucdavis.edu/ kapovich/EPR/kapovich drutu.pdf
[20] C. Druţu, M. Sapir, Tree-graded spaces and asymptotic cones of groups. With an
appendix by D. Osin and M. Sapir, Topology 44 (2005), 959-1058.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
41
[21] M. J. Dunwoody, The accessibility of finitely presented groups, Invent. Math. 81
(1985), 449–457.
[22] M. J. Dunwoody, An inaccessible group, Geometric group theory, Vol. 1 (Sussex,
1991) (G. A. Niblo and M. A. Roller, eds.), London Math. Soc. Lecture Note Ser.,
vol. 181, Cambridge Univ. Press, Cambridge, 1993, pp. 75–78.
[23] M. J. Dunwoody, M. Sageev, JSJ-splittings for finitely presented groups over slender
groups, Invent. Math. 135 (1999), no. 1, 25–44.
[24] A. Dyubina, Instability of the virtual solvability and the property of being virtually
torsion-free for quasi-isometric groups, Internat. Math. Res. Notices 21 (2000), 1097–
1101.
[25] A. Dyubina, I. Polterovich, Explicit constructions of universal R-trees and asymptotic
geometry of hyperbolic spaces, Bull. London Math. Soc. 33 (2001), no. 6, 727–734.
[26] A. Erschler, D. Osin, Fundamental groups of asymptotic cones, Topology 44 (2005),
827–843.
[27] A. Eskin, Quasi-isometric rigidity of nonuniform lattices in higher rank symmetric
spaces, J. Amer. Math. Soc. 11 (1998), no. 2, 321–361.
[28] A. Eskin, B. Farb, Quasi-flats and rigidity in higher rank symmetric spaces, J. Amer.
Math. Soc. 10 (1997), no. 3, 653–692.
[29] A. Eskin, D. Fisher, K. Whyte, Quasi-isometries and rigidity of solvable groups, Pure
Appl. Math. Q. 3 (2007), no. 4, part 1, 927–947.
[30] B. Farb, R. E. Schwartz, The large-scale geometry of Hilbert modular groups, J. Differential Geom. 44 (1996), no. 3, 435–478.
[31] B. Farb, The quasi-isometry classification of lattices in semisimple Lie groups, Math.
Res. Lett. 4 (1997), no. 5, 705–717.
[32] B. Farb, Relatively hyperbolic groups, Geom. Funct. Anal. 8 (1998), 810-840.
[33] M. Forester, On uniqueness of JSJ decompositions of finitely generated groups, Comment. Math. Helv. 78 (2003), no. 4, 740–751.
[34] R. Frigerio, Commensurability of hyperbolic manifolds with geodesic boundary, Geom.
Dedicata 118 (2006), 105–131.
[35] R. Frigerio, K. Kuzmin, Quasi-isometric rigidity of cusp decomposable manifolds, in
preparation.
[36] R. Frigerio, J. F. Lafont, A. Sisto, Rigidity of higher dimensional graph manifolds,
Astérisque No. 372 (2015), xxi+177 pp.
[37] K. Fujiwara, P. Papasoglu, JSJ-decompositions of finitely presented groups and complexes of groups, Geom. Funct. Anal. 16 (2006), no. 1, 70–125.
[38] D. Gabai, Convergence groups are Fuchsian groups, Ann. of Math. (2) 136 (1992),
no. 3, 447–510.
[39] S. M. Gersten, Divergence in 3-manifold groups, Geom. Funct. Anal. 4 (1994), no. 6,
633–647.
[40] M. Gromov, Groups of polynomial growth and expanding maps, Inst. Hautes Études
Sci. Publ. Math. 53 (1981), 53–73.
[41] M. Gromov, Infinite groups as geometric objects, Proceedings of the International
Congress of Mathematicians, Vol. 1, 2 (Warsaw, 1983), 385–392, PWN, Warsaw,
1984.
[42] M. Gromov, Hyperbolic groups, Essays in group theory, 75–263, Math. Sci. Res. Inst.
Publ., 8, Springer, New York, 1987.
[43] M. Gromov, Asymptotic invariants of infinite groups, Geometric group theory, Vol. 2
(Sussex, 1991), 1–295, London Math. Soc. Lecture Note Ser., 182, Cambridge Univ.
Press, Cambridge, 1993.
[44] V. Guirardel, G. Levitt, JSJ decompositions: definitions, existence, uniqueness. II.
Compatibility and acylindricity, arXiv:1002.4564.
[45] Y. Guivarc’h, Croissance polynomiale et pèriodes des fonctions harmoniques, Bull.
Soc. Math. France 101 (1973), 333–379.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
42
[46] J. Hempel, 3-Manifolds, Ann. of Math. Studies, No. 86. Princeton University Press,
Princeton, N. J.; University of Tokyo Press, Tokyo, 1976. xii+195 pp.
[47] E. Hrushovski, Stable group theory and approximate subgroups, J. Amer. Math. Soc.
25 (2012), no. 1, 189–243.
[48] W. Jaco, P. Shalen, A new decomposition theorem for irreducible sufficiently-large
3-manifolds, Algebraic and geometric topology (Proc. Sympos. Pure Math., Stanford
Univ., Stanford, Calif., 1976), Part 2, pp. 71–84, Proc. Sympos. Pure Math., XXXII,
Amer. Math. Soc., Providence, R.I., 1978.
[49] W. Jaco, P. Shalen, Seifert fibered spaces in 3-manifolds, Geometric topology (Proc.
Georgia Topology Conf., Athens, Ga., 1977), pp. 91–99, Academic Press, New YorkLondon, 1979.
[50] W. Jaco, P. Shalen, Seifert fibered spaces in 3-manifolds, Mem. Amer. Math. Soc. 21
(1979), no. 220, viii+192 pp.
[51] M.
Kapovich,
Lectures
on
quasi-isometric
rigidity,
available
at
https://www.math.ucdavis.edu/ kapovich/EPR/pc lectures3.pdf
[52] M. Kapovich, B. Leeb, On asymptotic cones and quasi-isometry classes of fundamental groups of 3-manifolds, Geom. Funct. Anal. 5 (1995), no. 3, 582-603.
[53] M. Kapovich, B. Leeb, Actions of discrete groups on non-positively curved spaces,
Math. Ann. 306 (1996), 341-352.
[54] M. Kapovich, B. Leeb, Quasi-isometries preserve the geometric decomposition of
Haken manifolds, Invent. Math. 128 (1997), 393-416.
[55] M. Kapovich, B. Leeb, 3-manifold groups and non-positive curvature, Geom. Funct.
Anal 8 (1998), 841-852.
[56] C. Kent, Asymptotic cones of HNN extensions and amalgamated products, Algebr.
Geom. Topol. 14 (2014), no. 1, 551–595.
[57] B. Kleiner, B. Leeb, Rigidity of quasi-isometries for symmetric spaces and Euclidean
buildings, Inst. Hautes Études Sci. Publ. Math. 86 (1997), 115-197.
[58] B. Kleiner, B. Leeb, Groups quasi-isometric to symmetric spaces, Comm. Anal. Geom.
9 (2001), 239-260.
[59] K. Johannson, Homotopy equivalences of 3-manifolds with boundaries, Lecture Notes
in Mathematics, 761. Springer, Berlin, 1979. ii+303 pp.
[60] B. Leeb, 3-manifolds with(out) metrics of non-positive curvature, Invent. Math. 122
(1995), 277-289.
[61] B. Leeb, P. Scott, A geometric characteristic splitting in all dimensions, Comment.
Math. Helv. 75 (2000), no. 2, 201–215.
[62] G. R. Livesay, Fixed point free involutions on the 3-sphere, Ann. of Math. (2) 72
(1960), 603–611.
[63] D. McCullough, A. Miller, Manifold covers of 3-orbifolds with geometric pieces, Topol.
Appl. 31 (1989), 169–185.
[64] J. Milnor, A unique factorization theorem for 3-manifolds, Amer. J. Math. 79 (1962),
1–7.
[65] L. Mosher, M. Sageev, K. Whyte, Quasi-actions on trees I. Bounded Valence, Ann.
of Math. (2) 158 (2003), no. 1, 115-164.
[66] L. Mosher, M. Sageev, K. Whyte, Quasi-actions on trees II: Finite depth Bass-Serre
trees, Mem. Amer. Math. Soc. 214 (2011), no. 1008.
[67] T. Nguyen Phan, Smooth (non)rigidity of cusp-decomposable manifolds, Comment.
Math. Helv. 87 (2012), no. 4, 789–804.
[68] T. Nguyen Phan, Smooth (non)rigidity of piecewise cusped locally symmetric manifolds, J. Topol. 5 (2012), no. 2, 485–502.
[69] A. Nicol, Quasi-isometries of graph manifolds do not preserve non-positive curvature,
Ph. D. Thesis, The Ohio State University, July 2014.
[70] A. Y. Ol’shanskii, M. Sapir, A finitely presented group with two non-homeomorphic
asymptotic cones, Internat. J. Algebra Comput. 17 (2007), no. 2, 421–426.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
43
[71] A. Y. Ol’shanskii, M. Sapir, Lacunary hyperbolic groups, Geom. Topol. 13 (2009),
2051–2140. With an appendix by Michael Kapovich and Bruce Kleiner.
[72] P. Pansu, Croissance des boules et des géodésiques fermées dan les nilvariétés, Ergodic
Th. Dyn. Sys. 3 (1983), 415–455.
[73] P. Pansu, Métriques de Carnot-Carathéodory et quasiisométries des espaces
symétriques de rang un, Ann. of Math. (2) 129 (1989), no. 1, 1–60.
[74] P. Papasoglu, Group splittings and asymptotic topology, J. Reine Angew. Math. 602
(2007), 1-16.
[75] P. Papasoglu, Quasi-isometry invariance of group splittings, Ann. of Math. (2) 161
(2005), no. 2, 759–830.
[76] P. Papasoglu, K. Whyte, Quasi-isometries between groups with infinitely many ends
Comment. Math. Helv. 77 (2002), no. 1, 133–144.
[77] M. S. Raghunathan, Discrete subgroups of Lie groups, Springer-Verlag, New York,
1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 68.
[78] E. G. Rieffel, Groups quasi-isometric to H2 × R, J. London Math. Soc. (2) 64 (2001),
no. 1, 44-60.
[79] E. Rips, Z. Sela, Cyclic splittings of finitely presented groups and the canonical JSJ
decomposition, Ann. of Math. (2) 146 (1997), no. 1, 53–109.
[80] M. Sapir, On groups with locally compact asymptotic cones, Internat. J. Algebra Comput. 25 (2015), no. 1-2, 37–40.
[81] R. E. Schwartz, The quasi-isometry classification of rank-one lattices, Inst. Hautes
Études Sci. Publ. Math. 82 (1995), 133-168.
[82] R. E. Schwartz, Quasi-isometric rigidity and Diophantine approximation, Acta Math.
177 (1996), no. 1, 75–112.
[83] P. Scott, The geometries of 3-manifolds, Bull. London Math. Soc. 15 (1983), 401–487.
[84] P. Scott, G. Swarup, Regular neighbourhoods and canonical decompositions for groups,
Astérisque No. 289 (2003), vi+233 pp.
[85] P. Scott, T. Wall Topological methods in group theory, Homological group theory
(Proc. Sympos., Durham, 1977), pp. 137–203, London Math. Soc. Lecture Note Ser.
36, Cambridge Univ. Press, Cambridge-New York, 1979.
[86] Z. Sela, Structure and rigidity in (Gromov) hyperbolic groups and discrete groups in
rank 1 Lie groups. II. Geom. Funct. Anal. 7 (1997), no. 3, 561–593.
[87] J. P. Serre, Trees. Springer monographs in Mathematics, Springer-Verlag, Berlin and
New York, 1980.
[88] A. Sisto, Separable and tree-like asymptotic cones of groups, Münster J. Math. 5
(2012), 233–248.
[89] J. R. Stallings, On torsion-free groups with infinitely many ends, Ann. of Math. (2)
88 (1968), 312–334.
[90] D. Sullivan, On the ergodic theory at infinity of an arbitrary discrete group of hyperbolic motions, Riemann surfaces and related topics: Proceedings of the 1978 Stony
Brook Conference (State Univ. New York, Stony Brook, N.Y., 1978), 465–496, Ann.
of Math. Stud. 97, Princeton Univ. Press, Princeton, N.J., 1981.
[91] S. Thomas, B. Velickovic, Asymptotic cones of finitely generated groups, Bull. London
Math. Soc. 32 (2000), no. 2, 203–208.
[92] W. P. Thurston,Three-dimensional geometry and topology. Vol. 1, edited by Silvio
Levy. Princeton Mathematical Series, 35. Princeton University Press, Princeton, NJ,
1997. x+311 pp.
[93] P. Tukia, On quasiconformal groups, J. Analyse Math. 46 (1986), 318–346.
[94] P. Tukia, Homeomorphic conjugates of Fuchsian groups, J. Reine Angew. Math. 391
(1988), 1–54.
[95] L. van den Dries, J. Wilkie, Gromov’s theorem on groups of polynomial growth and
elementary logic, J. Algebra 89 (1984), 349-374.
[96] R. Zimmer, Ergodic theory and semi-simple Lie groups. Birkhauser, Boston, 1984.
QI-RIGIDITY OF PIECEWISE GEOMETRIC MANIFOLDS
44
Dipartimento di Matematica, Università di Pisa, Largo B. Pontecorvo 5,
56127 Pisa, Italy
E-mail address: [email protected]
| 4math.GR
|
Electrical energy prediction study case based on
neural networks
Cristian Vasar, Octavian Prostean, Ioan Filip, Iosif Szeidert
Department of Control System Engineering,
“Politehnica” University of Timisoara, Faculty of Automation and Computers
1900 Timisoara, Bv. V. Parvan, No.2, ROMANIA, [email protected]
Abstract: This paper presents some considerations regarding the prediction of the
electrical energy consumption. It is well known that the central element of a microeconomic
analysis is represented by the economical agents actions, actions that follow their own
interest such as: the consumer – maximization of his satisfaction, the producer –
maximization of his profit. The study case is focused on the prediction of the sold energy in
Banat region. The goal of this study case is to optimize the electrical energy quantity
purchased from the producer by the energy distributor in Banat region. The prediction is
based on neural networks. There are used feed-forward and Elman type neural networks.
In order to enhance the prediction accuracy there have been used both linear and
nonlinear preprocessing units. The aspects considered in this paper can be extrapolated in
any general case of prediction-based application, not only in the already stated case of
electrical energy.
Keywords: prediction, neural network, electrical energy
1
Introduction
The main goal of this paper is to obtain an accurate sold electrical energy
prognosis by the main distributor in Banat region. This study is required for an
optimization of the purchase costs and implicit of the profit.
In a general case, through production costs is understood the monetary expression
of all expenses supported by an economical agent in order to produce, deliver and
sale its products or services.
A classification of the costs can be the following:
- transaction costs – represents by expenses that assure the market’s functionality;
- information costs – represented by transaction’s variables removal (assurance
fees, publicity, etc);
- book-keeping costs – defines as the monetary expression of resources used in the
production and sale process of a certain quantity of economic goods.
In an economy with limited resources, the problem of alternatives use has a major
importance so that the cost is only an alternative measure, an indicator that allows
choosing the optimal solution of resource usage.
The production costs can be predicted, traced and analyzed through many
criterions, but certainly the most used is one that refers to the analyzed time
period. So, there can be distinguished short-term costs and long-term costs.
However, there must be noticed that, in economical context, the short/long term is
not related to a certain time period. The short/long term is not measured in months
or years. Establishing the difference between the mentioned periods requires an
effective economical analysis in concordance with a precise situation.
The study case is focused on electrical energy consumption prediction. There is
known the tact that if the energetic system operates near the limit of his capacity,
the any additional energy request leads to an increase of the undelivered energy,
respectively to the increase of the costs due to undelivered energy.
The electric energy request shows significant fluctuations in concordance to
seasons and even with the daytime and this is the main reason, which leads to the
idea that the additional requests should by satisfied by different production units.
Therefore, results the necessity of using a scientific method that allows an easy
adaptation on current energetic power demand. [1]
On the basis, of the analysis of differences recorded between the purchased and
sold energy by the Banat region’s distributor, the energy losses can be predicted.
Figure 1. The evolution of the sold energy in Banat area
In Figure 1 the sold energy evolution can be noticed for the Banat region in the
last two years. The time (measured in month from Jan 2002 to Oct 2003) is
represented on the abscise axis, and the sold energy (measured in MWh)
corresponds to the ordinate.
Among the variety of prediction methods, neural networks have been widely used.
They can approximate any continuous multivariate function to a desired degree of
accuracy, provided that there are a sufficient number of hidden neurons [2].[3]
In this paper there are compared the results of the prediction based on a classic
neural network versus a special designed neural network which presents a
nonlinear preprocessing unit.
2
Prediction structures
There were used two kind of neural networks: feed-forward neural network and
Elman network. They were trained by using the gradient descent with momentum
backpropagation. The best network structure had two nonlinear layers with
hyperbolic tangent sigmoid transfer function and one output linear layer with
linear transfer function.
The training method involves that each variable is adjusted according to gradient
descent with momentum rule, [4]
dX = mc ∗ dX prev + lr ∗ (1 − mc) ∗
where
dperf
dX
(1)
dXprev is the previous change to the weight or bias.
lr is the learning rate
mc is the momentum constant.
Plant
Inputs
Outputs
+
z-1
z-1
Neural
Network
– Error
z-1
z-1
Figure 2. Series-parallel prediction structure
The prediction structure used for training the neural network is depicted in Figure
2. In the series-parallel prediction structure, the neural network receive as inputs
the past outputs of the real system [5]
To improve the accuracy of prediction the network inputs pass through a linear
preprocessing block which limits the variation of the neural network inputs in the
range of [0,1]. This is achieved by a simple mathematical multiplying operation
performed over the measured inputs. Also the network output was forced to be in
the same range, so a postprocessing unit is required.
summer(k)
k=1,2,…n
Fuzzy
preprocessor
Y(k-1)
Y(k-2)
Y(k-3)
Y(k-4)
winter(k)
Y (k-1)
Linear
preprocessor
Neural
Network
Y (k)
Y(k)
Linear
postprocessor
Y (k-2)
Y (k-3)
Y (k-4)
Figure 4. Neural Network with pre/post-processing units
In order to take into consideration of other factors, such as outdoor temperatures,
daytime, vacations periods etc, there has to be considered a supplementary set of
inputs to the network. All these perturbations are correlated with the four seasons.
To minimize the number of network inputs (and consequently the neural network
complexity), this supplementary set was reduced to only 2 supplementary inputs,
referred generically as “summer” and “winter”. This was accomplished by using a
fuzzy preprocessor unit. The obtained neural structure is depicted in Figure 4.
Figure 5. Preprocessor output functions
The input of the fuzzy preprocessor is the prediction month, and the outputs of this
represent the membership degree of the current month to the two “seasons”, which
are referred “winter” and “summer”. The preprocessor characteristics are depicted
in Figure 5. The dotted line represents “winter” and the solid line is for “summer”.
3
Simulation studies
There have been created two neural network categories. One consisted of different
sized feedforward networks and the other consisted of different sized Elman
recurrent networks. The results showed the same trend of prediction error
evolution for the entire category, and also proved, in this particular case, that using
large networks does not increase the accuracy of prediction.
Figure 6. Predicted Sold energy evolution (without “seasons”)
The first presented case is the one of the prediction performed by using the seriesparallel prediction structure depicted in Figure 2, and using a feedforward network
without fuzzy-preprocessing is presented in Figure 6. The first 14 months were
used as inputs for the training process, and the last 4 months were predicted. The
predicted values correspond to the time interval between May 2002 and October
2003. The real measured values for the sold energy are plotted with dotted line
while the predicted valued are depicted with solid line.
As it can be noticed, the outputs presented in Figure 6 and 7 are the scaled ones,
and the real predicted sold energy can be obtained by applying this values to the
linear postprocessor (which means multiplying by 5*105). The time (measured in
month from May 2002 to Oct 2003) is represented on the abscise axis.
In the second case, the same situation was simulated, but the fuzzy preprocessing
unit was added to the neural network as depicted in Figure4. The accuracy of
prediction increased considerably as it can be noticed in Figure 7. The time
(measured in month from May 2002 to Oct 2003) is represented on the abscise
axis, and on the ordinate there is represented the scaled sold energy.
Figure 7. Predicted Sold energy evolution (with “seasons”)
In order to obtain predictions for more than 1 step ahead, the trained network,
whose behavior was presented in Figure 7, was included into a multi-step
prediction structure as the one depicted in Figure 8.
In this situation, the predicted values for the sold energy are obtained by using
already predicted values. [6] The predicted results were more than the
expectations.
z-1
z-1
z
-1
Neural
Network
ouput
z-1
z-1
Figure 8. Multi-step prediction structure
The initial inputs for the neural network were the first four measured sold energy
values, corresponding to the time interval (Jan-Apr 2002) and the “summer” and
“winter” input corresponding for May 2002. After this point the network did not
received any measured sold energy values, and the prediction relies only on
already predicted values.
Figure 9. Multi-step simulation for sold energy
The simulation result is depicted in Figure 9. The time (measured in month from
May 2002 to Jul 2004) is represented on the abscise axis, and on the ordinate there
is represented the scaled sold energy. The first 14 predicted values are closed to
the real ones (plotted with dotted line), and this was expectable since they were
the source of training, but the network was able to predict further, beyond July
2003.
Figure 10. Multi-step simulation for purchased energy
The same experiments were conducted over purchased energy, and the results
were quite similar. The purchased energy is depicted in Figure 10. The first 14
months were use as inputs in the training process. The multi-step predicted
purchased energy is depicted in Figure 10. The time (measured in month from
May 2002 to Jul 2004) is represented on the abscise axis, and on the ordinate there
is represented the scaled purchased energy.
Conclusions
From the simulations presented in this paper it can be noticed the prediction
performance improvement by using the presented nonlinear preprocessor. This
unit can very easily implement specific situations as “early/late winter”,
“early/late summer”, by modifying the characteristics presented in Figure 5, in
order to fine-tune the prediction (even without network retraining).
In order to persuade this research, the authors have prompted several simulations
on selection of proper neural network structures, parameters and optimized length
of necessary data.
References
[1]
I. Szeidert, G. Prostean, C. Vasar: Some Considerations about Electrical
Energy Production Costs, Proceedings of Sesiunea de comunicari stiintifice ale
Universitatii Petru Maior, Targu Mures, Oct 2000. Vol.7, ISBN 973-8084-19-9
[2]
L. Zhang, P. B. Luh, and K. Kasiviswanathan: Energy Clearing Price
Prediction and Confidence Interval Estimation with Cascaded Neural Networks,
IEEE Transactions on Power Systems, Vol. 18, No. 1, February 2003, pp. 99-105.
[3]
Mituharu Hayashi, Bahman Kermanshahi: Application of Artificial
Nerural Network for Wind Speed Prediction and Determination of Wind Power
Generation Output, ICEE2001 in Xian JUL22-26.2001
[4]
***: Neural Network Toolbox, User’s Guide Version 4, September 2000.
[5]
D.T. Pham, X. Liu: Neural Networks for Identification, Prediction and
Control, Springer-Verlag London 1995.
[6]
A. G. Parlos, O. Rais, A. F. Atiya: Multi-Step-Ahead Prediction in
Complex Process Systems Using Dynamic Recurrent Neural Networks, Neural
Networks, Vol. 13, No. 7, pp 765-786, 2000
| 3cs.SY
|
DEEP LOGISMOS: DEEP LEARNING GRAPH-BASED 3D SEGMENTATION OF
PANCREATIC TUMORS ON CT SCANS
Zhihui Guo1 , Ling Zhang2∗ , Le Lu2 , Mohammadhadi Bagheri2 , Ronald M. Summers2 ,
Milan Sonka1∗ , and Jianhua Yao2
1
arXiv:1801.08599v1 [cs.CV] 25 Jan 2018
2
Iowa Institute for Biomedical Imaging, The University of Iowa, Iowa City IA
Radiology and Imaging Sciences Department, National Institutes of Health (NIH), Bethesda MD
ABSTRACT
This paper reports Deep LOGISMOS approach to 3D tumor segmentation by incorporating boundary information
derived from deep contextual learning to LOGISMOS – layered optimal graph image segmentation of multiple objects
and surfaces. Accurate and reliable tumor segmentation is
essential to tumor growth analysis and treatment selection.
A fully convolutional network (FCN), UNet, is first trained
using three adjacent 2D patches centered at the tumor, providing contextual UNet segmentation and probability map for
each 2D patch. The UNet segmentation is then refined by
Gaussian Mixture Model (GMM) and morphological operations. The refined UNet segmentation is used to provide the
initial shape boundary to build a segmentation graph. The
cost for each node of the graph is determined by the UNet
probability maps. Finally, a max-flow algorithm is employed
to find the globally optimal solution thus obtaining the final segmentation. For evaluation, we applied the method to
pancreatic tumor segmentation on a dataset of 51 CT scans,
among which 30 scans were used for training and 21 for
testing. With Deep LOGISMOS, DICE Similarity Coefficient (DSC) and Relative Volume Difference (RVD) reached
83.2±7.8% and 18.6±17.4% respectively, both are significantly improved (p<0.05) compared with contextual UNet
and/or LOGISMOS alone.
Index Terms— Deep learning, fully convolutional network, graph, tumor, 3D segmentation
1. INTRODUCTION
Monitoring the growth and spread of tumors at different time
points helps physicians differentiate tumor types and plan the
proper treatment [1]. To achieve this, accurate and reliable
segmentation of tumors is of great importance.
Variety of methods have been proposed for medical image segmentation, among which deep learning has recently
*[email protected]; [email protected]. This research was
supported – in part – by Intramural Research Program of National Institutes
of Health Clinical Center and NIH R01 EB004640. The authors thank Nvidia
for the TITAN X Pascal GPU donation.
become prevalent and reached new levels of the state-of-theart accuracy in many tasks [2]. Medical imaging data such as
CT and MRI are inherently 3D, but can be visualized as stacks
of 2D slices. Deep learning based segmentation methods can
be divided into 4 categories according to how data is input to
the network: convolutional neural networks (CNN) with 2D
convolutions [3]; CNN with 3D convolutions [4]; combination of 2D CNN and recurrent neural networks (RNN) for 3D
segmentation [2, 5]; and combination of 2D CNN and optimization algorithm for 2D or 3D segmentation [6]. UNet [3]
is a type of CNN with 2D convolutions that only takes intraslice context into account, leaving out the inter-slice context.
3D UNets [4] apply 3D convolutions to capture 3D spatial
context but are computationally expensive. An alternative
way to use contextual UNet is to stack three adjacent slices
in three RGB channels to leverage inter-slice context from
adjacent slices. LSTM [7] is a type of RNN that is designed
for sequential data and can be used to leverage spatial context
between adjacent slices. Chen et al. [2] combined a modified 2D UNet and LSTM to do 3D segmentation of neuron
and fungus. Tseng et al. [5] applied CNN and convolutional
LSTM to multi-modality data and achieved 3D segmentation.
In [6], FCN and graph-based method worked together where
FCN provided the cost for the 2D graph. Since objects are 3D
in nature, 3D spatial context are valuable in 3D segmentation.
LOGISMOS is a graph-based framework that translates
geometric constraints of interacting surfaces and objects into
graph arcs and likelihood of segmentation surface positioning into graph node/arc costs [8]. With LOGISMOS, the
globally optimal N-dimensional solution satisfying defined
smoothness constraints is obtained. The node-weighted LOGISMOS has been successfully used in difficult tasks such as
3D knee and brain segmentation [8, 9]. This graph segmentation framework is robust to image noise and weak boundaries
but requires a proper initial segmentation as the shape prior
to build the graph and to assign proper costs for each node
reflecting its likelihood to occur on the desired segmentation surface. Defining the initial segmentation is not trivial
and often requires manual intervention. Similarly, the graph
costs are frequently derived from hand-crafted task-specific
Fully Convolutional Network
LOGISMOS
3 64 64
128 64 64 1
Refinement
128
256 128
256
Input: 32x32x3
512 256
512
Conv 3x3x3, ReLu
MaxPooling 2x2x2
Up-Conv 2x2x2
Crop and Append
Conv 1x1x1
Contextual UNet
Cropped Slices Ground Truth
UNet Segmentation
1
0
Initial
Max Flow
Segmentation
Final
Graph Construction
Segmentation
Cost Design
Probability Map
Fig. 1. Schematic diagram of Deep LOGISMOS method illustrated on an example of 3D pancreatic tumor segmentation.
features and may not be generalizable to other problems.
We propose a method that combines UNet and LOGISMOS for 3D tumor segmentation. Specifically, we adopt
UNet to integrate intra-slice and adjacent-slice contexts, and
regulate the 3D shape by LOGISMOS. Different from our
previous FCN+graph method [6] which uses FCN to locate the object center for graph construction and combines
FCN-derived cost with hand-crafted costs, our method directly constructs the graph based on the UNet-derived object
boundaries and assigns UNet-derived probabilities as costs.
Pancreatic cancer is a major health problem that shows a
steady increase in incidence and death rate while also exhibiting a slight improvement in survival rates over the past 5 years
[10]. To our best knowledge, this is the first approach for automated 3D segmentation of pancreatic tumors. The proposed
method can be extended to any tumor segmentation tasks.
2. METHODS
We present a method called Deep LOGISMOS to segment tumors in 3D by combining contextual UNet and a graph-based
framework LOGISMOS. The work-flow is described in Fig.1.
First, the tumor ROI, defined as a square cube (32×32×32
voxels) by a single click of its center point, is cropped from
the whole image. For each 2D slice, the contextual UNet takes
itself and its two adjacent 2D slices as input patch and outputs
the probability map and segmentation. We apply a GMM to
remove false positives. After that, morphological opening and
closing are applied to retain only the largest region in the segmentation. To construct the graph, the refined UNet segmentation is set as the initial segmentation to build the graph. The
UNet probability maps is used as the cost for nodes in the
graph. The final segmentation is given by the global optimal
solution via a max-flow algorithm in graph search [11].
2.1. Contextual UNet
With the multi-scale training architecture, UNet meets the
need for biomedical image segmentation and has achieved
great success in various tasks [2]. We use the UNet described
in [3] for end-to-end training in this study, with the modifica-
tion that the lowest scale of feature maps is removed due to
small 2D image size (32 × 32). The input is 3-adjacent 2D
slices, leveraging the adjacent spatial contexts, namely contextual UNet. To increase training sample size, data augmentations including translation, rotation and scaling are applied
to each sample. The initial learning rate is set as 1e-6 with
momentum optimizer. We train UNet for around 30 epochs.
We test several batch size options (1, 3, 10, 100) in the verification, and the batch size of 1 gives the best accuracy.
2.2. Refinement
The UNet output needs to be further refined due to two reasons. First, the intensity distributions of tumors vary greatly
for different patients and different contrast phase. Since the
training set is small, the diverse intensity distributions may
compromise the performance of UNet. Second, the purpose
is to segment the center tumor inside the ROI. However, there
may be other tumors in the image that are detected by UNet
and should be excluded. We adopt a GMM with prior information about the relative intensity distributions of tumors and
background to subtract background from UNet segmentation.
Afterwards, morphological opening and closing are applied
to ensure that only the largest region in the center is retained.
For false positive reduction, only pixels that are segmented as tumors by UNet inside the ROI are considered. We
fit two Gaussian distributions with GMM from all pixel intensities. GMM is a clustering method that applies maximum
likelihood estimation with Gaussian conditional distribution and is solved by Expectation-Maximization algorithm.
The motivation to fit two Gaussian distributions (N(µ1 , σ12 ),
N((µ2 , σ22 )) for tumor and background respectively is based
on the prior information that pixels inside one tumor have
relatively homogeneous intensities, which are higher than
intensities in the background. Suppose µ1 is larger than µ2 ,
the condition to apply the false positive reduction by GMM
is µ2 < µ1 − σ1 . If the condition is satisfied, pixels with
intensities less than µ2 are marked as background and the
probabilities are set to be 0. Otherwise, no false positive
reduction will be applied. Then, two iterations of 3D morphological opening are applied and only the largest region is
kept. Afterwards, 3D morphological closing is performed.
2.3. LOGISMOS
There are two key factors that affect the performance of our
graph-based method, namely the initial segmentation and the
cost design. We take advantage of UNet to generate a reliable
initial segmentation and assign costs from deep features.
2.3.1. Graph construction
Table 1. Comparison of 2D UNet and contextual UNet.
Methods
DSC (%)
RVD (%)
2D UNet
72.8 ± 22.0 42.5 ± 32.8
Contextual UNet 75.6 ± 16.6 35.7 ± 29.7
Table 2. Performance of contextual UNet segmentation.
Methods
DSC (%)
RVD (%)
Without refinement 75.6 ± 16.6 35.7 ± 29.7
With refinement
81.6 ± 11.2 26.1 ± 20.9
The UNet segmentation after refinement can be regarded as
a coarse initial segmentation. This type of initial segmentation contains image-specific shape information of the tumor
on an unseen image, which is preferable to be the shape prior
compared with simple shape such ellipse or a mean shape
model. Based on the boundary of initial segmentation, a geometric node-weighted graph is established. A stack of graph
nodes (called a column) are connected with intra-column arcs
that ensure only one cut through the column. Besides, intercolumn arcs encode the smoothness constraints. The columns
are built starting from the normal directions of points on the
boundary under electric lines of force (ELF) [8] to avoid intersection. The length of the columns is set as 50 with node
spacing of 0.5 mm to cover the potential area of the tumor.
mm3 after resampling. A pancreatic tumor ROI with the size
of 32×32×32 voxels was extracted from a scan by a single
click at the approximate tumor center. 30 tumor ROIs from 8
patients were used for training and 21 tumor ROIs from other
7 patients were used for testing. We assessed the effect of 3
main aspects, which were the adjacent-slice context, refinement, and true 3D constraints of LOGISMOS. The evaluation
metrics include DSC and RVD (relative volume difference).
Statistical significance was estimated using paired t-test and
the significance level was set at 0.05. The UNet is implemented using the Caffe platform [12] on a Nvidia TITAN X
Pascal GPU with 12 GB of memory.
2.3.2. Cost design
3.1. Contextual UNet vs. 2D UNet
Contextual UNet outputs a probability map for each 2D slice.
The probability is a region-based likelihood that ranges from
0 to 1 with higher value indicating higher chance of the pixel
to be inside the tumor. LOGISMOS requires the cost to be the
likelihood of nodes being not on the boundary. To translate
the region-based probabilities to boundary-based cost, Eq.1
is used to decide the cost for node j on column k based on
the summation of the probabilities for interior nodes on the
same column. The -0.5 term corresponds to the probability
threshold (0.5) when generating UNet segmentation.
Besides training a contextual UNet, we also trained a 2D
UNet on the same training set. All the parameters were the
same. The performance of the two networks on the test set are
presented in Table 1. Contextual UNet achieved significantly
superior segmentation accuracy than 2D UNet for RVD.
cnk,j = −
j
X
(p(nk,i ) − 0.5)
(1)
3.2. Refinement
Next, we compared the refined UNet segmentation with the
original UNet segmentation (Table 2). DSC and RVD indices
demonstrated that segmentation with refinement was significantly better than that without refinement.
i=1
3.3. Deep LOGISMOS vs. Contextual UNet, LOGISMOS
2.3.3. Segmentation
The constructed graph integrates shape prior, geometric
smoothness constrains, deep features, and ensures globally
optimized true 3D segmentation. The final segmentation is
obtained by max-flow algorithm [11] in polynomial time.
3. EXPERIMENTAL RESULTS
Deep LOGISMOS was applied to a dataset of 51 arterial
phase CT scans from 15 patients with pancreatic tumors studied at multiple time points, patients were participating in a
clinical trial. The CT scans have a resolution of 1×1×1
Segmentation results from contextual UNet after refinement,
LOGISMOS and Deep LOGISMOS are presented in Fig. 2.
The initial segmentation for the original LOGISMOS
method was a sphere centered on the ROI with the radius
of 8 mm (a quarter of the ROI size), the costs were derived
from inverted gradients along the graph columns. Deep LOGISMOS segmentation performance was significantly better
than that of the contextual UNet with refinement and also of
the original LOGISMOS when considering either of the DSC
and/or RVD metrics (Table 3). Note that the LOGISMOS
method failed to detect 3 tumors altogether since the tumors
were too small. Excluding the 3 missing tumors, LOGISMOS
(a)
(b)
(c)
Fig. 3. Tumor segmentations in the same 3D view; (a), (b)
and (c) represent segmentations from contextual UNet, Deep
LOGISMOS and original LOGISMOS respectively. The tumor is the same as in Fig. 2.
Fig. 2. Qualitative comparison of the three methods. The
rows represent three adjacent slices from one tumor. Ground
truth is marked as green regions. Yellow, red and blue contours are segmentations from the contextual UNet, deep LOGISMOS, and original LOGISMOS methods, respectively.
Table 3. Comparison of segmentations from contextual UNet
after refinement, LOGISMOS and Deep LOGISMOS.
Methods
DSC (%)
RVD (%)
Contextual UNet
81.6 ± 11.2 26.1 ± 20.9
LOGISMOS
70.4 ± 27.7 35.4 ± 51.1
Deep LOGISMOS
83.2 ± 7.8 18.6 ± 17.4
method gave an average DSC of 78.7%. The segmentation
results demonstrated in a 3D view are shown in Fig. 3.
4. CONCLUSION
A hybrid fully convolutional network – FCN combined with
the graph-based LOGISMOS approach was reported. Its performance was evaluated in a 3D pancreatic tumor segmentation task. Resulting from this study, we have demonstrated
that 1) context information from adjacent slices significantly
improved the performance of a UNet, and that 2) our novel
Deep LOGISMOS method achieved significantly better performance than the UNet and/or LOGISMOS methods alone.
5. REFERENCES
[1] L. Zhang, L. Lu, R. M. Summers, E. Kebebew, and
J. Yao, “Personalized pancreatic tumor growth prediction via group learning,” in MICCAI. Springer, 2017,
pp. 422–432.
[2] J. Chen, L. Yang, Y. Zhang, M. Alber, and D. Chen,
“Combining fully convolutional and recurrent neural
networks for 3D biomedical image segmentation,” in
NIPS, 2016, pp. 3036–3044.
[3] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in MICCAI. Springer, 2015, pp. 234–241.
[4] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and
O. Ronneberger, “3D U-Net: learning dense volumetric segmentation from sparse annotation,” in MICCAI.
Springer, 2016, pp. 424–432.
[5] K. Tseng, Y. Lin, W. Hsu, and C. Huang, “Joint sequence learning and cross-modality convolution for 3D
biomedical segmentation,” in CVPR, July 2017.
[6] L. Zhang, M. Sonka, L. Lu, R. M. Summers, and J. Yao,
“Combining fully convolutional networks and graphbased approach for automated segmentation of cervical
cell nuclei,” in IEEE ISBI, 2017, pp. 406–409.
[7] S. Hochreiter and Jü. Schmidhuber, “Long short-term
memory,” Neural Comput., vol. 9, no. 8, pp. 1735–1780,
1997.
[8] Y. Yin, X. Zhang, R. Williams, X. Wu, D. D. Anderson,
and M. Sonka, “LOGISMOS-layered optimal graph image segmentation of multiple objects and surfaces: cartilage segmentation in the knee joint,” IEEE TMI, vol.
29, no. 12, pp. 2023–2037, 2010.
[9] Z. Guo, S. Kashyap, M. Sonka, and I. Oguz, “Machine
learning in a graph framework for subcortical segmentation,” in SPIE Medical Imaging, 2017, pp. 101330H–
101330H.
[10] R. Siegel, J. Ma, Z. Zou, and A. Jemal, “Cancer statistics, 2014,” CA Cancer J Clin., vol. 64, no. 1, pp. 9–29,
2014.
[11] Y. Boykov and V. Kolmogorov, “An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision,” IEEE TPAMI, vol. 26, no. 9, pp.
1124–1137, 2004.
[12] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long,
R. Girshick, S. Guadarrama, and T. Darrell, “Caffe:
Convolutional architecture for fast feature embedding,”
in ACM, 2014, pp. 675–678.
| 1cs.CV
|
Stochastic Calculus of Wrapped Compartments∗
Mario Coppo1 , Ferruccio Damiani1 , Maurizio Drocco1 , Elena Grassi1,2 and Angelo Troina1
1 Dipartimento
2 Molecular
di Informatica, Università di Torino
Biotechnology Center, Dipartimento di Genetica, Biologia e Biochimica, Università di Torino
{coppo,damiani,troina}@di.unito.it
[email protected]
[email protected]
The Calculus of Wrapped Compartments (CWC) is a variant of the Calculus of Looping Sequences
(CLS). While keeping the same expressiveness, CWC strongly simplifies the development of automatic tools for the analysis of biological systems. The main simplification consists in the removal of
the sequencing operator, thus lightening the formal treatment of the patterns to be matched in a term
(whose complexity in CLS is strongly affected by the variables matching in the sequences).
We define a stochastic semantics for this new calculus. As an application we model the interaction between macrophages and apoptotic neutrophils and a mechanism of gene regulation in E.Coli.
1
Introduction
In computer science, several formalisms have been proposed for the description of the behaviour of
biological systems. Automata-based models [2, 22] have the advantage of allowing the direct use of
many verification tools such as model checkers. Rewrite systems [13, 25, 5] usually allow describing
biological systems with a notation that can be easily understood by biologists. Compositionality allows
studying the behaviour of a system componentwise. Both automata-like models and rewrite systems
present, in general, problems from the point of view of compositionality, which, instead, is in general
ensured by process calculi, included those commonly used to describe biological systems [26, 24, 9].
The Calculus of Looping Sequences [5] (CLS) has been introduced to describe microbiological systems, such as cellular pathways, and their evolution. The terms of the calculus are constructed by basic
constituent elements and operators of sequencing, looping, containment and parallel composition. The
looping operator allows tying up the ends of a sequence, thus creating a circular sequence which can
represent a membrane. The presence of sequences and loops, however, makes more difficult to develop
a correct and efficient implementation of the associated simulation tools. The main difficulties come
from the complexity of the pattern matching algorithms which should take sequences (and all their subsequences) into account. Moreover, the use of sequences and sequence variables seems prone to subtle
semantic ambiguities in rules definition.
In this paper we propose the Calculus of Wrapped Compartments (CWC for short), a variant of
CLS suitable for efficient implementation. In CWC, starting from an alphabet of atomic elements, we
build multisets of elements and compartments. We are able to localise elements by compartmentalisation
and we allow to specify the structure of the compartment (e.g., detailing the elements of interest on a
membrane). The evolution of the system is driven by a set of rewrite rules modelling the reactions of
interest. The main difference with respect to CLS consists in removing the sequencing operator and
using multisets (instead of ordered sequences) of atoms to represent membranes. Such a choice, while
preserving the Turing completeness of the calculus, strongly simplifies the pattern matching algorithm
and the underlying basic theory.
∗ This research is funded by the BioBITs Project (Converging Technologies 2007, area: Biotechnology-ICT), Regione
Piemonte.
A. Di Pierro & G. Norman (Eds): 8th Workshop on
Quantitative Aspects of Programming Languages (QAPL’10)
EPTCS 28, 2010, pp. 82–98, doi:10.4204/EPTCS.28.6
c M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
This work is licensed under the
Creative Commons Attribution License.
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
83
In our previous experience with biological models we realized that sequences were not used in many
situations: cellular pathways are usually represented with a greater abstraction than the one which uses
amino acids and nucleotides to depict proteins and genes - moreover representing the primary sequence
of a protein without considering secondary and tertiary structure is seldom useful. CLS used looping
sequences to represent membranes, but we believe that multisets of atomic elements are a more natural
representation, which reflects the fluid mosaic model [28]. A similar approach is taken in the CLS+
model (see [6]), a variant of CLS in which the looping operator can be applied to a parallel composition
of sequences.
We provide CWC with a stochastic operational semantics from which a continuous time Markov
chain can be build. With respect to [4], in which a constant rate is associated to the rewrite rules, we
allow a more flexible treatment of the dynamics. In particular, we associate a function to each rewrite
rule according to which the rate of the transition can be computed depending on the initial and the target
terms. We show how the law of mass action can be encoded in our stochastic semantics thus following
the standard Gillespie’s approach [17].
Summary. In Section 2 we introduce the new formalism and show its Turing completeness. In Section 3 we provide some guidelines for the modelling of common biological events. We give some detail
on how the interaction between macrophages and apoptotic neutrophils, and a mechanism of gene regulation can be described in the new calculus. In Section 4 we present a stochastic semantics of the
calculus. In Section 5, we report some simulations regarding the gene regulation mechanism introduced
in Section 3. Finally, in Section 6 we draw some conclusions.
2
The Calculus of Wrapped Compartments
In this section we present the nondeterministic Calculus of Wrapped Compartments (CWC).
2.1
Terms and Structural Congruence
A term of the CWC calculus is intended to represent a biological system. A term is a sequence of simple
terms. Simple terms, ranged over by t, u, v, w, are built by means of the compartment constructor,
(− c −), from a set A of atomic elements (atoms for short), ranged over by a, b, c, d. The syntax of
simple terms is given at the top of Figure 1. We write t to denote a (possibly empty) sequence of simple
terms t1 · · ·tn . Similarly, with a we denote a (possibly empty) sequence of atoms. The set of simple terms
will be denoted by T . The set of terms (sequences of simple terms) and the set of sequences of atoms
will be denoted by T and A , respectively. Note that A ⊆ T .
A term t = t1 · · · tn should be understood as the multiset containing the simple terms t1 , . . . ,tn . Therefore, we introduce a relation of structural congruence, following a standard approach in process algebra.
The CWC structural congruence is the least equivalence relation on terms satisfying the rules given at
the bottom of Figure 1. From now on we will always consider terms modulo structural congruence.1
Then a simple term is either an atom or a compartment (a ct) consisting of a wrap (represented by the
multiset of atoms a) and a content (represented by the term t). We write the empty multiset as • and
denote the union of two multisets u and v as u v.
1 In the implementation of the derived tools it will be useful to consider a normalized representation of these terms suitable
for efficient manipulation. In the description of the calculus given here however we will ignore implementation issues.
84
Stochastic Calculus of Wrapped Compartments
Simple terms syntax
t
::=
a
(a ct)
Structural congruence
t u w v≡t w u v
if a ≡ b and t ≡ u then (a ct) ≡ (b c u)
Figure 1: CWC term syntax and structural congruence rules.
Figure 2: (i) represents (a b c c •);
(a b c c (d e c •) f g).
(ii) represents (a b c c (d e c •));
(iii) represents
An example of term is a b (c d c e f ) representing a multiset consisting of two atoms a and b (for
instance two molecules) and a compartment (c d c e f ) which, in turn, consists of a wrap (a membrane)
with two atoms c and d (for instance, two proteins) on its surface, and containing the atoms e (for
instance, a molecule) and f (for instance a DNA strand). See Figure 2 for some graphical representations.
2.2
Systems and Reduction Semantics
Rewrite rules are defined as pairs of terms, in which the left term characterizes the portion of the system
in which the event modelled by the rule can occur, and the right one describes how that portion of the
system is changed by the event.
In order to formally define the rewriting semantics, we introduce the notion of open term (a term
containing variables) and pattern (an open term that may be used as left part of a rewrite rule). In
order to respect the syntax of terms, we distinguish between “wrap variables” which may occur only in
compartment wraps (and can be replaced only by multisets of atoms) and “term variables” which may
only occur in compartment contents or at top level (and can be replaced by arbitrary terms). Therefore,
we assume a set of term variables, VT , ranged over by X,Y, Z, and a set of wrap variables, VA , ranged
over by x, y, z. These two sets are disjoint. We denote by V the set of all variables VT ∪ VA , and with ρ
any variable in V .
Definition 2.1 (Open Terms and Patterns)
(i) Open terms are terms which may contain occurrences of wrap variables in compartment wraps
and term variables in compartment contents or at top level. They can be seen as multisets of
simple open terms. More formally, open terms, ranged over by O and simple open terms, ranged
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
85
over by o, are defined in the following way:
O
o
::=
::=
o
a
X
(a x c o)
(ii) An open term is linear if each variable occurs at most once in it.
(iii) Simple patterns, ranged over by S, are the linear open terms defined in the following way:
S
::= t
(a x c S X)
where t is an element of T , a is an element of A , x is a variable in VA , S is a possibly empty
multiset of simple patterns and X is a variable in VT . We denote with S the set of simple patterns.
(iv) Patterns, ranged over by P, are the linear open terms defined in the following way:
P ::= S S X
where S S is a nonempty multiset of simple patterns and X is an element of VT . We denote with
P the set of patterns.
An instantiation is a partial function σ : V → T . An instantiation must preserve the type of variables,
thus for X ∈ VT and x ∈ VA we have σ (X) ∈ T and σ (x) ∈ A , respectively. Given O ∈ O, with Oσ we
denote the term obtained by replacing each occurrence of each variable ρ ∈ V appearing in O with the
corresponding term σ (ρ). With Σ we denote the set of all the possible instantiations, and, given O ∈ O,
with Var(O) we denote the set of variables appearing in O. Now we can define rewrite rules.
Definition 2.2 (Rewrite Rules) A rewrite rule is a pair of a pattern P and an open term O, denoted with
P 7→ O, such that Var(O) ⊆ Var(P).
A rewrite rule P 7→ O states that a term Pσ , obtained by instantiating variables in P by some instantiation function σ , can be transformed into the term Oσ . A set of rewrite rules define a notion of
reduction between terms of CWC which are intended to represent the (nondeterministic) behaviour of
the represented system.
Definition 2.3 A CWC system over a set A of atoms is represented by a set FA (F for short when A
is understood) of rewrite rules over A .
We define the semantics of CWC as a transition system in which states correspond to terms and
transitions correspond to rule applications. To this aim we need to define the notion of contexts (i.e.
terms containing a hole).
Definition 2.4 (Contexts) Contexts C are defined as:
C ::=
(a cC) t
where a ∈ A and t ∈ T . The context is called the empty context. We denote with C the set of contexts.
86
Stochastic Calculus of Wrapped Compartments
By definition, every context contains a single hole . Let us assume C,C0 ∈ C . With C[t] we denote
the term obtained by replacing with t in C; with C[C0 ] we denote context composition, whose result is
the context obtained by replacing with C0 in C. For example, given C = (a b c ) i l, C0 = (c d c ) g h
and t = e f , we get C[C0 [t]] = (a b c (c d c e f ) g h) i l. Note that context holes take the place either of
the whole term or of the whole content of a compartment. This allows to make context unambiguous in
the following sense. Let’s extend the notion of subset (denoted as usual as ⊆) between terms interpreted
as multisets.
Proposition 2.5 (Uniqueness) For any term t if the term u occurs in t within a compartment content or
at top level, then there are, modulo ≡, a unique context C and a unique term t 0 such that t = C[t 0 ] and
u ⊆ t 0.
Given a CWC system F the associated notion of reduction is then defined in the following way.
Definition 2.6 (Reduction) Given a CWC system F , the F -reduction relation between terms of CWC
is defined by the following rule:
P 7→ O ∈ F
σ ∈Σ
t = Pσ
C[t ] 7→F C[u]
u = Oσ
C∈C
closed under structural congruence, i.e.:
u ≡ u0 and u0 7→F v0 and v0 ≡ v
⇒
u 7→F v
We will write simply 7→ instead of 7→F when F is understood.
Remark 2.7
(i) Taken a pattern P = S S X (representing the reactants of the reaction that will be simulated) the
crucial point for determining an application of P to the whole biological ambient t is the choice
of the occurrences of simple terms matching with S S (determining the compartment in which
this reaction will take place). By Proposition 2.5 and the linearity condition both the context C
(such that t ≡ C[t 0 ]) and the term replacing X (the part of t 0 that does not match S S) are uniquely
determined. This will play a role in the implementation of the stochastic calculus.
(ii) There can be, in general, many different substitution such that Pσ ≡ t but not all of them produce
the same term Oσ (modulo ≡). This is not particularly relevant for the case of qualitative systems
but we will need to take it into account in the stochastic formulation.
Notation 2.8
(i) We write S S ⇒ O as short for S S X 7→ O X, where the variable X does not occur in Var(S S) ∪
Var(O).
(ii) We write a a → b as short for the rewrite rule (a a x cY ) Z 7→ (b x cY ) Z, where the variables
x,Y, Z are distinct.
The CWC is Turing Complete. In the following we sketch how Turing Machines can be simulated
by CWC models.
Theorem 2.9 (Turing Completeness) The class of CWC models is Turing complete.
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
Biomolecular Entity
Elementary object (genes, domains, etc...)
Molecular population
Membrane
87
CWM term
Atoms
Term multiset
Atom multiset
Table 1: Guidelines for the abstraction of biomolecular entities into CWC.
Proof: (Sketch). A Turing machine T over an alphabet Σ ∪ {]} (where ] represent the blank) with a set
Q of states can be simulated by a system FT of CWC in the following way.
Take A = Q ∪ Σ ∪ {], l, r}, where l and r are special symbols to represent the left and right ends of the
tape. The tape of the Turing machine can be represented by a sequence of nested compartments whose
wraps consist of a single atom (representing a symbol of the tape). The content of each compartment
defined in this way represents a right suffix of the written portion of the tape, while the atom on the
wrap represents its initial (w.r.t. the suffix) symbol. In each term representing a tape there is exactly
one compartment which contains a state (the present state). For example, the term (l c (a c q (b c (r c •))))
represents the configuration in which the tape is ...], a, b, ].... (the remaining positions are blank), the
machine is in state q and the head is positioned on b. Rewrite rules are then used to model the machine
evolution. We can define rules creating new blanks when needed (at the tape ends) to mimic a possibly
infinite tape. For instance a transition (q, b) → (q0 , c, right) (meaning that in state q with a b on the head
the machine goes in state q0 writing c on the tape and moving the head right) is represented by the rules2 :
(1) q (b y c (x c Z) Y ) X 7→ (c y c q0 (x c Z) Y ) X
(2) q (b y c (r c Z) Y ) X 7→ (c y c q0 (] c (r c Z)) Y ) X
The second rule represents the case that b is the rightmost non blank symbol and so a new blank must be
introduced in the simulation. 3 By construction, the system FT correctly represents the behaviour of T .
3
Modelling Guidelines
In this section we will give some explanations and general hints about how CWC could be used to
represent the behaviour of various biological systems.
In rewrite systems, such as CWC, entities are usually represented by terms of the rewrite system,
and events by rewrite rules. We have already introduced the biological interpretation of our terms in
Section 2.1. Here we will go deeper and describe the most straightforward abstraction that can be used
to define terms and rules.
First of all, we should select the biomolecular entities of interest. Since we want to describe cells,
we consider molecular populations and membranes. Molecular populations are groups of molecules that
are in the same compartment of the cells and inside them. As we have said before, molecules can be of
many types: we classify them as proteins, chemical moieties and other molecules.
Membranes are considered as elementary objects: we do not describe them at the level of the phospholipids they are made of. The only interesting properties of a membrane are that it may have a content
2 In
these rules, in a Turing Machine simulation, the variables X and Y are always instantiated with the empty multiset
that rule (1) could be applied also in this case but it would lead the system in a configuration (containing a subterm
q0 (r ct) for some t) from which no further move could be possible. However situations of this kind can be easily avoided with
a little complication of the encoding.
3 Note
88
Stochastic Calculus of Wrapped Compartments
Biomolecular Event
State change
Complexation
Decomplexation
Catalysis
State change on membrane
Complexation
on membrane
Decomplexation
on membrane
Catalysis on membrane
Membrane crossing
Catalyzed
membrane crossing
Membrane joining
Catalyzed
membrane joining
Membrane fusion
Vesicle dynamics
Examples of CWC Rewrite Rules
a⇒b
a b⇒c
c⇒a b
c S S ⇒ c O where S S ⇒ O is the catalyzed event
a→b
a b→c
a (b x c X) ⇒ (c x c X)
(b x c a X) ⇒ (c x c X)
c→a b
(c x c X) ⇒ a (b x c X)
(c x c X) ⇒ (b x c a X)
c a a → c b where a a → b is the catalyzed event
a (x c X) ⇒ (x c a X)
(x c a X) ⇒ a (x c X)
a (b x c X) ⇒ (b x c a X)
(b x c a X) ⇒ a (b x c X)
a (x c X) ⇒ (a x c X)
(x c a X) ⇒ (a x c X)
a (b x c X) ⇒ (a b x c X)
(b x c a X) ⇒ (a b x c X)
(x c a b X) ⇒ (a x c b X)
(a x c X) (b y cY ) ⇒ (a x b y c X Y )
(x c X (a y cY )) ⇒ Y (a y x c X)
Table 2: Guidelines for the abstraction of biomolecular events into CWC.
(hence, create a compartment) and that in its phospholipid bilayer various proteins are embedded, which
act for example as transporters and receptors. Since membranes are represented as multisets of the
embedded structures, we are modelling a fluid mosaic in which the membranes become similar to a
two-dimensional liquid where molecules can diffuse more or less freely [28].
Now, we select the biomolecular events of interest. The simplest kind of event is the change of
state of an elementary object. Then, there are interactions between molecules: in particular complexation, decomplexation and catalysis. Interactions could take place between simple molecules, depicted
as single symbols, or between membranes and molecules: for example a molecule may cross or join a
membrane. Finally, there are also interactions between membranes: in this case there may be many kinds
of interactions (fusion, vesicle dynamics, etc. . . ).
The guidelines for the abstraction of biomolecular entities and events into CWC are given in Table 1
and Table 2, respectively. Biomolecular events are associated with rewrite rules. In the table we give
some examples of rewrite rules for different kinds of events. The list has not the ambition of being
complete but to give some general hints on rules definition.
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
89
Figure 3: Neutrophils phagocytosis: a) a macrophage distinguishes between apoptotic and living neutrophils, b) the death neutrophile is “eaten” and embedded in a phagosome, c) the phagosome fuses
with a lysosome.
3.1
Modelling Examples
We will now present two examples of possible CWC simulations, in order to testify that our calculus is
able to express a broad range of biological situations even if the possibility to represent sequences, which
is a computational burden, is removed from its syntax.
Macrophages. All the biological problems involving cell interactions are feasible. One could for example represent macrophages and their interaction with the apoptotic cells: in order to “clean up” the
remains of death cells they have to recognize them amongst the alive ones and then remove them by a
process called phagocytosis, which consists in the internalization of the death cell in a membrane compartment, called phagosome. The phagosome inside the macrophage will then join another membrane
compartment called the lysosome in order to start the degradation of the ingested cellular components.
To give the rules that could represent this process in CWC we decided to concentrate on the removal of neutrophils that have undergone apoptosis: removing them before their membrane collapses is
crucial because they contain some lytic enzymes that, if released in the organism, would lead to tissue
damages [15, 12]. The mechanism which mediates the recognition of apoptotic neutrophils is not perfectly clear but we decided to reproduce the one described in [29, 8] which suggest that neutrophils and
macrophages interact via CD31 and that somehow the CD31 of vial neutrophils is capable of “escaping”
the phagocytosis.
We will represent a macrophage with a lysosome and a CD31 exposed on its membrane in this
way (where M is a marker to distinguish macrophages from other cells and lyso represents a lysosome
containing some lytic enzymes):4
(CD31 M c (lyso c lyticEnz) innerM)
A living neutrophile (where N is a marker for neutrophils and V indicates that this is a viable neutrophile)
is represented by:
(CD31 V N c innerL)
An apoptotic one (where A indicates its apoptotic state) is represented by:
(CD31 A N c innerA)
4 Our calculus allows to represent with simple atoms some higher level information such as that represented by marker
names.
90
Stochastic Calculus of Wrapped Compartments
Figure 4: The Pho regulation system.
where innerM, innerL and innerA denote the internal material of a macrophage, a living neutrophile and
an apoptotic one, respectively.
Now a fairly simple set of rules is sufficient to represent the whole process.
The binding of a neutrophile and a macrophage mediated by their CD31 molecules is represented by
encapsulating the interacting cells within a new compartment wrapped with the marker I:
(CD31 M x c X) (CD31 N y cY ) ⇒ (I c (CD31 M x c X) (CD31 N y cY ))
After the contact the viable neutrophils are released:
(I z c (CD31 M x c X) (CD31 N V y cY ) Z) ⇒ (CD31 M x c X) (CD31 N V y cY )
While the apoptotic ones are phagocyted (phago represents the phagosome which contains the phagocyted neutrophile):
(I z c (CD31 M x c X) (CD31 N A y cY ) Z) ⇒ (CD31 M x c (phago c (CD31 N A y cY )) X)
The phagosome can join the lysosome to start hydrolysis (phagolyso represents the phagolysosome
derived from the fusion):
(lyso x c lyticEnz X) (phago z c (CD31 N A y cY ) Z) ⇒ (phagolyso x z c (CD31 N A y cY ) lyticEnz X Z)
This should show that interactions involving cells of the immune system can be represented in CWC
with little effort at different levels of detail: modelling some aspects of the complex network that drives
immunological defence could be useful to study many human pathologies and involved processes, like
in this example about inflammation and how it stops.
Phosphate regulation. Switching to a completely different field we would like to describe a mechanism of gene regulation fairly well defined in E.Coli and other bacteria, which is based on a two component system: (i) the sensor, which can react to changes in the environment and interact with the other
component; (ii) the response regulator, which is capable of determining, when activated by the sensor,
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
91
the reactions necessary to cope with the new environmental conditions. This schema is common to drive
gene regulation in bacteria: the activated response regulator will act as a transcriptional activator.
We will model the system related to environmental phosphate concentration, where the sensor is a
transmembrane protein located on the inner membrane of the bacteria, called PhoR, and the response
regulator is a cytoplasmatic protein which can be activated by phosphorylation and is called PhoB. Environmental phosphate will freely diffuse in the periplasmic space (between the two bacterial membranes)
thanks to the outer membrane pores and there it will bind to PhoR. When bound by phosphate PhoR will
be silent, while it shows a kinase activity on PhoB when it is unbound; therefore low levels of phosphate
will leave many PhoR unbound and able to activate PhoB, while high levels will lead to a smaller fraction of transmembrane unbound PhoR and thus to little activation of PhoB. When phosphorylated, PhoB
activates the transcription of many genes aimed at helping the bacteria survive in low phosphate conditions [21]. These mechanisms have been recently described as related to the pathogenic potential [16] of
some bacteria and therefore studying them could also have some importance in medicine.
We will now present the set of rules needed to model this regulation system; in the following we
will refer to the outer membrane pores as pore, to phosphate as Pi, to the phosphate bound and unbound
PhoR as PhoR and PhoRP, to the inactive and active PhoB as PhoB and PhoBP, to its target genes as
PhoGenes and to the proteins that they encode as PhoProt.
We may represent an E.Coli with the following CWC term:
(pore c (PhoR PhoRP c PhoB PhoBP PhoGenes))
where, here, a sequence PhoR represents a multiset of PhoR atoms (similar for the other atoms). Two
rules describe the phosphate diffusion in the periplasmic space:
Pi (pore x c X) ⇒ (pore x c Pi X)
(pore x c Pi X) ⇒ Pi (pore x c X)
Two rules for the PhoR phosphate binding:
Pi (PhoR x c X) ⇒ (PhoRP x c X)
(PhoRP x c X) ⇒ Pi (PhoR x c X)
PhoB phosphorylation:
(PhoR x c PhoB X) ⇒ (PhoR x c PhoBP X)
High level rules that represent the translation and traduction to proteins of the PhoB target genes and
these proteins degradation:
PhoBP PhoGenes ⇒ PhoBP PhoGenes PhoProt
PhoProt ⇒ •
One last rule represents the bound form of PhoR phosphatase activity on PhoBP [10], which is a negative
regulating loop of the whole mechanism:
(PhoRP x c PhoBP X) ⇒ (PhoRP x c PhoB X)
In Section 5 we will give some details about simulations of this model.
92
4
Stochastic Calculus of Wrapped Compartments
The Stochastic Calculus of Wrapped Compartments
In order to make the formal framework suitable to model quantitative aspects of biological systems we
must associate to each rule a numerical parameter (the rule rate) which determines (in a stochastic sense)
the time spent between two successive interactions and the frequency with which each interaction will
take place in order to represent faithfully the system evolution. The operational semantics will be defined
by incorporating the rule rates in a stochastic framework along the lines of the one presented by Gillespie
in [17]. The same approach has been incorporated, for example, in other simulators as the ones for the
stochastic π-calculus [23, 24].
In this framework the rule rate is used as the parameter of an exponential distribution modelling
the time spent between two occurrences of the considered chemical reaction. The use of exponential
distributions to represent the (stochastic) time spent between two occurrences of chemical reactions
allows describing the system as a Continuous Time Markov Chain (CTMC), and consequently allows
verifying properties of the described system analytically and by means of stochastic model checkers.
The basic choice for determining the rule rate of a reaction (suggested also by Gillespie’s algorithm)
is to obtain it by multiplying the kinetic constant of the reaction by the number of possible combinations
of reactants that may occur in the system, representing in this way the law of mass action.
In our calculus we will represent a reaction rate as a function of the context in which the reaction takes
place. This allows to tailor the reaction rates on the specific characteristics of the system, as for instance
when representing nonlinear reactions as it happens for Michaelis–Menten kinetics. The standard mass
action rate can be describes as a particular choice of the rate function (see below). A similar approach is
used in [14] to model reactions with inhibitors and catalysers in a single rule.
Obviously some care must be taken in the choice of the rate function: for instance it must be complete
(defined on the domain of the application) and nonnegative. This properties are enjoyed by the function
representing the law of mass action.
f
Definition 4.1 A stochastic rewrite rule is a triple (P, R, f ), denoted P 7−→ R, where (P, R) is a rewrite
rule and f : T × T → R≥0 is the rate function associated to the rule.5
We can then define, as in Definition 2.6, the “stochastic” reduction relation for CWC. In this case,
we must also take into account that, by Remark 2.7, different instantiations that allow the l.h.s. P of a
rule to match a term t can produce different outcomes which could determine different rates. This is
shown in Example 4.2. So, in the definition of the rate function, we must also take into account the term
produced by the reaction after the application of the rule.
f
Example 4.2 Consider the rewrite rule a (b x c X) 7−→ (a b x c X) modelling a catalyzed membrane
joining. In this case, the initial term a (b b c c) (b c c) results in two different terms depending on which
membrane will be joined by the element a; namely (a b b c c) (b c c) and (b b c c) (a b c c). The application of the rule, moreover, should take into account that the first case is catalyzed by the presence of two
elements b on the membrane, while the second one has only one catalyzer.
A Stochastic CWC systems RA (R for short) can be defined as in Definition 2.6 from a set R of
stochastic reduction rules over A .
5 The value 0 in the codomain of f models the situations in which the given rule cannot be applied, for example when the
particular environment conditions forbid the application of the rule.
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
93
Definition 4.3 (Stochastic Reduction) Given a stochastic CWC system R, the induced stochastic reduction relation between terms of CWC is defined by the following rule:
f
P 7−→ O ∈ R
σ ∈Σ
C[t]
f
t = Pσ
f (t,u)
7−→ R
u = Oσ
C∈C
C[u]
f
We will write simply 7−→ instead of 7−→R when R is understood.
Chemical–like Reactions. To define the “canonical” function that represents the law of mass action the
application rate of a reduction rule should depend on the kinetic constant of the reaction and the number
of possible reactants in the considered compartment in which the reaction will take place. Following an
approach similar to that of [4, 18], this can be obtained counting the number of different ways in which
the rule representing a reaction can be applied in the considered context producing the same outcome. In
our setting, however, we can formalize the canonical function in a simpler way.
To this aim we need a few definitions. Assume to have an infinite set of labels (e.g. the natural
numbers) that can be attached to the atomic elements in A to create, for each a ∈ A an infinite set of
distinguished versions a1 , a2 , a3 , ... of the element. If ai is a labelled atom, we write a = |ai | for the
support of ai .
Definition 4.4
(i) A labelled term of CWC is a term in which some labelled atomic element can occur.
(ii) A complete labelled term is a term in which all atomic elements are labelled and no two labelled
atoms having the same support have the same label.
(iii) If t is a labelled term, its erasure |t| is the term obtained from t by erasing all the labels.
Labels can be applied also to atoms occurring in open terms, but no label can be associated to a
variable.
If t is a labelled (possibly open) term a labeling t + of t is obtained by assigning labels to the non
labelled atoms of t such that t + is a completely labelled term.6 For example, taking t = a1 a a3 b a
possible labeling of t is t + = a1 a2 a3 b1 .
f
Given a reduction rule P 7−→ O and terms t, u such that for some instantiation σ it holds Pσ ≡ t and
Oσ ≡ u, we define the canonical rate function f (t, u) by the following steps:
1. take any completely labeled version t + of t;
2. let n be the number of distinct instantiations of variables with completely labeled terms such that
P0 σ ≡ t + , for some labeling P0 of P, and |Rσ | ≡ u;7
3. define the function f as f (t, u) = k · n where k is the kinetic constant of the associated reaction.
We denote with fkcan the function obtained in this way to compute the evolution rate following the
law of mass action with kinetic constant k.
Note that since the labels of the atomic elements are all different, by Proposition 2.5 and Remark 2.7
(i) each instantiation σ such that P0 σ ≡ t + corresponds to a distinct combination of the elements that
6 Note
that there are many possible labellings of t.
7 Note that the labeling of the atoms in P is taken only in order to allow matching.
factor.
It is the choice of the variables the relevant
94
Stochastic Calculus of Wrapped Compartments
determine the reaction described by the rule P 7−→ O. So the number n of substitutions defined in point
2. corresponds exactly to the number of possible distinct reactants for the considered reaction.
This number can often be calculated in a rather simple way. For instance in a rule X a b 7−→ X c
each instantiation of X that matches any term containing at least one a and one b corresponds exactly
to a pair of occurrences of a and b (the ones not included in the variable’s instantiation). So the total
number of distinct instantiations of X (i.e. of possible reactants) is given by the product of the number
of occurrences of a times the number of occurrences of b.
Example 4.5 Consider the completely labeled term t + = a1 a2 a3 b1 defined above and the qualitative
reduction rule a a X 7→ a c X. The l.h.s. of the rule can be matched by the instantiations X = a1 b1 ,
X = a2 , b1 and X = a3 b1 . In the first case, the match is obtained by taking P0 = a2 a3 X and so on. So,
we have that fkcan (a a a b, a a c b) = 3 · k where k is the kinetic constant of the reaction.
f
Example 4.6 Consider again the term t = a (b b c c) (b c c) and the rewrite rule a (b x c X) 7−→ (a b x c X)
2·k
described in Example 4.2. If the function f is defined as the canonical function fkcan , we get t 7−→
1·k
(a b b c c) (b c c) and t 7−→ (b b c c) (a b c c).
Remark 4.7 Note that the canonical function f can has the same meaning of the function O defined in
[4].
5
An application
We used a prototype simulator for CWC to run some simulations for the Pho regulation system described
k
in Section 3.1. In figure 5) we use ⇒ to indicate that we use the canonical function with kinetic constant
k. The rates here are chosen manually following these considerations: diffusion of phosphate across the
outer membrane pores has the same speed in both directions, PhoRP phosphatase activity is slow with
respect to its kinase activity in the unbound form, both these mechanism are slower than the simpler
binding and unbinding of phosphate to PhoR and the rules that describe translation and traduction and
protein degradation should be rather slow.
0.1
Pi (pore x c X) ⇒ (pore x c Pi X)
0.1
(pore x c Pi X) ⇒ Pi (pore x c X)
0.01
Pi (PhoR x c X) ⇒ (PhoRP x c X)
0.005
(PhoRP x c X) ⇒ Pi (PhoR x c X)
0.001
(PhoR x c PhoB X) ⇒ (PhoR x c PhoBP X)
0.0001
PhoBP PhoGenes ⇒ PhoBP PhoGenes PhoProt
0.00008
PhoProt ⇒ •
0.0001
(PhoRP x c PhoBP X) ⇒ (PhoRP x c PhoB X)
Figure 5: Stochastic rules for the phosphate regulation mechanism.
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
95
Figure 6: Levels of PhoB target proteins over time with different extracellular phosphate levels.
Figure 7: Levels of PhoB target proteins and periplasmic phosphate in the low extracellular phosphate
conditions.
We performed two simulations: both of them started with 10 PhoB, 5 PhoR, 5 PhoRP and a PhoGenes
element in a single bacterial cell. To reflect high and low extracellular phosphate we started with 20 and 5
extracellular Pi, respectively. In Figure 6 we show the comparison of the PhoProt levels obtained in these
simulations: as expected in the low phosphate condition the average level of the PhoB target proteins is
higher. Figure 7 shows the relation between PhoProt levels and periplasmic phosphate for the low phosphate simulation: oscillations in the protein levels are in sync with the levels of periplasmic phosphate:
when the latter increases the former decreases and viceversa. Figure 8 shows the beginning of the same
simulation but focuses on the relationships between PhoRP, PhoPB and periplasmic phosphate: PhoRP
levels follows with a little delay the phophate ones, while PhoBP shows a clearly specular behaviour
96
Stochastic Calculus of Wrapped Compartments
Figure 8: Levels of phosphorilated PhoB and phosphate bound PhoR in the low extracellular phosphate
conditions.
with respect to PhoRP, as expected.
In the future we will like to compare the results of these simulations with available experimental
data [3].
6
Conclusions
As we have seen, CWC allows to model cellular interaction, localisation and membrane structures. Other
formalisms were developed to describe membrane systems. Among them we cite Brane Calculi [9] and
P-Systems [25].
CWC can describe situations that cannot be easily captured by the above mentioned formalisms,
which consider membranes as atomic objects. Representing the membrane structure as a multiset of
the elements of interest allows the definition of different functionalities depending on the type and the
number of elements on the membrane itself.
From a qualitative point of view, our calculus is essentially a variant of CLS [5], while, from a
quantitative point of view, the possibility to model the speed of reactions using functions instead of
constant rates makes the CWC stochastic semantics more general with respect to the one of SCLS [4].
In this sense, CWC stochastic semantics is more similar to the one in [14].
The current simulator for CWC is implemented in OCaml (see [1]). Since we do not need to analyse
sequence structures and variables, the pattern matching algorithm has been simplified with respect to the
simulator for SCLS described in [27].
We are also considering to enrich our analysis framework by taking into account statistical model
checking as done in [7] for stochastic CLS models, and with techniques to approximate the value of
unknown kinetics in the stile of [19, 20, 11].
M. Coppo, F. Damiani, M. Drocco, E. Grassi & A. Troina
97
References
[1] M. Aldinucci, M. Coppo, F. Damiani, M. Drocco, E. Giovannetti, E. Grassi, and A. Troina. SCWC-biosimulator. Dipartimento di Informatica, Università di Torino, 2010. http://scwc-sim.sourceforge.
net/.
[2] R. Alur, C. Belta, and F. Ivancic. Hybrid modeling and simulation of biomolecular networks. In HSCC,
volume 2034 of LNCS, pages 19–32. Springer, 2001.
[3] JH. Baek, YJ. Kang, and SY. Lee. Transcript and protein level analyses of the interactions among phob, phor,
phou and crec in response to phosphate starvation in escherichia coli. FEMS Microbiol Lett., 277:254–259,
2007.
[4] R. Barbuti, A. Maggiolo-Schettini, P. Milazzo, P. Tiberi, and A. Troina. Stochastic calculus of looping
sequences for the modelling and simulation of cellular pathways. Transactions on Computational Systems
Biology, IX:86–113, 2008.
[5] R. Barbuti, A. Maggiolo-Schettini, P. Milazzo, and A. Troina. A calculus of looping sequences for modelling
microbiological systems. Fundam. Inform., 72(1-3):21–35, 2006.
[6] Roberto Barbuti, Andrea Maggiolo-Schettini, Paolo Milazzo, and Angelo Troina. The calculus of looping
sequences for modeling biological membranes. In Workshop on Membrane Computing, volume 4860 of
Lecture Notes in Computer Science, pages 54–76. Springer, 2007.
[7] T. A. Basuki, A. Cerone, and P. Milazzo. Translating stochastic cls into maude. Electr. Notes Theor. Comput.
Sci., 227:37–58, 2009.
[8] S. Brown, I. Heinisch, E. Ross, K. Shaw, CD. Buckley, and J. Savill. Apoptosis disables CD31-mediated cell
detachment from phagocytes promoting binding and engulfment. Nature, 418:200–203, 2002.
[9] L. Cardelli. Brane calculi. In CMSB, volume 3082 of LNCS, pages 257–278. Springer, 2005.
[10] DO. Carmany, K. Hollingsworth, and WR. McCleary. Genetic and biochemical studies of phosphatase activity of phor. J Bacteriol., 185:1112–1115, 2003.
[11] A. Coletta, R. Gori, and F. Levi. Approximating probabilistic behaviors of biological systems using abstract
interpretation. Electr. Notes Theor. Comput. Sci., 229(1):165–182, 2009.
[12] G. Cox, J. Crossley, and Z. Xing. Macrophage engulfment of apoptotic neutrophils contributes to the resolution of acute pulmonary inflammation in vivo. Am J Respir Cell Mol Biol., 12:232–237, 1995.
[13] V. Danos and C. Laneve. Formal molecular biology. Theor. Comput. Sci., 325(1):69–110, 2004.
[14] M. Dezani-Ciancaglini, P. Giannini, and A. Troina. A type system for a stochastic cls. In MeCBIC, volume 11
of EPTCS, pages 91–106, 2009.
[15] B. Fadeel and VE. Kagan. Apoptosis and macrophage clearance of neutrophils: regulation by reactive oxygen
species. Redox Rep., 8:143–150, 2003.
[16] GM. Ferreira and B. Spira. The pst operon of enteropathogenic escherichia coli enhances bacterial adherence
to epithelial cells. Microbiology, 154:2025–2036, 2008.
[17] D. Gillespie. Exact stochastic simulation of coupled chemical reactions. J. Phys. Chem., 81:2340–2361,
1977.
[18] J. Krivine, R. Milner, and A. Troina. Stochastic bigraphs. Electron. Notes Theor. Comput. Sci., 218:73–96,
2008.
[19] R. Lanotte, A. Maggiolo-Schettini, and A. Troina. Decidability results for parametric probabilistic transition
systems with an application to security. In SEFM’04, pages 114–121. IEEE Computer Society, 2004.
[20] R. Lanotte, A. Maggiolo-Schettini, and A. Troina. Parametric probabilistic transition systems for system
design and analysis. Formal Asp. Comput., 19(1):93–109, 2007.
[21] H. Lodish, A. Berk, L.S. Zipursky, P. Matsudaira, D. Baltimore, and J. Darnell. Many Bacterial Responses
Are Controlled by Two-Component Regulatory Systems Molecular Cell Biology. W.H. Freeman, 2000.
98
Stochastic Calculus of Wrapped Compartments
[22] H. Matsuno, A. Doi, M. Nagasaki, and S. Miyano. Hybrid petri net representation of gene regulatory network.
In Prooceedings of Pacific Symposium on Biocomputing, pages 341–352. World Scientific Press, 2000.
[23] C. Priami. Stochastic pi-calculus. Comput. J., 38(7):578–589, 1995.
[24] C. Priami, A. Regev, E. Y. Shapiro, and W. Silverman. Application of a stochastic name-passing calculus to
representation and simulation of molecular processes. Inf. Process. Lett., 80(1):25–31, 2001.
[25] G. Pǎun. Membrane computing. An introduction. Springer, 2002.
[26] A. Regev and E. Shapiro. Cells as computation. Nature, 419:343, 2002.
[27] G. Scatena. Development of a stochastic simulator for biological systems based on the calculus of looping
sequences. Master’s thesis, University of Pisa, 2007.
[28] SJ. Singer and GL. Nicolson. The fluid mosaic model of the structure of cell membranes. Science, 175:720–
731, 1972.
[29] EF. Vernon-Wilson, F. Aurad, L. Tian, IC. Rowe, MJ. Shipston, J. Savill, and SB. Brown. CD31 delays
phagocyte membrane repolarization to promote efficient binding of apoptotic cells. J Leukoc Biol., 82:1278–
1288, 1995.
| 5cs.CE
|
arXiv:1405.3883v1 [cs.PL] 15 May 2014
Under consideration for publication in Theory and Practice of Logic Programming
1
Analysis and Transformation Tools for Constrained
Horn Clause Verification∗
John P. Gallagher
Roskilde University, Denmark and IMDEA Software Institute, Madrid, Spain
(e-mail: [email protected])
Bishoksan Kafle
Roskilde University, Denmark
(e-mail: [email protected])
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
Several techniques and tools have been developed for verification of properties expressed as
Horn clauses with constraints over a background theory (CHC). Current CHC verification tools
implement intricate algorithms and are often limited to certain subclasses of CHC problems.
Our aim in this work is to investigate the use of a combination of off-the-shelf techniques from
the literature in analysis and transformation of Constraint Logic Programs (CLPs) to solve
challenging CHC verification problems. We find that many problems can be solved using a
combination of tools based on well-known techniques from abstract interpretation, semanticspreserving transformations, program specialisation and query-answer transformations. This gives
insights into the design of automatic, more general CHC verification tools based on a library of
components.
KEYWORDS: Constraint Logic Program, Constrained Horn Clause, Abstract Interpretation,
Software Verification.
1 Introduction
CHCs provide a suitable intermediate form for expressing the semantics of a variety
of programming languages (imperative, functional, concurrent, etc.) and computational
models (state machines, transition systems, big- and small-step operational semantics,
Petri nets, etc.). As a result it has been used as a target language for software verification.
Recently there is a growing interest in CHC verification from both the logic programming
and software verification communities, and several verification techniques and tools have
been developed for CHC.
Pure CLPs are syntactically and semantically the same as CHC. The main difference is that sets of constrained Horn clauses are not necessarily intended for execution,
∗ The research leading to these results has received funding from the European Union 7th Framework
Programme under grant agreement no. 318337, ENTRA - Whole-Systems Energy Transparency and
the Danish Natural Science Research Council grant NUSA: Numerical and Symbolic Abstractions for
Software Model Checking.
2
John P. Gallagher and Bishoksan Kafle
but rather as specifications. From the point of view of verification, we do not distinguish between CHC and pure CLP. Much research has been carried out on the analysis
and transformation of CLP programs, typically for synthesising efficient programs or for
inferring run-time properties of programs for the purpose of debugging, compile-time
optimisations or program understanding. In this paper we investigate the application of
this research to the CHC verification problem.
In Section 2 we define the CHC verification problem. In Section 3 we define basic
transformation and analysis components drawn from or inspired by the CLP literature.
Section 4 discusses the role of these components in verification, illustrating them on an
example problem. In Section 5 we construct a tool-chain out of these components and test
it on a range of CHC verification benchmark problems. The results reported represent
one of the main contributions of this work. In Section 6 we propose possible extensions
of the basic tool-chain and compare them with related work on CHC verification tool
architectures. Finally in Section 7 we summarise the conclusions from this work.
2 Background: The CHC Verification Problem
A CHC is a first order predicate logic formula of the form ∀(φ ∧ B1 (X1 ) ∧ . . . ∧ Bk (Xk ) →
H(X)) (k ≥ 0), where φ is a conjunction of constraints with respect to some background
theory, Xi , X are (possibly empty) vectors of distinct variables, B1 , . . . , Bk , H are predicate symbols, H(X) is the head of the clause and φ ∧ B1 (X1 ) ∧ . . . ∧ Bk (Xk ) is the
body. Sometimes the clause is written H(X) ← φ ∧ B1 (X1 ), . . . , Bk (Xk ) and in concrete
examples it is written in the form H :- φ, B1 (X1 ),. . .,Bk (Xk ). In examples, predicate
symbols start with lowercase letters while we use uppercase letters for variables.
We assume here that the constraint theory is linear arithmetic with relation symbols
≤, ≥, >, < and = and that there is a distinguished predicate symbol false which is
interpreted as false. In practice the predicate false only occurs in the head of clauses;
we call clauses whose head is false integrity constraints, following the terminology of
deductive databases. Thus the formula φ1 ← φ2 ∧ B1 (X1 ), . . . , Bk (Xk ) is equivalent to
the formula false ← ¬φ1 ∧ φ2 ∧ B1 (X1 ), . . . , Bk (Xk ). The latter might not be a CHC
but can be converted to an equivalent set of CHCs by transforming the formula ¬φ1 and
distributing any disjunctions that arise over the rest of the body. For example, the formula
X=Y :- p(X,Y) is equivalent to the set of CHCs false :- X>Y, p(X,Y) and false :X<Y, p(X,Y). Integrity constraints can be viewed as safety properties. If a set of CHCs
encodes the behaviour of some system, the bodies of integrity constraints represent unsafe
states. Thus proving safety consists of showing that the bodies of integrity constraints
are false in all models of the CHC clauses.
The CHC verification problem. To state this more formally, given a set of CHCs P , the
CHC verification problem is to check whether there exists a model of P . We restate this
property in terms of the derivability of the predicate false.
Lemma 2.1
P has a model if and only if P 6|= false.
Tools for Constrained Horn Clause Verification
3
Proof
Let us write I(F ) to mean that interpretation I satisfies F (I is a model of F ).
P 6|= false
≡ ∃I.(I(P ) and ¬I(false))
≡ ∃I.I(P ) (since ¬I(false) is true by defn. of false)
≡ P has a model.
Obviously any model of P assigns false to the bodies of integrity constraints.
The verification problem can be formulated deductively rather than model-theoretically.
Let the relation P ⊢ A denote that A is derivable from P using some proof procedure.
If the proof procedure is sound and complete then P 6|= A if and only if P 6⊢ A. So the
verification problem is to show (using CLP terminology) that the computation of the
goal ← false in program P does not succeed using a complete proof procedure. Although
in this work we follow the model-based formulation of the problem, we exploit the equivalence with the deductive formulation, which underlies, for example, the query-answer
transformation and specialisation techniques to be presented.
2.1 Representation of Interpretations
An interpretation of a set of CHCs is represented as a set of constrained facts of the
form A ← C where A is an atomic formula p(Z1 , . . . , Zn ) where Z1 , . . . , Zn are distinct
variables and C is a constraint over Z1 , . . . , Zn . If C is true we write A ← or just A.
The constrained fact A ← C is shorthand for the set of variable-free facts Aθ such that
Cθ holds in the constraint theory, and an interpretation M denotes the set of all facts
denoted by its elements; M assigns true to exactly those facts. M1 ⊆ M2 if the set of
denoted facts of M1 is contained in the set of denoted facts of M2 .
Minimal models. A model of a set of CHCs is an interpretation that satisfies each clause.
There exists a minimal model with respect to the subset ordering, denoted M [[P ]] where
P is the set of CHCs. M [[P ]] can be computed as the least fixed point (lfp) of an immediate consequences operator, TPC , which is an extension of the standard TP operator from
logic programming, extended to handle constraints (Jaffar and Maher 1994). Furthermore lfp(TPC ) can be computed as the limit of the ascending sequence of interpretations
∅, TPC (∅), TPC (TPC (∅)), . . .. For more details, see (Jaffar and Maher 1994). This sequence
provides a basis for abstract interpretation of CHC clauses.
Proof by over-approximation of the minimal model. It is a standard theorem of CLP that
the minimal model M [[P ]] is equivalent to the set of atomic consequences of P . That is,
P |= p(v1 , . . . , vn ) if and only if p(v1 , . . . , vn ) ∈ M [[P ]]. Therefore, the CHC verification
problem for P is equivalent to checking that false 6∈ M [[P ]]. It is sufficient to find a set of
constrained facts M ′ such that M [[P ]] ⊆ M ′ , where false 6∈ M ′ . This technique is called
proof by over-approximation of the minimal model.
3 Relevant tools for CHC Verification
In this section, we give a brief description of some relevant tools borrowed from the
literature in analysis and transformation of CLP.
4
John P. Gallagher and Bishoksan Kafle
Unfolding. Let P be a set of CHCs and c0 ∈ P be H(X) ← B1 , p(Y ), B2 where B1 , B2
are possibly empty conjunctions of atomic formulas and constraints. Let {c1 , . . . , cm } be
the set of clauses of P that have predicate p in the head, that is, ci = p(Zi ) ← Di ,
where the variables of these clauses are standardised apart from the variables of c0 and
from each other. Then the result of unfolding c0 on p(Y ) is the set of CHCs P ′ =
P \ {c0 } ∪ {c′1 , . . . , c′m } where c′i = H(X) ← B1 , Y = Zi , Di , B2 . The equality Y = Zi
stands for the conjunction of the equality of the respective elements of the vectors Y
and Zi . It is a standard result that unfolding a clause in P preserves P ’s minimal model
(Pettorossi and Proietti 1999). In particular, P |= false ≡ P ′ |= false.
Specialisation. A set of CHCs P can be specialised with respect to a query. Assume A is
an atomic formula; then we can derive a set PA such that P |= A ≡ PA |= A. PA could
be simpler than P , for instance, parts of P that are irrelevant to A could be omitted in
PA . In particular, the CHC verification problem for Pfalse and P are equivalent. There
are many techniques in the CLP literature for deriving a specialised program PA . Partial
evaluation is a well-developed method (Gallagher 1993; Leuschel 1999).
We make use a form of specialisation know as forward slicing, more specifically redundant argument filtering (Leuschel and Sørensen 1996), in which predicate arguments
can be removed if they do not affect a computation. Given a set of CHCs P and a
query A, denote by PAraf the program obtained by applying the RAF algorithm from
(Leuschel and Sørensen 1996) with respect to the goal A. We have the property that
raf
P |= A ≡ PAraf |= A and in particular that P |= false ≡ Pfalse
|= false.
Query-answer transformation. Given a set of CHCs P and an atomic query A, the queryanswer transformation of P with respect to A is a set of CHCs which simulates the
computation of the goal ← A in P , using a left-to-right computation rule. Query-answer
transformation is a generalisation of the magic set transformations for Datalog. For each
predicate p, two new predicates pans and pquery are defined. For an atomic formula A,
Aans and Aquery denote the replacement of A’s predicate symbol p by pans and pquery
respectively. Given a program P and query A, the idea is to derive a program PAqa with
the following property P |= A iff PAqa |= Aans . The Aquery predicates represent calls in
the computation tree generated during the execution of the goal. For more details see
(Debray and Ramakrishnan 1994; Gallagher and de Waal 1993; Codish and Demoen 1993).
qa
In particular, Pfalse
|= falseans ≡ P |= false, so we can transform a CHC verification problem to an equivalent CHC verification problem on the query-answer program generated
with respect to the goal ← false.
Predicate splitting. Let P be a set of CHCs and let {c1 , . . . , cm } be the set of clauses in P
having some given predicate p in the head, where ci = p(X) ← Di . Let C1 , . . . , Ck be some
partition of {c1 , . . . , cm }, where Cj = {cj1 , . . . , cjnj }. Define k new predicates p1 . . . pk ,
where pj is defined by the bodies of clauses in partition Cj , namely Q j = {pj (X) ←
Dj1 , . . . , pj (X) ← Djnj }. Finally, define k clauses Cp = {p(X) ← p1 (X), . . . , p(X) ←
pk (X)}. Then we define a splitting transformation as follows.
1. Let P ′ = P \ {c1 , . . . , cm } ∪ Cp ∪ Q 1 ∪ . . . ∪ Q k .
2. Let P split be the result of unfolding every clause in P ′ whose body contains p(Y )
with the clauses Cp .
Tools for Constrained Horn Clause Verification
5
In our applications, we use splitting to create separate predicates for clauses for a
given predicate whose constraints are mutually exclusive. For example, given the clauses
new3(A,B) :- A=<99, new4(A,B) and new3(A,B) :- A>=100, new5(A,B), we produce
two new predicates, since the constraints A=<99 and A>=100 are disjoint. The new predicates are defined by clauses new31(A,B) :- A=<99, new4(A,B) and new32 (A,B) :A>=100, new5(A,B), and all calls to new3 throughout the program are unfolded using
these new clauses. Splitting has been used in the CLP literature to improve the precision
of program analyses, for example in (Serebrenik and De Schreye 2001). In our case it
improves the precision of the convex polyhedron analysis discussed below, since separate
polyhedra will be maintained for each of the disjoint cases. The correctness of splitting can
be shown using standard transformations that preserve the minimal model of the program
(with respect to the predicates of the original program) (Pettorossi and Proietti 1999).
Assuming that the predicate false is not split, we have that P |= false ≡ P split |= false.
Convex polyhedron approximation. Convex polyhedron analysis (Cousot and Halbwachs 1978)
is a program analysis technique based on abstract interpretation (Cousot and Cousot 1977).
When applied to a set of CHCs P it constructs an over-approximation M ′ of the minimal model of P , where M ′ contains at most one constrained fact p(X) ← C for each
predicate p. The constraint C is a conjunction of linear inequalities, representing a convex polyhedron. The first application of convex polyhedron analysis to CLP was by
Benoy and King (1996). Since the domain of convex polyhedra contains infinite increasing chains, the use of a widening operator is needed to ensure convergence of the abstract
interpretation. Furthermore much research has been done on improving the precision of
widening operators. One technique is known as widening-upto, or widening with thresholds (Halbwachs et al. 1994).
Recently, a technique for deriving more effective thresholds was developed (Lakhdar-Chaouch et al. 2011),
which we have adapted and found to be effective in experimental studies. The thresholds
are computed by the following method. Let TPC be the standard immediate consequence
operator for CHCs, that is, TPC (I) is the set of constrained facts that can be derived
in one step from a set of constrained facts I. Given a constrained fact p(Z̄) ← C, define atomconstraints(p(Z̄) ← C) to be the set of constrained facts {p(Z̄) ← Ci | C =
C1 ∧ . . . ∧ Ck , 1 ≤ i ≤ k)}. The function atomconstraints is extended to interpretations
S
by atomconstraints(I) = p(Z̄)←C∈I {atomconstraints(p(Z̄) ← C)}.
Let I⊤ be the interpretation consisting of the set of constrained facts p(Z̄) ← true
for each predicate p. We perform three iterations of TPC starting with I⊤ (the first three
elements of a “top-down” Kleene sequence) and then extract the atomic constraints.
That is, thresholds is defined as follows.
C(3)
thresholds(P ) = atomconstraints(TP
(I⊤ ))
A difference from the method in (Lakhdar-Chaouch et al. 2011) is that we use the concrete semantic function TPC rather than the abstract semantic function when computing
thresholds. The set of threshold constraints represents an attempt to find useful predicate
properties and when widening they help to preserve invariants that might otherwise be
lost during widening. See (Lakhdar-Chaouch et al. 2011) for further details. Threshold
constraints that are not invariants are simply discarded during widening.
6
new6(A,B)
new5(A,B)
new5(A,B)
new4(A,B)
John P. Gallagher and Bishoksan Kafle
::::-
B=<99.
B>=101.
B=<100, new6(A,B).
C=1+A, A=<49, new3(C,B).
new4(A,B) :- C=1+A,D=1+B,A>=50,new3(C,D).
new3(A,B) :- A=<99, new4(A,B).
new3(A,B) :- A>=100, new5(A,B).
false :- A=0, B=50, new3(A,B).
Fig. 1. The example program MAP-disj.c.map.pl
4 The role of CLP tools in verification
The techniques discussed in the previous section play various roles. The convex polyhedron analysis, together with the helper tool to derive threshold constraints, constructs
an approximation of the minimal model of a CHC theory. If false (or falseans ) is not in
the approximate model, then the verification problem is solved. Otherwise the problem is
not solved; in effect a “don’t know” answer is returned. We have found that polyhedron
analysis alone is seldom precise enough to solve non-trivial CHC verification problems;
in combination with the other tools, it is very effective.
Unfolding can improve the structure of a program, removing some cases of mutual
recursion, or propagating constraints upwards towards the integrity constraints, and can
improve the precision and performance of convex polyhedron analysis.
Specialisation can remove parts of theories not relevant to the verification problem,
and can also propagate constraint downwards from the integrity constraints. Both of
these have a beneficial effect on performance and precision of polyhedron analysis.
Analysis of a query-answer program (with respect to false) is in effect the search for a
derivation tree for false. Its effectiveness in CHC verification problems is variable. It can
sometimes worsen performance since the query-answer transformed program is larger and
contains more recursive dependencies than the original. On the other hand, one seldom
loses precision and it is often more effective in allowing constraints to be propagated
upwards (through the ans predicates) and downwards (through the query predicates).
4.1 Application of the tools
We illustrate the tools on a running example (Figure 1), one of the benchmark suite of the
VeriMAP system De Angelis et al. (2013). The result of applying unfolding is shown in
Figure 2 (omitting the definitions of the unfolded predicates new4, new5 and new6, which
are no longer reachable from false). The unfolding strategy we adopt is the following:
the predicate dependency graph of a program consists of the set of edges (p, q) such that
there is clause where p is the predicate of the head and q is a predicate occurring in the
body. We perform a depth-first search of the predicate dependency graph, starting from
false, and identify the backward edges, namely those edges (p, q) where q is an ancestor
of p in the depth-first search. We then unfold every body call whose predicate is not at
the end of a backward edge. In Figure 1, we thus unfold calls to new4, new5 and new6.
The query-answer transformation is applied to the program in Figure 2, with respect
to the goal false resulting in the program shown in Figure 3. The model of the predicate
new3 query corresponds to those calls to new3 that are reachable from the call in the
integrity constraint. Explicit representation of the query predicates permits more effective
propagation of constraints from the integrity clauses during model approximation.
The splitting transformation is now applied to the program in Figure 3. We do not
Tools for Constrained Horn Clause Verification
7
false :- A=0, B=50, new3(A,B).
new3(A,B) :- A=<99, C = 1+A, A=<49, new3(C,B).
new3(A,B) :- A=<99, C = 1+A, D = 1+B, A>=50, new3(C,D).
new3(A,B) :- A>=100, B>=101.
new3(A,B) :- A>=100, B=<100, B=<99.
Fig. 2. Result of unfolding MAP-disj.c.map.pl
false ans :- false query, A=0, B=50, new3 ans(A,B).
new3 ans(A,B) :- new3 query(A,B), A=<99, C = 1+A, A=<49, new3 ans(C,B).
new3 ans(A,B) :- new3 query(A,B),A=<99,C is 1+A,D is 1+B, A>=50, new3 ans(C,D).
new3 ans(A,B) :- new3 query(A,B), A>=100, B>=101.
new3 ans(A,B) :- new3 query(A,B), A>=100, B=<100, B=<99.
new3 query(A,B) :- false query, A=0, B=50.
new3 query(A,B) :- new3 query(C,B), C=<99, A = 1+C, C=<49.
new3 query(A,B) :- new3 query(C,D), C=<99, A = 1+C, B = 1+D, C>=50.
false query.
Fig. 3. The query-answer transformed program for program of Figure 2
show the complete program, which contains 30 clauses. Figure 4 shows the split definition
of new3 query, which is split since the last two clauses for new3 query in Figure 3 have
mutually disjoint constraints, when projected onto the head variables.
A convex polyhedron approximation is then computed for the split program, after
computing threshold constraints for the predicates. The resulting approximate model is
shown in Figure 5 as a set of constrained facts. Since the model does not contain any
constrained fact for false ans we conclude that false ans is not a consequence of the
split program. Hence, applying the various correctness results for the unfolding, queryanswer and splitting transformations, false is not a consequence of the original program.
Discussion of the example. Application of the convex polyhedron tool to the original, or
the intermediate programs, does not solve the problem; all the transformations are needed
in this case, apart from redundant argument filtering, which only affects efficiency. The
ordering of the tool-chain can be varied somewhat, for instance switching query-answer
transformation with splitting or unfolding. In our experiments we found the ordering in
Figure 6 to be the most effective.
The model of the query-answer program is finite for this example. However, the problem is essentially the same if the constants are scaled; for instance we could replace 50
by 5000, 49 by 4999, 100 by 10000 and 101 by 10001, and the problem is essentially
unchanged. We noted that some CHC verification tools applied to this example solve
the problem, but essentially by enumeration of the finite set of values encountered in the
new3
new3
new3
new3
new3
query
query
query
query
query
1(A,B)
1(A,B)
1(A,B)
2(A,B)
2(A,B)
:::::-
false query
new3 query
new3 query
new3 query
new3 query
1, A=0, B=50.
1(C,B), C=<99,
2(C,B), C=<99,
1(C,D), C=<99,
2(C,D), C=<99,
A
A
A
A
=
=
=
=
Fig. 4. Part of the split program for the program in Figure 3
1+C,
1+C,
1+C,
1+C,
C=<49.
C=<49.
B = 1+D, C>=50.
B = 1+D, C>=50.
8
John P. Gallagher and Bishoksan Kafle
false query 1 :- []
new3 query 1(A,B) :- [1*A>=0,-1*A>= -50,1*B=50]
new3 query 2(A,B) :- [1*A>=51,-1*A>= -100,1*A+ -1*B=0]
Fig. 5. The convex polyhedral approximate model for the split program
search. Such a solution does not scale well. On the other hand the polyhedral abstraction shown above is not an enumeration; an essentially similar polyhedron abstraction
is generated for the scaled version of the example, in the same time. The VeriMAP tool
(De Angelis et al. 2013) also handles the original and scaled versions of the example in
the same time.
CHC Program P
RAF
FU
RAF – Redundant Argument Filtering
FU – Forward Unfolding
QA – Query Answer Transformation
QA
PS
TC
CHA
PS – Predicate Splitting
TC – Threshold Constraint
CHA – Convex Hull Analyzer
Safe
unknown
Fig. 6. The basic tool chain for CHC verification.
5 Combining off-the-shelf tools: Experiments
The motivation for our tool-chain, summarised in Figure 6, comes from our example program, which is a simple yet challenging program. We applied the tool-chain to a number of
benchmarks from the literature, taken mainly from the repository of Horn clause benchmarks in SMT-LIB2 (https://svn.sosy-lab.org/software/sv-benchmarks/trunk/clauses/)
and other sources including (Gange et al. 2013) and some of the VeriMap benchmarks
(De Angelis et al. 2013). We selected these examples because many of them are considered challenging because they cannot be solved by one or more of the state-of-the-artverification tools discussed below. Programs taken from the SMT-LIB2 repository are
first translated to CHC form. The results are summarised in Table 1.
In Table 1, columns Program and Result respectively represent the benchmark program and the results of verification using our tool combination. Problems marked with
(*) could not be handled by our tool-chain since they contain numbers which do not
fit in 32 bits, the limit of our Ciao Prolog implementation. whereas problems marked
with (**) are solvable by simple ad hoc modification of the tool-chain, which we are
currently investigating (see Section 7). Problems such as systemc-token-ring.01-safeil.c
contain complicated loop structure with large strongly connected components in the
predicate dependency graph and our convex polyhedron analysis tool is unable to derive the required invariant. However overall results show that our simple tool-chain begins to compete with advanced tools like HSF (Grebenshchikov et al. 2012), VeriMAP
(De Angelis et al. 2013), TRACER (Jaffar et al. 2012), etc. We do not report timings,
Tools for Constrained Horn Clause Verification
9
though all these results are obtained in a matter of seconds, since our tool-chain is not
at all optimised, relying on file input-output and the individual components are often
prototypes.
Table 1. Experiments results on CHC benchmark program
SN
Program
Result
SN
Program
Result
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
MAP-disj.c.map.pl
MAP-disj.c.map-scaled.pl
t1.pl
t1-a.pl
t2.pl
t3.pl
t4.pl
t5.pl
pldi12.pl
INVGEN-id-build
INVGEN-nested5
INVGEN-nested6
INVGEN-nested8
INVGEN-svd-some-loop
INVGEN-svd1
INVGEN-svd4
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
17
18
19
20
21
22
23
24
25
26**
27**
28*
29
30*
31
32*
MAP-forward.c.map.pl
tridag.smt2
qrdcmp.smt2
choldc.smt2
lop.smt2
pzextr.smt2
qrsolv.smt2
INVGEN-apache-escape-absolute
TRACER-testabs15
amebsa.smt2
DAGGER-barbr.map.c
sshsimpl-s3-srvr-1a-safeil.c
sshsimpl-s3-srvr-1b-safeil.c
bandec.smt2
systemc-token-ring.01-safeil.c
crank.smt2
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
verified
NOT
NOT
NOT
NOT
NOT
CHC Program P
PA – Predicate Abstraction
RAF
FU
QA
PS
TC
CHA
PA
Safe
CEx.
unknown
props
Fig. 7. Future extension of our tool-chain.
6 Discussion and Related Work
The most similar work to ours is by De Angelis et al. (2013) which is also based on CLP
program transformation and specialisation. They construct a sequence of transformations
of P , say, P, P1 , P2 , . . . , Pk ; if Pk contains no clause with head false then the verification
problem is solved. A proof of unsafety is obtained if Pk contains a clause false ←. Both
our approach and theirs repeatedly apply specialisations preserving the property to be
10
John P. Gallagher and Bishoksan Kafle
proved. However the difference is that their specialisation techniques are based on unfoldfold transformations, with a sophisticated control procedure controlling unfolding and
generalisation. Our specialisations are restricted to redundant argument filtering and the
query-answer transformation, which specialises predicate answers with respect to a goal.
Their test for success or failure is a simple syntactic check, whereas ours is based on an
abstract interpretation to derive an over-approximation. Informally one can say that the
hard work in their approach is performed by the specialisation procedure, whereas the
hard work in our approach is done by the abstract interpretation. We believe that our
tool-chain-based approach gives more insight into the role of each transformation.
Work by Gange et al. (2013) is a top-town evaluation of CLP programs which records
certain derivations and learns only from failed derivations. This helps to prune further
derivations and helps to achieve termination in the presence of infinite executions. Duality
(http://research.microsoft.com/en-us/projects/duality/) and HSF(C) (Grebenshchikov et al. 2012)
are examples of the CEGAR approach (Counter-Example-Guided Abstraction Refinement). This approach can be viewed as property-based abstract interpretation based on
a set of properties that is refined on each iteration. The refinement of the properties is
the key problem in CEGAR; an abstract proof of unsafety is used to generate properties
(often using interpolation) that prevent that proof from arising again. Thus, abstract
counter-examples are successively eliminated. The relatively good performance of our
tool-chain, without any refinement step at all, suggests that finding the right invariants
is aided by a tool such as the convex polyhedron solver and the pre-processing steps we
applied. In Figure 7 we sketch possible extensions of our basic tool-chain, incorporating
a refinement loop and property-based abstraction.
It should be noted that the query-answer transformation, predicate splitting and unfolding may all cause an blow-up in the program size. The convex polyhedron analysis
becomes more effective as a result, but for scalability we need more sophisticated heuristics controlling these transformations, especially unfolding and splitting, as well as lazy or
implicit generation of transformed programs, using techniques such as a fixpoint engine
that simulates query-answer programs (Codish 1999).
7 Concluding remarks and future work
We have shown that a combination of off-the-shelf tools from CLP transformation and
analysis, combined in a sensible way, is surprisingly effective in CHC verification. The
component-based approach allowed us to experiment with the tool-chain until we found
an effective combination. This experimentation is continuing and we are confident of
making improvements by incorporating other standard techniques and by finding better heuristics for applying the tools. Further we would like to investigate the choice of
chain suitable for each example since more complicated problems can be handled just
by altering the chain. We also suspect from initial experiments that an advanced partial
evaluator such as ECCE (Leuschel et al. 2006) will play a useful role. Our results give
insights for further development of automatic CHC verification tools. We would like to
combine our program transformation techniques with abstraction refinement techniques
and experiment with the combination.
Tools for Constrained Horn Clause Verification
11
References
Benoy, F. and King, A. 1996. Inferring argument size relationships with CLP(R). In LogicBased Program Synthesis and Transformation (LOPSTR’96), J. P. Gallagher, Ed. Lecture
Notes in Computer Science, vol. 1207. Springer, 204–223.
Codish, M. 1999. Efficient goal directed bottom-up evaluation of logic programs. J. Log.
Program. 38, 3, 355–370.
Codish, M. and Demoen, B. 1993. Analysing logic programs using “Prop”-ositional logic
programs and a magic wand. In Proceedings of the 1993 International Symposium on Logic
Programming, Vancouver, D. Miller, Ed. MIT Press.
Cousot, P. and Cousot, R. 1977. Abstract interpretation: A unified lattice model for static
analysis of programs by construction or approximation of fixpoints. In POPL, R. M. Graham,
M. A. Harrison, and R. Sethi, Eds. ACM, 238–252.
Cousot, P. and Halbwachs, N. 1978. Automatic discovery of linear restraints among variables of a program. In Proceedings of the 5th Annual ACM Symposium on Principles of
Programming Languages. ACM Press, 84–96.
De Angelis, E., Fioravanti, F., Pettorossi, A., and Proietti, M. 2013. Verification of
imperative programs by transforming constraint logic programs. In CILC, D. Cantone and
M. N. Asmundo, Eds. CEUR Workshop Proceedings, vol. 1068. CEUR-WS.org, 83–98.
Debray, S. and Ramakrishnan, R. 1994. Abstract Interpretation of Logic Programs Using
Magic Transformations. Journal of Logic Programming 18, 149–176.
Gallagher, J. P. 1993. Specialisation of logic programs: A tutorial. In Proceedings PEPM’93,
ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based Program Manipulation. ACM Press, Copenhagen, 88–98.
Gallagher, J. P. and de Waal, D. 1993. Deletion of redundant unary type predicates from
logic programs. In Logic Program Synthesis and Transformation, K. Lau and T. Clement,
Eds. Workshops in Computing. Springer-Verlag, 151–167.
Gange, G., Navas, J. A., Schachte, P., Søndergaard, H., and Stuckey, P. J. 2013.
Failure tabled constraint logic programming by interpolation. TPLP 13, 4-5, 593–607.
Grebenshchikov, S., Gupta, A., Lopes, N. P., Popeea, C., and Rybalchenko, A. 2012.
HSF(C): A software verifier based on Horn clauses - (competition contribution). In TACAS,
C. Flanagan and B. König, Eds. LNCS, vol. 7214. Springer, 549–551.
Halbwachs, N., Proy, Y. E., and Raymound, P. 1994. Verification of linear hybrid systems
by means of convex approximations. In Proceedings of the First Symposium on Static Analysis.
Lecture Notes in Computer Science, vol. 864. Springer, 223–237.
Jaffar, J. and Maher, M. 1994. Constraint Logic Programming: A Survey. Journal of Logic
Programming 19/20, 503–581.
Jaffar, J., Murali, V., Navas, J. A., and Santosa, A. E. 2012. TRACER: A symbolic
execution tool for verification. In CAV, P. Madhusudan and S. A. Seshia, Eds. Lecture Notes
in Computer Science, vol. 7358. Springer, 758–766.
Lakhdar-Chaouch, L., Jeannet, B., and Girault, A. 2011. Widening with thresholds for
programs with complex control graphs. In ATVA 2011, T. Bultan and P.-A. Hsiung, Eds.
Lecture Notes in Computer Science, vol. 6996. Springer, 492–502.
Leuschel, M. 1999. Advanced logic program specialisation. In Partial Evaluation - Practice
and Theory, J. Hatcliff, T. Æ. Mogensen, and P. Thiemann, Eds. Lecture Notes in Computer
Science, vol. 1706. Springer, 271–292.
Leuschel, M., Elphick, D., Varea, M., Craig, S.-J., and Fontaine, M. 2006. The Ecce
and Logen partial evaluators and their web interfaces. In PEPM 2006, J. Hatcliff and F. Tip,
Eds. ACM, 88–94.
Leuschel, M. and Sørensen, M. H. 1996. Redundant argument filtering of logic programs. In
Logic Programming Synthesis and Transformation, 6th International Workshop, LOPSTR’96,
12
John P. Gallagher and Bishoksan Kafle
Stockholm, Sweden, August 28-30, 1996, Proceedings, J. P. Gallagher, Ed. Lecture Notes in
Computer Science, vol. 1207. Springer, 83–103.
Pettorossi, A. and Proietti, M. 1999. Synthesis and transformation of logic programs using
unfold/fold proofs. J. Log. Program. 41, 2-3, 197–230.
Serebrenik, A. and De Schreye, D. 2001. Inference of termination conditions for numerical
loops in Prolog. In LPAR 2001, R. Nieuwenhuis and A. Voronkov, Eds. Lecture Notes in
Computer Science, vol. 2250. Springer, 654–668.
| 6cs.PL
|
Modeling and identification of uncertain-input systems
Riccardo Sven Risuleo
arXiv:1709.03421v1 [cs.SY] 11 Sep 2017
ACCESS Linnaeus Center, KTH Royal Institute of Technology, Sweden
[email protected]
Giulio Bottegal
Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands
[email protected]
Håkan Hjalmarsson
ACCESS Linnaeus Center, KTH Royal Institute of Technology, Sweden
[email protected]
September 12, 2017
Abstract
In this work, we present a new class of models, called uncertain-input models, that allows
us to treat system-identification problems in which a linear system is subject to a partially
unknown input signal. To encode prior information about the input or the linear system, we
use Gaussian-process models. We estimate the model from data using the empirical Bayes
approach: the input and the impulse responses of the linear system are estimated using the
posterior means of the Gaussian-process models given the data, and the hyperparameters
that characterize the Gaussian-process models are estimated from the marginal likelihood of
the data. We propose an iterative algorithm to find the hyperparameters that relies on the
EM method and results in simple update steps. In the most general formulation, neither the
marginal likelihood nor the posterior distribution of the unknowns is tractable. Therefore, we
propose two approximation approaches, one based on Markov-chain Monte Carlo techniques
and one based on variational Bayes approximation. We also show special model structures
for which the distributions are treatable exactly. Through numerical simulations, we study
the application of the uncertain-input model to the identification of Hammerstein systems
and cascaded linear systems. As part of the contribution of the paper, we show that this
model structure encompasses many classical problems in system identification such as classical PEM, Hammerstein models, errors-in-variables problems, blind system identification,
and cascaded linear systems. This allows us to build a systematic procedure to apply the
algorithms proposed in this work to a wide class of classical problems.
1
Introduction
In most system identification problems, the input signal—that is, the independent variable—is
perfectly known [24]. Often, the input signal is the result of an identification experiment, where a
signal with certain characteristics is designed and applied to the system to measure its response.
However, in some applications, the hypothesis that the input signal is known may be too restrictive.
In this work, we propose a new model structure that accounts for partial knowledge about the input
signal and we show how many classical system identification problems can be seen as problems of
identifying instances of this model structure.
The proposed model structure, which we call uncertain-input model, is composed of a linear
time-invariant dynamical system (the linear system) and of a signal of which partial information
is available (the unknown input). In the next section, we characterize formally the unknown
1
input; before that, we give some examples of classical models that can be seen as uncertain-input
models. The Hammerstein model is a cascade composition of a static nonlinear function followed
by a linear time-invariant dynamical system [3, 20, 39]. In the Hammerstein model, the (perfectly
known) input signal passes through the unknown static nonlinear function. After the nonlinear
transformation, the signal which is fed to the linear system, is completely unknown. However, some
characteristics of the signal may be known; for instance, we may known that the nonlinear function
is smooth or we may have a set of candidate basis functions to choose among. Another instance of
a model where the input is not perfectly known is the errors-in-variables model [45]. In the errorsin-variables formulation, the input in known up to noisy measurements. The noise in the input
introduces many difficulties and special techniques have been developed to deal with it [43, 44].
Closely related to errors-in-variables models, blind system identification methods are used when the
input signal is completely unknown [1]. These are particularly useful in telecommunications, image
reconstruction, and biomedical applications [31, 30, 27]. Blind system identification problems are
generally ill posed, and certain assumptions on the input signal are needed to recover a solution [2].
Similar to blind problems are the problems of system identification with missing data. In these
cases, the missing data are estimated, together with a description of the system, by making
hypotheses on the mechanism that generated the missing data [47, 26, 34, 42, 23].
In all the applications we have outlined, we can identify the common thread of a linear system
fed by a signal about which we have limited prior information. This leads us naturally to consider
a Bayesian framework where we can use prior distributions to encode beliefs about the unknown
quantities [8, Section 2.4]. Within the vast framework of Bayesian methods, we concentrate on
Gaussian processes [38]. These enable us to compute many quantities in closed form and to reason about identification in terms of a limited number of sufficient statistics. For these reasons,
Gaussian-process modeling has become a popular approach in system identification [36, 17, 46].
Although Gaussian processes are typically analytically convenient, the structure of the uncertaininput problem leads to an intractable inference problem: even though we model the system and the
input as Gaussian processes, the output of the system depends on their convolution and therefore
does not admit a Gaussian description. To perform the inference—that is find the posterior distribution of the unknowns given the observations—we need approximation methods. We propose
two different approximation methods for the posterior distribution of the unknowns: one Markov
Chain Monte Carlo (MCMC, see [19]) method and one variational approximation [6] method. In
the MCMC method, we use the Gibbs sampler [18] to draw particles from the posterior distribution and we approximate expectations as averages computed with the particles. In the variational
method, we find the factorized distribution that best approximates the posterior distribution in
Kullback-Leibler distance.
To give flexibility to the model, we allow the Gaussian priors to depend on certain parameters
(called hyperparameters) that need to be estimated from data together with the measurement
noise variances. To estimate these parameters, we use the empirical Bayes method which requires
maximizing the marginal distribution of the data (sometimes called evidence, see [25]). To this
end, we propose an iterative algorithm based on the Expectation-Maximization (EM) method [14].
The EM method alternates between the computation of the expected value of the joint likelihood
of the data, of the unknown system, and of the input (E-step), and the maximization of this
expected value with respect to the unknown parameters (M-step). We show that the E-step can
be computed using the same approximations of the posterior distributions that are used in the
inference and that the M-step consists in a series of simple and independent optimization problems
that can be solved easily.
As mentioned above, the uncertain-input model encompasses several classical model structures that have been object of research in the system-identification community for decades. Two
important contributions of this work are as follows.
1. We unify the problems of identifying systems that are usually regarded as belonging to
different model classes into a single identification framework.
2. We formalize a method to apply the new tools of Gaussian processes and Bayesian inference
to classical system identification problems.
2
To support the validity of the proposed methods, we present identification experiments on
synthetic datasets of cascaded linear systems and of Hammerstein systems.
1.1
Notation
The notation [A]i,j indicates the element of matrix A in position i, j (single subscripts are used for
vectors). “TN ×n (v)” denotes the N by n lower-triangular Toeplitz matrix of the m dimensional
vector v:
(
vi−j+1 0 ≤ i − j + 1 ≤ m
TN ×n (v) i,j =
0
otherwise
If v is a vector, then V is the N by N Toeplitz matrix whose elements are given by v. The notation
“kak2M ” is shorthand for aT M a. The notation “N (α, Σ)” indicates the Gaussian distribution with
mean vector α and covariance matrix Σ. The notation “GP(µ, Σ)” indicate a Gaussian process
with mean function µ and covariance function Σ. Random variables and their realizations have the
same symbol. The notation “x; θ” indicates that the random variable x depends on the parameter
θ. If x is a random variable, p(x) denotes its density. The symbol “∼
=” indicates equality up to an
additive constant and “δ” is the Dirac density.
2
Uncertain-input systems
In this work, we propose a new model structure called the uncertain-input model. Consider the
block scheme in Figure 1. Many system identification tasks can be formulated as the identification
of a linear system S, subject to an input sequence {wt }. In this work, we consider problems in
which we have partial information about the input sequence, and this partial information depends
on the specific problem at hand.
information
wt
ηt
εt
+
S
+
yt
vt
Figure 1: A block scheme of the general uncertain input system.
We assume that the linear system S is time invariant, stable, and causal. Therefore, it is
uniquely described by the sequence {gt } of its impulse response samples, and the output of the
system generated by an input {wt } can be represented as the discrete convolution of the system
impulse response with the input signal—that is, at time t, the measurements of the output can be
written as the noise-corrupted discrete convolution
yt = (w ∗ g)t + εt ,
(1)
where {εt } is a stochastic process that describes additive measurement noise, and where “∗”
denotes the discrete time convolution
(w ∗ g)t =
∞
X
gk wt−k .
(2)
k=1
In the uncertain-input model, we consider that the input signal is measured with additive white
noise described by a stochastic processes {ηt }. This assumption allows us to write, for the input
measurements, the model
vt = wt + ηt .
(3)
3
We assume that the noise processes {ηt } and {εt } are independent Gaussian white-noise processes. This means that every noise sample has a Gaussian distribution,
ηt ∼ N (0, σv2 ),
εt ∼ N (0, σy2 ),
(4)
and that εt is independent of εs , for s 6= t, and of ηs for any s. To allow for models where some
observations are missing, we assume infinite variance for those noise components that correspond
to the missing samples.
To encode the prior information we have about the input signal and about the linear system,
we use Gaussian process models. We model the unknown input signal and the impulse response of
the linear system as a realization of a joint Gaussian processes with suitable mean and covariance
functions,
"
" #
#
#"
T
w
µ
(·;
θ)
Kw (·; θ)
Kgw (·, ·; ρ, θ)
.
(5)
∼ GP w
,
g
µg (·; ρ) Kgw (·, ·; ρ, θ)
Kg (·, ·; ρ)
The mean functions of the Gaussian processes, µg (· ; θ) and µw (· ; ρ), may depend on the parameter
vectors θ and ρ, called hyperparameter vectors, which can be used to shape the prior information
to the specific application. The same goes for the covariance functions Kw ( · , · ; θ), Kg ( · , · ; ρ),
and Kgw ( · , · ; ρ, θ) which may depend on (possibly different) hyperparameters.
For notational convenience, we present the explicit computations in the case of independent
Gaussian process models for g and w—that is, we consider the case where
w ∼ GP µw ( · ; θ), Kw ( · , · ; θ) ,
(6)
g ∼ GP µg ( · ; ρ), Kg ( · , · ; ρ) ,
and the cross-covariance of processes is zero. However, all results we show hold also in the more
general case.
We assume that we have collected N measurements of the processes {vt } and {yt } and, for
sake of simplicity, we also assume that wt = 0 for t < 0 (see [41] for a way to extend the proposed
framework to unknown initial conditions). From (1), we see that the output measurements only
depend on the values of the impulse response at the discrete time instants t = 1, 2, . . . , N ; therefore,
we can consider the joint distribution of the samples gt for t = 1, 2, . . . , N . From the Gaussian
process model (6), we have that, if we collect the samples of {gt } into an N -dimensional column
vector g, this vector has a joint Gaussian distribution given by
g ∼ N µg (ρ), Kg (ρ) ,
(7)
where we have defined the mean vector and the covariance matrix induced by (6) as
µg (ρ) j := µg (j ; ρ),
Kg (ρ) i,j := Kg (i, j ; ρ).
From (3) and (1), we have that the N measurements of the input and output only depend on
the samples wt for t = 1, . . . , N ; therefore, we can consider the joint distribution of these samples,
collected in an N -dimensional vector w. This distribution is Gaussian, and it is given by
w ∼ N µw (θ), Kw (θ) ,
(8)
where we have defined the mean vector and the covariance matrix induced by (6) as
µw (θ) j := µw (j ; θ),
Kw (θ) i,j := Kw (i, j ; θ).
Assembling the models for the different components, given by (1), (3), (4), (7), and (8), we
4
arrive at the following definition of the uncertain-input model:
y = Wg + ε
v =w+η
g ∼ N µg (ρ), Kg (ρ)
w ∼ N µw (θ), Kw (θ)
ε ∼ N 0, σy2 IN
η ∼ N 0, σv2 IN
g, w, ε, η mutually independent
(9)
where we have collected the output measurements {yt } in a vector y and where ε and η are the
vectors of the first N input and output noise samples. The matrix W is the N ×N Toeplitz matrix
of the input, W := TN ×N (w), which represents the discrete-time convolution (2) as the product
W g. If we define the N × N Toeplitz matrix of the impulse response samples, G := TN ×N (g),
then we have the property
W g = Gw.
(10)
In the next section, we give examples of some classical system identification problems that can be
cast as uncertain-input identification problems.
3
Examples of uncertain-input models
The uncertain-input framework is a generalization of many classical system-identification problems. All these classical problems can be analyzed using the tools of uncertain-input models;
furthermore, under the right conditions, the identification approach that we propose for uncertaininput models reduces to classical system-identification approaches.
3.1
Linear predictor model
Consider the output-error transfer-function model [24],
yt =
B(q; ρ)
ut + εt ,
F (q; ρ)
where B(q; ρ) and F (q; ρ) are polynomials in the one-step shift operator q and εt is Gaussian white
noise. If we consider the parametric predictor of the output-error model, we can write it as
ŷt|t−1 =
B(q; ρ)
ut = g(ρ) ∗ u t ,
F (q; ρ)
where gt (ρ) is the impulse response of the predictor transfer function. We can see this model as a
degenerate uncertain-input model with
[µg (ρ)]i = gi (ρ),
[Kg (ρ)]i,j = 0,
[µw (θ)]i = ui ,
[Kw (θ)]i,j = 0.
We can also incorporate the framework of Bayesian identification of finite impulse-response models
with first order stable-spline kernels (for a survey, see [35]) with the choice
[µg (ρ)]i = 0,
max(i,j)
[Kg (ρ)]i,j = ρ1 ρ2
[µw (θ)]i = ui , [Kw (θ)]i,j = 0,
,
(11)
where ρ1 ≥ 0 is a scaling parameter and ρ2 ∈ [0, 1] regulates the decay rate of g (see, [37]). Note
that, in this formulation, any kernel can be used to model g (see, for instance, [13, 15]).
5
3.2
Errors-in-variables system identification
Errors-in-variables models are often described by the set of equations [44],
yt = (g ∗ w)t + ηt ,
vt = wt + εt .
It is clear that this type of models naturally fit into the uncertain-input framework of (9). In
particular, we can consider the classical errors-in-variables problem of identifying a parametric
model of S when wt is the realization of a stationary stochastic signal with a rational spectrum [12].
In this case, we can write {wt } as the filtered white noise process
wt =
C(q; θ)
et ,
D(q; θ)
where et is unitary variance Gaussian white noise, and C(q; θ) and D(q; θ) are complex polynomials
in the one-step shift operator q.
From this expression, we see that w is a Gaussian process with zero mean and covariance
matrix Σw (ρ) that depends on the parameterization of the input filter. Using a parametric model
for the system, we obtain an uncertain-input system with
[µg (ρ)]i = gi (ρ),
[Kg (ρ)]i,j = 0,
[µw (ρ)]i = 0,
[Kw (ρ)]i,j = Σw (ρ).
Alternatively, we could estimate all samples of the input signal with the choice [µw (ρ)]i = θi and
[Kw (ρ)] = 0, even though this may lead to nonidentifiability of the model [43, 51, 42].
3.3
Blind system identification
Blind system identification can also be cast as the problem of identifying an uncertain-input
model by setting the input noise variance to σv2 = ∞ (this indicates that no input measurements
are available). In this case, different parameterizations of the input lead to different models for the
input process. For instance, we can consider the parameterization of the input as a switching signal
with known switching instants T0 < T1 < · · · < Tp ; in this case we can choose [µw (ρ)]i = hti θ,
where hi is a selection vector that is nonzero in the ith interval:
(
[hi ]j = 1 if Ti−1 < t ≤ Ti ,
[hi ]j = 0 otherwise.
Models similar to this one were used, for instance, in [33] and [10].
3.4
Cascaded system identification
In cascaded linear systems, the output of one linear system is used as the input to a second linear
system (see Figure 2).
εt
ut
wt
S1
ηt
S2
+
vt
+
Figure 2: Cascaded linear systems.
6
yt
For sake of argument, we consider nonparametric models for both linear systems (the reasoning
also holds for parametric models):
g1 ∼ N 0, K1 (θ) ,
g2 ∼ N 0, K2 (ρ) .
Because g1 is a Gaussian vector, the intermediate variable w is also a Gaussian vector, with zero
mean and covariance matrix given by
Kw (θ) = U K1 (θ)U T ,
(12)
where U := TN ×N (u) is the Toeplitz matrix of the input signal ut . Therefore, we can model
the linear cascade as an uncertain-input model with input modeled as a zero-mean process with
covariance matrix given by (12) where, for instance, we use the first-order stable spline kernel
introduced in (11). The same choice of kernel can be made for K2 (ρ).
3.5
Hammerstein model identification
The Hammerstein model is a cascade of a static nonlinear function followed by a linear dynamical
system (see Figure 3).
εt
ut
f (·)
wt
S
+
yt
Figure 3: The Hammerstein model.
In the Hammerstein model, the intermediate variable wt is not observed (which, symbolically,
corresponds to an infinite σv2 ). If we consider models for the input block that are combinations of
known basis functions [3], according to
wt =
p
X
θj ϕj (ut )
j=1
we can collect the values of the unknown input w and the parameters in a vector such that
w = Φθ,
[Φ]i,j = ϕj (ui ).
This can be modeled as an uncertain-input model with µw (θ) = Φθ and Kw (θ) = 0.
The uncertain-input framework also encompasses nonparametric models for the input nonlinearity. For instance, we can model the Hammerstein cascade as the uncertain-input model with
the Gaussian radial-basis-function kernel as input model:
1
2
[Kw (θ)]i,j = θ1 exp − (ui − uj ) .
(13)
θ2
As for the linear system, we can use either parametric or nonparametirc modeling approaches
(see [3, 40]).
4
Estimation of uncertain-input models
As discussed in Section 2, we suppose that we have collected N samples of the output yt and,
possibly, N samples of the noisy input signal vt (in some applications, such as Hammerstein models
and blind system identification, these samples are not available). Whenever present, we assume
that the external input ut is completely known. We consider the following identification problem.
7
Problem 1. Given the N -dimensional vectors of measurements y and v, generated according
to (9), estimate the impulse response g, the unknown input w, and the hyperparameters τ =
{ρ, θ, σy2 , σv2 }.
Because we are using the Gaussian process model (6), we have natural candidates for the
estimates of g and w. Interpreting (7) and (8) as prior distributions of the unknowns, we know that
the best estimates given the data (in the minimum mean-square error sense) are the conditional
expectations
g ? = E g y, v , w? = E w y, v .
However, these conditional expectations depend on the value of the hyperparameter vector τ .
Because this value is not available, we follow an empirical Bayes approach [25] and we approximate
the true conditional expectations—that correspond to the true values of the hyperparameters τ —
with the conditional expectations
Z
Z
ĝ := g p(g|y, v ; τ̂ ) dg,
ŵ := w p(w|y, v ; τ̂ ) dw,
(14)
where we are using estimated values τ̂ of the hyperparameters. In the empirical Bayes approach,
the estimates of the hyperparameters are chosen by maximizing the marginal likelihood of the
data,
τ̂ := arg max log p(y, v ; τ ),
(15)
τ
where p(y, v ; τ ) is the marginal distribution of the measurements according to the model in (9).
Solving (15) yields the marginal likelihood estimate of the hyperparameters that can be used
to find the empirical Bayesian estimates of g and w in (14). However, this approach requires
distributions that, in general, are not available in closed form. Furthermore, (15) is possibly a
high-dimensional optimization problem that does not admit an analytical expression. To address
this last problem, we use the EM method to derive an iterative algorithm that solves (15). We
start by rewriting the marginal likelihood as
Z
p(y, v ; τ ) = p(y, v, g, w ; τ ) dg dw.
With this observation, we can see (15) as a maximum likelihood problem with latent variables,
where the latent variables are g and w. Appealing to the theory of the EM method, we have that
iterating the two steps
E-step: Given an estimate τ̂ (k) of τ , construct the following lower bound of the marginal likelihood
Z
Q(τ, τ̂ (k) ) = log p(y, v, g, w ; τ )p(g, w|y, v ; τ̂ (k) ) dg dw;
(16)
M-step: Update the hyperparameter estimates as
τ̂ (k+1) = arg max Q(τ, τ̂ (k) );
τ
from an arbitrary initial condition τ̂ (0) , we obtain a sequence of estimates {τ̂ (k) } of increasing
likelihood, which converges to a stationary point of the marginal likelihood of the data. In practice,
this stationary point will always be a local maximum: saddle points are numerically unstable and
minimal perturbations will drive the sequence of updates away from them [28].
Using the EM method, we have transformed the problem of maximizing the marginal likelihood
into a sequence of optimization problems. The whole point of the EM method is that these
problems should be simpler to solve than the original optimization problem.
In addition to using the EM method to solve the marginal likelihood problem, we can rewrite (14)
as
Z
Z
ĝ := g p(g, w|y, v ; τ̂ ) dw dg,
ŵ := w p(g, w|y, v ; τ̂ ) dg dw.
(17)
8
Comparing (17) and the Q function in the E-step, we see that the solution of Problem 1 using the procedure we have described depends on expectations with respect to the distribution
p(g, w |y, v ; τ ). This distribution is, in general, not available in closed form. In the next section
we present three special cases when this distribution can be computed in closed form and we present
the resulting estimation algorithms. In Section 6, we show two different ways to approximate this
joint posterior distribution in the general case.
5
Cases with degenerate prior distributions
There are cases where the integrals (16) and (17), required to estimate uncertain-input systems,
admit closed-form solutions. This happens when either the prior for g or for w (or both) are
degenerate distributions. This means that, symbolically, we let the covariances Kg (ρ) and Kw (θ)
go to zero and, respectively,
p(g; ρ) → δ g − µg (ρ) , p(w; θ) → δ w − µw (θ) .
From these expressions, we see that the models of the unknown quantities g and w are uniquely
determined by the parameter vector τ (there is no uncertainty or variability): therefore, we refer
to these kind of models as parametric models. We now present three cases of parametric models
that admit closed form expressions for the EM algorithm.
5.1
Semiparametric model
The first model is called semiparametric. It is obtained when Kw (θ) → 0. This effectively means
that the prior density (8) collapses into the Dirac density centered around the mean function, and
the posterior distributions of the unknowns admit closed form expressions:
Lemma 1. Consider the uncertain-input
system (9). In the limit when Kw (θ) → 0, we have that
p(w|y, v; τ ) = δ w−µw (θ) .
Proof. When Kw (θ) → 0, the prior density becomes the degenerate normal distribution p(w; θ) =
δ w − µw (θ) . From the law of conditional expectation, we have
p(w|y, v; τ ) =
p(y, v|w; τ )δ(w − µw (θ))
;
p(y, v; τ )
(18)
in addition, the evidence becomes
Z
p(y, v; τ ) = p(y, v|w; τ )p(w; θ) dw = p(y, v|µw (θ); τ ).
Plugging this expression into (18) we have the result.
Lemma 2. Consider the uncertain-input system (9). In the limit when Kw (θ) → 0, the posterior
distribution p(g|y, v; τ ) is Gaussian with covariance matrix and mean vector given by
!−1
!
1
1
T
−1
T
−1
Pg =
Mw (θ) Mw (θ) + Kg (ρ)
,
mg = Pg
Mw (θ) y + Kg (ρ) µg (ρ) , (19)
σy2
σy2
where Mw (θ) := TN ×N µw (θ) .
Proof. Note that y|g, w, v; τ is an affine transformation of the Gaussian random variable ε; hence,
it is Gaussian. By the law of conditional expectation and ignoring terms independent of g, we
have that
∼ log p(y|g, w; τ ) + log p(g; ρ) =
∼ − 1 y − Mw (θ)g 2 − 1 kgk2
log p(g|w, y, v; τ ) =
Kg (ρ)−1
2σy2
2
1
1
2
∼
= − kgkPg−1 + g T mg ∼
= − g − mg Pg −1
2
2
9
where Pg and mg are defined in (19). Because it is quadratic, the posterior distribution of g is
Gaussian, with the indicated covariance matrix and mean vector.
Thanks to Lemma 1 and Lemma 2, the E-step can be computed analytically when Kw (θ) = 0,
and the function Q(τ, τ̂ (k) ) admits a closed-form expression. To this end, let ∆k be the N by N
matrix given by
(
1 if i + j − 1 = k
k
[∆ ]i,j =
0 otherwise
and let
h
R = ∆1
∆2
∆3
i
∆N .
···
(20)
Then, we have the following result.
Theorem 1. Consider a semiparametric uncertain-input model with Kw (θ) = 0. Let τ̂ (k) be
(k)
estimates of the hyperparameters at the kth iteration of the EM method and let ĝ (k) and P̂g be
(k)
the moments in (19) when τ = τ̂ . Define
2
R̂y (θ) = y − Ĝ(k) µw (θ)
R̂v (θ) = v − µw (θ)
Ŝg(k)
=R
T
IN ⊗
2
,
,
P̂g(k)
R;
where Ĝ(k) = TN ×N (ĝ (k) ). Then, the function Q(τ, τ̂ (k) ) is given by
N
1
N
1
2
2
−
R̂
(θ)
−
R̂
(θ)
+
µ
(θ)
log
σ
log σy2
−
(k)
v
y
w
v
Ŝg
2σv2
2
2σy2
2
o 1
2
1 (k)
1 n
−1 (k)
P̂
−
ĝ − µg (ρ)
−
Tr
K
(ρ)
− log det Kg (ρ).
g
g
2
2
2
Kg (ρ)−1
Q(τ, τ̂ (k) ) = −
(21)
Proof. See Appendix A.1.
From (21), we see that the optimization with respect to θ is not independent of σy2 and σv2 .
Therefore, to update the hyperparameter θ, we use a conditional-maximization step [29], where we
keep the noise variances fixed to their values at the previous iterations. The use of the conditional
maximization step allows us to write the updates of the EM method in closed form:
Corollary 1. At the kth iteration of the EM method, the parameters can be updated as
−1
ρ̂(k+1) = arg min Tr Kg (ρ) P̂g(k) + log det Kg (ρ)
ρ
+ ĝ
(k)
− µg (ρ)
θ̂(k+1) = arg min
θ
2
Kg−1 (ρ)
R̂v (θ)
(k) 2
2σ̂v
+
1
R̂y (θ)+ µw (θ)
(k) 2
2σ̂y
1
R̂y (θ̂(k+1) ) + µw (θ̂(k+1) )
N
1
= R̂v (θ̂(k+1) ) .
N
σ̂y(k+1) =
σ̂v(k+1)
,
2
(k)
Ŝg
,
2
(k)
Sg
,
Proof. Follows from the two-step maximization of (21): first, maximize with respect to θ and ρ
keeping σy2 and σv2 fixed to their values at the previous iteration; then, maximize with respect to
σy2 and σv2 using the updated values of the hyperparameters.
10
Thanks to Corollary 1, we have a simple way to compute the EM estimates of the kernel
hyperparameters and of the noise variances for semiparametric models with Kw (θ) = 0: starting
from an initial value of the unknown parameters, we first update the hyperparameters ρ and
θ, then we use the new values to update the noise variances σy2 and σv2 . Under mild regularity
conditions, this procedure yields a sequence of estimates that converges to a local maximum of
the marginal likelihood (it is a Generalized EM sequence, see [50]).
Remark 1. In this section, we have presented the case when Kw (θ) → 0. However, thanks to the
symmetry of the model assured by (10), the same kind of algorithm works when Kg (ρ) → 0 (by
exchanging the roles of g and w).
5.2
Parametric model
In case we let both the input and the system covariance matrices go to zero, all the variability in
the model is removed, and we are left with classical parametric models. In this case, the marginal
likelihood of the data collapses into the likelihood where the impulse response and the input are
replaced with the parametric models µg (ρ) and µw (θ)
Z
p(y, v; ρ, θ, σ 2 ) = p(y, v|g, w; σ 2 )p(g; ρ)p(w; θ) dg dw = p(y, v| g = µg (ρ), w = µw (θ); σ 2 ).
In other words, the marginal likelihood of the data is the distribution of the data conditioned on
the events g = µg (ρ) and w = µw (θ). This distribution is given in closed form by
log p(y, v|µg (ρ), µw (θ); σ 2 ) = −
1
y − Mw (θ)µg (ρ)
2σy2
2
−
1
N
log σy2 − 2 v − µw (θ)
2
2σv
2
−
N
log σv2 .
2
(22)
where Mw (θ) is the Toeplitz matrix of µw (θ).
In this parametric-model case, we have that the posterior means reduce to the prior means
and the maximum marginal-likelihood criterion collapses into the classical maximum-likelihood or
prediction-error estimation method. To estimate the system, we first maximize (22) to find the
parameter values τ̂ ; then, we estimate the system with
ĝ = µg (ρ̂),
ŵ = µw (θ̂).
The strategy to maximize (22) depends on the specific structure of the problem. In some
applications, concentrated-likelihood or integrated-likelihood approaches have been proposed (for
a review, see [7]). An interesting consistent approach, for the parametric EIV case, has been
proposed in [51]. In [4], the authors show that if g and w are linearly parameterized, alternating
between estimation of g and of w leads to the minimum of (22).
Remark 2. The EM based algorithm presented in Section 4 cannot be used in the parametric
model case because of the impulsive posterior distributions: during the M-step, the method is
overconfident in the current value of the parameters and no update occurs. However, the EM
method can be used in the parametric case by considering a covariance matrix that shrinks toward
zero at every iteration.
6
Approximations of the joint posterior distribution
In the previous section, we have shown three cases in which the collapse of the prior distribution
allows us to express the marginal likelihood of the data and the posterior distributions in closed
form. In general, however, these distributions do not have a closed form expression. Therefore, in
this section, we present two ways to approximate the joint posterior distribution p(g, w |y, v ; τ ). In
the first, we make a particle approximation. The particles are drawn from the joint posterior using
an MCMC method. In the second, we make a variational approximation of the joint posterior.
11
6.1
Markov Chain Monte Carlo integration
Monte Carlo methods are built around the concept of particle approximation. In a particle approximation method, a density with a complicated functional form is approximated with a set of
point probabilities—that is, we approximate a density p(x) according to
p(x) ≈
M
1 X
δ(x − xj ).
M j=1
If the particle locations xj are drawn from p(x), and the number of particles M is large enough,
the expectation of any measurable function f (x) over any set can be approximated as
Z
E{f (x)} =
f (x) p(x) dx ≈
M
1 X
f (xj ),
M j=1
(23)
where {xj } are drawn from p(x). This result comes directly from the sampling property of the
Dirac density δ( · ). From a different perspective, we can see (23) as an estimation of the true
expectation. With this interpretation, we have that this estimator is unbiased,
(
)
M
1 X
E
f (xj ) = E{f (x)},
M j=1
and its covariance is inversely proportional to the number of samples used,
(
)
M
1 X
1
cov
f (xj ) =
cov{f (x)}.
M j=1
M
In practice, the number of samples needed depends on the specific application: in certain applications, few particles (say 10 or 20) may suffice; in other applications, we might need a much larger
number of particles (in the order of thousands; for a complete treatment, see [9, Chapter 11]).
When implementing Monte Carlo integrations, a common approach is MCMC. In these methods, we set up a Markov chain whose stationary distribution is the distribution we want to approximate and we run it to collect samples [19].
One convenient way to create a Markov chain is Gibbs sampling. Using this method, we obtain
a particle approximation of a joint distribution (called the target distribution) by sampling from all
the full conditional distributions—the distribution of one random variable conditioned on all other
variables—in sequence. This procedure results in a Markov chain that has the target distribution
as its stationary distribution. Contrary to many other sampling methods, Gibbs sampling does
not include a rejection step; this means that the samples proposed at every step are accepted as
samples from the chain. This may lead to faster mixing and decorrelation of the chain compared
to other MCMC methods [9, Chapter 11].
The main drawback with Gibbs sampling is that we must sample the full conditional distributions of all variables. Therefore, it is only applicable if these distributions have a functionally
convenient form. In the case at hand, we have the following results.
Lemma 3. Consider the uncertain-input model (9). The density p(g|y, w; τ ) is Gaussian with
covariance matrix and mean vector given by
Pg =
WTW
−1
+ Kg (ρ)
σy2
!−1
,
m g = Pg
!
WTy
−1
+ Kg (ρ) µg (ρ) .
σy2
Proof. The proof follows along the same line of reasoning as the proof of Lemma 2.
12
(24)
Lemma 4. Consider the uncertain-input model (9). The density p(w|y, v, g; τ ) is Gaussian with
covariance matrix and mean given by
!−1
!
GT G IN
v
GT y
−1
−1
Pw =
+ 2 + Kw (θ)
,
m w = Pw
+ 2 + Kw (θ) µw (θ) .
(25)
σy2
σv
σy2
σv
Proof. Because y and v are conditionally independent given w and g, we have that
log p(w|y, v, g; τ ) ∼
= log p(y|g, w; σy2 )p(v|w; σv2 )p(w; θ)
1
1
2
2 1
∼
= − 2 ky−Gwk − 2 kv−wk − w−µw (θ)
2σy
2σv
2
1
1
2
2
∼
= − kwkPw−1 + wT mw ∼
= − kw − mw kPw−1
2
2
2
Kw (θ)−1
where Pw and mw are given in (25). The log-density of w|y, v, g, w; τ is quadratic and, hence, it
is Gaussian with the indicated mean vector and covariance matrix.
Remark 3. In case we consider the more general Gaussian process model (5), where g and w are
a priori dependent, Lemma 3 and Lemma 4 still hold with slightly modified expressions for the
mean vectors and covariance matrices (to account for the prior correlation). For instance, the
conditional density of g is Gaussian with covariance matrix and mean vector given by
!−1
!
WTy
WTW
Pg =
+ Λg (ρ, θ)
,
mg = Pg
+ Λg (ρ, θ)µg (ρ) + Λgw (ρ, θ)(w−µw (θ)) ,
σy2
σy2
where Λgw (ρ, θ) and Λg (ρ, θ) are, respectively, the lower left and right blocks of the inverse of the
prior covariance matrix.
In view of Lemma 3 and Lemma 4, we can easily set up the Gibbs sampler to draw from the
joint posterior distribution: from any initialization of the impulse response g (0) and of the input
signal w(0) , we sample
(j)
g (j+1) |w(j) , y, v; τ ∼ N (m(j)
g , Pg ),
(26)
(j)
w(j+1) |g (j+i) , y, v; τ ∼ N (m(j)
w , Pw ).
(j)
(j)
(j)
where mg and Pg are the mean and covariance in (24) when w = w(j) , and where mw and
(j)
Pw are the mean and covariance in (25) when g = g (j+1) .
Because it is a Markov chain, the samples drawn using (26) are correlated, and subsequent
samples have memory about the initial conditions and are far away from the stationary distribution
(which is equal to the target distribution). Therefore, we discard the first samples of the Markov
chain, and we only retain the M samples after a burn-in of B samples:
ḡ (j) = g (j+B) ,
w̄(j) = w(j+B) ,
j = 1, . . . , M.
(27)
If the burn-in is large enough, the Markov chain has lost its memory about the initial conditions
and is producing samples that come form the stationary distribution. The choice of the length
of the burn-in is a difficult problem, and some heuristic algorithms have been proposed (see [19,
Section 1.4.6]).
When we have drawn enough samples from the Markov chain, we compute the Monte Carlo
estimate of the function Q; in other words, we replace the E-step in the EM method with a
Monte Carlo E-step (this is sometimes known as the MCEM method; see [49]). We create the
approximate lower bound (at the kth iteration of the EM method) by setting
Qmc (τ, τ̂ (k) ) =
Mk
1 X
log p(y, v, ḡ (j,k) , w̄(j,k) ; τ )
Mk j=1
13
where ḡ (j,k) and w̄(j,k) are samples from the stationary distribution of (26) at the kth iteration
of the EM method. In the uncertain-input case, the function Qmc is available in closed form as a
function of the sample moments of g and w.
Mk
Mk
Theorem 2. Let ḡ (j,k) j=1 and w̄(j,k) j=1 be samples from the stationary distribution of the
Gibbs sampler (26) at the kth iteration of the EM method and define
ĝ (k) =
Mk
1 X
ḡ (j,k) ,
Mk j=1
ŵ(k) =
Mk
1 X
w̄(j,k) ,
Mk j=1
P̂g(k) =
Mk
T
1 X
ḡ (j,k) − ĝ (k) ḡ (j,k) − ĝ (k)
Mk j=1
P̂w(k) =
Mk
T
1 X
w̄(j,k) − ŵ(k) w̄(j,k) − ŵ(k)
Mk j=1
R̂v(k) =
Mk
1 X
v − w̄(j,k)
Mk j=1
R̂y(k)
2
(28)
,
Mk
1 X
=
y − Ḡ(j,k) w̄(j,k)
Mk j=1
2
.
Then, the function Qmc (τ, τ̂ (k) ) is given by
(k)
(k)
o
N
R̂y
N
1 n
R̂v
−1 (k)
2
2
−
log
σ
−
−
log
σ
−
Tr
K
(ρ)
P̂
g
v
y
g
2σv2
2
2σy2
2
2
n
o 1
2
1
1
−1 (k)
P̂
−
Tr
K
(θ)
− ŵ(k) −µw (θ)
− ĝ (k) −µg (ρ)
w
w
2
2
2
Kg (ρ)−1
1
1
− log det Kg (ρ)− log det Kw (θ).
2
2
Qmc (τ, τ̂ (k) ) = −
2
(29)
Kw (θ)−1
Proof. See Appendix A.2.
In the M-step, we update the hyperparameters τ̂ (k) by maximizing the approximate lower
bound of the marginal likelihood, Qmc . Because of the closed form expression in Theorem 2, the
M-step splits into the decoupled optimization problems for the kernel hyperparameters and the
noise variances according to the following:
Corollary 2. At the kth iteration of the EM method, the kernel hyperparameters can be updated
as
n
o
2
−1 (k)
ρ̂(k+1) = arg min ĝ (k) − µg (ρ)
+
Tr
K
(ρ)
P̂
+ log det Kg (ρ),
g
g
−1
ρ
θ̂(k+1) = arg min ŵ(k) − µw (θ)
Kg (ρ)
2
Kw (θ)−1
θ
o
n
−1
+ Tr Kw (θ) P̂w(k) + log det Kw (θ),
and the noise variances can be updated as
(k)
σ̂v2 (k+1) =
(k)
R̂v
,
N
σ̂y2 (k+1) =
R̂y
.
N
Proof. Follows from direct maximization of (29).
Thanks to Theorem 2 and Corollary 2, we have a simple way to compute the MCEM estimates
of the kernel hyperparameters and of the noise variances; starting from an initial value of the
hyperparameters, we iterate the following three steps:
14
1. Run a Gibbs sampler according to (26).
2. Collect the samples according to (27) and compute the moments according to (28).
3. Update the parameters according to Corollary 2.
Under mild regularity conditions, these iterations yield a sequence of parameter estimates that
converges to a stationary point of the marginal likelihoodP
of the data (under the condition that
∞
the number of particles Mk at iteration k is such that that k=1 Mk−1 = ∞; see [32]). Then, using
the estimated hyperparameters, we can run a new Gibbs sampler and approximate the integrals
in (17) with averages over the samples:
ĝ ≈
6.2
M
1 X (j)
ḡ ,
M j=1
ŵ ≈
M
1 X (j)
w̄ .
M j=1
Variational Bayes approximation
The second method we present is a variational approximation method. Instead of approximating
the unknown joint posterior density using sampling, we propose an analytically tractable family
of distributions and we look for the best approximation of the unknown posterior density within
that family.
The variational Bayes method hinges on the fact that
log p(y, v, g, w; τ ) = log(g, w|y, v; τ ) + log p(y, v; τ ).
Hence, for any proposal distribution q in some family of distributions Q, we can write
log p(y, v; τ ) = log
p(y, v, g, w; τ )
p(g, w|y, v; τ )
− log
.
q(g, w)
q(g, w)
Taking the expectation with respect to q and observing that the left hand side is independent of
g and w, we get that
log p(y, v; τ ) = L(q) + KL(q),
(30)
where we have defined the functional
Z
p(y, v, g, w; τ )
q(g, w) dg dw,
L(q) = log
q(g, w)
and the Kullback-Leibler (KL) distance [22]
Z
q(g, w)
KL(q) = log
q(g, w) dg dw.
p(g, w|y, v; τ )
Although the KL distance is not a metric—it is not symmetric and it does not satisfy the triangle inequality—it is a useful measure of similarity between probability distributions (see [9,
Section 1.6.1]).
Because the left hand side of (30) is independent of q, we can find the distribution q? with
minimum distance (in the KL sense) to the target distribution by maximizing the functional L(q)
with respect to q ∈ Q,
q? (g, w) = arg min KL(q) = arg max L(q).
(31)
q∈Q
q∈Q
This technique allows us to use the known functional L(q) to find the q with minimum KL distance
to the unknown joint posterior distribution.
To use the variational approximation, we need to fix a family of distributions Q among which
to look for q? . In this work, we use a mean-field approximation, meaning that we look for an
15
approximation of the posterior distribution where g and w are independent given the data; in other
words we consider proposal distributions that factorize into two independent factors according to
q(g, w) = qg (g)qw (w).
After choosing the family of proposal distributions, we need to find the best approximation
q? in terms of KL distance to the unknown posterior distribution; in view of (31), the solution is
given by
q? (g, w) = arg max L(qg qw ).
qg , qw
Consider first the factor qg . We have that
Z
L(qg qw ) =
∼
=
Z
log
p(y, v, g, w; τ )
qg (g)qw (w)
!
qg (g)qw (w) dg dw,
log p(y, v, g, w; τ )qw (w) dw − log qg (g) qg (g) dg,
ignoring terms independent of qg . If we define the distribution pw (y, v, g; τ ) such that
Z
log pw (y, v, g; τ ) = log p(y, v, g, w; τ )qw (w) dw,
we have that, again ignoring terms independent of qg (g),
L(qg qw ) ∼
=−
Z
log
pw (y, v, g; τ )
qg (g)
!
qg (g) dg,
which is the negative KL distance between the factor qg and the density pw (y, v, g; τ ). Because
the KL distance is nonnegative, by choosing q?g (g) = pw (y, v, g; τ ) (where the KL distance is zero)
we are maximizing the functional L with respect to qg . Considering now qw (w), we can trace the
same argument and find that the optimal choice is
Z
log q?w (w) = log p(y, v, g, w; τ )q?g (g) dg,
(32)
where q?g (g) is the solution of
log q?g (g) =
Z
log p(y, v, g, w; τ )q?w (w) dw.
(33)
The maximum of L(qg qw ) is, therefore, the simultaneous solution of (32) and (33). The solution
(0)
(0)
can be found with the following iterative procedure: from an initialization qg and qw of the
densities, compute
Z
log q(j+1)
(w) =
w
log q(j+1)
(g)
g
log p(y, v, g, w; τ )q(j)
g (g) dg,
Z
=
(34)
log p(y, v, g, w; τ )q(j+1)
(w) dw.
w
This iterative procedure will converge to the simultaneous solution of (33) and (32) (see [9, Chapter 10]; see also [11]).
As was the case for the Gibbs sampler, which can be used only if it easy to sample from the
full conditional distributions, the variational approximation of the joint posterior is only useful if
it is possible to compute the expectations in (32) and (33). In the uncertain-input case, we have
the following result.
Theorem 3. Let q?g q?w be the factorized density with minimum KL distance to posterior density
p(g, w|y, v; τ ), for a fixed value of the hyperparameters. Then, q?g and q?w are Gaussian distributions.
16
Proof. See Appendix A.3.
Theorem 3 allows us to compute expectations with respect to q?g and q?w easily. In addition, at
every iteration of (34) the approximating densities remain Gaussian. This allows us to write the
update (34) in terms of the first and second moments of the approximating densities:
(j)
(j)
Corollary 3. Let w(j) and g (j) be the mean vectors of qw and qg at the jth iteration of (34)
(j)
(j)
(j+1)
(j+1)
and let Pw and Pg be the covariance matrices. Let g (j+1) , w(j+1) , Pg
, and Pw
be the
mean vectors and covariance matrices at the (j + 1)th iteration. Let
h
i
(j)
(j)
(j) (j) T
Tg = R IN ⊗ Pg + g g
RT ,
i
h
(j+1)
(j+1)
(j+1) (j+1) T
RT .
Tw
= R IN ⊗ Pw
+w
w
where the matrix R is defined in (20). Then,
−1
1
1 (j)
−1
T + 2 In + Kw (θ)
,
σy2 g
σv
(j) T
G
1
−1
w(j+1) = Pw(j+1)
y
+
v
+
K
(θ)
µ
(θ)
,
w
w
σy2
σv2
−1
1 (j+1)
−1
Pg(j+1) =
T
+
K
(ρ)
,
g
σy2 w
(j+1) T
W
−1
(j+1)
(j+1)
g
= Pg
y + Kg (ρ) µg (ρ) .
σy2
Pw(j+1) =
(35)
Proof. See Appendix A.4.
Thanks to Corollary 3, we can iteratively update the moments of the Gaussian factors, and the
iterations will converge to the moments of optimal variational approximation of the joint posterior
distribution.
Remark 4. In case we consider the more general Gaussian process model (5), the results of Theorem 3 and of Corollary 3 still hold with minor modifications (similarly to what is presented in
Remark 3). However, the approximation of posterior independence may not make sense when
using a-priori dependent Gaussian process models.
Using the factorized approximation of the joint distribution, we can approximate the E-step
in the EM method with a variational E-step (this is sometimes known as the VBEM method,
see [6]). We create the variational approximation of the lower bound (at the kth iteration of the
EM method) by setting
Z
(k)
Qvb (τ, τ̂ (k) ) := log p(y, v, g, w; τ )q̂(k)
g (g)q̂w (w) dw dg,
(k)
(k)
where q̂g and q̂w are the limits of the variational Bayes iterations with the hyperparameters set
to τ̂ (k) .
Because the complete-data likelihood is quadratic in g and w, the approximation Qvb admits
the closed form expression in function of the moments of g and w.
(k)
(k)
Theorem 4. Let ĝ (k) and ŵ(k) be the mean vectors of q̂g and of q̂w , respectively, and let P̂ (k)
and P̂ (k) be their covariance matrices. Define
(k)
(k)
Ŝw
= R In ⊗ P̂g(k) RT , T̂w(k) = Ŝw
+ Ŵ (k)T Ŵ (k) ,
R̂v(k) = v − ŵ(k)
2
R̂y(k) = y − Ŵ (k) ĝ (k)
,
17
2
,
where R is defined in (20). Then,
(k)
n
o
R̂v
N
N
1
(k)
(k) 2
(k) (k)
2
2
Q (τ, τ̂ ) = − 2 −
log σv −
log σy − 2 R̂y − ĝ
(k) − Tr T̂w Pg
Ŝw
2σv
2
2
2σy
o 1
o
2
1 n
1 n
−1
−1 (k)
(36)
−
ĝ (k) − µg (ρ)
− Tr Kg (ρ) P̂g(k) −
Tr
K
(θ)
P̂
w
w
2
2
2
Kg (ρ)−1
2
1
1 (k)
1
− log det Kg (ρ) − log det Kw (θ).
ŵ − µw (θ)
−
2
2
2
Kw (θ)−1
vb
(k)
Proof. See Appendix A.5.
Thanks to the structure of the function Qvb (τ, τ̂ (k) ), the M-step splits into decoupled optimization problems for the kernel hyperparameters and for the noise variances.
Corollary 4. At the kth iteration of the EM method, the kernel hyperparameters can be updated
as
o
n
2
−1 (k)
ρ̂(k+1) = arg min ĝ (k) − µg (ρ)
P̂
+ log det Kg (ρ),
+
Tr
K
(ρ)
g
g
−1
ρ
θ̂(k+1) = arg min ŵ(k) − µw (θ)
Kg (ρ)
2
θ
n
o
−1
+ Tr Kw (θ) P̂w(k) + log det Kw (θ),
Kw (θ)−1
and the noise variances can be updated as
(k)
σ̂v(k+1) =
R̂v
,
N
(k)
σ̂y(k+1) =
R̂y + ĝ (k)
2
(k) +Tr
Ŝw
o
n
(k) (k)
T̂w Pg
N
.
Proof. Follows from direct maximization of (36).
Thanks to Theorem 4 and Corollary 4, we have a simple iterative proceduce to compute the
VBEM estimates of the kernel hyperparameters and of the noise variances; starting from an inital
value of the hyperparameters, we iterate the following two steps:
1. Compute the moments of the variational approximation according to Corollary 3.
2. Update the hyperparameters according to Corollary 4.
Under mild regularity conditions, these iterations yield a sequence of parameter estimates that
converges to a stationary point of the marginal likelihood of the data (see [6, Section 2.2]). Then,
we can run the iterations in Corollary 3 again to find the posterior mean estimates of g and w.
7
Simulations
In this section, we evaluate the methods proposed on some problems that can be cast as problems
of identifying uncertain-input systems.
7.1
Cascaded linear systems
In this numerical experiment, we estimate cascaded systems with the structure presented in Section 3.4. We perform a Monte Carlo experiment consisting of 500 runs. In each run, we generate
two systems by randomly sampling 40 poles and 40 zeros, in complex conjugate pairs, using the
following technique. We sample the poles randomly, with magnitudes uniformly between 0.4 and
0.8 and phases uniformly between 0 and π. We sample the zeros randomly, with magnitudes uniformly between 0 and 0.92 and phases uniformly between 0 and π. All systems are generated with
unitary static gain. The noise variances on the input and output measurements are 1, respectively
18
1/100, times the variance of the corresponding noiseless signals; this means that the sensor at the
output of S2 is considerably more accurate than the sensor at the output of S1 .
We simulate the responses of the systems with a Gaussian white-noise input with variance
1. We collect N = 200 samples of the output, from zero initial conditions, and we estimate the
samples of the impulse responses of the two systems.
As described in Section 3.4, the systems are modeled as zero-mean Gaussian processes with
first order stable-spline kernels. All the methods are initialized with the choices ρ1 = θ1 = 1 and
ρ2 = θ2 = 0.6. The noise variances are initialized from the sample variances of the errors of the
linear least squares estimates of g1 and g2 from the noisy data.
In the experiment, we compare the following estimators.
C-MCEM The method described in Section 6.1. It uses an MCMC approximation of the joint
posterior with B = 400 and M = 2000. The EM iterations are stopped once the relative
change in the parameter values is below 10−2 .
C-VBEM The method described in Section 6.2. It uses a variational approximation of the joint
posterior. The EM iterations are stopped once the relative change in the parameter values
is below 10−2 .
C-2Stage A kernel-based two-stage method. First, it estimates the first system in the cascade
from u and v. Then, it simulates the intermediate signal ŵ as the response of the estimated
system to u and uses ŵ and y to estimate the second system in the cascade.
C-Naive A naive kernel-based estimation method. It estimates the first system in the cascade
from u and vt and the second system from vt and yt . It corresponds to using the noisy signal
vt as if it were the noiseless input to the second system in the cascade.
To evaluate the performance of the estimators, we use the following goodness-of-fit metric
Fitgj = 1 −
gj − ĝj
gj − mean(gj )
(37)
where gj is the impulse response of the system at the jth Monte Carlo run, and ĝj is an estimate
of the same impulse response.
The results of the experiment are presented in Figure 4. The figure shows the boxplots of the
fit of the estimated impulse responses of the two blocks in the cascade over the systems in the
dataset.
Fitgj
0.8
.786
.782
.741
.756
.741
.768
.709
0.6
.459
0.4
0.2
g1
C-MCEM
g2
C-VBEM
C-2Stage
C-Naive
Figure 4: Results of the estimation of cascaded linear systems.
From the figure, it appears that the proposed approximation methods are able to reconstruct
the cascaded model with higher accuracy than the alternative approaches we have considered.
Furthermore, there seems to be no clear disadvantage in using the variational Bayes approximation
as compared to the, more correct, sampling-based approximation. Regarding the performance of
the methods in estimating g1 , we see that the methods C-MCEM and C-VBEM perform better
19
than the other methods (which give the same result). Both C-2Stage and C-Naive only use the
information in v to estimate g1 , whereas C-MCEM and C-VBEM use the full joint distribution of
v and y to estimate g1 . Given that in our setting the noise on y is much lower than the noise on
v, there is information in y that the joint methods are able to leverage to improve the estimate
of g1 (similar phenomena were already observed in [21], and in [16]). This allows C-MCEM and
C-VBEM to better estimate g1 .
7.2
Hammerstein systems
In this numerical experiment, we estimate Hammerstein systems with the structure presented in
Section 3.5. We perform four Monte Carlo experiments consisting of 500 runs. In each run, we
generate a stable transfer-function model by sampling poles and zeros in the complex plane. We
sample the poles, uniformly in magnitude and phase, in the annulus of radii 0.4 and 0.8. We
sample the zeros uniformly in the disk of radius 0.92. We generate the nonlinear transformation
as a finite combination of Legendre polynomials defined as
j
ϕj (x) = 2 ·
j
X
k=0
j+k−1
j
2
.
x
j
x
k
We sample the coefficients of the combination independently and uniformly in the interval [−1, 1].
In each Monte Carlo experiment, we consider Hammerstein systems with different orders for
both the nonlinear system and the polynomial nonlinearity. In Table 1, we present the orders of
the systems considered in the various experiments.
Table 1: Orders of the Hammerstein systems used in the simulations.
Dataset
S
f (·)
LOLO (Low-Low)
HILO (High-Low)
LOHI (Low-High)
HIHI (High-High)
{3, . . . , 5}
{9, . . . , 20}
{3, . . . , 5}
{9, . . . , 20}
{5, . . . , 10}
{5, . . . , 10}
{15, . . . , 20}
{15, . . . , 20}
We simulate the responses of the systems in the datasets to a uniform white noise input in the
interval [−1, 1]. We collect N = 200 samples of the output, from zero initial conditions, and we
estimate the static nonlinearity and the impulse response.
As described in Section 3.5, the linear blocks are modeled as zero-mean Gaussian processes
with first order stable-spline kernels. We consider both a parametric model and a nonparametric
model for the static nonlinearity. All the methods are initialized with ρ1 = 1, ρ2 = 0.6. The noise
variances are initialized from the prediction error of an overparameterized least-squares estimate
(see [3, 40]).
In the simulation, we compare the performance of the following estimators:
H-P A semiparametric model for the Hammerstein system. It uses the Legendre polynomial basis
to construct a linear parameterization (with the correct order) of the input:
µw (θ) = Φθ,
Φ i,j = ϕj (ui ).
The dynamical system is modeled as a zero-mean Gaussian process with covariance matrix
given by the first order stable-spline kernel.
H-MCEM A nonparametric model for the Hammerstein system with Gibbs sampling from the
joint posterior with B = 200 and M = 500. It uses the radial-basis-function kernel (13) to
model the input nonlinearity. Note that, because the Hammerstein system is not identifiable,
we fix θ1 = 1 in the algorithm.
20
H-VBEM A nonparametric model for the Hammerstein system with variational-Bayes approximation of the joint posterior. It uses the same kernel as H-MCEM to model the input
nonlinearity.
NLHW The parametric model in Matlab with the default parameters. It corresponds to the
maximum-likelihood estimator of the model with the correct parameterization.
In all methods, the EM iterations are stopped once the relative change in the parameter values is
below 10−2 .
To evaluate the performance of the methods, we use the standard goodness-of-fit criterion (37)
for the impulse response of the linear system. For the input nonlinearity, we compute the estimated
value ŵj on a uniform grid of 300 values between -1 and 1 and we compare it to the true value wj
according to
wj − ŵj
,
Fitfj = 1 −
wj − mean(wj )
where wj is the vector of values of the true nonlinearity ot the jth Monte Carlo run, and ŵj is an
estimate of the same vector of values.
.863 .863 .864
.845
.85
Fitgj
0.8
.874 .866 .868
.853
.837 .839
.85
.846 .847
0.6
.435
0.4
.353
1
.864
0.8
.871
.845
.805 .801
.843
.789 .787
Fitfj
.702 .692
.675 .671
0.6
0.4
.387
.342
0.2
.159
LOLO
.136
HILO
H-P
LOHI
H-MCEM
H-VBEM
HIHI
NLHW
Figure 5: Boxplot of the estimation result
The result of the experiment are presented in Figure 5. The figure shows the boxplots of the
fit of the estimated impulse responses (upper pane) and of the static nonlinearities (lower pane)
over the systems in the datasets.
From this simulation, it appears that the proposed nonparametric models are capable of recovering the system better than the fully parametric NLHW. In addition, it appears that using the
correct parametric model for the input nonlinearity is beneficial in terms of accuracy. As was the
case in the cascaded-system estimation problem, the two approximation methods have comparable
performance.
21
8
Conclusions
In this work, we have proposed a new model structure, which we have called the uncertain-input
model. Uncertain-input models describe linear systems subject to inputs about which we have
limited information. To encode the information we have available about the input and the system,
we have used Gaussian-process models.
We have shown how classical problems in system identification can be seen as uncertain-input
estimation problems. Among these applications we find classical PEM, errors-in-variables and
blind system-identification problems, identification of cascaded linear systems, and identification
of Hammerstein models.
We have proposed an iterative algorithm to estimate the uncertain-input model. We estimate
the impulse response of the linear system and the input nonlinearity as the posterior means of the
Gaussian-process models given the data. The hyperparameters of the Gaussian-process models
are estimated using the marginal-likelihood method. To solve the related optimization problem,
we have proposed an iterative method based on the EM method.
In the general formulation, the model depends on the convolution of two Gaussian processes.
Therefore, the joint distribution of the data is not available in closed form. To circumvent this issue,
we have proposed specialized models, namely the semiparametric and the parametric models, for
which the integrals defining the posterior distributions are available. In the more general case, we
have proposed two approximation methods for the joint posterior distribution. In the first method,
we have used a particle approximation of the posterior distribution. The particles are drawn using
the Gibbs sampler from Gaussian full-conditional distributions. In the second method, we have
used the variational-Bayes approach to approximate the posterior distribution. Using a mean-field
approximation, we have found that the posterior distribution can be approximated as a product
of two independent Gaussian random variables.
We have tested the proposed model on two problems: the estimation of cascaded linear systems
and of Hammerstein models. In both cases, the proposed uncertain-input formulation is able to
capture the systems and to provide good estimates.
Although hinged on the EM method (which is guaranteed to converge under certain smoothness assumptions) the approximate methods we have proposed do not have general convergence
guarantees: in the formulation given by (9), there may instances of uncertain-input models for
which the assumptions required for convergence may not hold. In future publications, we plan
to analyze whether there exists general conditions on the uncertain-input model such that the
algorithms are guaranteed to converge to optimal solutions.
In addition, the uncertain-input model can be nonidentifiable in certain configurations (for
instance, consider the general errors-in-variables problem). We plan to further explore this nonidentifiability. Connections with other problems sharing the same bilinear structure [5, 48] outside
of the system identification framework are also under investigation.
A
A.1
Proofs of the main results
Proof of Theorem 1
We consider the complete-data likelihood p(y, w, g; τ ) where g acts as latent variables. We have
that
Z
Z
p(y, v, g; τ ) = p(y, v, g, w; τ ) dw = p(y|g, w; σy2 )p(v|w; σv2 )p(g; ρ)p(w; θ) dw
= p(y|g, w = µw (θ); σy2 )p(v|w = µw (θ); σv2 )p(g; ρ),
where we have used the sampling property of the Dirac density. Hence,
N
1
N
1
2
2
ky − Gµw k −
log σy2 − 2 kv − µw k −
log σv2
2σy2
2
2σv
2
1
1
2
−
g − µg Kg −1 − log det Kg ,
2
2
log p(y, v, g; τ ) = −
22
(38)
where we have dropped explicit dependencies
on the hyperparameters. Taking expectations
with
2
respect to p(g|y, v; θ̂(k) ), we have that E ky − Gµw k = y T y − 2y T E {G} µw + µTw E GT G µw .
The matrix R in (20) is such that GT = R(IN ⊗ g); hence, we have that
E{GT } = E{R(IN ⊗ g)} = R(IN ⊗ ĝ (k) ) = Ĝ(k)T ,
n
o
T
E G G = E R(IN ⊗ g)(IN ⊗ g T )R = R(IN ⊗ E{gg T })RT = R(IN ⊗ P̂g(k) + ĝ (k) ĝ (k)T )RT
= R(IN ⊗ P̂g(k) )R + R(IN ⊗ ĝ (k) ĝ (k)T )RT = Ŝg(k) + R(IN ⊗ ĝ (k) )(IN ⊗ ĝ (k)T )RT
= Ŝg(k) + Ĝ(k)T Ĝ(k) ;
2
2
2
hence, E ky − Gµw k = y − Ĝ(k) µw +kµw kŜ (k) .
g
Similarly,
io
h
n
o
n
2
E g − µg
= Tr Kg−1 E{gg T }−2µg E{g T }+µg µTg
n
h
io
= Tr Kg−1 P̂g(k) + ĝ (k) g (k)T − 2µg ĝ (k)T + µg µTg
n
o
2
= ĝ (k) − µg −1 + Tr Kg−1 P̂g(k) .
Kg
Plugging these expressions into the expectation of (38) we find (21).
A.2
Proof of Theorem 2
Let ḡ (j) and w̄(j) be samples draw from the stationary distribution of the Gibbs sampler with
hyperparameters τ̂ . Now, consider the complete-data likelihood
1
N
1
N
2
2
kv − wk −
log σv2 − 2 ky − W gk −
log σy2
2σv2
2
2σy
2
1
1
1
2
− log det Kg − kw − µw kKw − log det Kw ,
2
2
2
log p(y, v|w, g; τ )p(g; ρ)p(w; θ) = −
−
1
g − µg
2
2
Kg
(39)
where we have dropped the explicit dependencies on the hyperparameters. We have that
Qmc (τ, τ̂ ) := −
M
1 X
v − w̄(j)
2M σv2 j=1
M
1 X (j)
ḡ − µg
−
2M j=1
2
Kg −1
2
−
M
1 X
N
log σv2 −
y − W̄ (j) ḡ (j)
2
2M σy2 j=1
M
1
1 X (j)
− log det Kg −
w̄ − µw
2
2M j=1
2
Kw −1
Using the definitions in (28), we have that
M
X
ḡ
j=1
=
(j)
−µg
M
X
j=1
=
M
X
j=1
2
Kg −1
=
M
X
j=1
T
−1 (j)
(j)
Tr Kg (ḡ −µg )(ḡ −µg )
T
−1 (j)
(j)
Tr Kg (ḡ −ĝ+ĝ−µg )(ḡ −ĝ+ĝ−µg )
n
o
Tr Kg−1 (ḡ (j) −ĝ)(ḡ (j) −ĝ) + M ĝ−µg
M
X
(ḡ (j) −ĝ)(ḡ (j) −ĝ) + M ĝ−µg
= Tr Kg−1
j=1
n
o
2
= M Tr Kg−1 P̂g + M ĝ−µg K −1 ;
g
23
2
Kg−1
2
Kg−1
2
−
N
log σy2
2
1
− log det Kw .
2
(40)
similarly,
M
X
j=1
w̄(j) −µw
2
Kw −1
n
o
2
−2
= M Tr Kw
P̂w + M kŵ−µw kKw−1 ;
Plugging these expressions into (40) we obtain (28).
A.3
Proof of Theorem 3
Consider the complete-data likelihood (39). From (32) we have that log q?w = E log p(y, v, g, w; τ ) ,
where the expectation is taken with respect to q?g . Then, disregarding terms independent of w,
we have that
)
(
2
2
kv − wk
ky − Gwk
2
? ∼
∼ − 1 kwk2 −1 + wT mw ,
−
−kw − µw kKw −1 =
log qw = E −
Pw
2σy2
2σv2
2
where
Pw =
1
1
−1
E{GT G} + 2 In + Kw
σy2
σv
!−1
,
mw = Pw
1
1
−1
E{GT }y + 2 v + Kw
µw
σy2
σv
!
. (41)
Because it is quadratic in w, qw is a Gaussian distribution. Similarly,
)
(
2
1
ky
−
W
gk
2
2
− g − µg Kg −1 ∼
log q?g ∼
= − kgkPg−1 + g T mg ,
=E −
2σy2
2
where
Pg =
1
E{W T W } + Kg−1
σy2
!−1
,
mg = Pg
1
E{W T }y + Kg−1 µg
σy2
!
.
(42)
and where all expectations are taken with respect to q?w . Because it is quadratic in g, qg is also a
Gaussian distribution.
A.4
Proof of Corollary 3
(j+1)
is a Gaussian distribution with covariance
Tracing the proof of Theorem 3, we have that qw
(j)
matrix and mean given by (41) where the expectations are taken with respect to qg . Using the
matrix R in (20), we have
E{GT G} = E{R(IN ⊗ g)(IN ⊗ g T )RT } = R(IN ⊗ E{gg T })RT = Tg(j) ,
E{GT } = E{R(IN ⊗ g)} = R(IN ⊗ g (j+1) ) = G(j)T .
(j+1)
Similarly, qg
is a Gaussian distribution with covariance matrix and mean given by (42), where
(j+1)
the expectations are taken with respect to qw . We have that
E{W T W } = E{R(IN ⊗ w)(IN ⊗ wT )RT } = R(IN ⊗ E{wwT })RT = Tw(j+1) ,
E{W T } = R(IN ⊗ w(j+1) ) = W (j+1)T .
Plugging these expectations into (42) and (41) we obtain (35).
24
A.5
Proof of Theorem 4
We consider again the complete-data likelihood (39). Taking the expectation with respect to the
(k)
(k)
independent Gaussian densities qg and qw , we have that
n
o
2
E kv − wk = R̂v(k) + Tr{P̂ (k) },
n
o
2
2
E g − µg Kg −1 = ĝ (k) − µg Kg −1 + Tr{Kg−1 P̂g(k) },
n
o
2
2
−1 (k)
E kw − µw kKw −1 = ŵ(k) − µw Kw −1 + Tr{Kw
P̂w }.
(k)
(k)
Note that E{W T W } = RT (IN ⊗ P̂w + ŵ(k) ŵ(k)T )R = RT (IN ⊗ P̂w )R + RT (IN ⊗ ŵ(k) ŵ(k)T )R =
(k)
(k)
(k)
2
(k)
Ŝw + Ŵ (k)T Ŵ (k) = T̂w , and thatnE{g T W T Wog} = Tr{T̂w (Pg + ĝ (k) ĝ (k)T ) = Ŵ (k) ĝ (k) +
2
(k) (k)
2
ĝ (k) S (k) + Tr{T̂w P̂g }, Hence, E ky − W gk = y T y − 2y T E{W g} + E{g T W T W g} = R̂y (k) +
w
2
ĝ (k) S (k)
w
get (36).
B
(k)
(k)
+ Tr{T̂w P̂g }. Plugging the terms in the expression of the complete likelihood, we
Acknowledgment
This work was supported by the Swedish Research Council via the projects NewLEADS (contract
number: 2016-06079) and System identification: Unleashing the algorithms (contract number:
2015-05285), and by the European Research Council under the advanced grant LEARN (contract
number: 267381).
References
[1] K. Abed-Meraim, W. Qiu, and Y. Hua. Blind system identification. Proc. IEEE, 85(8):1310–1322,
1997.
[2] A. Ahmed, B. Recht, and J. Romberg. Blind deconvolution using convex programming. IEEE Trans.
Inform. Theory, 60(3):1711–1732, 2014.
[3] E. W. Bai. An optimal two-stage identification algorithm for Hammerstein–Wiener nonlinear systems.
Automatica, 34(3):333–338, 1998.
[4] E. W. Bai and D. Li. Convergence of the iterative Hammerstein system identification algorithm.
IEEE Trans. Autom. Control, 49(11):1929–1940, 2004.
[5] E. W. Bai and Y. Liu. On the least squares solutions of a system of bilinear equations. In Proc.
IEEE Conf. Decis. Control (CDC). IEEE, 2005.
[6] M. J. Beal. Variational Algorithms for Approximate Bayesian Inference. PhD thesis, Gatsby Computational Neuroscience Unit, University College London, 2003.
[7] J. O. Berger, B. Liseo, and R. L. Wolpert. Integrated likelihood methods for eliminating nuisance
parameters. Statist. Sci., 14(1):1–28, 1999.
[8] J. M. Bernardo and A. F. M. Smith. Bayesian Theory. JOHN WILEY & SONS INC, 2000.
[9] C. M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[10] G. Bottegal, R. S. Risuleo, and H. Hjalmarsson. Blind system identification using kernel-based
methods. In Proc. IFAC Symp. System Identification (SYSID), volume 48, pages 466–471, 2015.
[11] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
[12] P. Castaldi and U. Soverini.
32(4):631–636, 1996.
Identification of dynamic errors-in-variables models.
Automatica,
[13] T. Chen and L. Ljung. Constructive state space model induced kernels for regularized system identification. In Proc. IFAC World Cong., volume 19, pages 1047–1052, 2014.
25
[14] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the
em algorithm. J. R. Stat. Soc. Ser. B. Stat. Methodol., pages 1–38, 1977.
[15] F. Dinuzzo. Kernels for linear time invariant system identification.
53(5):3299–3317, 2015.
SIAM J. Control Optim.,
[16] N. Everitt, C. Rojas, and H. Hjalmarsson. A geometric approach to variance analysis of cascaded
systems. In Proc. IEEE Conf. Decis. Control (CDC), 2013.
[17] R. Frigola, F. Lindsten, T. B. Schn, and C. E. Rasmussen. Identification of gaussian process statespace models with particle stochastic approximation EM. IFAC Proc. Vol., 47(3):4097–4102, 2014.
[18] S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration
of images. IEEE Trans. Pattern Anal. Mach. Intell., (6):721–741, 1984.
[19] W. R. Gilks, S. Richardson, and D. J. Spiegelhalter. Markov Chain Monte Carlo in Practice. Chapman
and Hall London, 1996.
[20] F. Giri and E. W. Bai. Block-oriented nonlinear system identification. Springer, 2010.
[21] H. Hjalmarsson. System identification of complex and structured systems. Eur. J. Control, 15(34):275–310, 2009.
[22] S. Kullback and R. A. Leibler. On information and sufficiency. Ann. Math. Statist., 22(1):79–86,
1951.
[23] J. Linder and M. Enqvist. Identification of systems with unknown inputs using indirect input measurements. International Journal of Control, 90(4):729–745, 2017.
[24] L. Ljung. System Identification, Theory for the User. Prentice Hall, 1999.
[25] J. Maritz and T. Lwin. Empirical bayes methods. Chapman and Hall London, 1989.
[26] I. Markovsky and K. Usevich. Structured low-rank approximation with missing data. SIAM J. Matrix
Anal. & Appl., 34(2):814–830, 2013.
[27] D. B. McCombie, A. T. Reisner, and H. H. Asada. Laguerre-model blind system identification:
Cardiovascular dynamics estimated from multiple peripheral circulatory signals. IEEE Trans. Biomed.
Eng., 52(11):1889–1901, 2005.
[28] G. McLachlan and T. Krishnan. The EM algorithm and extensions, volume 382. John Wiley and
Sons, 2007.
[29] X. L. Meng and D. B. Rubin. Maximum likelihood estimation via the ECM algorithm: A general
framework. Biometrika, 80(2):267–278, 1993.
[30] E. Moulines, P. Duhamel, J. F. Cardoso, and S. Mayrargue. Subspace methods for the blind identification of multichannel FIR filters. IEEE Trans. Signal Process., 43(2):516–525, 1995.
[31] N. Nakajima. Blind deconvolution using the maximum likelihood estimation and the iterative algorithm. Opt. Commun., 100(1-4):59–66, 1993.
[32] R. C. Neath. On convergence properties of the Monte Carlo EM algorithm. In Advances in Modern
Statistical Theory and Applications: A Festschrift in honor of Morris L. Eaton, pages 43–62. Institute
of Mathematical Statistics, 2013.
[33] H. Ohlsson, L. J. Ratliff, R. Dong, and S. S. Sastry. Blind identification via lifting. In Proc. IFAC
World Cong., 2014.
[34] G. Pillonetto and A. Chiuso. A Bayesian learning approach to linear system identification with
missing data. In Proc. IEEE Conf. Decis. Control (CDC), pages 4698–4703, 2009.
[35] G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and L. Ljung. Kernel methods in system identification, machine learning and function estimation: A survey. Automatica, 50(3):657–682, 2014.
[36] G. Pillonetto, M. H. Quang, and A. Chiuso. A new kernel-based approach for nonlinear system
identification. IEEE Trans. Autom. Control, 56(12):2825–2840, 2011.
[37] Gianluigi Pillonetto and Giuseppe De Nicolao. A new kernel-based approach for linear system identification. Automatica, 46(1):81–93, 2010.
[38] C. Rasmussen and C. Williams. Gaussian processes for machine learning. the MIT Press, 2006.
[39] R. S. Risuleo, G. Bottegal, and H. Hjalmarsson. A kernel-based approach to Hammerstein system
identication. In Proc. IFAC Symp. System Identification (SYSID), volume 48, pages 1011–1016, 2015.
26
[40] R. S. Risuleo, G. Bottegal, and H. Hjalmarsson. A new kernel-based approach to overparameterized
Hammerstein system identification. In Proc. IEEE Conf. Decis. Control (CDC), pages 115–120, 2015.
[41] R. S. Risuleo, G. Bottegal, and H. Hjalmarsson. On the estimation of initial conditions in kernel-based
system identification. In Proc. IEEE Conf. Decis. Control (CDC), pages 1120–1125, 2015.
[42] R. S. Risuleo, G. Bottegal, and H. Hjalmarsson. Kernel-based system identification from noisy and
incomplete input-output data. In Proc. IEEE Conf. Decis. Control (CDC). Institute of Electrical
and Electronics Engineers (IEEE), 2016.
[43] T. Söderström. Why are errors-in-variables problems often tricky? In Proc. European Control Conf.
(ECC), pages 802–807, 2003.
[44] T. Söderström. Errors-in-variables methods in system identification. Automatica, 43(6):939–958,
2007.
[45] T. Söderström. System identification for the errors-in-variables problem. In Proc. UKACC Int. Conf.
Control, pages 1–14, 2010.
[46] A. Svensson and T. B. Schön. A flexible state–space model for learning nonlinear dynamical systems.
Automatica, 80:189–199, 2017.
[47] R. Wallin and A. Hansson. Maximum likelihood estimation of linear SISO models subject to missing
output data and missing input data. Int. J. Control, pages 1–11, 2014.
[48] J. Wang, Q. Zhang, and L. Ljung. Revisiting the two-stage algorithm for hammerstein system
identification. In Proc. IEEE Conf. Decis. Control (CDC). IEEE, 2009.
[49] G. C. G. Wei and M. A. Tanner. A Monte Carlo implementation of the EM algorithm and the poor
man’s data augmentation algorithms. Journal of the American Statistical Association, 85(411):699–
704, 1990.
[50] C. F. J. Wu. On the convergence properties of the EM algorithm. Ann. Statist., 11(1):95–103, 1983.
[51] E. Zhang and R. Pintelon. Errors-in-variables identification of dynamic systems in general cases. In
Proc. IFAC Symp. System Identification (SYSID), volume 48, pages 309–313, 2015.
27
| 3cs.SY
|
1
Adaptive Least Mean Squares Estimation
of Graph Signals
arXiv:1602.05703v3 [cs.LG] 11 Jul 2016
Paolo Di Lorenzo1, Member, IEEE, Sergio Barbarossa2, Fellow, IEEE,
Paolo Banelli1 , Member, IEEE, and Stefania Sardellitti2 , Member, IEEE
Abstract—The aim of this paper is to propose a least mean
squares (LMS) strategy for adaptive estimation of signals defined
over graphs. Assuming the graph signal to be band-limited,
over a known bandwidth, the method enables reconstruction,
with guaranteed performance in terms of mean-square error,
and tracking from a limited number of observations over a
subset of vertices. A detailed mean square analysis provides
the performance of the proposed method, and leads to several
insights for designing useful sampling strategies for graph signals.
Numerical results validate our theoretical findings, and illustrate
the performance of the proposed method. Furthermore, to cope
with the case where the bandwidth is not known beforehand, we
propose a method that performs a sparse online estimation of the
signal support in the (graph) frequency domain, which enables
online adaptation of the graph sampling strategy. Finally, we
apply the proposed method to build the power spatial density
cartography of a given operational region in a cognitive network
environment.
Index Terms—Least mean squares estimation, graph signal
processing, sampling on graphs, cognitive networks.
I. I NTRODUCTION
In many applications that are of current interest, like social
networks, vehicular networks, big data or biological networks,
the signals of interest are defined over the vertices of a graph
[1]. This has motivated the development of tools for analyzing
signals defined over a graph, or graph signals for short [1]–
[3]. Graph signal processing (GSP) aims at extending classical
discrete-time signal processing tools to signals defined over a
discrete domain whose elementary units (vertices) are related
to each other through a graph. This framework subsumes as a
very simple special case discrete-time signal processing, where
the vertices are associated to time instants and edges link
consecutive time instants. A peculiar aspect of GSP is that,
since the signal domain is dictated by the graph topology, the
analysis tools come to depend on the graph topology as well.
This paves the way to a plethora of methods, each emphasizing
different aspects of the problem. An important feature to have
in mind about graph signals is that the signal domain is not
a metric space, as in the case, for example, of biological
networks, where the vertices may be genes, proteins, enzymes,
Di Lorenzo and Banelli are with the Dept. of Engineering,
University of Perugia, Via G. Duranti 93, 06125, Perugia, Italy; Email:
[email protected], [email protected].
Barbarossa and Sardellitti are with the Dept. of Information
Engineering,
Electronics,
and
Telecommunications,
Sapienza University
of Rome, Via Eudossiana 18, 00184,
Rome,
Italy;
E-mail:
[email protected],
[email protected]. This work has been
supported by TROPIC Project, Nr. ICT-318784. The work of Paolo Di
Lorenzo was supported by the “Fondazione Cassa di Risparmio di Perugia”.
etc, and the presence of an edge between two molecules means
that those molecules undergo a chemical reaction. This marks
a fundamental difference with respect to time signals where the
time domain is inherently a metric space. Processing signals
defined over a graph has been considered in [2], [4]–[6]. A
central role in GSP is of course played by spectral analysis
of graph signals, which passes through the introduction of
the so called Graph Fourier Transform (GFT). Alternative
definitions of GFT have been proposed, depending on the
different perspectives used to extend classical tools [7], [8],
[1], [9], [2]. Two basic approaches are available, proposing
the projection of the graph signal onto the eigenvectors of
either the graph Laplacian, see, e.g., [7], [1], [9] or of the
adjacency matrix, see, e.g. [2], [10]. The first approach applies
to undirected graphs and builds on the spectral clustering
properties of the Laplacian eigenvectors and the minimization
of the ℓ2 norm graph total variation; the second approach was
proposed to handle also directed graphs and it is based on
the interpretation of the adjacency operator as the graph shift
operator, which lies at the heart of all linear shift-invariant
filtering methods for graph signals [11], [12]. A further very
recent contribution proposes to build the graph Fourier basis
as the set of orthonormal signals that minimize the (directed)
graph cut size [13].
After the introduction of the GFT, an uncertainty principle
for graph signals was derived in [14], [15], [16], [17], [18].
with the aim of assessing the link between the spread of a
signal on the vertices of the graph and on its dual domain, as
defined by the GFT. A simple closed form expressions for the
fundamental tradeoff between the concentrations of a signal
in the graph and the transformed domains was given in [18].
One of the basic problems in GSP is the development of a
graph sampling theory, whose aim is to recover a band-limited
(or approximately band-limited) graph signal from a subset of
its samples. A seminal contribution was given in [7], later
extended in [19] and, very recently, in [10], [18], [20], [21],
[22]. Dealing with graph signals, the recovery problem may
easily become ill-conditioned, depending on the location of
the samples. Hence, for any given number of samples enabling
signal recovery, the identification of the sampling set plays a
key role in the conditioning of the recovery problem. It is
then particularly important to devise strategies to optimize the
selection of the sampling set. Alternative signal reconstuction
methods have been proposed, either iterative as in [23], [20],
[24], or single shot, as in [10], [18]. Frame-based approaches
to reconstruct signals from subsets of samples have been
proposed in [7], [20], [18].
2
The theory developed in the last years for GSP was then applied to solve specific learning tasks, such as semi-supervised
classification on graphs [25]–[27], graph dictionary learning
[28], [29], learning graphs structures [30], smooth graph signal
recovery from random samples [31], [32], inpainting [33],
denoising [34], and community detection on graphs [35].
Finally, in [36], [37], the authors proposed signal recovery
methods aimed to recover graph signals that are assumed to
be smooth with respect to the underlying graph, from sampled,
noisy, missing, or corrupted measurements.
Contribution: The goal of this paper is to propose LMS
strategies for the adaptive estimation of signals defined on
graphs. To the best of our knowledge, this is the first attempt
to merge the well established theory of adaptive filtering [38],
[39], with the emerging field of signal processing on graphs.
The proposed method hinges on the graph structure describing
the observed signal and, under a band-limited assumption,
it enables online reconstruction and tracking from a limited
number of observations taken over a subset of vertices. An
interesting feature of our proposed strategy is that this subset
is allowed to vary over time, in adaptive manner. A detailed
mean square analysis illustrates the role of the sampling
strategy on the reconstruction capability, stability, and meansquare performance of the proposed algorithm. Based on these
results, we also derive adaptive sampling strategies for LMS
estimation of graph signals. Several numerical results confirm
the theoretical findings, and assess the performance of the
proposed strategies. Furthermore, we consider the case where
the graph signal is band-limited but the bandwidth is not
known beforehand; this case is critical because the selection
of the sampling strategy fundamentally depends on such prior
information. To cope with this issue, we propose an LMS
method with adaptive sampling, which estimates and tracks
the signal support in the (graph) frequency domain, while at
the same time adapting the graph sampling strategy. Numerical
results illustrate the tracking capability of the aforementioned
method in the presence of time-varying graph signals. As
an example, we apply the proposed strategy to estimate and
track the spatial distribution of the electromagnetic power in a
cognitive radio framework. The resulting graph signal turns out
to be smooth, i.e. the largest part of its energy is concentrated
at low frequencies, but it is not perfectly band-limited. As
a consequence, recovering the overall signal from a subset
of samples is inevitably affected by aliasing [22]. Numerical
results show the tradeoff between complexity, i.e. number of
samples used for processing, and mean-square performance of
the proposed strategy, when applied to such cartography task.
Intuitively, processing with a larger bandwidth and a (consequent) larger number of samples, improves the performance
of the algorithm, at the price of a larger complexity.
The paper is organized as follows. In Sec. II, we introduce
some basic graph signal processing tools, which will be useful
for the following derivations. Sec. III introduces the proposed
LMS algorithm for graph signals, illustrates its mean-square
analysis, and derives useful graph sampling strategies. Then, in
Sec. IV we illustrate the proposed LMS strategy with adaptive
sampling, while Sec. V considers the application to power
density cartography. Finally, Sec. VI draws some conclusions.
II. G RAPH S IGNAL P ROCESSING T OOLS
We consider a graph G = (V, E) consisting of a set of N
nodes V = {1, 2, ..., N }, along with a set of weighted edges
E = {aij }i,j∈V , such that aij > 0, if there is a link from node
j to node i, or aij = 0, otherwise. The adjacency matrix A of
a graph is the collection of all the weights aij , i, j = 1, . . . , N .
PN
The degree of node i is ki := j=1 aij . The degree matrix K
is a diagonal matrix having the node degrees on its diagonal.
The Laplacian matrix is defined as:
L = K − A.
(1)
If the graph is undirected, the Laplacian matrix is symmetric
and positive semi-definite, and admits the eigendecomposition
L = UΛUH , where U collects all the eigenvectors of L in
its columns, whereas Λ is a diagonal matrix containing the
eigenvalues of L. It is well known from spectral graph theory
[40] that the eigenvectors of L are well suited for representing
clusters, since they minimize the ℓ2 norm graph total variation.
A signal x over a graph G is defined as a mapping from the
vertex set to the set of complex numbers, i.e. x : V → C. In
many applications, the signal x admits a compact representation, i.e., it can be expressed as:
x = Us
(2)
where s is exactly (or approximately) sparse. As an example,
in all cases where the graph signal exhibits clustering features,
i.e. it is a smooth function within each cluster, but it is
allowed to vary arbitrarily from one cluster to the other, the
representation in (2) is compact, i.e. the only nonzero (or
approximately nonzero) entries of s are the ones associated
to the clusters.
The GFT x̂ of a signal x is defined as the projection onto
the orthogonal set of vectors {ui }i=1,...,N [1], i.e.
GFT:
x̂ = UH x.
(3)
The GFT has been defined in alternative ways, see, e.g., [1],
[9], [2], [10]. In this paper, we basically follow the approach
based on the Laplacian matrix, assuming an undirected graph
structure, but the theory could be extended to handle directed
graphs with minor modifications. We denote the support of x̂
in (2) as F = {i ∈ {1, . . . , N } : x̂i 6= 0}, and the bandwidth
of the graph signal x is defined as the cardinality of F , i.e.
|F |. Clearly, combining (2) with (3), if the signal x exhibits
a clustering behavior, in the sense specified above, computing
its GFT x̂ is the way to recover the sparse vector s in (2).
Localization Operators: Given a subset of vertices S ⊆ V,
we define a vertex-limiting operator as the diagonal matrix
DS = diag{1S },
(4)
where 1S is the set indicator vector, whose i-th entry is equal
to one, if i ∈ S, or zero otherwise. Similarly, given a subset of
frequency indices F ⊆ V, we introduce the filtering operator
BF = UΣF UH ,
(5)
where ΣF is a diagonal matrix defined as ΣF = diag{1F }.
It is immediate to check that both matrices DS and BF are
self-adjoint and idempotent, and so they represent orthogonal
3
projectors. The space of all signals whose GFT is exactly
supported on the set F is known as the Paley-Wiener space
for the set F [7]. We denote by B F ⊆ L2 (G) the set of all
finite ℓ2 -norm signals belonging to the Paley-Wiener space
associated to the frequency subset F . Similarly, we denote by
DS ⊆ L2 (G) the set of all finite ℓ2 -norm signals with support
on the vertex subset S. In the rest of the paper, whenever there
will be no ambiguities,, we will drop the subscripts referring
to the sets. Finally, given a set S, we denote its complement
set as S, such that V = S ∪ S and S ∩ S = ∅. Similarly,
we define the complement set of F as F . Thus, we define
the vertex-projector onto S as D and, similarly, the frequency
projector onto the frequency domain F as B.
Exploiting the localization operators in (4) and (5), we say
that a vector x is perfectly localized over the subset S ⊆ V if
Dx = x,
(6)
with D defined as in (4). Similarly, a vector x is perfectly
localized over the frequency set F if
Bx = x,
(7)
with B given in (5). As previously stated, |F | represents the
(not necessarily contiguous) bandwidth of the graph signal.
The localization properties of graph signals were studied in
[22] and later extended in [18] to derive the fundamental tradeoff between the localization of a signal in the graph and on its
dual domain. An interesting consequence of that theory is that,
differently from continuous-time signals, a graph signal can
be perfectly localized in both vertex and frequency domains.
The conditions for having perfect localization are stated in the
following theorem, which we report here for completeness of
exposition; its proof can be found in [22].
Theorem 1: There is a vector x, perfectly localized over
both vertex set S and frequency set F (i.e. x ∈ B F ∩ DS ) if
and only if the operator BDB (or DBD) has an eigenvalue
equal to one; in such a case, x is an eigenvector of BDB
associated to the unitary eigenvalue.
Equivalently, the perfect localization properties can be expressed in terms of the operators BD and DB. Indeed, since
the operators BD and DB have the same singular values [18],
perfect localization onto the sets S and F can be achieved if
and only if
kBDk2 = kDBk2 = 1.
(8)
Building on these previous results on GSP, in the next section
we introduce the proposed LMS strategy for adaptive estimation of graph signals.
III. LMS E STIMATION OF G RAPH S IGNALS
The least mean square algorithm, introduced by Widrow and
Hoff [41], is one of the most popular methods for adaptive
filtering. Its applications include echo cancelation, channel
equalization, interference cancelation and so forth. Although
there exist algorithms with faster convergence rates such as
the Recursive Least Square (RLS) methods [38], LMS-type
methods are popular because of their ease of implementation,
low computational costs and robustness. For these reasons,
a huge amount of research was produced in the last decades
focusing on improving the performance of LMS-type methods,
exploiting in many cases some prior information that is
available on the observed signals. For instance, if the observed
signal is known to be sparse in some domain, such prior
information can help improve the estimation performance, as
demonstrated in many recent efforts in the area of compressed
sensing [42], [43]. Some of the early works that mix adaptation
with sparsity-aware reconstruction include methods that rely
on the heuristic selection of active taps [44], and on sequential
partial updating techniques [45]; some other methods assign
proportional step-sizes to different taps according to their magnitudes, such as the proportionate normalized LMS (PNLMS)
algorithm and its variations [46]. In subsequent studies, motivated by the LASSO technique [47] and by connections with
compressive sensing [43], [48], several algorithms for sparse
adaptive filtering have been proposed based on LMS [49],
RLS [50], and projection-based methods [51]. Finally, sparsity
aware distributed methods were proposed in [52]–[56].
In this paper, we aim to exploit the intrinsic sparsity that is
present in band-limited graph signals, thus designing proper
sampling strategies that guarantee adaptive reconstruction of
the signal, with guaranteed mean-square performance, from a
limited number of observation sampled from the graph. To this
aim, let us consider a signal x0 ∈ CN defined over the graph
G = (V, E). The signal is initially assumed to be perfectly
band-limited, i.e. its spectral content is different from zero
only on a limited set of frequencies F . Later on, we will
relax such an assumption. Let us consider partial observations
of signal x0 , i.e. observations over only a subset of nodes.
Denoting with S the sampling set (observation subset), the
observed signal at time n can be expressed as:
y[n] = D (x0 + v[n]) = DBx0 + Dv[n]
(9)
where D is the vertex-limiting operator defined in (4), which
takes nonzero values only in the set S, and v[n] is a zeromean, additive noise with covariance matrix Cv . The second
equality in (9) comes from the bandlimited assumption, i.e.
Bx0 = x0 , with B denoting the operator in (5) that projects
onto the (known) frequency set F . We remark that, differently
from linear observation models commonly used in adaptive
filtering theory [38], the model in (9) has a free sampling
parameter D that can be properly selected by the designer,
with the aim of reducing the computational/memory burden
while still guaranteeing theoretical performance, as we will
illustrate in the following sections. The estimation task consists
in recovering the band-limited graph signal x0 from the noisy,
streaming, and partial observations y[n] in (9). Following an
LMS approach [41], the optimal estimate for x0 can be found
as the vector that solves the following optimization problem:
min E ky[n] − DBxk2
(10)
x
s.t.
Bx = x,
where E(·) denotes the expectation operator. The solution of
problem (10) minimizes the mean-squared error and has a
bandwidth limited to the frequency set F . For stationary y[n],
the optimal solution of (10) is given by the vector x̂ that
4
Algorithm 1: LMS algorithm for graph signals
Start with x[0] ∈ B F chosen at random. Given a sufficiently
small step-size µ > 0, for each time n > 0, repeat:
x[n + 1] = x[n] + µ BD (y[n] − x[n])
(12)
satisfies the normal equations [38]:
BDB x̂ = BD E{y[n]}.
(11)
Exploiting (9), this statement can be readily verified noticing
that x̂ minimizes the objective function of (10) and is bandlimited, i.e. it satisfies Bx̂ = x̂. Nevertheless, in many linear
regression applications involving online processing of data,
the expectation E{y[n]} may be either unavailable or timevarying, and thus impossible to update continuously. For this
reason, adaptive solutions relying on instantaneous information
are usually adopted in order to avoid the need to know the
signal statistics beforehand. A typical solution proceeds to
optimize (10) by means of a steepest-descent procedure. Thus,
letting x[n] be the instantaneous estimate of vector x0 , the
LMS algorithm for graph signals evolves as illustrated in
Algorithm 1, where µ > 0 is a (sufficiently small) stepsize, and we have exploited the fact that D is an idempotent
operator, and Bx[n] = x[n] (i.e., x[n] is band-limited) for
all n. Algorithm 1 starts from an initial signal that belongs
to the Paley-Wiener space for the set F , and then evolves
implementing an alternating orthogonal projection onto the
vertex set S (through D) and the frequency set F (through B).
The properties of the LMS recursion in (12) crucially depend
on the choice of the sampling set S, which defines the structure
of the operator D [cf. (4)]. To shed light on the theoretical
behavior of Algorithm 1, in the following sections we illustrate
how the choice of the operator D affects the reconstruction
capability, mean-square stability, and steady-state performance
of the proposed LMS strategy.
A. Reconstruction Properties
It is well known from adaptive filters theory [38] that the
LMS algorithm in (12) is a stochastic approximation method
for the solution of problem (10), which enables convergence
in the mean-sense to the true vector x0 (if the step-size µ is
chosen sufficiently small), while guaranteing a bounded meansquare error (as we will see in the sequel). However, since the
existence of a unique band-limited solution for problem (12)
depends on the adopted sampling strategy, the first natural
question to address is: What conditions must be satisfied by
the sampling operator D to guarantee reconstruction of signal
x0 from the selected samples? The answer is given in the
following Theorem, which gives a necessary and sufficient
condition to reconstruct graph signals from partial observations
using Algorithm 1.
Theorem 2: Problem (10) admits a unique solution, i.e. any
band-limited signal x0 can be reconstructed from its samples
taken in the set S, if and only if
DB
2
< 1,
(13)
i.e. if the matrix BDB does not have any eigenvector that is
perfectly localized on S and bandlimited on F .
Proof: From (11), exploiting the relation D = I − D, it
holds
(14)
I − BDB x0 = BD E{y[n]}.
Hence, it is possible to recover x0 from (14) if I − BDB is
invertible. This happens if the sufficient condition (13) holds
true. Conversely, if kDBk2 = 1 (or, equivalently, kBDk2 =
1), from (8) we know that there exist band-limited signals that
are perfectly localized over S. This implies that, if we sample
one of such signals over the set S, we get only zero values and
then it would be impossible to recover x0 from those samples.
This proves that condition (13) is also necessary.
A necessary condition that enables reconstruction, i.e. the
non-existence of a non-trivial vector x satisfying DBx = 0,
is that |S| ≥ |F |. However, this condition is not sufficient,
because matrix DB in (9) may loose rank, or easily become
ill-conditioned, depending on the graph topology and sampling
strategy defined by D. This suggests that the location of
samples plays a key role in the performance of the LMS
reconstruction algorithm in (12). For this reason, in Section
III.D we will consider a few alternative sampling strategies
satisfying different optimization criteria.
B. Mean-Square Analysis
When condition (13) holds true, Algorithm 1 can reconstruct
the graph signal from a subset of samples. In this section, we
study the mean-square behavior of the proposed LMS strategy,
illustrating how the sampling operator D affects its stability
and steady-state performance. From now on, we view the
estimates x[n] as realizations of a random process and analyze
the performance of the LMS algorithm in terms of its meane [n] = x[n] − x0 be the error vector at
square behavior. Let x
time n. Subtracting x0 from the left and right hand sides of
e [n], we obtain:
(12), using (9) and relation Be
x[n] = x
e [n + 1] = (I − µ BDB) x
e [n] + µ BDv[n].
x
(15)
Applying a GFT to each side of (15) (i.e., multiplying by UH ),
and exploiting the structure of matrix B in (5), we obtain
e
s[n + 1] = (I − µ ΣUH DUΣ) e
s[n] + µ ΣUH Dv[n], (16)
e [n] is the GFT of the error x
e [n]. From
where e
s[n] = UH x
(16) and the definition of Σ in (5), since sei [n] = 0 for all
i ∈
/ F , we can analyze the behavior of the error recursion
(16) only on the support of e
s[n], i.e. b
s[n] = {e
si [n], i ∈ F } ∈
C|F| . Thus, letting UF ∈ CN ×|F | be the matrix having as
columns the eigenvectors of the Laplacian matrix associated
to the frequency indices F , the error recursion (16) can be
rewritten in compact form as:
b
s[n + 1] = (I − µ UH
s[n] + µ UH
F DUF ) b
F Dv[n].
(17)
e [n] in the compact
The evolution of the error b
s[n] = UH
Fx
transformed domain is totally equivalent to the behavior of
e [n] from a mean-square error point of view. Thus, using
x
energy conservation arguments [57], we consider a general
s[n]H Φb
s[n],
weighted squared error sequence kb
s[n]k2Φ = b
5
where Φ ∈ C|F|×|F| is any Hermitian nonnegative-definite
matrix that we are free to choose. In the sequel, it will be clear
the role played by a proper selection of the matrix Φ. Then,
from (17) we can establish the following variance relation:
Ekb
s[n + 1]k2Φ = Ekb
s[n]k2Φ′ + µ2 E{v[n]H DUF ΦUH
F Dv[n]}
= Ekb
s[n]k2Φ′ + µ2 Tr(ΦUH
F DCv DUF )
(18)
where Tr(·) denotes the trace operator, and
H
.
DU
Φ
I
−
µ
U
DU
Φ′ = I − µ UH
F
F
F
F
(19)
Let ϕ = vec(Φ) and ϕ′ = vec(Φ′ ), where the notation vec(·)
stacks the columns of Φ on top of each other and vec−1 (·) is
the inverse operation. We will use interchangeably the notation
sH Φb
s. Exploiting
kb
sk2Φ and kb
sk2ϕ to denote the same quantity b
the Kronecker product property
vec(XΦY) = (Y H ⊗ X)vec(Φ),
and we conclude that Ekb
s[n]k2ϕ will converge to a steadystate value. From (22), we deduce that Q is stable if matrix
I − µUH
F DUF is stable as well. This holds true under the
two following conditions. The first condition is that matrix
H
UH
F DUF = I− UF DUF must have full rank, i.e. (13) holds
true, where we have exploited the relation
kDUF k = kDUΣk = kDBk.
Now recalling that, for any Hermitian matrix X, it holds
kXk = ρ(X) [58], with ρ(X) denoting the spectral radius
of X, the second condition guaranteing the stability of Q is
that kI−µUH
F DUF k < 1, which holds true for any step-sizes
satisfying (23). This concludes the first part of the Theorem.
We now prove the second part. Selecting ϕ = vec(I) in
(25), we obtain the following bound
Ekb
s[n]k2 ≤ Ekb
s[0]k2Qn vec(I) + µ2 c
n
X
kQkl
(26)
l=0
and the trace property
where c = krkkvec(I)k. Taking the limit of (26) as n → ∞,
since kQk < 1 if conditions (13) and (23) hold, we obtain
Tr(ΦX) = vec(XH )T vec(Φ),
in the relation (18), we obtain:
Ekb
s[n + 1]k2ϕ = Ekb
s[n]k2Qϕ + µ2 vec(G)T ϕ
lim Ekb
s[n]k2 ≤
n→∞
(20)
µ2 c
.
1 − kQk
(27)
From (22), we have
where
G = UH
F DCv DUF
(21)
H
Q = (I − µ UH
F DUF ) ⊗ (I − µ UF DUF ).
(22)
The following theorem guarantees the asymptotic mean-square
stability (i.e., convergence in the mean and mean-square error
sense) of the LMS algorithm in (12).
Theorem 3: Assume model (9) holds. Then, for any bounded
initial condition, the LMS strategy (12) asymptotically converges in the mean-square error sense if the sampling operator
D and the step-size µ are chosen to satisfy (13) and
2
,
0<µ<
(23)
λmax UH
F DUF
with λmax (A) denoting the maximum eigenvalue of the symmetric matrix A. Furthermore, it follows that, for sufficiently
small step-sizes:
lim sup Ekb
s[n]k2 = O(µ).
(24)
n→∞ n
Proof: Letting r = vec(G), recursion (20) can be
equivalently recast as:
Ekb
s[n]k2ϕ
=
Ekb
s[0]k2Qn ϕ
2 T
+µ r
n−1
X
Ql ϕ
(25)
l=0
where Ekb
s[0]k denotes the initial condition. We first note that
if Q is stable, Qn → 0 as n → ∞. In this way, the first
term on the RHS of (25) vanishes asymptotically. At the same
time, the convergence of the second term on the RHS
of (25)
Pn−1
depends only on the geometric series of matrices l=0
Ql ,
which is known to be convergent to a finite value if the
matrix Q is a stable matrix [58]. In summary, if Q is stable,
the RHS of (25) asymptotically converges to a finite value,
2
H
2
kQk = kI − µUH
DU
k
=
ρ
I
−
µU
DU
F
F
F
F
2
2
≤ max (1 − µδ) , (1 − µν)
(a)
≤ 1 − 2µν + µ2 δ 2
(28)
H
where δ = λmax (UH
F DUF ), ν = λmin (UF DUF ), and in (a)
we have exploited δ ≥ ν. Substituting (28) in (27), we get
µc
lim Ekb
s[n]k2 ≤
.
(29)
n→∞
2ν − µδ 2
It is easy to check that the upper bound (29) does not exceed
µc/ν for any stepsize 0 < µ < ν/δ 2 . Thus, we conclude that
(24) holds for sufficiently small step-sizes.
C. Steady-State Performance
Taking the limit of (20) as n → ∞ (assuming conditions
(13) and (23) hold true), we obtain:
lim Ekb
s[n]k2(I−Q)ϕ = µ2 vec(G)T ϕ.
n→∞
(30)
Expression (30) is a useful result: it allows us to derive several
performance metrics through the proper selection of the free
weighting parameter ϕ (or Φ). For instance, let us assume that
one wants to evaluate the steady-state mean square deviation
(MSD) of the LMS strategy in (12). Thus, selecting ϕ = (I −
Q)−1 vec(I) in (30), we obtain
MSD = lim Eke
x[n]k2 = lim Ekb
s[n]k2
n→∞
n→∞
= µ2 vec(G)T (I − Q)−1 vec(I).
(31)
If instead one is interested in evaluating the mean square
deviation obtained by the LMS algorithm in (12) when reconstructing the value of the signal associated to k-th vertex
6
Sampling strategy 1: Minimization of MSD
Input Data : M , the number of samples.
Output Data : S, the sampling set.
Function : initialize S ≡ ∅
while |S| < M
s = arg min vec(G(DS∪{j} ))T (I − Q(DS∪{j} ))† vec(I);
j
S ← S ∪ {s};
end
Function :
initialize S ≡ ∅
while |S| < M
s = arg max
UH
F DS∪{j} UF
S ← S ∪ {s};
end
=
lim Ekb
s[n]k2UH Ek UF
F
n→∞
= µ2 vec(G)T (I − Q)−1 vec(UH
F Ek UF ),
(32)
where Ek = diag{ek }, with ek ∈ RN denoting the k-th
canonical vector. In the sequel, we will confirm the validity
of these theoretical expressions by comparing them with
numerical simulations.
D. Sampling Strategies
As illustrated in the previous sections, the properties of the
proposed LMS algorithm in (12) strongly depend on the choice
of the sampling set S, i.e. on the vertex limiting operator D.
Indeed, building on the previous analysis, it is clear that the
sampling strategy must be carefully designed in order to: a)
enable reconstruction of the signal; b) guarantee stability of
the algorithm; and c) impose a desired mean-square error at
convergence. In particular, we will see that, when sampling
signals defined on graphs, besides choosing the right number
of samples, whenever possible it is also fundamental to have a
strategy indicating where to sample, as the samples’ location
plays a key role in the performance of the reconstruction
algorithm in (12). To select the best sampling strategy, one
should optimize some performance criterion, e.g. the MSD in
(31), with respect to the sampling set S, or, equivalently, the
vertex limiting operator D. However, since this formulation
translates inevitably into a selection problem, whose solution
in general requires an exhaustive search over all the possible
combinations, the complexity of such procedure becomes
intractable also for graph signals of moderate dimensions.
Thus, in the sequel we will provide some numerically efficient,
albeit sub-optimal, greedy algorithms to tackle the problem of
selecting the sampling set.
Greedy Selection - Minimum MSD: This strategy aims at
minimizing the MSD in (31) via a greedy approach: the
method iteratively selects the samples from the graph that lead
to the largest reduction in terms of MSD. Since the proposed
greedy approach starts from an initially empty sampling set,
when |S| < |F |, matrix I − Q in (31) is inevitably rank
deficient. Then, in this case, the criterion builds on the pseudoinverse of the matrix I − Q in (31), denoted by (I − Q)† ,
which coincides with the inverse as soon as |S| ≥ |F |.
The resulting algorithm is summarized in the table entitled
“Sampling strategy 1”, where we made explicit the dependence
+
Input Data : M , the number of samples.
Output Data : S, the sampling set.
j
of the graph, selecting ϕ = (I−Q)−1 vec(UH
F Ek UF ) in (30),
we obtain
MSDk = lim Eke
x[n]k2Ek
n→∞
Sampling strategy 2: Maximization of UH
F DUF
+
;
of matrices G and Q on the sampling operator D. In the
sequel, we will refer to this method as the Min-MSD strategy.
Greedy Selection - Maximum |UH
F DUF |+ : In this case, the
strategy aims at maximizing the volume of the parallelepiped
build with the selected rows of matrix UF . The algorithm
starts including the row with the largest norm in UF , and then
it adds, iteratively, the rows having the largest norm and, at the
same time, are as orthogonal as possible to the vectors already
in S. The rationale underlying this strategy is to design a well
suited basis for the graph signal that we want to estimate. This
criterion coincides with the maximization of the the pseudodeterminant of the matrix UH
F DUF (i.e. the product of all
nonzero eigenvalues), which is denoted by UH
. In
F DUF
+
the sequel, we motivate the rationale underlying this strategy.
Let us consider the eigendecomposition Q = VΛVH . From
(31), we obtain:
MSD = µ2 vec(G)T (I − Q)−1 vec(I)
= µ2 vec(G)T V(I − Λ)−1 VH vec(I)
|F|2
= µ2
X
i=1
pi · qi
1 − λi (Q)
(33)
where p = {pi } = VH vec(G), q = {qi } = VH vec(I). From
(33), we notice how the MSD of the LMS algorithm in (12)
strongly depends on the values assumed by the eigenvalues
λi (Q), i = 1, . . . , |F |2 . In particular, we would like to design
matrix Q in (22) such that its eigenvalues are as far as possible
from 1. From (22), it is easy to understand that
λi (Q) = 1 − µλk (UH
1 − µλl (UH
F DUF )
F DUF )
k, l = 1, . . . , |F |. Thus, requiring λi (Q), i = 1, . . . , |F |2 ,
to be as far as possible from 1 translates in designing the
|F |×|F |
matrix UH
such that its eigenvalues are as
F DUF ∈ C
far as possible from zero. Thus, a possible surrogate criterion
for the approximate minimization of (33) can be formulated
as the selection of the sampling set S (i.e. operator D) that
maximizes the determinant (i.e. the product of all eigenvalues)
H
of the matrix UH
F DUF . When |S| < |F |, matrix UF DUF
is inevitably rank deficient, and the strategy builds on the
pseudo-determinant of UH
F DUF . Of course, when |S| ≥ |F |,
the pseudo determinant coincides with the determinant. The
resulting algorithm is summarized in the table entitled “Sampling strategy 2”. In the sequel, we will refer to this method
as the Max-Det sampling strategy.
7
H
Sampling strategy 3: Maximization of λ+
min UF DUF
1.2
Input Data : M , the number of samples.
Output Data : S, the sampling set.
Function :
1
0.8
initialize S ≡ ∅
while |S| < M
H
s = arg max λ+
min UF DS∪{j} UF ;
0.6
0.4
j
0.2
S ← S ∪ {s};
end
0
−0.2
−0.4
−0.6
Fig. 1: Example of graph signal and sampling.
−26
−28
Steady−state MSD (dB)
H
Greedy Selection - Maximum λ+
min (UF DUF ): Finally,
using similar arguments as before, a further surrogate criterion for the minimization of (33) can be formulated as the
maximization of the minimum nonzero eigenvalue
of the
ma
H
H
+
trix UF DUF , which is denoted by λmin UF DUF . This
greedy strategy exploits the same idea of the sampling method
introduced in [10] in the case of batch signal reconstruction.
The resulting algorithm is summarized in the table entitled
“Sampling strategy 3”. We will refer to this method as the
Max-λmin sampling strategy.
In the sequel, we will illustrate some numerical results
aimed at comparing the performance achieved by the proposed
LMS algorithm using the aforementioned sampling strategies.
−30
−32
−34
−36
−38
−40
−42
0
E. Numerical Results
In this section, we first illustrate some numerical results
aimed at confirming the theoretical results in (31) and (32).
Then, we will illustrate how the sampling strategy affects the
performance of the proposed LMS algorithm in (12). Finally,
we will evaluate the effect of a graph mismatching in the
performance of the proposed algorithm.
Performance: Let us consider the graph signal shown in Fig.
1 and composed of N = 50 nodes, where the color of each
vertex denotes the value of the signal associated to it. The
signal has a spectral content limited to the first ten eigenvectors
of the Laplacian matrix of the graph in Fig. 1, i.e. |F | = 10.
The observation noise in (9) is zero-mean, Gaussian, with
a diagonal covariance matrix, where each element is chosen
uniformly random between 0 and 0.01. An example of graph
sampling, obtained selecting |S| = 10 vertexes using the MaxDet sampling strategy, is also illustrated in Fig. 1, where the
sampled vertexes have thicker marker edge. To validate the
theoretical results in (32), in Fig. 2 we report the behavior
of the theoretical MSD values achieved at each vertex of
the graph, comparing them with simulation results, obtained
averaging over 200 independent simulations and 100 samples
of squared error after convergence of the algorithm. The stepsize is chosen equal to µ = 0.5 and, together with the selected
sampling strategy D, they satisfy the reconstruction and stability conditions in (13) and (23). As we can notice from Fig.
2, the theoretical predictions match well the simulation results.
Effect of sampling strategies: It is fundamental to assess
the performance of the LMS algorithm in (12) with respect
to the adopted sampling set S. As a first example, using the
Max-Det sampling strategy, in Fig. 3 we report the transient
MSD − Simulation
MSD − Theory
10
20
30
Node Index
40
50
Fig. 2: Comparison between theoretical MSD in (30) and
simulation results, at each vertex of the graph. The theoretical
expressions match well with the numerical results.
behavior of the MSD, considering different number of samples
taken from the graph, i.e. different cardinalities |S| of the
sampling set. The results are averaged over 200 independent
simulations, and the step-sizes are tuned in order to have the
same steady-state MSD for each value of |S|. As expected,
from Fig. 3 we notice how the learning rate of the algorithm
improves by increasing the number of samples. Finally, in Fig.
4 we illustrate the steady-state MSD of the LMS algorithm in
(12) comparing the performance obtained by four different
sampling strategies, namely: a) the Max-Det strategy; b) the
Max-λmin strategy; c) the Min-MSD strategy; and d) the
random sampling strategy, which simply picks at random |S|
nodes. We consider the same parameter setting of the previous
simulation. The results are averaged over 200 independent
simulations. As we can notice from Fig. 4, the LMS algorithm
with random sampling can perform quite poorly, especially at
low number of samples. This poor result of random sampling
emphasizes that, when sampling a graph signal, what matters is
not only the number of samples, but also (and most important)
where the samples are taken. Comparing the other sampling
strategies, we notice from Fig. 4 that the Max-Det and Maxλmin strategies perform well also at low number of samples
(|S| = 10 is the minimum number of samples that allows
signal reconstruction). As expected, the Max-Det strategy
outperforms the Max-λmin strategy, because it considers all
8
10
0
|S| = 10
|S| = 20
|S| = 30
|S| = 50
Transient MSD (dB)
0
−4
−5
−10
−15
−6
−8
−10
−12
−14
−20
−16
−25
−18
0
50
100
150
200
250
Iteration index
300
350
Link 1 removed
Link 2 removed
Link 3 removed
Link 4 removed
Theoretical MSD
−2
Transient MSD (dB)
5
400
Fig. 3: Transient MSD, for different number of samples |S|.
Increasing the number of samples, the learning rate improves.
−20
0
50
100
150
200
250
Iteration index
300
350
400
Fig. 5: Transient MSD versus iteration index, for different
links removed from the original graph in Fig. 1.
−10
Min−MSD strategy
Max−Det strategy
Max−λmin strategy
−11
Steady−state MSD (dB)
−12
Random strategy
−13
−14
−15
−16
−17
−18
−19
−20
10
15
20
25
30
35
Number of samples
40
45
50
Fig. 4: Steady-state MSD versus number of samples, for
different sampling strategies.
the modes of the MSD in (33), as opposed to the single mode
associated to the minimum eigenvalue considered by the Maxλmin strategy. It is indeed remarkable that, for low number
of samples, Max-Det outperforms also Min-MSD, even if
the performance metric is MSD. There is no contradiction
here because we need to remember that all the proposed
methods are greedy strategies, so that there is no claim of
optimality in all of them. However, as the number of samples
increases above the limit |S| = |F | = 10, the Min-MSD
strategy outperforms all other methods. This happens because
the Min-MSD strategy takes into account information from
both graph topology and spatial distribution of the observation
noise (cf. (31)). Thus, when the number of samples is large
enough to have sufficient degrees of freedom in selecting the
samples’ location, the Min-MSD strategy has the capability
of selecting the vertexes in a good position to enable a wellconditioned signal recovery, with possibly low additive noise,
thus improving the overall performance of the LMS algorithm
in (12). Conversely, when the number of samples is very
close to its minimum value, the Min-MSD criterion may
give rise to ill-conditioning of the signal recovery strategy
because the low noise samples may be in sub-optimal positions
with respect to signal recovery. This explains its losses with
respect to Max-Det and Max-λmin strategies, for low values
of the number of samples. This analysis suggests that an
optimal design of the sampling strategy for graph signals
should take into account processing complexity (in terms of
number of samples), prior knowledge (e.g., graph structure,
noise distribution), and achievable performance.
Effect of graph mismatching: In this last example, we aim
at illustrating how the performance of the proposed method
is affected by a graph mismatching during the processing. To
this aim, we take as a benchmark the graph signal in Fig.
1, where the signal bandwidth is set equal to |F | = 10.
The bandwidth defines also the sampling operator D, which
is selected through the Max-Det strategy, introduced in Sec.
III.D, using |S| = 10 samples. Now, we assume that the LMS
processing in (12) is performed keeping fixed the sampling
operator D, while adopting an operator B in (5) that uses
the same bandwidth as for the benchmark case (i.e., the
same matrix ΣF ), but different GFT operators U, which are
generated as the eigenvectors of Laplacian matrices associated
to graphs that differs from the benchmark in Fig. 1 for one
(removed) link. The aim of this simulation is to quantify the
effect of a single link removal on the performance of the
LMS strategy in (12). Thus, in Fig. 5, we report the transient
MSD versus the iteration index of the proposed LMS strategy,
considering four different links that are removed from the
original graph. The four removed links are those shown in Fig.
1 using thicker lines; the colors and line styles are associated to
the four behaviors of the transient MSD in Fig. 5. The results
are averaged over 100 independent simulations, using a stepsize µ = 0.5. The theoretical performance in (31) achieved
by the ideal LMS, i.e. the one perfectly matched to the graph,
are also reported as a benchmark. As we can see from Fig.
5, the removal of different links from the graph leads to very
different performance obtained by the algorithm. Indeed, while
removing Link 1 (i.e., the red one), the algorithm performs as
in the ideal case, the removal of links 2, 3, and 4, progressively
9
determine a worse performance loss. This happens because the
structure of the eigenvectors of the Laplacian of the benchmark
graph is more or less preserved by the removal of specific
links. Some links have almost no effects (e.g., Link 1), whereas
some others (e.g., Link 4) may lead to deep modification
of the structure of such eigenvectors, thus determining the
mismatching of the LMS strategy in (12) and, consequently, its
performance degradation. This example opens new theoretical
questions that aim at understanding which links affect more
the graph signals’ estimation performance in situations where
both the signal and the graph are jointly time-varying. We plan
to tackle this exciting case in future work.
IV. LMS E STIMATION WITH A DAPTIVE G RAPH S AMPLING
The LMS strategy in (12) assumes perfect knowledge of
the support where the signal is defined in the graph frequency
domain, i.e. F . Indeed, this prior knowledge allows to define
the projector operator B in (5) in a unique manner, and to
implement the sampling strategies introduced in Sec. III.D.
However, in many practical situations, this prior knowledge is
unrealistic, due to the possible variability of the graph signal
over time at various levels: the signal can be time varying
according to a given model; the signal model may vary over
time, for a given graph topology; the graph topology may vary
as well over time. In all these situations, we cannot always
assume that we have prior information about the frequency
support F , which must then be inferred directly from the
streaming data y[n] in (9). Here, we consider the important
case where the graph is fixed, and the spectral content of the
signal can vary over time in an unknown manner. Exploiting
the definition of GFT in (3), the signal observations in (9) can
be recast as:
y[n] = DUs0 + Dv[n].
(34)
The problem then translates in estimating the coefficients of
the GFT s0 , while identifying its support, i.e. the set of indexes
where s0 is different from zero. The support identification
is deeply related to the selection of the sampling set. Thus,
the overall problem can be formulated as the joint estimation
of sparse representation s and sampling strategy D from the
observations y[n] in (34), i.e.,
min Eky[n] − DUsk2 + λ f (s),
(35)
s,D∈D
where D is the (discrete) set that constraints the selection of
the sampling strategy D, f (·) is a sparsifying penalty function
(typically, ℓ0 or ℓ1 norms), and λ > 0 is a parameter that
regulates how sparse we want the optimal GFT vector s.
Problem (35) is a mixed integer nonconvex program, which is
very complicated to solve, especially in the adaptive context
considered in this paper. Thus, to favor low complexity online
solutions for (35), we propose an algorithm that alternates
between the optimization of the vector s and the selection of
the sampling operator D. The rationale behind this choice is
that, given an estimate for the support of vector s, i.e. F ,
we can select the sampling operator D in a very efficient
manner through one of the sampling strategies illustrated in
Sec. III.D. Then, starting from a random initialization for
Algorithm 2: LMS with Adaptive Graph Sampling
Start with s[0] chosen at random, D[0] = I, and F [0] = V.
Given µ > 0, for each time n > 0, repeat:
1) s[n + 1] = Tλµ s[n] + µ UH D[n] (y[n] − Us[n]) ;
2) Set F [n + 1] = {i ∈ {1, . . . , N } : si [n + 1] 6= 0};
3) Given UF [n+1] , select D[n + 1] according to one of the
criteria proposed in Sec. III.D;
s and a full sampling for D (i.e., D = I), the algorithm
iteratively proceeds as follows. First, fixing the value of the
sampling operator D[n] at time n, we update the estimate of
the GFT vector s using an online version of the celebrated
ISTA algorithm [59], [60], which proceeds as:
s[n + 1] = T λµ s[n] + µ UH D[n] (y[n] − Us[n]) , (36)
n ≥ 0, where µ > 0 is a (sufficiently small) step-size, and
T γ (s) is a thresholding function that depends on the sparsityinducing penalty f (·) in (35). Several choices are possible, as
we will illustrate in the sequel. The aim of recursion (36)
is to estimate the GFT s0 of the graph signal x0 in (9),
while selectively shrinking to zero all the components of s0
that are outside its support, i.e., which do not belong to the
bandwidth of the graph signal. Then, the online identification
of the support of the GFT s0 enables the adaptation of the
sampling strategy, which can be updated using one of the
strategies illustrated in Sec. III.D. Intuitively, the algorithm
will increase (reduce) the number of samples used for the estimation, depending on the increment (reduction) of the current
signal bandwidth. The main steps of the LMS algorithm with
adaptive graph sampling are listed in Algorithm 2.
Thresholding functions : Several different functions can
be used to enforce sparsity. A commonly used thresholding
function comes directly by imposing an ℓ1 norm constraint
in (35), which is commonly known as the Lasso [47]. In this
case, the vector threshold function T γ (s) is the componentwise thresholding function Tγ (sm ) applied to each element of
vector s, with
sm − γ, sm > γ;
Tγ (sm ) =
0,
−γ ≤ sm ≤ γ;
(37)
sm + γ, sm < −γ.
The function T γ (s) in (37) tends to shrink all the components
of the vector s and, in particular, sets to zero the components
whose magnitude are within the threshold γ. Since the Lasso
constraint is known for introducing a large bias in the estimate,
the performance would deteriorate for vectors that are not
sufficiently sparse, i.e. graph signals with large bandwidth.
To reduce the bias introduced by the Lasso constraint, several
other thresholding functions can be adopted to improve the
performance also in the case of less sparse systems. A potential
improvement can be made by considering the non-negative
Garotte estimator as in [61], whose thresholding function is
defined as a vector whose entries are derived applying the
10
0
Lasso
Garotte
Hard Thresholding
−5
Transient NMSD (dB)
−10
−15
−20
−25
−30
−35
−40
−45
0
50
100
150
200
Iteration index
250
300
Fig. 8: Optimal Sampling at iteration n = 80.
Fig. 6: LMS with Adaptive Sampling: NMSD versus iteration
index, for different thresholding functions.
40
Lasso
Garotte
Hard Thresholding
True bandwidth
35
Cardinality of F
30
25
20
15
10
Fig. 9: Optimal Sampling at iteration n = 180.
5
0
0
50
100
150
200
250
300
Iteration Index
Fig. 7: LMS with Adaptive Sampling: |F | versus iteration
index, for different thresholding functions.
threshold
Tγ (sm ) =
(
sm (1 − γ 2 /s2m ),
|sm | > γ;
0,
|sm | ≤ γ;
(38)
m = 1, . . . , M . Finally, to completely remove the bias over
the large components, we can implement a hard thresholding
mechanism, whose function is defined as a vector whose
entries are derived applying the threshold
(
sm ,
|sm | > γ;
Tγ (sm ) =
(39)
0,
|sm | ≤ γ;
In the sequel, numerical results will illustrate how different
thresholding functions such as (37), (38), and (39), affect the
performance of Algorithm 2.
Numerical Results
In this section, we illustrate some numerical results aimed
at assessing the performance of the proposed LMS method
with adaptive graph sampling, i.e. Algorithm 2. In particular,
Fig. 10: Optimal Sampling at iteration n = 280.
to illustrate the adaptation capabilities of the algorithm, we
simulate a scenario with a time-varying graph signal with
N = 50 nodes, which has the same topology shown in Fig.
1, and spectral content that switches between the first 5, 15,
and 10 eigenvectors of the Laplacian matrix of the graph.
The elements of the GFT s0 inside the support are chosen
to be equal to 1. The observation noise in (9) is zero-mean,
Gaussian, with a diagonal covariance matrix Cv = σv2 I, with
σv2 = 4 × 10−4 . In Fig. 6 we report the transient behavior of
11
the normalized Mean-Square Deviation (NMSD), i.e.
NMSD[n] =
ks[n] − s0 k2
,
ks0 k2
versus the iteration index, considering the evolution of Algorithm 2 with three different thresholding functions, namely:
a) the Lasso threshold in (37), the Garotte threshold in (38),
and the hard threshold in (39). Also, in Fig. 7, we illustrate
the behavior of the estimate of the cardinality of F versus the
iteration index (cf. Step 2 of Algorithm 2), obtained by the
three aforementioned strategies at each iteration. The value
of the cardinality of F of the true underlying graph signal
is also reported as a benchmark. The curves are averaged
over 100 independent simulations. The step-size is chosen to
be µ = 0.5, the sparsity parameter λ = 0.1, and thus the
threshold is equal to γ = µλ = 0.05 for all strategies. The
sampling strategy used in Step 3 of Algorithm 2 is the MaxDet method introduced in Sec. III.D, where the number of
samples M [n] to be selected at each iteration is chosen to be
equal to the current estimate of cardinality of the set F [n]. As
we can notice from Fig. 6, the LMS algorithm with adaptive
graph sampling is able to track time-varying scenarios, and its
performance is affected by the adopted thresholding function.
In particular, from Fig. 6, we notice how the algorithm based
on the hard thresholding function in (39) outperforms the other
strategies in terms of steady-state NMSD, while having the
same learning rate. The Garotte based algorithm has slightly
worse performance with respect to the method exploiting hard
thresholding, due to the residual bias introduced at large values
by the thresholding function in (38). Finally, we can notice
how the LMS algorithm based on Lasso may lead to very
poor performance, due to misidentifications of the true graph
bandwidth. This can be noticed from Fig. 7 where, while
the Garotte and hard thresholding strategies are able to learn
exactly the true bandwidth of the graph signal (thus leading
to very good performance in terms of NMSD, see Fig. 6), the
Lasso strategy overestimates the bandwidth of the signal, i.e.
the cardinality of the set F (thus leading to poor estimation
performance, see Fig. 6). Finally, to illustrate an example of
adaptive sampling, in Figs. 8, 9, and 10 we report the samples
(depicted as black nodes) chosen by the proposed LMS
algorithm based on hard thresholding at iterations n = 80,
n = 180, and n = 280. As we can notice from Figs. 6, 7
and 8, 9, and 10, the algorithm always selects a number of
samples equal to the current value of the signal bandwidth,
while guaranteeing good reconstruction performance.
V. A PPLICATION TO P OWER S PATIAL D ENSITY
E STIMATION IN C OGNITIVE N ETWORKS
The advent of intelligent networking of heterogeneous devices such as those deployed to monitor the 5G networks,
power grid, transportation networks, and the Internet, will
have a strong impact on the underlying systems. Situational
awareness provided by such tools will be the key enabler
for effective information dissemination, routing and congestion control, network health management, risk analysis, and
security assurance. The vision is for ubiquitous smart network
devices to enable data-driven statistical learning algorithms
−10
1.1
−8
1
−6
0.9
−4
0.8
−2
0.7
0
0.6
2
0.5
4
0.4
6
0.3
8
10
−10
0.2
−5
0
5
10
Fig. 11: PSD cartography: spatial distribution of primary users’
power, small cell base stations deployment, graph topology,
and graph signal.
for distributed, robust, and online network operation and
management, adaptable to the dynamically evolving network
landscape with minimal need for human intervention. In
this context, the unceasing demand for continuous situational
awareness in cognitive radio (CR) networks calls for innovative signal processing algorithms, complemented by sensing
platforms to accomplish the objectives of layered sensing
and control. These challenges are embraced in the study of
power cartography, where CRs collect data to estimate the
distribution of power across space, namely the power spatial
density (PSD). Knowing the PSD at any location allows
CRs to dynamically implement a spatial reuse of idle bands.
The estimated PSD map need not be extremely accurate, but
precise enough to identify idle spatial regions.
In this section, we apply the proposed framework for
LMS estimation of graph signals to spectrum cartography
in cognitive networks. We consider a 5G scenario, where a
dense deployment of radio access points (RAPs) is envisioned
to provide a service environment characterized by very low
latency and high rate access. Each RAP collects streaming data
related to the spectrum utilization of primary users (PU’s) at
its geographical position. This information can then be sent to
a processing center, which collects data from the entire system,
through high speed wired links. The aim of the center is then
to build a spatial map of the spectrum usage, while processing
the received data on the fly and envisaging proper sampling
techniques that enable a proactive sensing of the system from
only a limited number of RAP’s measurements. As we will
see in the sequel, the proposed approach hinges on the graph
structure of the signal received from the RAP’s, thus enabling
real-time PSD estimation from a small set of observations that
are smartly sampled from the graph.
Numerical examples: Let us consider an operating region
where 100 RAPs are randomly deployed to produce a map of
the spatial distribution of power generated by the transmissions
of two active primary users. The PU’s emit electromagnetic
radiation with power equal to 1 Watt. For simplicity, the
12
0
0
|F| = 10
|F| = 20
|F| = 40
|F| = 60
−5
Transient NMSD (dB)
Steady−state NMSD (dB)
−5
−10
−15
−20
−10
−15
−20
−25
−30
0
|F| = 10 ; |S| = 20
|F| = 20 ; |S| = 30
|F| = 30 ; |S| = 40
20
40
60
Number of Samples
80
100
0
50
100
150
200
250
Iteration index
300
350
400
Fig. 12: PSD cartography: Steady-state NMSD versus number
of samples taken from the graph, for different bandwidths used
for processing.
Fig. 13: PSD cartography: Transient NMSD versus iteration
index, for different number of samples and bandwidths used
for processing.
propagation medium is supposed to introduce a free-space path
loss attenuation on the PU’s transmissions. The graph among
RAPs is built from a distance based model, i.e. stations that
are sufficiently close to each other are connected through a
link (i.e. aij = 1, if nodes i and j are neighbors). In Fig. 11,
we illustrate a pictorial description of the scenario, and of the
resulting graph signal. We assume that each RAP is equipped
with an energy detector, which estimates the received signal
using 100 samples, considering an additive white Gaussian
noise with variance σv2 = 10−4 . The resulting signal is
not perfectly band-limited, but it turns out to be smooth
over the graph, i.e. neighbor nodes observe similar values.
This implies that sampling such signals inevitably introduces
aliasing during the reconstruction process. However, even if
we cannot find a limited (lower than N ) set of frequencies
where the signal is completely localized, the greatest part
of the signal energy is concentrated at low frequencies. This
means that if we process the data using a sufficient number
of observations and (low) frequencies, we should still be able
to reconstruct the signal with a satisfactory performance.
To illustrate an example of cartography based on the LMS
algorithm in (12), in Fig. 12 we report the behavior of the
steady-state NMSD versus the number of samples taken from
the graph, for different bandwidths used for processing. The
step-size is chosen equal to 0.5, while the adopted sampling
strategy is the Max-Det method introduced in Sec. III.D.
The results are averaged over 200 independent simulations.
As expected, from Fig. 12, we notice that the steady-state
NMSD of the LMS algorithm in (12) improves by increasing
the number of samples and bandwidths used for processing.
Interestingly, in Fig. 12 we can see a sort of threshold
behavior: the NMSD is large for |S| < |F |, when the signal is
undersampled, whereas the values become lower and stable as
soon as |S| > |F |. Finally, we illustrate an example that shows
the tracking capability of the proposed method in time-varying
scenarios. In particular, we simulate a situation the two PU’s
switch between idle and active modes: for 0 ≤ n < 133 only
the first PU transmits; for 133 ≤ n < 266 both PU’s transmit;
for 266 ≤ n ≤ 400 only the second PU’s transmits. In Fig. 13
we show the behavior of the transient NMSD versus iteration
index, for different number of samples and bandwidths used
for processing. The results are averaged over 200 independent
simulations. From Fig. 13, we can see how the proposed
technique can track time-varying scenarios. Furthermore, its
steady-state performance improves with increase in the number
of samples and bandwidths used for processing. These results,
together with those achieved in Fig. 12, illustrate an existing
tradeoff between complexity, i.e. number of samples used for
processing, and mean-square performance of the proposed
LMS strategy. In particular, using a larger bandwidth and a
(consequent) larger number of samples for processing, the
performance of the algorithm improves, at the price of a larger
computational complexity.
VI. C ONCLUSIONS
In this paper we have proposed LMS strategies for adaptive
estimation of signals defined over graphs. The proposed strategies are able to exploit the underlying structure of the graph
signal, which can be reconstructed from a limited number of
observations properly sampled from a subset of vertexes, under
a band-limited assumption. A detailed mean square analysis
illustrates the deep connection between sampling strategy
and the properties of the proposed LMS algorithm in terms
of reconstruction capability, stability, and mean-square error
performance. From this analysis, some sampling strategies
for adaptive estimation of graph signals are also derived.
Furthermore, to cope with time-varying scenarios, we also
propose an LMS method with adaptive graph sampling, which
estimates and tracks the signal support in the (graph)frequency
domain, while at the same time adapting the graph sampling
strategy. Several numerical simulations confirm the theoretical
findings, and illustrate the potential advantages achieved by
these strategies for adaptive estimation of band-limited graph
signals. Finally, we apply the proposed method to estimate and
track the spatial distribution of power transmitted by primary
users in a cognitive network environment, thus illustrating
13
the existing tradeoff between complexity and mean-square
performance of the proposed strategy.
We expect that such processing tools will represent a key
technology for the design and proactive sensing of Cyber
Physical Systems, where a proper adaptive control mechanism
requires the availability of data driven sampling strategies
able to control the overall system by only checking a limited
number of nodes, in order to collect correct information at the
right time, in the right place, and for the right purpose.
R EFERENCES
[1] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular
domains,” IEEE Signal Proc. Mag., vol. 30, no. 3, pp. 83–98, 2013.
[2] A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on
graphs,” IEEE Trans. on Signal Processing, vol. 61, pp. 1644–1656,
2013.
[3] ——, “Big data analysis with signal processing on graphs: Representation and processing of massive data sets with irregular structure,” IEEE
Signal Proc. Mag., vol. 31, no. 5, pp. 80–90, 2014.
[4] A. Sandryhaila and J. M. Moura, “Discrete signal processing on graphs:
Frequency analysis,” IEEE Transactions on Signal Processing, vol. 62,
no. 12, pp. 3042–3054, 2014.
[5] S. K. Narang and A. Ortega, “Perfect reconstruction two-channel wavelet
filter banks for graph structured data,” IEEE Transactions on Signal
Processing, vol. 60, no. 6, pp. 2786–2799, 2012.
[6] ——, “Compact support biorthogonal wavelet filterbanks for arbitrary
undirected graphs,” IEEE Transactions on Signal Processing, vol. 61,
no. 19, pp. 4673–4685, 2013.
[7] I. Z. Pesenson, “Sampling in Paley-Wiener spaces on combinatorial
graphs,” Trans. of the American Mathematical Society, vol. 360, no. 10,
pp. 5603–5627, 2008.
[8] I. Z. Pesenson and M. Z. Pesenson, “Sampling, filtering and sparse
approximations on combinatorial graphs,” Journal of Fourier Analysis
and Applications, vol. 16, no. 6, pp. 921–942, 2010.
[9] X. Zhu and M. Rabbat, “Approximating signals supported on graphs,” in
IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP),
March 2012, pp. 3921–3924.
[10] S. Chen, R. Varma, A. Sandryhaila, and J. Kovačević, “Discrete signal
processing on graphs: Sampling theory,” IEEE Trans. on Signal Proc.,
vol. 63, pp. 6510–6523, Dec. 2015.
[11] M. Püschel and J. M. F. Moura, “Algebraic signal processing theory:
Foundation and 1-D time,” IEEE Trans. Signal Process., vol. 56, pp.
3572–3585, 2008.
[12] ——, “Algebraic signal processing theory: 1-D space,” IEEE Trans. on
Signal Processing, pp. 3586–3599, 2008.
[13] S. Sardellitti and S. Barbarossa, “On the graph fourier transform for
directed graphs,” available at: http://arxiv.org/abs/1601.05972, 2016.
[14] A. Agaskar and Y. M. Lu, “A spectral graph uncertainty principle,” IEEE
Trans. on Inform. Theory, vol. 59, no. 7, pp. 4338–4356, 2013.
[15] B. Pasdeloup, R. Alami, V. Gripon, and M. Rabbat, “Toward an
uncertainty principle for weighted graphs,” online: arXiv:1503.03291,
2015.
[16] J. J. Benedetto and P. J. Koprowski, “Graph theoretic uncertainty principles,” http://www.math.umd.edu/ jjb/graph theoretic UP April 14.pdf,
2015.
[17] P. J. Koprowski,
“Finite
frames
and
graph
theoretic
uncertainty principles,” Ph.D. dissertation, 2015. [Online]. Available:
http://hdl.handle.net/1903/16666
[18] M. Tsitsvero, S. Barbarossa, and P. Di Lorenzo, “Signals on graphs:
Uncertainty principle and sampling,” to appear on the IEEE Trans. on
Signal Processing; available at http://arxiv.org/abs/1507.08822, 2015.
[19] S. Narang, A. Gadde, and A. Ortega, “Signal processing techniques for
interpolation in graph structured data,” in IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), May 2013, pp.
5445–5449.
[20] X. Wang, P. Liu, and Y. Gu, “Local-set-based graph signal reconstruction,” IEEE Trans. on Signal Processing, vol. 63, no. 9, pp. 2432–2444,
2015.
[21] A. G. Marquez, S. Segarra, G. Leus, and A. Ribeiro, “Sampling of graph
signals with successive local aggregations,” To appear on IEEE Trans.
Signal Process., 2016.
[22] M. Tsitsvero and S. Barbarossa, “On the degrees of freedom of signals
on graphs,” in 2015 European Signal Proc. Conf. (Eusipco 2015), Sep.
2015, pp. 1521–1525.
[23] S. K. Narang, A. Gadde, E. Sanou, and A. Ortega, “Localized iterative
methods for interpolation in graph structured data,” in IEEE Global
Conference on Signal and Information Processing, 2013.
[24] X. Wang, M. Wang, and Y. Gu, “A distributed tracking algorithm for
reconstruction of graph signals,” IEEE Journal of Selected Topics in
Signal Processing, vol. 9, no. 4, pp. 728–740, 2015.
[25] S. Chen, F. Cerda, P. Rizzo, J. Bielak, J. H. Garrett, and J. Kovacevic,
“Semi-supervised multiresolution classification using adaptive graph
filtering with application to indirect bridge structural health monitoring,”
IEEE Trans. on Signal Processing, vol. 62, no. 11, pp. 2879–2893, 2014.
[26] A. Sandryhaila and J. M. Moura, “Classification via regularization on
graphs.” in Proc. of IEEE Global conference on Signal and Information
Processing, 2013, pp. 495–498.
[27] V. N. Ekambaram, G. Fanti, B. Ayazifar, and K. Ramchandran, “Waveletregularized graph semi-supervised learning,” in IEEE Global Conference
on Signal and Information Processing, 2013, pp. 423–426.
[28] X. Zhang, X. Dong, and P. Frossard, “Learning of structured graph
dictionaries,” in IEEE International Conference on Acoustics, Speech
and Signal Processing, 2012, pp. 3373–3376.
[29] D. Thanou, D. I. Shuman, and P. Frossard, “Parametric dictionary
learning for graph signals,” in IEEE Global Conference on Signal and
Information Processing, 2013, pp. 487–490.
[30] X. Dong, D. Thanou, P. Frossard, and P. Vandergheynst, “Learning
graphs from signal observations under smoothness prior,” submitted to:
IEEE Transacctions on Signal Processing, 2014.
[31] D. Zhou and B. Schölkopf, “A regularization framework for learning
from graph data,” in ICML workshop on statistical relational learning
and Its connections to other fields, vol. 15, 2004, pp. 67–68.
[32] M. Belkin, P. Niyogi, and V. Sindhwani, “Manifold regularization: A
geometric framework for learning from labeled and unlabeled examples,”
The J. of Machine Learning Research, vol. 7, pp. 2399–2434, 2006.
[33] S. Chen, A. Sandryhaila, G. Lederman, Z. Wang, J. M. Moura, P. Rizzo,
J. Bielak, J. H. Garrett, and J. Kovacevic, “Signal inpainting on graphs
via total variation minimization,” in IEEE International Conference on
Acoustics, Speech and Signal Processing, 2014, pp. 8267–8271.
[34] S. Chen, A. Sandryhaila, J. M. Moura, and J. Kovacevic, “Signal
denoising on graphs via graph filtering,” in IEEE Global Conference
on Signal and Information Processing, 2014, pp. 872–876.
[35] P.-Y. Chen and A. O. Hero, “Local fiedler vector centrality for detection
of deep and overlapping communities in networks,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2014,
pp. 1120–1124.
[36] S. Chen, A. Sandryhaila, J. M. Moura, and J. Kovacevic, “Signal
recovery on graphs: Variation minimization,” IEEE Transactions on
Signal Processing, vol. 63, no. 17, pp. 4609–4624, 2015.
[37] S. Chen, R. Varma, A. Singh, and J. Kovačević, “Signal recovery
on graphs: Fundamental limits of sampling strategies,” arXiv preprint
arXiv:1512.05405, 2015.
[38] A. H. Sayed, Adaptive filters. John Wiley & Sons, 2011.
[39] ——, “Adaptation, learning, and optimization over networks,” Foundations and Trends R in Machine Learning, vol. 7, no. 4-5, pp. 311–801,
2014.
[40] F. R. K. Chung, Spectral Graph Theory.
American Mathematical
Society, 1997.
[41] B. Widrow and S. D. Stearns, “Adaptive signal processing,” Englewood
Cliffs, NJ, Prentice-Hall, Inc.,, vol. 1, 1985.
[42] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, 2006.
[43] R. G. Baraniuk, “Compressive sensing,” IEEE signal processing magazine, vol. 24, no. 4, 2007.
[44] J. Homer, I. Mareels, R. R. Bitmead, B. Wahlberg, and F. Gustafsson,
“LMS estimation via structural detection,” IEEE Transactions on Signal
Processing, vol. 46, no. 10, pp. 2651–2663, 1998.
[45] M. Godavarti and A. O. Hero III, “Partial update LMS algorithms,”
IEEE Trans. on Signal Processing, vol. 53, no. 7, pp. 2382–2399, 2005.
[46] D. L. Duttweiler, “Proportionate normalized least-mean-squares adaptation in echo cancelers,” IEEE Transactions on Speech and Audio
Processing, vol. 8, no. 5, pp. 508–518, 2000.
[47] R. Tibshirani, “Regression shrinkage and selection via the lasso,”
Journal of the Royal Statistical Society. Series B (Methodological), pp.
267–288, 1996.
[48] E. J. Candes, M. B. Wakin, and S. P. Boyd, “Enhancing sparsity
by reweighted l 1 minimization,” Journal of Fourier analysis and
applications, vol. 14, no. 5-6, pp. 877–905, 2008.
14
[49] Y. Chen, Y. Gu, and A. O. Hero III, “Sparse LMS for system identification,” in IEEE International Conference on Acoustics, Speech and
Signal Processing, 2009, pp. 3125–3128.
[50] D. Angelosante, J. A. Bazerque, and G. B. Giannakis, “Online adaptive
estimation of sparse signals: Where RLS meets the-norm,” IEEE Transactions on Signal Processing, vol. 58, no. 7, pp. 3436–3447, 2010.
[51] Y. Kopsinis, K. Slavakis, and S. Theodoridis, “Online sparse system
identification and signal reconstruction using projections onto weighted
balls,” IEEE Transactions on Signal Processing, vol. 59, no. 3, pp. 936–
952, 2011.
[52] S. Chouvardas, K. Slavakis, Y. Kopsinis, and S. Theodoridis, “A sparsity
promoting adaptive algorithm for distributed learning,” IEEE Transactions on Signal Processing, vol. 60, no. 10, pp. 5412–5425, 2012.
[53] P. Di Lorenzo and A. H. Sayed, “Sparse distributed learning based on
diffusion adaptation,” IEEE Transactions on Signal Processing, vol. 61,
no. 6, pp. 1419–1433, 2013.
[54] P. Di Lorenzo, S. Barbarossa, and A. H. Sayed, “Distributed spectrum
estimation for small cell networks based on sparse diffusion adaptation,”
IEEE Signal Processing Letters, vol. 20, no. 12, pp. 1261–1265, 2013.
[55] P. Di Lorenzo, “Diffusion adaptation strategies for distributed estimation
[56]
[57]
[58]
[59]
[60]
[61]
over gaussian markov random fields,” IEEE Transactions on Signal
Processing, vol. 62, no. 21, pp. 5748–5760, 2014.
S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, Distributed Detection
and Estimation in Wireless Sensor Networks. Academic Press Library
in Signal Processing, 2014, vol. 2, pp. 329–408.
A. H. Sayed and V. H. Nascimento, Energy conservation and the
learning ability of LMS adaptive filters. John Wiley, New York, 2003.
R. A. Horn and C. R. Johnson, Eds., Matrix Analysis. New York, NY,
USA: Cambridge University Press, 1986.
I. Daubechies, M. Defrise, and C. De Mol, “An iterative thresholding
algorithm for linear inverse problems with a sparsity constraint,” Communications on pure and applied mathematics, vol. 57, no. 11, pp. 1413–
1457, 2004.
A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM journal on imaging sciences,
vol. 2, no. 1, pp. 183–202, 2009.
M. Yuan and Y. Lin, “On the non-negative garrotte estimator,” Journal of
the Royal Statistical Society: Series B (Statistical Methodology), vol. 69,
no. 2, pp. 143–161, 2007.
| 3cs.SY
|
arXiv:1506.07898v4 [cs.DS] 20 Jun 2017
Efficient computation of middle levels Gray codes
Torsten Mütze†
Jerri Nummenpalo
School of Mathematics
Georgia Institute of Technology
30332 Atlanta GA, USA
[email protected]
Department of Computer Science
ETH Zürich
8092 Zürich, Switzerland
[email protected]
∗
Abstract. For any integer n ≥ 1 a middle levels Gray code is a cyclic listing of all
bitstrings of length 2n+1 that have either n or n+1 entries equal to 1 such that any two
consecutive bitstrings in the list differ in exactly one bit. The question whether such
a Gray code exists for every n ≥ 1 has been the subject of intensive research during
the last 30 years, and has been answered affirmatively only recently [T. Mütze. Proof
of the middle levels conjecture. Proc. London Math. Soc., 112(4):677–713, 2016]. In
this work we provide the first efficient algorithm to compute a middle levels Gray code.
For a given bitstring, our algorithm computes the next ` bitstrings in the Gray code in
time O(n`(1 + n` )), which is O(n) on average per bitstring provided that ` = Ω(n).
Keywords: Gray code, middle levels conjecture
1. Introduction
Efficiently generating all objects in a particular combinatorial class (e.g. permutations, combinations,
partitions or trees) in such a way that each object is generated exactly once is one of the oldest and
most fundamental problems in the area of combinatorial algorithms, and such generation algorithms
are used as core building blocks in a wide range of practical applications. The survey [Sav97] lists
numerous references. A classical example is the so-called binary Gray code [Gra53], which lists
all 2n bitstrings of length n such that any two consecutive bitstrings differ in exactly one bit. A
straightforward implemention of this algorithm takes time O(n) to compute from a given bitstring
the next one in the list (see Algorithm G in [Knu11, Section 7.2.1.1]), which can be improved to
O(1) [Ehr73, BER76] (see Algorithm L in [Knu11, Section 7.2.1.1]). The space requirement of both
algorithms is O(n). Similar minimum-change generation algorithms have been developed for various
other combinatorial classes. We exemplarily cite four examples from the excellent survey [Sav97]
on this topic: (1) listing all permutations of {1, 2, . . . , n} so that consecutive permutations differ
only by the swap of one pair of adjacent elements [Joh63, Tro62] (see also [Sed77]), (2) listing all kelement subsets of an n-element set such that consecutive sets differ only by exchanging one element
[Ehr73, BER76, EHR84, EM84, Rus88], (3) listing all binary trees with n vertices so that consecutive
trees differ only by one rotation operation [Luc87, LvBR93], (4) listing all spanning trees of a graph
such that consecutive trees differ only by exchanging one edge [Cum66, HH72].
Coming back to Gray codes, we say that a bitstring of length n has weight k, if it has exactly k entries
equal to 1 and n − k entries equal to 0. Furthermore, for any integer n ≥ 1 we define a middle levels
Gray code as a cyclic listing of all bitstrings of length 2n + 1 that have weight n or n + 1 such that
any two consecutive bitstrings in the list differ in exactly one bit. The name ‘middle levels’ becomes
clear when considering the relevant bitstrings as subsets in the Hasse diagram of the subset inclusion
∗
An extended abstract of this work has appeared in the proceedings of the European Symposium on Algorithms (ESA)
2015.
†The author was supported by a fellowship of the Swiss National Science Foundation. This work was initiated when
the author was at the Department of Computer Science at ETH Zürich.
1
2
2n+1
lattice. Clearly, a middle levels Gray code has to visit N := 2n+1
+ n+1 = 2 2n+1
= 2Θ(n) many
n
n
bitstrings in total, and the weight of the bitstrings will alternate between n and n + 1 in every step.
The existence of a middle levels Gray code for every value of n is asserted by the infamous middle
levels conjecture, which originated probably with Havel [Hav83] and Buck and Wiedemann [BW84],
but has also been attributed to Dejter, Erdős, Trotter [KT88] and various others. It also appears
as Exercise 56 in Knuth’s book [Knu11, Section 7.2.1.3] and in the popular book [DG12]. This
conjecture also became known as revolving door conjecture for the following reason: Imagine a set of
2n + 1 people, split into two subsets of size n and n + 1 that are separated by a revolving door. The
conjecture asks whether it is possible to move in each step one person from the larger group through
the revolving door to the other side, such that every way to split the people into two subsets of size
n and n + 1 is encountered exactly once. It may come as a surprise that establishing the existence
of a middle levels Gray code appears to be a difficult problem, given that by item (2) above one can
easily come up with a listing of all bitstrings of length 2n + 1 with weight exactly n such that any
two consecutive bitstrings differ in two bits.
The middle levels conjecture has attracted considerable attention over the last 30 years [Sav93,
FT95, SW95, Joh04, DSW88, KT88, DKS94, HKRR05, GŠ10]. Until recently, middle levels Gray
codes had only been found with brute-force computer searches for n ≤ 19 [SS99, SSS09, SA11]. For
n = 19 this Gray code already consists of N = 137.846.528.820 bitstrings. A complete proof of the
conjecture has only been announced very recently.
Theorem 1 ([Müt16]). A middle levels Gray code exists for every n ≥ 1.
1.1. Our results. Even though the proof of Theorem 1 given in [Müt16] is constructive, a straightforward implementation takes exponential (in n) time and space to compute for a given bitstring the
next one in the middle levels Gray code. Essentially, we need to compute and store the entire list of
N = 2Θ(n) bitstrings. The main contribution of this paper is a time- and space-efficient algorithm to
compute a middle levels Gray code for every n ≥ 1, which can be considered an algorithmic proof of
Theorem 1. Specifically, given any bitstring of length 2n + 1 with weight n or n + 1, our algorithm
computes the next ` bitstrings in the Gray code in time O(n`(1 + n` )), which is O(n) on average per
bitstring provided that ` = Ω(n). For most bitstrings the worst-case running time is O(n), for few
it is O(n2 ). The space requirement of our algorithm is O(n).
An implementation of this algorithm in C++ can be found on the authors’ websites [www], and we
invite the reader to experiment with this code. We used it to compute a middle levels Gray code for
n = 19 in less than a day on an ordinary desktop computer. For comparison, the above-mentioned
intelligent brute-force computation for n = 19 from [SA11] took about 164 days using comparable
hardware.
Remark 2. Clearly, the ultimate goal would be a generation algorithm with a worst-case running
time of O(1) per bitstring, but this would require substantial new ideas that would in particular
yield a much simpler proof of Theorem 1 than the one presented in [Müt16]. In fact, in a recent
paper [MN17] that appeared after the submission of this manuscript, we were able to extend the
ideas and techniques presented in this paper, and to improve our algorithm to achieve an optimal
O(1) worst-case running time per generated bitstring. This improved algorithm builds on top of the
algorithm presented in this paper.
Remark 3. It was shown in [Müt16] that there are in fact double-exponentially (in n) many different
middle levels Gray codes, which is easily seen to be best possible. This raises the question whether
our algorithm can be parametrized to compute any of these Gray codes. While this is possible
in principle, choosing between doubly-exponentially many different Gray codes would require a
parameter of exponential size, spoiling the above-mentioned runtime and space bounds. Moreover,
it would introduce a substantial amount of additional complexity in the description and correctness
3
proof of the algorithm. To avoid all this, our algorithm computes only one particular ‘canonical’
middle levels Gray code. We will briefly come back to this point at the end of Section 3 (Remark 10).
1.2. Outline of this paper. In Section 2 we present the pseudocode of our middle levels Gray
code algorithm. In Section 3 we prove the correctness of this algorithm, and in Section 4 we discuss
how to implement it to achieve the claimed runtime and space bounds.
2. The algorithm
It is convenient to reformulate our problem in graph-theoretic language: To this end we define the
middle levels graph, denoted by Q2n+1 (n, n + 1), as the graph whose vertices are all bitstrings of
length 2n + 1 that have weight n or n + 1, with an edge between any two bitstrings that differ in
exactly one bit. In other words, bitstrings that may appear consecutively in the middle levels Gray
code correspond to neighboring vertices in the middle levels graph. Clearly, computing a middle
levels Gray code is equivalent to computing a Hamilton cycle in the middle levels graph. Throughout
the rest of this paper we talk about middle levels Gray codes using this graph-theoretic terminology.
Our algorithm to compute a Hamilton cycle in the middle levels graph, i.e., to compute a middle
levels Gray code, is inspired by the constructive proof of Theorem 1 given in [Müt16]. Efficiency
is achieved by reformulating this inductive construction as a recursive procedure. Even though the
description of the algorithm in the present paper is completely self-contained and illustrated with
several figures that highlight the main ideas, the reader may find it useful to first read an informal
overview of the proof of Theorem 1, which can be found in [Müt16, Section 1.2].
Roughly speaking, our algorithm consists of a lower level function that computes sets of disjoint
paths in the middle levels graph and several higher level functions that combine these paths to form
a Hamilton cycle. In the following we explain these functions from bottom to top. Before doing so
we introduce some notation that will be used throughout this paper.
2.1. Basic definitions. Composition of mappings. We write the composition of mappings f, g as
f • g, where (f • g)(x) := f (g(x)).
Reversing/inverting and concatenating bitstrings. We define 0 := 1 and 1 := 0. For any bitstring x we let rev(x) denote the bitstring obtained from x by reversing the order of the bits
and inverting every bit. Moreover, for any bitstring x = (x1 , x2 , . . . , x2n ) we define π(x) :=
(x1 , x3 , x2 , x5 , x4 , . . . , x2n−1 , x2n−2 , x2n ), i.e., except the first and last bit, all adjacent pairs of bits
are swapped. E.g., for x = 11001110 we have rev(x) = 10001100 and π(x) = 10110110. Here and
throughout this paper, we omit commas and brackets when writing strings of 0’s and 1’s. Note that
the mappings rev, π and consequently also rev • π are self-inverse, and that rev • π = π • rev. For two
bitstrings x and y we denote by x ◦ y the concatenation of x and y. For any graph G whose vertices
are bitstrings and any bitstring y we denote by G ◦ y the graph obtained from G by attaching y to
every vertex of G.
Layers of the cube. For n ≥ 0 and k with 0 ≤ k ≤ n, we denote by Bn (k) the set of all bitstrings of
length n with weight k. For n ≥ 1 and k with 0 ≤ k ≤ n − 1, we denote by Qn (k, k + 1) the graph
with vertex set Bn (k) ∪ Bn (k + 1), with an edge between any two bitstrings that differ in exactly
one bit.
Oriented paths, first/second/last vertices. An oriented path P in a graph is a path with a particular
orientation, i.e., we distinguish its first and last vertex. For an oriented path P = (v1 , v2 , . . . , v` ) we
define its first, second and last vertex as F (P ) := v1 , S(P ) := v2 and L(P ) := v` , respectively.
Bitstrings and Dyck paths. We often identify a bitstring x with a lattice path in the integer lattice
Z2 as follows, see the left hand side of Figure 2: Starting at the coordinate (0, 0), we read the bits
4
of x from left to right and interpret every 1-bit as an upstep that changes the current coordinate by
(+1, +1) and every 0-bit as a downstep that changes the current coordinate by (+1, −1). Specifically,
we define the following sets of lattice paths:
• For any n ≥ 0 and k ≥ 0 we denote by Dn (k) the set of lattice paths with k upsteps and n − k
downsteps (n steps in total) that never move below the line y = 0.
• For n ≥ 1 and k ≥ 0 we define Dn>0 (k) ⊆ Dn (k) as the set of lattice paths that have no point of
the form (x, 0), 1 ≤ x ≤ n, and Dn=0 (k) ⊆ Dn (k) as the set of lattice paths that have at least one
point of the form (x, 0), 1 ≤ x ≤ n. For n = 0 we define D0=0 (0) := {()}, where () denotes the
empty lattice path, and D0>0 (0) := ∅. We clearly have Dn (k) = Dn=0 (k) ∪ Dn>0 (k).
• Furthermore, for n ≥ 1 and k ≥ 0 we let Dn− (k) denote the set of lattice paths with k upsteps and
n − k downsteps (n steps in total) that have exactly one point of the form (x, −1), 1 ≤ x ≤ n.
Note that all these lattice paths end at the coordinate (n, 2k − n).
For example, we have
D6=0 (3) = {111000, 110100, 110010, 101010, 101100} ,
D6>0 (4) = {111100, 111010, 111001, 110110, 110101} ,
D6− (3) = {110001, 101001, 100110, 011100, 011010} .
=0 (n)| = |D >0 (n+1)| = |D − (n)| and that the size of these sets is given
It is well known that in fact |D2n
2n
2n
by the n-th Catalan number (see [Che08] and [Sta99]). Depending on the values of n and k the above>0
(n) = ∅. Observe furthermore
mentioned sets of lattice paths might be empty, e.g., we have D2n
=0 (n) and D − (n)
that the mappings rev, π and therefore also rev • π map each of the sets D2n
2n
onto itself (see [Müt16, Lemma 11] for a formal proof). E.g., we have rev(π(111000)) = 111000,
rev(π(110100)) = 101010, rev(π(110010)) = 110010, rev(π(101010)) = 110100, rev(π(101100)) =
101100 and therefore rev(π(D6=0 (3))) = D6=0 (3).
2.2. Computing paths in Q2n (n, n + 1). The algorithm Paths() is at the core of our Hamilton
cycle algorithm. Its description is given in Algorithm 1. For simplicity let us ignore for the moment
the parameter flip ∈ {true, false} and assume that it is set to false. Then for every n and k with
1 ≤ n ≤ k ≤ 2n − 1 the algorithm Paths() constructs a set of disjoint oriented paths P2n (k, k + 1)
in the graph Q2n (k, k + 1) in the following way: Given a vertex x ∈ Q2n (k, k + 1) and the parameter
dir ∈ {prev, next}, the algorithm computes a neighbor of x on the path that contains the vertex
x. The parameter dir controls the search direction, so for dir = prev we obtain the neighbor of
x that is closer to the first vertex of the path, and for dir = next the neighbor that is closer to
the last vertex of the path. If x is a first vertex or a last vertex of one of the paths — we will see
momentarily what these are — then the result of a call to Paths() with dir = prev or dir = next,
respectively, is undefined and such calls will not be made from the higher level functions.
The algorithm Paths() works recursively: For the base case of the recursion n = 1 (lines P1–P2) it
computes neighbors for the set of paths
P2 (1, 2) := {(10, 11, 01)} ,
(1)
which consists only of a single path on three vertices. For example, the result of Paths(1, 1, 10, next, false)
is 11 and the result of Paths(1, 1, 11, prev, false) is 10. For the recursion step the algorithm considers the last two bits of the current vertex x (see line P5) and, depending on their values (see lines P9,
P11, P14 and P19), either flips one of these two bits (see lines P12, P15, P16, P17, P20, P21), or
recurses to flip one of the first 2n − 2 bits instead, leaving the last two bits unchanged (see lines P7,
P10, P13, P18, P22). As already mentioned, the algorithm Paths() is essentially a recursive formulation of the inductive construction of paths described in [Müt16] (and recapitulated in Section 3.1),
and the different cases in the algorithm reflect the different cases in this construction. The recursion
5
step in line P22 is where the mappings rev and π introduced in Section 2.1 come into play (recall that
rev−1 = rev and π −1 = π). To give another example, the result of Paths(3, 3, 111000, next, false)
is 111001, and the result of Paths(3, 3, 111001, prev, false) is 111000, so the set of paths P6 (3, 4) in
the graph Q6 (3, 4) defined by the algorithm Paths() contains a path with the edge (111000, 111001).
Algorithm 1: Paths(n, k, x, dir, flip)
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
P11
P12
P13
P14
P15
P16
P17
P18
P19
P20
P21
P22
Input: Integers n and k with 1 ≤ n ≤ k ≤ 2n − 1, a vertex x ∈ Q2n (k, k + 1), parameters
dir ∈ {prev, next} and flip ∈ {true, false}
Output: A neighbor of x in Q2n (k, k + 1)
if n = 1 then
/* base cases */
depending on x and dir, return the previous/next neighbor of x on P2 (1, 2) defined in (1)
else if n = k = 2 and flip = true then
e4 (2, 3) defined in (3)
depending on x and dir, return the previous/next neighbor of x on P
Split x = (x1 , x2 , . . . , x2n ) into x− := (x1 , x2 , . . . , x2n−2 ) and x+ := (x2n−1 , x2n )
if k ≥ n + 1 then
return Paths(n − 1, k − x2n−1 − x2n , x− , dir, flip) ◦ x+
else
if x+ = 10 then
return Paths(n − 1, n − 1, x− , dir, flip) ◦ x+
/* k = n */
else if x+ = 00 then
>0
(n) then return x− ◦ 01
if x− ∈ D2n−2
else return Paths(n − 1, n, x− , dir, flip) ◦ x+
else if x+ = 01 then
=0 (n − 1) then return x− ◦ 11
if x− ∈ D2n−2
−
(n − 1) and dir = next then return x− ◦ 11
else if x− ∈ D2n−2
>0
−
else if x ∈ D2n−2 (n) and dir = prev then return x− ◦ 00
else return Paths(n − 1, n − 1, x− , dir, false) ◦ x+
else if x+ = 11 then
=0 (n − 1) and dir = next return x− ◦ 01
if x− ∈ D2n−2
−
(n − 1) and dir = prev then return x− ◦ 01
else if x− ∈ D2n−2
else return rev π Paths(n − 1, n − 1, rev−1 (π −1 (x− )), dir, flip) ◦ x+ where
dir := prev if dir = next and dir := next otherwise
We will later prove that the set of paths P2n (n, n + 1) computed by the algorithm Paths() (with
parameters k = n and flip = false) has the following properties:
(i) All paths in P2n (n, n+1) are disjoint, and together they visit all vertices of the graph Q2n (n, n+1).
=0 (n) and D − (n), respectively.
(ii) The sets of first and last vertices of the paths in P2n (n, n+1) are D2n
2n
We only state these properties here; the proof will be provided in Section 3. From (ii) we conclude
that the number of paths in P2n (n, n + 1) equals the n-th Catalan number. E.g., the entire set of
paths P4 (2, 3) computed by the algorithm Paths() is
P4 (2, 3) = {P, P 0 } ,
P := (1100, 1101, 0101, 0111, 0011, 1011, 1001) ,
0
P := (1010, 1110, 0110) .
(2a)
(2b)
(2c)
6
P1
P2
P3
P4
P5
111000 110100 110010 101010 101100
111001 110101 110110 111010 111100
011001 010101 010110 011010 011100
011011 011101 011110
010011 001101 001110
010111 101101 101110
000111 100101 100110
001111 100111
001011 100011
101011 110011
101001 110001
Table 1. The five paths in the set P6 (3, 4) computed by the algorithm Paths()
with parameter flip = false. Each path is listed in a separate column, the first
vertex from the set D6=0 (3) is at the top, and the last vertex from the set D6− (3) is
at the bottom. These paths are disjoint and together visit all vertices of the graph
Q6 (3, 4).
R1
R2
R3
R4
111000 110100 110010 101010
111001 110101 111010 101110
011001 010101 011010 001110
011011 011101
011110
001011 001101
010110
001111 101101
110110
000111 100101
100110
010111 100111
010011 100011
110011 101011
110001 101001
e6 (3, 4) computed by the algorithm Paths()
Table 2. The four paths in the set P
with parameter flip = true. Comparing those paths to Table 1, note that (P1 , P2 )
and (P3 , P4 ) are flippable pairs of paths and the corresponding flipped pairs of paths
e6 (3, 4) do not visit
are (R1 , R2 ) and (R3 , R4 ), respectively. Note that the paths in P
all vertices of Q6 (3, 4) (unlike the paths in P6 (3, 4)), as the vertices of P5 from Table 1
are not covered.
To give another example, the set of paths P6 (3, 4) computed by the algorithm Paths() consists of
the five paths in Table 1. We remark that the length of each of the paths in P2n (n, n + 1) depends
only on the length of the first ‘hill’ of the lattice path corresponding to the first vertex. In particular,
not all paths have the same length. The precise length formula will be given in Section 4, when
we analyze the running time of our algorithm. For the correctness of the algorithm the lengths of
the paths are irrelevant. We will see in the next section how to combine the paths P2n (n, n + 1)
in the graph Q2n (n, n + 1) computed by our algorithm Paths() (called with flip = false) to
compute a Hamilton cycle in the middle levels graph Q2n+1 (n, n + 1). One crucial ingredient we
e2n (n, n + 1), computed by calling the algorithm Paths() with
need for this is another set of paths P
flip = true. We shall see that the core ‘intelligence’ of our Hamilton cycle algorithm consists of
e2n (n, n + 1) to a Hamilton
cleverly combining some paths from P2n (n, n + 1) and some paths from P
7
cycle in the middle levels graph. We will see that from the point of view of all top-level routines that
e2n (k, k + 1) for k = n are used, but
call the algorithm Paths() only the paths P2n (k, k + 1) and P
the recursion clearly needs to compute these sets also for all other values k = n + 1, n + 2, . . . , 2n − 1.
e2n (k, k + 1)
Specifically, calling the algorithm Paths() with flip = true yields a set of paths P
that differs from P2n (k, k + 1) as follows: First note that for certain inputs of the algorithm the
value of the parameter flip is irrelevant, and the computed paths are the same regardless of its
value. This is true whenever the recursion ends in the base case in lines P1-P2. One such example
is the path P5 in Table 1. Computing previous or next neighbors for any vertex on this path yields
e2n (k, k + 1), and we
the same result regardless of the value of flip. Such paths are not part of P
ignore them in the following. The remaining paths from P2n (k, k + 1) can be grouped into pairs
e2n (k, k + 1) contains two disjoint paths R and R0 that visit
and for every such pair (P, P 0 ) the set P
0
the same set of vertices as P and P , but that connect the end vertices of the paths the other way:
Formally, denoting by V (G) the vertex set of any graph G, these last two conditions can be captured
as V (P ) ∪ V (P 0 ) = V (R) ∪ V (R0 ), F (P ) = F (R), F (P 0 ) = F (R0 ), L(P ) = L(R0 ) and L(P 0 ) = L(R).
We refer to a pair (P, P 0 ) satisfying these conditions as a flippable pair of paths, and to the corresponding paths (R, R0 ) as a flipped pair. This notion of flippable/flipped pairs is extremely valuable
for designing our Hamilton cycle algorithm, as it allows the algorithm to decide independently for
each flippable pair of paths, whether to follow one of the original paths or the flipped paths in the
graph Q2n (n, n + 1).
e2n (k, k + 1) is the set of paths P
e4 (2, 3) in the graph
The base case for the recursive computation of P
Q4 (2, 3) defined by
e4 (2, 3) := {R, R0 } ,
P
R := (1100, 1110, 0110) ,
0
R := (1010, 1011, 0011, 0111, 0101, 1101, 1001)
(3a)
(3b)
(3c)
(see lines P3–P4). Observe that (P, P 0 ) defined in (2) and (R, R0 ) satisfy precisely the conditions for
flippable/flipped pairs of paths, i.e., (P, P 0 ) is a flippable pair and (R, R0 ) is a corresponding flipped
e6 (3, 4) computed by our
pair of paths. To continue also the previous example, the set of paths P
algorithm is shown in Table 2.
2.3. Computing a Hamilton cycle in the middle levels graph. The algorithm HamCycleNext()
e2n (n, n + 1) in the graph Q2n (n, n + 1) computed by the algorithm
uses the paths P2n (n, n + 1) and P
Paths() as described in the previous section to compute for a given vertex x the next vertex on a
Hamilton cycle in the middle levels graph Q2n+1 (n, n + 1). To simplify the exposition of the algorithm, let us ignore for the moment the parameter flip ∈ {true, false} and assume that it is set to
false, and let us also ignore line N3. With these simplifications only paths from P2n (n, n + 1) are
e2n (n, n + 1) are ignored. Then instead of computing a Hamilton cycle
considered, and those from P
the algorithm HamCycleNext() computes several smaller cycles that together visit all vertices of
the middle levels graph. We will correct this later by setting flip accordingly.
The algorithm HamCycleNext() is based on the following decomposition of the middle levels graph
Q2n+1 (n, n + 1), see Figure 1: By partitioning the vertices of the graph Q2n+1 (n, n + 1) according
to the value of the last bit, we observe that it consists of a copy of the graph Q2n (n, n + 1) ◦ 0 and a
copy of the graph Q2n (n − 1, n) ◦ 1 plus the set of edges M2n+1 = {(x ◦ 0, x ◦ 1) | x ∈ B2n (n)} along
which the last bit is flipped. Observe furthermore that the graphs Q2n (n, n+1) and Q2n (n−1, n) are
isomorphic, and that the mapping rev is an isomorphism between these graphs. It is easy to check
that this isomorphism preserves the sets of end vertices of the paths P2n (n, n+1): Using property (ii)
=0 (n)) = D =0 (n) and rev(D − (n)) = D − (n) (in
mentioned in the previous section, we have rev(D2n
2n
2n
2n
8
Q2n+1 (n, n + 1)
C2n+1
P2n (n, n + 1)
Q2n (n, n + 1) ◦ (0)
Q2n (n − 1, n) ◦ (1)
B2n (n) ◦ (1)
B2n (n + 1) ◦ (0)
rev(P2n (n, n + 1))
B2n (n) ◦ (0)
M2n+1
B2n (n − 1) ◦ (1)
Figure 1. The top part of the figure shows the decomposition of the middle levels
graph and the definition (4). The 2-factor consists of three disjoint cycles that together visit all vertices of the graph. The first and last vertices of the paths are drawn
in black and white, respectively. The bottom part of the figure shows a simplified
drawing that helps analyzing the cycle structure of the 2-factor (it has two short
cycles and one long cycle).
Figure 1, these sets are the black and white vertices). By property (i) of the paths P2n (n, n + 1),
we conclude that
0
C2n+1 := P2n (n, n + 1) ◦ 0 ∪ rev(P2n (n, n + 1)) ◦ 1 ∪ M2n+1
(4)
−
(n)}
D2n
0
=0 (n) ∪
⊆ M2n+1 is a so-called 2-factor of the middle
with M2n+1
:= {(x ◦ 0, x ◦ 1) | x ∈ D2n
levels graph, i.e., a set of disjoint cycles that together visit all vertices of the graph. Note that along
each of the cycles in the 2-factor, the paths from P2n (n, n + 1) ◦ 0 are traversed in forward direction,
and the paths from rev(P2n (n, n + 1)) ◦ 1 in backward direction. The algorithm HamCycleNext()
(called with flip = false) computes exactly this 2-factor C2n+1 : Given a vertex x of the middle
levels graph, it computes the next vertex on one of the cycles from the 2-factor by checking the value
of the last bit (line N2), and by returning either the next vertex on the corresponding path from
P2n (n, n+1)◦0 (line N5) or the previous vertex on the corresponding path from rev(P2n (n, n+1))◦1
0
(line N8; recall that rev−1 = rev). The cases that the next cycle edge is an edge from M2n+1
receive
special treatment: In these cases the last bit is flipped (lines N4 and N7).
Algorithm 2: HamCycleNext(n, x, flip)
N1
N2
N3
N4
N5
N6
N7
N8
Input: An integer n ≥ 1, a vertex x ∈ Q2n+1 (n, n + 1), state variable flip ∈ {true, false}
Output: Starting from x the next vertex on a Hamilton cycle in Q2n+1 (n, n + 1), updated
state variable flip ∈ {true, false}
Split x = (x1 , x2 , . . . , x2n+1 ) into x− := (x1 , x2 , . . . , x2n ) and the last bit x2n+1
if x2n+1 = 0 then
=0 (n) then return Paths(n, n, x− , next, a) ◦ x
if x− ∈ D2n
2n+1 , a where
a := IsFlipVertex(n, x− )
−
else if x− ∈ D2n
(n) then return (x− ◦ 1, false)
else return Paths(n, n, x− , next, flip) ◦ x2n+1 , flip
else
/* x2n+1 = 1 */
=0 (n) then return (x− ◦ 0, false)
if x− ∈ D2n
else return rev Paths(n, n, rev−1 (x− ), prev, false) ◦ x2n+1 , false
9
As mentioned before, calling the algorithm HamCycleNext() with flip = false yields the 2factor in the middle levels graph defined in (4) that consists of more than one cycle. However, we
will show that by replacing in (4) some of the flippable pairs of paths from P2n (n, n + 1) ◦ 0 by
e2n (n, n + 1) ◦ 0, which are computed by calling the
the corresponding flipped paths from the set P
algorithm Paths() with flip = true, we obtain a 2-factor that has only one cycle, i.e., a Hamilton
cycle. The key insight is that replacing a pair of flippable paths that are contained in two different
cycles of the 2-factor C2n+1 by the corresponding flipped paths joins these two cycles to one cycle.
This is an immediate consequence of the definition of flippable/flipped pairs of paths. This flipping
is controlled by the parameter flip of the algorithm HamCycleNext(): It decides whether to
e2n (n, n + 1) ◦ 0
compute the next vertex on a path from P2n (n, n + 1) ◦ 0 (if flip = false) or from P
(if flip = true). Note that these modifications do not affect the paths rev(P2n (n, n + 1)) ◦ 1 in the
union (4): For the corresponding instructions in lines N6–N8, the value of flip is irrelevant.
e2n (n, n + 1) ◦ 0 is computed at
The decision whether to follow a path from P2n (n, n + 1) ◦ 0 or from P
=0
the vertices x ◦ 0, x ∈ D2n (n), by calling the function IsFlipVertex(n, x) (line N3). This decision
is returned to the caller and maintained until the last vertex of the corresponding path in the graph
Q2n (n, n + 1) ◦ 0 is reached. Recall that by the definition of flippable pairs of paths, this decision
can be made independently for each flippable pair. Of course it has to be consistent for both paths
in a flippable pair: either both are flipped or none of them are.
2.4. The top-level algorithm. The algorithm HamCycle(n, x, `) takes as input a vertex x of the
middle levels graph Q2n+1 (n, n + 1) and computes
the next ` vertices that follow x on a Hamilton
,
every
vertex appears at most once in the output,
cycle in this graph. I.e., for ` ≤ N = 2 2n+1
n
and for ` = N every vertex appears exactly once and the vertex x comes last. In terms of Gray
codes, the algorithm takes a bitstring x of length 2n + 1 that has weight n or n + 1 and outputs the
` subsequent bitstrings in a middle levels Gray code. A single call HamCycle(n, x, `) yields the
same output as ` subsequent calls xi+1 := HamCycle(n, xi , 1), i = 0, 1, . . . , ` − 1, with x0 := x.
However, with respect to running times, the former is faster than the latter.
Algorithm 3: HamCycle(n, x, `)
H1
H2
H3
H4
H5
H6
H7
H8
H9
H10
H11
Input: An integer n ≥ 1, a vertex x ∈ Q2n+1 (n, n + 1), an integer ` ≥ 1
Output: Starting from x, the next ` vertices on a Hamilton cycle in Q2n+1 (n, n + 1)
Split x = (x1 , x2 , . . . , x2n+1 ) into x− := (x1 , x2 , . . . , x2n ) and the last bit x2n+1
flip := false
if x2n+1 = 0 then
/* initialize state variable flip */
y := x−
=0 (n) do
while y ∈
/ D2n
/* move backwards to first path vertex */
y := Paths(n, n, y, prev, false)
flip := IsFlipVertex(n, y)
y := x
for i := 1 to ` do
(y, flip) := HamCycleNext(n, y, flip)
output y
/* Hamilton cycle computation */
The algorithm HamCycle() consists of an initialization phase (lines H2–H7) in which the initial
value of the state variable flip is computed. This is achieved by following the corresponding
path from P2n (n, n + 1) ◦ 0 backwards to its first vertex (lines H5–H6) and by calling the function
IsFlipVertex() (line H7). The actual Hamilton cycle computation (lines H8–H11) repeatedly
10
lattice path
=0 (5)
from D10
bitstring
x = (1, 1, 0, 1, 1, 0, 1, 0, 0, 0)
0
ordered rooted
tree from T5∗
10
Figure 2. Bijections between bitstrings and lattice paths (left hand side), and be=0 (n) and ordered rooted trees from T ∗ (right hand side).
tween lattice paths from D2n
n
computes the subsequent cycle vertex and updates the state variable flip by calling the function
HamCycleNext() discussed in the previous section (line H10).
2.5. Flip vertex computation. To complete the description of our Hamilton cycle algorithm, it
remains to specify the auxiliary function IsFlipVertex(n, x). As mentioned before, this function
=0 (n) whether our Hamilton cycle algorithm should follow the path
decides for each vertex x ∈ D2n
from P2n (n, n + 1) that starts with this vertex (return value false) or the corresponding flipped
e2n (n, n + 1) (return value true) in the graph Q2n (n, n + 1) ◦ 0 (recall (4)). This function
path from P
therefore nicely encapsulates the core ‘intelligence’ of our algorithm so that it produces a 2-factor
consisting only of a single cycle and not of several smaller cycles. Admittedly, the definition of this
function is rather technical, and we postpone the formal correctness proof until Section 3. We begin
by introducing several new concepts related to trees.
Ordered rooted trees. An ordered rooted tree is a rooted tree where the children of each vertex have a
specified left-to-right ordering. We think of an ordered rooted tree as a tree embedded in the plane
with the root on top, with downward edges leading from any vertex to its children, and the children
appear in the specified left-to-right ordering. This is illustrated on the right hand side of Figure 2,
where the root vertex is drawn boldly. We denote by Tn∗ the set of all ordered rooted trees with n
edges. It is well known that the number of trees in Tn∗ is given by the n-th Catalan number (see
[Sta99]).
Bijection between lattice paths and ordered rooted trees. We identify each lattice path (=bitstring)
=0 (n) with an ordered rooted tree from the set T ∗ as follows
x = (x1 , x2 , . . . , x2n ) from the set D2n
n
(see the right hand side of Figure 2 and [Sta99]): Starting with a tree that has only a root vertex,
we read the bits of x from left to right, and for every 1-bit we add a new rightmost child to the
current vertex and move to this child, for every 0-bit we add no edge but simply move back to
the parent of the current vertex (we clearly end up back at the root). This construction defines a
=0 (n) and the ordered rooted trees T ∗ , and in the following we
bijection between the lattice paths D2n
n
will repeatedly switch between these two representations.
Thin/thick leaves, clockwise/counterclockwise-next leaves. We call a leaf u of a tree thin or thick, if
the vertex adjacent to u has degree exactly 2 or at least 3, respectively. Clearly, for any tree with at
least two edges, every leaf is either thin or thick. Given two leaves u and v of an ordered rooted tree
T , we say that v is the clockwise-next leaf from u, or equivalently, u is the counterclockwise-next leaf
from v, if all edges of T not on the path p from u to v lie to the right of the path p when traversing
p from u to v in the given embedding of T , respectively. Equivalently, in a (cyclic) depth-first search
starting at the root, u is the first leaf encountered after the leaf v. This definition is illustrated in
Figure 3.
∗ , T ∗ ⊆ T ∗ and transforTree transformations τ1 and τ2 . We define sets of ordered rooted trees Tn,1
n
n,2
∗
mations τ1 , τ2 operating on these trees as follows: The set Tn,1 contains all ordered rooted trees T
with at least three edges for which the leftmost child u0 of the root v has exactly one child u and u
11
T
τ1 (T )
v
u0
v
deg = 2 u
u0
T
u0
u
τ2 (T )
deg ≥ 3
u0
u
deg = 2
v
v is clockwise-next leaf from u
u is counterclockwise-next leaf from v
v
u
Figure 3. Definition of the tree transformations τ1 (left) and τ2 (right). The edges
in which the trees differ are drawn bold, and the grey areas represent arbitrary
subtrees.
is a leaf. In other words, the leftmost subtree of the root is a path on two edges, see the left hand
side of Figure 3. For such a tree T , we define τ1 (T ) as the tree obtained by replacing the edge (u, u0 )
by (u, v) so that u becomes the leftmost child of the root. In the bitstring representation, we have
∗ , and τ (T ) = 1010 ◦ T 0 . The set T ∗ contains all ordered rooted
T = 1100 ◦ T 0 for some T 0 ∈ Tn−2
1
n,2
trees T for which the rightmost child u of the leftmost child u0 of the root is a thick leaf (so u0 has
more children to the left of u), and the clockwise-next leaf v of u is thin. Clearly, such a tree must
have at least four edges, see the right hand side of Figure 3. For such a tree T , we define τ2 (T ) as
the tree obtained by replacing the edge (u, u0 ) by (u, v). In the bitstring representation, we have
T = 1 ◦ T1 ◦ 1 ◦ T2 ◦ 1 ◦ · · · ◦ Tk+1 ◦ 1100 ◦ 0k ◦ 10 ◦ 0 ◦ T0
for some k ≥ 0 and Ti ∈ Tn∗i for all i = 0, 1, . . . , k + 1, and
τ2 (T ) = 1 ◦ T1 ◦ 1 ◦ T2 ◦ 1 ◦ · · · ◦ Tk+1 ◦ 111000 ◦ 0k ◦ 0 ◦ T0 .
∗ , T ∗ , τ (T ∗ ) and τ (T ∗ ) are all disjoint.
Note that the four sets Tn,1
2 n,2
n,2 1 n,1
=0 (n) we inductively define a lattice path h(x) ∈ D =0 (n)
The mapping h. For any lattice path x ∈ D2n
2n
as follows: If x = () then we define
h(x) := () .
(5a)
=0
Otherwise there is a unique partition x = 1 ◦ x` ◦ 0 ◦ xr such that x` ∈ D2k (k) for some k ≥ 0 and
we define
h(x) := 1 ◦ π(h(x` )) ◦ 0 ◦ h(xr ) ,
(5b)
where π is the permutation defined in Section 2.1. It is easy to check that h defines a bijection
=0 (n), and that the definition of the inverse mapping h−1 can be derived from (5) simply by
on D2n
replacing π(h(x` )) in (5b) by h−1 (π(x` )) and h(xr ) by h−1 (xr ). Clearly, via the bijection between
lattice paths and ordered rooted trees discussed before, the mapping h can be interpreted as a
bijection on the set Tn∗ . As an example, Figure 5 shows the effect of applying h to all 14 trees in the
set T4∗ . The preimages under h appear at the top of the figure, the corresponding images below.
With these definitions at hand, consider now the definition of the function IsFlipVertex(n, x).
It first computes the ordered rooted tree T obtained by applying h−1 to the bitstring/lattice path
x (line T1). On lines T2–T3 the function checks whether T is a preimage or image under τ1
and calls an auxiliary function IsFlipTree1 () (defined below) with T or τ1−1 (T ) as an argument,
12
respectively. A similar check is performed in lines T4–T5 with respect to τ2 , using the auxiliary
function IsFlipTree2 () (defined below).
Algorithm 4: IsFlipVertex(n, x)
T1
T2
T3
T4
T5
T6
=0 (n)
Input: An integer n ≥ 1, a vertex x ∈ D2n
Output: {true, false}
T := h−1 (x)
/* compute corresponding rooted tree */
∗
if T ∈ Tn,1 then return IsFlipTree1 (T )
∗ ) then return IsFlipTree (τ −1 (T ))
else if T ∈ τ1 (Tn,1
1 1
∗
else if T ∈ Tn,2 then return IsFlipTree2 (T )
∗ ) then return IsFlipTree (τ −1 (T ))
else if T ∈ τ2 (Tn,2
2 2
else return false
We proceed to define the auxiliary functions IsFlipTree1 () and IsFlipTree2 () that have return
values {true, false}. For brevity we define them in textual form, not as pseudocode. For the
definitions we need to introduce the notion of tree rotation.
Rotation rot() of ordered rooted trees. For an ordered rooted tree T ∈ Tn∗ we let rot(T ) denote the
tree that is obtained from T by shifting the root to the leftmost child of the root. In terms of lattice
=0 (k) for some k ≥ 0,
paths, considering the unique decomposition T = 1 ◦ T` ◦ 0 ◦ Tr with T` ∈ D2k
we have rot(T ) = T` ◦ 1 ◦ Tr ◦ 0. Note that we obtain the same tree again after 2n/s such rotations,
where the value of s depends on the symmetries of the tree. As an example, the top part of Figure 5
shows the effect of applying the rotation operation to all 14 trees in the set T4∗ .
∗ , compute all rotated versions of it that are
The function IsFlipTree1 (T ): Given a tree T ∈ Tn,1
∗
also in the set Tn,1 . Return true if T is lexicographically smallest among them, and false otherwise.
∗ , if T = 1n−1 ◦ 0n−2 ◦ 100 or if T has more
The function IsFlipTree2 (T ): Given a tree T ∈ Tn,2
than one thin leaf, return false. Otherwise, let v be the unique thin leaf of T , v 0 the parent of
v, v 00 the parent of v 0 , and w the clockwise-next leaf of v, and let T 0 be the tree obtained from T
by replacing the edge (v, v 0 ) by (v, v 00 ) so that v becomes a child of v 00 to the left of v 0 (so T 0 has
only thick leaves), and by rotating it such that the leaf w becomes the root. Let d be the distance
between the root w and the leftmost leaf v of T 0 . Compute all other rotated versions of T 0 for which
the root is a leaf and the parent of the leftmost leaf x has another leaf as its child to the right of x
(initially, v and v 0 are these two leaves). For each of them compute the distance d0 between the root
and the leftmost leaf. Return true if d0 ≤ d for all of them and if T 0 is lexicographically smallest
among all rotated trees with d0 = d, and false otherwise.
When referring to lexicographic comparisons between ordered rooted trees in the previous definitions
we mean lexicographic comparisons between the corresponding bitstrings.
3. Correctness of the algorithm
In this section we prove the correctness of the algorithm presented in the previous section. Before
en (k, k +1) described
doing this, we recapitulate the inductive construction of paths Pn (k, k +1) and P
in [Müt16] in Sections 3.1 and 3.2 below. These sections are essentially duplicate material from
[Müt16], and readers familiar with this construction may want to skip them. The main goal of
Section 3.3 is to prove that those paths are the same as the ones produced by our recursive algorithm
en (k, k + 1),
Paths(), so this will justify denoting both with the same symbol Pn (k, k + 1) or P
respectively. Once this is established, we apply several lemmas from [Müt16] to prove the correctness
of the algorithm HamCycle() in Section 3.4. When looking up those lemmas the reader should be
13
aware of some minor differences in notation between [Müt16] and our paper: We use bitstrings to also
denote the corresponding lattice paths and ordered rooted trees, implicitly using the corresponding
bijections. In the paper [Müt16], those bijections are made explicit and denoted ϕ and ψ, and
upsteps and downsteps of a lattice path are denoted by % and & rather than 1 and 0.
3.1. Recap: Inductive construction of paths from [Müt16]. The construction is parametrized
by some sequence (α2i )i≥1 , α2i ∈ {0, 1}i−1 . Given this sequence, we inductively construct a set
P2n (k, k + 1) of disjoint oriented paths in Q2n (k, k + 1) that start and end in B2n (k) for all n ≥ 1
and all k = n, n + 1, . . . , 2n − 1 such that the following conditions hold:
(a) The paths in P2n (n, n + 1) visit all vertices in the sets B2n (n + 1) and B2n (n).
(b) For k = n+1, . . . , 2n−1, the paths in P2n (k, k +1) visit all vertices in the set B2n (k +1), and the
only vertices not visited in the set B2n (k) are exactly the elements in the set S(P2n (k − 1, k)).
Induction basis n = 1 (Q2 ): The induction basis is the set of paths P2 (1, 2) := {(10, 11, 01)},
consisting only of a single oriented path on three vertices. It is easily checked that this set of paths
satisfies properties (a) and (b) (property (b) is satisfied trivially).
Induction step n → n + 1 (Q2n → Q2n+2 ), n ≥ 1: The inductive construction consists of two
intermediate steps.
First intermediate step: Construction of a 2-factor in the middle levels graph Q2n+1 (n, n + 1). Using
only the paths in the set P2n (n, n + 1) and the parameter α2n = (α2n (1), . . . , α2n (n − 1)) ∈ {0, 1}n−1
we first construct a 2-factor in the middle levels graph Q2n+1 (n, n + 1).
The parameter α2i is used to generalize the permutation π introduced before: Specifically, we let
πα2n denote the permutation on the set {0, 1}2n that swaps any two adjacent bits at positions 2i and
2i + 1 for all i = 1, 2, . . . , n − 1, if and only if α2n (i) = 1, and that leaves the bits at position 1 and
2n unchanged. Note that if α2n = (1, 1, . . . , 1), then πα2n equals the mapping π from before. On
the other hand, if α2n = (0, 0, . . . , 0), then no bits are swapped and πα2n = id is simply the identity
mapping. For any bitstring x ∈ {0, 1}2n we then define
fα2n (x) := rev(πα2n (x)) .
(6)
Note that we have fα2n = rev • π for α2n = (1, 1, . . . , 1) and fα2n = rev for α2n = (0, 0, . . . , 0).
It was proved in [Müt16, Lemma 3] that for any α2n ∈ {0, 1}n−1 the mapping fα2n maps each of the
sets F (P2n (n, n + 1)) and L(P2n (n, n + 1)) onto itself. As explained in Section 2.3, the middle layer
graph Q2n+1 (n, n + 1) can be decomposed into the graphs Q2n (n, n + 1) ◦ 0 and Q2n (n − 1, n) ◦ 1
plus the edges from the matching M2n+1 = {(x ◦ 0, x ◦ 1) | x ∈ B2n (n)} along which the last
FL
bit is flipped. Denoting by M2n+1
the edges from M2n+1 that have one end vertex in the set
F (P2n (n, n + 1)) ∪ L(P2n (n, n + 1)) ◦ 0 ⊆ B2n (n) ◦ 0 (and the other in the set F (P2n (n, n + 1)) ∪
L(P2n (n, n + 1)) ◦ 1 ⊆ B2n (n) ◦ 1), the union
FL
C2n+1 := P2n (n, n + 1) ◦ 0 ∪ fα2n (P2n (n, n + 1)) ◦ 1 ∪ M2n+1
(7)
therefore yields a 2-factor in the middle levels graph, with the property that on every cycle of C2n+1 ,
every edge of the form (F (P ), S(P )) ◦ 0 for some P ∈ P2n (n, n + 1) is oriented the same way.
Note that (4) is the special case of (7) with α2n = (0, 0, . . . , 0). In the correctness proof for our
algorithm we will later also consider the case α2n = (1, 1, . . . , 1). Even though we are eventually only
interested in the 2-factor C2n+1 defined in (7), we need to specify how to proceed with the inductive
construction of the sets of paths P2n+2 (k, k + 1).
Second intermediate step: Splitting up the 2-factor into paths. We proceed by describing how the
sets of paths P2n+2 (k, k + 1) for all k = n + 1, n + 2, . . . , 2n + 1 satisfying properties (a) and (b) are
defined, using the previously constructed sets P2n (k, k + 1) and the 2-factor C2n+1 defined in the
first intermediate step.
14
Consider the decomposition of Q2n+2 into Q2n ◦00, Q2n ◦10, Q2n ◦01 and Q2n ◦11 plus the two perfect
0
matchings M2n+2 and M2n+2
along which the last and second to last bit are flipped, respectively.
For all k = n + 2, . . . , 2n + 1 we define
P2n+2 (k, k + 1) := P2n (k, k + 1) ◦ 00 ∪ P2n (k − 1, k) ◦ 10
(8)
∪ P2n (k − 1, k) ◦ 01 ∪ P2n (k − 2, k − 1) ◦ 11 ,
where we use the convention P2n (2n, 2n + 1) := ∅ and P2n (2n + 1, 2n + 2) := ∅ to unify treatment
of the sets of paths P2n+2 (2n, 2n + 1) and P2n+2 (2n + 1, 2n + 2) between the uppermost levels of
0
Q2n+2 . Note that so far none of the edges from the matchings M2n+2 or M2n+2
is used.
The definition of the set P2n+2 (n+1, n+2) is slightly more involved. Note that the graph Q2n+2 (n+
1, n + 2) can be decomposed into Q2n+1 (n + 1, n + 2) ◦ 0 and Q2n+1 (n, n + 1) ◦ 1 plus the edges from
M2n+2 that connect the vertices in the set B2n+1 (n + 1) ◦ 0 to the vertices in the set B2n+1 (n + 1) ◦ 1.
The first graph can be further decomposed into Q2n (n+1, n+2)◦00 and Q2n (n, n+1)◦10 plus some
matching edges that are not relevant here. The second graph is the middle levels graph of Q2n+1 ◦ 1.
−
Let C2n+1
denote the graph obtained from the 2-factor C2n+1 defined in (7) by removing every edge
of the form (F (P ), S(P )) ◦ 0 for some P ∈ P2n (n, n + 1). As on every cycle of C2n+1 every such
−
is a set of paths (visiting all vertices of the middle levels graph
edge is oriented the same way, C2n+1
Q2n+1 (n, n + 1)), with the property that each of those paths starts at a vertex of the form S(P ) ◦ 0
S
and ends at a vertex of the form F (P 0 ) ◦ 0 for two paths P, P 0 ∈ P2n (n, n + 1). Denoting by M2n+2
the edges from M2n+2 that have one end vertex in the set S(P2n (n, n + 1)) ◦ 00 ⊆ B2n (n + 1) ◦ 00
(and the other in the set S(P2n (n, n + 1)) ◦ 01 ⊆ B2n (n + 1) ◦ 01), it follows that
−
∗
S
◦1
P2n+2
:= M2n+2
∪ C2n+1
(9)
is a set of oriented paths that start and end in B2n+2 (n + 1), where we choose the orientation of
S
each path such that the edge from the set M2n+2
is the first edge. Note that we have
∗
F (P2n+2
) = S(P2n (n, n + 1)) ◦ 00 .
(10)
∗
P2n+2 (n + 1, n + 2) := P2n (n + 1, n + 2) ◦ 00 ∪ P2n (n, n + 1) ◦ 10 ∪ P2n+2
,
(11)
We then define
where in the case n = 1 we use the convention P2 (2, 3) := ∅.
It was argued in [Müt16, Section 2.2] that the sets of paths P2n+2 (k, k+1), k = n+1, n+2, . . . , 2n+1,
defined in (8) and (11) have properties (a) and (b). We omit those arguments here.
3.2. Recap: Inductive construction of flippable pairs from [Müt16]. Let P2n (k, k + 1) be the
sets of paths defined in Section 3.1 for an arbitrary parameter sequence (α2i )i≥1 , α2i ∈ {0, 1}i−1 .
In the following we show how to inductively construct a set of flippable pairs X2n (k, k + 1) for the
set P2n (k, k + 1) for all n ≥ 2 and all k = n, n + 1, . . . , 2n − 1, and the corresponding flipped paths
e2n (k, k +1). So the set X2n (k, k +1) contains pairs of paths (P, P 0 ), P, P 0 ∈ P2n (k, k +1), that form
P
e2n (k, k + 1) is the
a flippable pair, where every path appears in at most one such pair, and the set P
0
0
0
union of all paths R, R that form a corresponding flipped pair (R, R ) for (P, P ). This construction
arises very naturally from the inductive construction of the sets P2n (k, k + 1) given in the previous
section.
Induction basis n = 2 (Q4 ): It is easy to check that the set of paths P4 (2, 3) arising from the
inductive construction described in the previous section after one step satisfies P4 (2, 3) = {P, P 0 }
with P and P 0 as defined in (2) (as α2 = () this step does not involve any parameter choices
yet). The set X4 (2, 3) := {(P, P 0 )} is a set of flippable pairs for P4 (2, 3), which can be seen by
considering the flipped paths (R, R0 ) with R and R0 as defined in (3). For completeness we also
define X4 (3, 4) := ∅, which is trivially a set of flippable pairs for the set of paths P4 (3, 4) constructed
e4 (2, 3) := {R, R0 } and P
e4 (3, 4) := ∅.
before. The corresponding flipped paths are therefore P
15
Induction step n → n+1 (Q2n → Q2n+2 ), n ≥ 2: Consider the sets of flippable pairs X2n (k, k +1),
k = n, n + 1, . . . , 2n − 1, for the sets of paths P2n (k, k + 1), and the corresponding flipped paths
e2n (k, k + 1). In the following we describe how to use them to construct sets of flippable pairs
P
X2n+2 (k, k + 1), k = n + 1, n + 2, . . . , 2n + 1, for the sets P2n+2 (k, k + 1) in Q2n+2 (k, k + 1), and
e2n (k, k + 1).
corresponding flipped paths P
For all k = n + 2, . . . , 2n + 1 we define, in analogy to (8),
X2n+2 (k, k + 1) := X2n (k, k + 1) ◦ 00 ∪ X2n (k − 1, k) ◦ 10
∪ X2n (k − 1, k) ◦ 01 ∪ X2n (k − 2, k − 1) ◦ 11 ,
(12)
where we use the convention X2n (2n, 2n + 1) := ∅ and X2n (2n + 1, 2n + 2) := ∅ to unify treatment
of the sets of flippable pairs X2n+2 (2n, 2n + 1) and X2n+2 (2n + 1, 2n + 2) between the uppermost
levels of Q2n+2 . The sets of flippable pairs on the right hand side of (12) clearly lie in four disjoint
subgraphs of Q2n+2 (k, k + 1), so by induction X2n+2 (k, k + 1) is indeed a set of flippable pairs for
P2n+2 (k, k + 1).
The corresponding flipped paths are defined as
e2n+2 (k, k + 1) := P
e2n (k, k + 1) ◦ 00 ∪ P
e2n (k − 1, k) ◦ 10
P
e2n (k − 1, k) ◦ 01 ∪ P
e2n (k − 2, k − 1) ◦ 11 ,
∪P
e2n (2n, 2n + 1) := ∅ and P
e2n (2n + 1, 2n + 2) := ∅.
where we use the convention P
∗
defined in (9) (recall
To define the set X2n+2 (n + 1, n + 2), we consider the oriented paths P2n+2
that these paths originate from splitting up the 2-factor C2n+1 defined in (7)). By (7) and (9),
∗
every path P + ∈ P2n+2
has the following structure: There are two paths P, Pb ∈ P2n (n, n + 1) with
fα2n (L(Pb)) = L(P ) such that P + contains all edges except the first one from P ◦ 01 and all edges
F L ◦ 1 and one from the
from fα2n (Pb) ◦ 11 (P + has three more edges, two from the matching M2n+1
S
∗
matching M2n+2
). By this structural property of paths from P2n+2
and the fact that fα2n is an
isomorphism between the graphs Q2n (n, n + 1) and Q2n (n − 1, n), the set
0
0
∗
∗
X2n+2
:= (P + , P + ) | P + , P + ∈ P2n+2
and there is a flippable pair (Pb, Pb0 ) ∈ X2n (n, n + 1)
(13)
0
with fα2n (Pb) ◦ 11 ⊆ P + and fα2n (Pb0 ) ◦ 11 ⊆ P + .
∗
. We now define, in analogy to (11),
is a set of flippable pairs for P2n+2
∗
X2n+2 (n + 1, n + 2) := X2n (n + 1, n + 2) ◦ 00 ∪ X2n (n, n + 1) ◦ 10 ∪ X2n+2
.
(14)
The sets of flippable pairs on the right hand side of (14) lie in three disjoint subgraphs of Q2n+2 (n +
1, n + 2), so by induction X2n+2 (n + 1, n + 2) is indeed a set of flippable pairs for P2n+2 (n + 1, n + 2).
The corresponding flipped paths are defined as
∗
e2n+2 (n + 1, n + 2) := P
e2n (n + 1, n + 2) ◦ 00 ∪ P
e2n (n, n + 1) ◦ 10 ∪ P
e2n+2
P
,
(15)
+
+0
e∗
where P
2n+2 is the union of all paths obtained by considering every flippable pair (P , P ) in the
0
∗
set X2n+2
defined in (13) and by modifying P + and P + as follows: Let (Pb, Pb0 ) ∈ X2n (n, n + 1) be
0
such that fα2n (Pb) ◦ 11 ⊆ P + and fα2n (Pb0 ) ◦ 11 ⊆ P + (see the right hand side of (13)) and let P and
0
e2n (n, n + 1) so that F (Pb) = F (P ), F (Pb0 ) = F (P 0 ),
P be the corresponding flipped paths from P
0
0
L(Pb) = L(P ), L(Pb0 ) = L(P ). We replace the subpath fα2n (Pb)◦11 of P + by the subpath fα2n (P )◦11
0
0
and the last edge (F (Pb)◦11, F (Pb)◦01) of P + by the edge (F (P )◦11, F (P )◦01). Similarly, we replace
0
the subpath fα2n (Pb0 )◦11 of P + by the subpath fα2n (P )◦11 and the last edge (F (Pb0 )◦11, F (Pb0 )◦01)
0
of P + by the edge (F (P ) ◦ 11, F (P ) ◦ 01). Clearly, the resulting paths are a flipped pair of paths
0
∗
corresponding to the flippable pair of paths (P + , P + ) that was added to X2n+2
.
16
3.3. Correctness of the algorithm Paths(). In this section we show that the paths P2n (k, k +1)
e2n (k, k + 1) in the graph Q2n (k, k + 1) computed by the algorithm Paths() that form the basic
and P
building blocks of our Hamilton cycle algorithm (recall (4)) are exactly the same as the ones defined in
Sections 3.1 and 3.2. We shall see that essentially, the algorithm Paths() is a recursive reformulation
of this inductive construction. This allows us to exploit all the properties proved in [Müt16] about
these paths. In particular, we establish that the paths P2n (n, n + 1) satisfy properties (i) and (ii)
mentioned in Section 2.2. Recall that all paths we consider are oriented from the first vertex to
the last, and the notions of previous and next vertices on a path are defined with respect to this
orientation.
Lemma 4. Let n ≥ 1 and k ∈ {n, n + 1, . . . , 2n − 1} be fixed, and let P2n (k, k + 1) be the set
of paths in Q2n (k, k + 1) defined in Section 3.1 for the parameter sequence α2i = (1, 1, . . . , 1) ∈
{0, 1}i−1 , i = 1, 2, . . . , n − 1. For any path P ∈ P2n (k, k + 1) and any two consecutive vertices u
and v on P (i.e., P has the form P = (. . . , u, v, . . .)) we have v = Paths(n, k, u, next, false) and
u = Paths(n, k, v, prev, false).
In the following we will repeatedly use that by [Müt16, Lemma 11] the sets of first, second and last
vertices of the paths in P2n (k, k + 1) satisfy the relations
=0
F (P2n (k, k + 1)) = D2n
(k) ,
S(P2n (k, k + 1)) =
L(P2n (k, k + 1)) =
>0
D2n
(k + 1)
−
D2n
(k) .
(16a)
,
(16b)
(16c)
Proof. Note that when calling the algorithm Paths() with the parameter flip = false, this parameter retains its value in all recursive calls, so we can ignore lines P3–P4 of the algorithm for this
proof. We prove the lemma by induction on n. To settle the base case n = 1 observe that the set of
paths P2 (1, 2) defined in Section 3.1 and containing only a single path is the same as the one defined
in (1) and returned in lines P1–P2.
For the induction step n → n + 1 we assume that the lemma holds for n ≥ 1 and all k = n, n +
1, . . . , 2n − 1, and prove it for n + 1 and all k = n + 1, n + 2, . . . , 2n + 1. We distinguish the cases
k = n + 1 and k ∈ {n + 2, . . . , 2n + 1}.
For the case k ∈ {n + 2, . . . , 2n + 1} let P2n+2 (k, k + 1) be the set of paths defined in (8), let
P be a path from this set and let u and v be two consecutive vertices on P . Note that for both
calls Paths(n + 1, k, u, next, false) and Paths(n + 1, k, v, prev, false) the condition in line P6 is
satisfied, entailing the recursive call in line P7. The claim therefore follows from the definition (8)
and by induction: The quantity k − u2n − u2n+1 computed in line P7 evaluates to k, k − 1 or k − 2,
depending on the values of the last two bits of u and an analogous statement holds for v.
For the case k = n + 1 let P2n+2 (n + 1, n + 2) be the set of paths defined in (11), let P be a path from
this set and let u and v be two consecutive vertices on P . We distinguish three cases depending on
which of the three sets on the right hand side of (11) the path P is contained in.
Case 1: If P is contained in the second set on the right hand side of (11), i.e., in the set P2n (n, n+1)◦
10, then the last two bits of u and v are equal to 10, so for both calls Paths(n + 1, k, u, next, false)
and Paths(n + 1, k, v, prev, false) the condition in line P9 is satisfied, entailing the recursive call
in line P10. In this case the claim follows by induction.
Case 2: If P is contained in the first set on the right hand side of (11), i.e., in the set P2n (n+1, n+2)◦
00, then the last two bits of u and v are equal to 00, so for both calls Paths(n + 1, k, u, next, false)
and Paths(n + 1, k, v, prev, false) the condition in line P11 is satisfied. However, neither u nor
>0
v satisfies the condition in line P12, as by (10) and (16b), the vertices in D2n
(n + 1) ◦ 00 are all
∗
visited by the paths P2n+2
that form the third set on the right hand side of set of (11), so P does
17
not contain any of these vertices. Consequently, both cases entail the recursive call in line P13 and
the claim follows by induction.
∗
Case 3: If P is contained in the third set on the right hand side of (11), i.e., in the set P2n+2
, then by
n−1
the definitions (7) and (9) and the fact that for α2n = (1, 1, . . . , 1) ∈ {0, 1}
the mapping defined
in (6) satisfies fα2n = rev • π, the sequence of edges of the path P when traversing it from its first to
its last vertex has the form (e1 , E2 , e3 , E4 , e5 ), where e1 , e3 and e5 are single edges and E2 and E4 are
sequences of edges that satisfy the following conditions: There are two paths P 0 , P 00 ∈ P2n (n, n + 1)
such that
e1 = S(P 0 ) ◦ 00, S(P 0 ) ◦ 01 ,
e3 = L(P 0 ) ◦ 01, L(P 0 ) ◦ 11 = L(P 0 ) ◦ 01, rev(π(L(P 00 ))) ◦ 11 ,
e5 = rev(π(F (P 00 ))) ◦ 11, rev(π(F (P 00 ))) ◦ 01 ,
E2 is given by traversing the edges of P 0 ◦01 starting at the vertex S(P 0 )◦01 and ending at the vertex
L(P 0 ) ◦ 01, and E4 is given by the traversing all the edges of rev(π(P 00 )) ◦ 11 in reverse order starting
at the vertex rev(π(L(P 00 ))) ◦ 11 and ending at the vertex rev(π(F (P 00 ))) ◦ 11. We distinguish five
subcases depending on which part of P the edge (u, v) belongs to.
Case 3.1: If (u, v) = e1 = S(P 0 ) ◦ 00, S(P 0 ) ◦ 01 , then the vertices u and v differ in the last bit.
It follows from (16b) that for the call Paths(n + 1, k, u, next, false) the conditions in line P11 and
P12 are both satisfied, so the algorithm correctly returns v which is obtained from u by flipping the
last bit. Similarly, for the call Paths(n + 1, k, v, prev, false) the conditions in line P14 and P17 are
both satisfied, so the algorithm correctly returns u which is obtained from v by flipping the last bit.
Case 3.2: If (u, v) is an edge from E2 , then the last two bits of u and v are equal to 01, so for both
calls Paths(n + 1, k, u, next, false) and Paths(n + 1, k, v, prev, false) the condition in line P14
is satisfied. However, the conditions in line P15, P16 and P17 are not satisfied: To see this recall
(16a), (16b) and (16c), and that by the definition of E2 , u and v are different from F (P 0 ) ◦ 01, u
is different from L(P 0 ) ◦ 01, and v is different from S(P 0 ) ◦ 01. Consequently, both cases entail the
recursive call in line P18 and the claim follows by induction.
Case 3.3: If (u, v) = e3 = L(P 0 ) ◦ 01, L(P 0 ) ◦ 11 , then the vertices u and v differ in the second
to last bit. It follows from (16c) that for the call Paths(n + 1, k, u, next, false) the conditions in
line P14 and P16 are both satisfied, so the algorithm correctly returns v which is obtained from u by
flipping the second to last bit. Similarly, for the call Paths(n + 1, k, v, prev, false) the conditions
in line P19 and P21 are both satisfied, so the algorithm correctly returns u which is obtained from
v by flipping the second to last bit.
Case 3.4: If (u, v) is an edge from E4 , then the last two bits of u and v are equal to 11, so for both
calls Paths(n + 1, k, u, next, false) and Paths(n + 1, k, v, prev, false) the condition in line P19
is satisfied. However, the conditions in line P20 and P21 are not satisfied: To see this recall (16a)
=0 (n) and
and (16c) and that the mapping rev • π maps each of the sets F (P2n (n, n + 1)) = D2n
−
L(P2n (n, n + 1)) = D2n (n) onto itself, and note that by the definition of E4 , u is different from
rev(π(F (P 00 ))) ◦ 11 and v is different from rev(π(L(P 00 ))) ◦ 11. Consequently, both cases entail the
recursive call in line P22. The claim follows by induction, noting that inverting the value of the
variable dir accounts for the fact that in E4 the edges of rev(π(P 00 )) ◦ 11 are traversed in reverse
order.
Case 3.5: If (u, v) = e5 = rev(π(F (P 00 ))) ◦ 11, rev(π(F (P 00 ))) ◦ 01 , then the vertices u and v
=0 (n), so for the
differ in the second to last bit. It follows from (16a) that rev(π(F (P 00 ))) ∈ D2n
call Paths(n + 1, k, u, next, false) the conditions in line P19 and P20 are both satisfied and the
algorithm correctly returns v which is obtained from u by flipping the second to last bit. Similarly,
18
for the call Paths(n + 1, k, v, prev, false) the conditions in line P14 and P15 are both satisfied, so
the algorithm correctly returns u which is obtained from v by flipping the second to last bit.
This completes the proof.
Lemma 5. Let n ≥ 2 and k ∈ {n, n + 1, . . . , 2n − 1} be fixed, let P2n (k, k + 1) be the set of paths
e2n (k, k + 1)
defined in Section 3.1 and X2n (k, k + 1) the set of flippable pairs for P2n (k, k + 1) and P
the corresponding set of flipped paths in Q2n (k, k + 1) defined in Section 3.2 for the parameter
e2n (k, k + 1) and
sequence α2i = (1, 1, . . . , 1) ∈ {0, 1}i−1 , i = 1, 2, . . . , n − 1. For any path P ∈ P
any two consecutive vertices u and v on P (i.e., P has the form P = (. . . , u, v, . . .)) we have
v = Paths(n, k, u, next, true) and u = Paths(n, k, v, prev, true).
From (16) and from the definition of flippable/flipped pairs of paths we obtain that
=0
e2n (k, k + 1)) = D2n
F (P
(k) ,
e2n (k, k + 1)) = D− (k) .
L(P
2n
(17a)
(17b)
Proof. We prove the lemma by induction on n, in an analogous way as the proof of Lemma 4. One
crucial difference, however, is the following: When calling the algorithm Paths() with the parameter
flip = true, this parameter retains its value in all recursive calls except in line P18, where it is set
to false.
e4 (2, 3) defined in Section 3.2 (consisting
To settle the base case n = 2 observe that the set of paths P
of two paths on three and seven vertices, respectively) is the same as the one defined in (3) and
e4 (3, 4) defined in Section 3.2 is empty, so the claim is
returned in lines P3–P4. The set of paths P
trivially true.
The induction step n → n + 1 proceeds in an analogous way as in the proof of Lemma 4. We only
give the details for case 3, which is proved in a slightly different way:
e∗ , then
Case 3’: If P is contained in the third set on the right hand side of (15), i.e., in the set P
2n+2
n−1
by the definition given after (15) and the fact that for α2n = (1, 1, . . . , 1) ∈ {0, 1}
the mapping
defined in (6) satisfies fα2n = rev • π, the sequence of edges of the path P when traversing it from its
first to its last vertex has the form (e1 , E2 , e3 , E4 , e5 ), where e1 , e3 and e5 are single edges and E2 and
E4 are sequences of edges that satisfy the following conditions: There is a path P 0 ∈ P2n (n, n + 1)
e2n (n, n + 1) (this is the crucial difference to before, where P 0 and P 00 where both
and a path P 00 ∈ P
from the set P2n (n, n + 1)) such that
e1 = S(P 0 ) ◦ 00, S(P 0 ) ◦ 01 ,
e3 = L(P 0 ) ◦ 01, L(P 0 ) ◦ 11 = L(P 0 ) ◦ 01, rev(π(L(P 00 ))) ◦ 11 ,
e5 = rev(π(F (P 00 ))) ◦ 11, rev(π(F (P 00 ))) ◦ 01 ,
E2 is given by traversing the edges of P 0 ◦ 01 starting at the vertex S(P 0 ) ◦ 01 and ending at the
vertex L(P 0 ) ◦ 01, and E4 is given by the traversing all the edges of rev(π(P 00 )) ◦ 11 in reverse order
starting at the vertex rev(π(L(P 00 ))) ◦ 11 and ending at the vertex rev(π(F (P 00 ))) ◦ 11.
As in the proof of Lemma 4, we distinguish five cases depending on which part of P the edge (u, v)
belongs to. The last two cases are treated analogously to cases 3.4 and 3.5 in the proof of Lemma 4,
using (17a) and (17b) instead of (16a) and (16c). The first three cases are exactly the same (not just
analogous) as cases 3.1, 3.2 and 3.3. In particular, in case 3.2 where (u, v) is an edge from E2 note
that the edges in E2 originate from a path P 0 ∈ P2n (n, n + 1), which is accounted for by setting the
value of the variable flip to false in the recursive call in line P18 of the algorithm Paths().
19
e2n (k, k +1)
Given Lemma 4 and Lemma 5, we may from now on use the notations P2n (k, k +1) and P
interchangeably for the sets of paths defined in Sections 3.1 and 3.2 and for the sets of paths computed
by the algorithm Paths() called with parameter flip = false or flip = true, respectively.
3.4. Correctness of the algorithm HamCycle().
3.4.1. HamCycle() computes a 2-factor. By combining Lemma 4, property (a) from Section 3.1
and the relations (16a) and (16c), we obtain that the paths P2n (n, n + 1) computed by the algorithm
Paths() indeed have properties (i) and (ii) claimed in Section 2.2. Consequently, assuming for
a moment that the auxiliary functions IsFlipTree1 () and IsFlipTree2 () always return false
(i.e., IsFlipVertex() always returns false), the arguments given in Sections 2.3 and 2.4 show
that the algorithm HamCycle() correctly computes the 2-factor C2n+1 defined in (4) in the middle
levels graph Q2n+1 (n, n + 1). We proceed to show that the algorithm HamCycle() computes a
different 2-factor (but still a 2-factor) for each possible choice of boolean functions IsFlipTree1 ()
and IsFlipTree2 () that are called from within IsFlipVertex(). Later we prove that the functions
IsFlipTree1 () and IsFlipTree2 () specified in Section 2.5 yield a 2-factor that consists only of
a single cycle, i.e., a Hamilton cycle. As already mentioned in Section 2.3, the different 2-factors
are obtained by replacing some flippable pairs of paths from P2n (n, n + 1) in the first set on the
e2n (n, n + 1). There are
right hand side of (4) by the corresponding flipped pairs of paths from P
two potential problems that the function IsFlipVertex() as it is called from HamCycleNext()
and HamCycle() (see line N3 and line H7, respectively) could cause in this approach, and the next
lemma shows that none of these problems occurs. First, the function might return true for a vertex
=0 (n) for which the path P ∈ P (n, n + 1) that starts with this vertex (i.e., F (P ) = x) is
x ∈ D2n
2n
not contained in a flippable pair, so the subsequent calls to Paths() with parameter flip = true
would produce undefined output. Second, given a flippable pair of paths P, P 0 ∈ P2n (n, n + 1), the
results of the calls IsFlipVertex(n, x) and IsFlipVertex(n, x0 ) with x := F (P ) and x0 := F (P 0 )
might be different/inconsistent, so our algorithm would not compute a valid 2-factor.
Lemma 6. Let n ≥ 1, and let the sets P2n (n, n+1) and X2n (n, n+1) be as in Lemma 5. Furthermore,
let IsFlipTree1 () and IsFlipTree2 () be arbitrary boolean functions on the sets of ordered rooted
∗ and T ∗ defined in Section 2.5, respectively, and let IsFlipVertex() be as defined in
trees Tn,1
n,2
=0 (n) with IsFlipVertex(n, x) = true there is another
Algorithm 4. For any lattice path x ∈ D2n
0
=0
lattice path x ∈ D2n (n) with IsFlipVertex(n, x0 ) = true and two paths P, P 0 ∈ P2n (n, n + 1) with
(F (P ), F (P 0 )) = (x, x0 ) that form a flippable pair (P, P 0 ) ∈ X2n (n, n + 1).
Proof. By the definitions in lines T1–T5, the tree T := h−1 (x) is contained in exactly one of the
∗ , τ (T ∗ ), T ∗ , τ (T ∗ ). All these sets of trees are disjoint. We consider four cases
four sets Tn,1
1 n,1
2 n,2
n,2
depending on which of the sets the tree T is contained in.
∗ , then by the instructions in line T2 we have IsFlipTree (T ) = true. Now
Case 1: If T ∈ Tn,1
1
0
0
0
=0 (n). Clearly,
consider the tree T := τ1 (T ) and the lattice path x := h(T ) = h(τ1 (T )) ∈ D2n
0
∗
the instructions in line T3 ensure that IsFlipVertex(n, x ) = true. By the definition of Tn,1
0
and τ1 , the lattice paths corresponding to the trees T and T satisfy the preconditions of [Müt16,
Lemma 21], implying that the pair of lattice paths (h(T ), h(T 0 )) = (x, x0 ) is contained in the set
H1 defined in [Müt16, Lemma 24], so by this lemma there are two paths P, P 0 ∈ P2n (n, n + 1) with
(F (P ), F (P 0 )) = (x, x0 ) that form a flippable pair (P, P 0 ) ∈ X2n (n, n + 1).
∗ ), then by the instructions in line T3 we have IsFlipTree (τ −1 (T )) = true.
Case 2: If T ∈ τ1 (Tn,1
1 1
=0 (n).
Now consider the tree T 0 := τ1−1 (T ) and the lattice path x0 := h(T 0 ) = h(τ1−1 (T )) ∈ D2n
Clearly, the instructions in line T2 ensure that IsFlipVertex(n, x0 ) = true. From here the proof
continues as in case 1, however, the roles of T and T 0 are interchanged.
20
∗ and T ∈ τ (T ∗ ) are dealt with analogously to case 1 and case 2, respectively,
The cases T ∈ Tn,2
2 n,2
replacing the application of [Müt16, Lemma 21] by [Müt16, Lemma 22] and the set H1 by H2 defined
in [Müt16, Lemma 24].
The following simple but powerful lemma follows immediately from the definition of flippable/flipped
pairs of paths given in Section 2.2.
Lemma 7. Let G be a graph, C a 2-factor in G, C and C 0 two cycles in C, (P, P 0 ) a flippable pair of
paths such that P is contained in C and P 0 is contained in C 0 , and (R, R0 ) a corresponding flipped
pair of paths. Replacing the paths P and P 0 in the 2-factor C by R and R0 yields a 2-factor in which
the two cycles C and C 0 are joined to a single cycle.
The Hamilton cycle computed by our algorithm HamCycle() is obtained from the 2-factor C2n+1
defined in (4) by repeatedly applying the flipping operation from Lemma 7. Specifically, we replace
several flippable pairs of paths P, P 0 ∈ P2n (n, n + 1) by the corresponding flipped paths R, R0 ∈
e2n (n, n+1) in the first set on the right hand side of (4) such that all cycles of the 2-factor C2n+1 are
P
joined to a single cycle. This of course requires that all flippable pairs of paths are disjoint, which is
guaranteed by the definitions of P2n (n, n + 1) and X2n (n, n + 1): The paths in P2n (n, n + 1) are all
disjoint, and each path from this set appears in at most one flippable pair in the set X2n (n, n + 1).
To reasonably choose which flippable pairs of paths to replace by the corresponding flipped paths,
we need to understand the cycle structure of the 2-factor C2n+1 .
3.4.2. Structure of cycles in the 2-factor C2n+1 . To describe the structure of the cycles in the 2-factor
C2n+1 defined in (4), we introduce the concept of plane trees.
Plane trees. A plane tree is a tree with a cyclic ordering of all neighbors of each vertex. We think
of a plane tree as a tree embedded in the plane such that for each vertex v the order in which
its neighbors are encountered when walking around v in counterclockwise direction is precisely the
specified cyclic ordering, see the middle of Figure 4. We denote by Tn the set of all plane trees with n
edges. The notions of thin/thick leaves and of clockwise/counterclockwise-next leaves introduced in
Section 2.5 for ordered rooted trees can be defined for plane trees in a completely analogous fashion.
Transformations plane() and Root() between ordered rooted trees and plane trees. For any ordered
rooted tree T ∗ ∈ Tn∗ , we define a plane tree T = plane(T ∗ ) ∈ Tn as follows: The underlying
(abstract) tree of T ∗ and T is the same. For the root r of T ∗ , the cyclic ordering of neighbors of r
in T is given by the left-to-right ordering of the children of r in T ∗ . For any other vertex v of T ∗ , if
(u1 , u2 , . . . , uk ) is the left-to-right ordering of the children of v in T ∗ and w is the parent of v, then
we define (w, u1 , u2 , . . . , uk ) as the cyclic ordering of neighbors of v in T .
For any plane tree T ∈ Tn and any edge (r, u) of T , we define an ordered rooted tree T ∗ =
Root(T, (r, u)) ∈ Tn∗ as follows: The underlying (abstract) tree of T and T ∗ is the same. The
vertex r is the root of T ∗ , and if (u1 , u2 , . . . , uk ) with u1 = u is the cyclic ordering of neighbors of
r in T , then (u1 , u2 , . . . , uk ) is the left-to-right ordering of the children of r in T ∗ . For any other
vertex v of T , if w is the vertex on the path from v to r and (u0 , u1 , . . . , uk ) with u0 = w is the cyclic
ordering of neighbors of v in T , then (u1 , u2 , . . . , uk ) is the left-to-right ordering of the children of v
in T ∗ .
Informally speaking, plane(T ∗ ) is obtained from T ∗ by ‘forgetting’ the root vertex, and Root(T, (r, u))
is obtained from T by ‘pulling out’ the vertex r as root such that u is the leftmost child of the root,
see Figure 4.
For any cycle C of the 2-factor C2n+1 defined in (4) we let F (C) denote the cyclic sequence of all
vertices of the form F (P ), P ∈ P2n (n, n+1), for which the path P ◦0 is contained in C (when walking
along C in the direction such that P ◦ 0 is traversed starting at the first vertex and ending at the
21
T1∗
plane(T1∗ )
r1
plane(T2∗ )
T
u1
∗
r2 T2
u2
u1 r1
u2
root(T, (r1 , u1 ))
r2
root(T, (r2 , u2 ))
Figure 4. Two ordered rooted trees T1∗ , T2∗ ∈ T7∗ (left and right), a plane tree T ∈ T7
(middle), and the transformations between them.
rot
Tb1
rot
rot
Tb2
Tb3
Tb4
Tb5
Tb6
Tb7
Tb8
Tb9
Tb10
Tb11
Tb12
Tb13
Tb14
T2
T3
T4
T5
T6
T7
T8
T9
T10
T11
T12
T13
T14
T1
F (P1 ) F (P2 ) F (P3 ) F (P4 ) F (P5 ) F (P6 ) F (P7 ) F (P8 )
C1
C2n+1 = {C1 , C2 , C3 }
F (P9 ) F (P10 ) F (P11 ) F (P12 ) F (P13 ) F (P14 )
C2
C3
P2n (n, n + 1)
rev(P2n (n, n + 1))
h
Figure 5. Structure of the 2-factor C2n+1 for n = 4 (cf. the bottom part of Figure 1).
The figure shows the cyclic sequences F (C) of ordered rooted trees for all cycles
C ∈ C2n+1 (bottom) and their preimages under the mapping h (top). Note that trees
rotate along each of the sequences h−1 (F (C)), C ∈ C2n+1 . The boxes show the plane
∗ and T ∗ defined in
trees associated with each cycle C ∈ C2n+1 . The sets of trees Tn,1
n,2
∗ = {T
b7 , Tb11 } and T ∗ = {Tb6 } and we have τ1 (Tb7 ) = Tb14 ,
Section 2.5 are given by Tn,1
n,2
τ1 (Tb11 ) = Tb3 and τ2 (Tb6 ) = Tb9 .
last vertex). It turns out to be very useful to consider the ordered rooted trees corresponding to the
=0 (n) and
elements of the sequence F (C) (recall (16a) and the bijection between lattice paths in D2n
ordered rooted trees in Tn∗ ). The following lemma is a slight reformulation of [Müt16, Lemma 20].
For the reader’s convenience, the lemma is illustrated in Figure 5.
Lemma 8. For any n ≥ 1 and any cycle C of the 2-factor C2n+1 defined in (4), considering the
cyclic sequence of ordered rooted trees (T1 , T2 , . . . , Tk ) := F (C) we have h−1 (Ti+1 ) = rot(h−1 (Ti ))
for all i = 1, 2, . . . , k (indices are considered modulo k) with h−1 and rot as defined in Section 2.5.
I.e., we can associate the cycle C with the plane tree plane(h−1 (T1 )) = · · · = plane(h−1 (Tk )) ∈ Tn .
Moreover, for any plane tree T ∈ Tn there is a cycle C ∈ C2n+1 to which the tree T is associated. In
particular, the number of cycles of the 2-factor is |C2n+1 | = |Tn |.
22
3.4.3. HamCycle() computes a Hamilton cycle. Lemmas 6, 7 and 8 motivate the following definitions of two directed multigraphs Gn and Hn ⊆ Gn that are illustrated in Figure 6.
Graphs Gn and Hn . The nodes of Gn are sets of ordered rooted trees from Tn∗ such that all trees that
differ by the rotation operation rot are in the same set. The edges of Gn are defined as follows: For
∗ and T 0 = τ (T ) we add an edge directed from
i ∈ {1, 2} and any pair of trees (T, T 0 ) with T ∈ Tn,i
i
the set of trees containing T to the set of trees containing T 0 . We refer to the edges of Gn induced
by the mappings τ1 and τ2 as τ1 -edges or τ2 -edges, respectively. The graph Hn has the same nodes
as Gn , but only a subset of its edges. Specifically, for i ∈ {1, 2} and any pair of trees (T, T 0 ) with
∗ and T 0 = τ (T ) we only add an edge directed from the set of trees containing T to the set
T ∈ Tn,i
i
of trees containing T 0 if IsFlipTreei (T ) = true, where IsFlipTree1 () and IsFlipTree2 () are as
defined in Section 2.5.
As each node of Gn and Hn is a set {T1 , T2 , . . . , Tk } of ordered rooted trees that differ by rotation,
i.e., we have Ti+1 = rot(Ti ) for i = 1, 2, . . . , k, we can identify this node with the plane tree
plane(T1 ) = · · · = plane(Tk ) ∈ Tn , see the example in Figure 6 for n = 6. Also, we can interpret the
result of the operations τ1 and τ2 that define the edges of Gn and Hn in terms of the corresponding
plane trees. Note e.g. that τ1 increases the number of leaves by one, and τ2 decreases the number of
leaves by one, implying that neither Gn nor Hn has any loops. This effect can be seen in Figure 6,
where all τ1 -edges go from right to left, and all τ2 -edges from left to right. The graph Gn may have
multiple edges between nodes, see Figure 6. The τ1 - and τ2 -edges added to the graph Gn are also
described in Figure 5 for the smaller example n = 4.
By Lemma 8, the nodes of Gn and Hn correspond to the cycles of the 2-factor C2n+1 defined in (4).
Applying Lemma 6 with boolean functions IsFlipTree1 () and IsFlipTree2 () that always return
true shows that each edge of Gn corresponds to a flippable pair of paths from the set X2n (n, n+1) for
which the paths are contained in the corresponding cycles. In other words, the edges of Gn capture
all potential flipping operations that could be performed to modify the 2-factor C2n+1 . Applying
Lemma 6 with the boolean functions IsFlipTree1 () and IsFlipTree2 () defined in Section 2.5
shows that each edge of Hn corresponds to a flippable pair of paths from P2n (n, n + 1) that are
e2n (n, n + 1) in the first set on the right
actually replaced by the corresponding flipped paths from P
hand side of (4) by our Hamilton cycle algorithm HamCycle(). By Lemma 7 the edges of Hn
therefore indicate which pairs of cycles from the 2-factor C2n+1 the algorithm joins to a single cycle.
Consequently, to complete the correctness proof for the algorithm HamCycle(), it suffices to show
the following lemma.
Lemma 9. For any n ≥ 1, the graph Hn is a spanning tree of Gn .
Proof. For n = 1 and n = 2 the graph Gn consists only of a single node and no edges, and we have
Gn = Hn , so in these cases the claim is trivially true. For n = 3 the graph Gn consists of two nodes,
connected by a τ1 -edge, and we again have Gn = Hn , proving the lemma also in this case. For the
rest of the proof we assume that n ≥ 4.
Let Hn0 denote the graph obtained from Hn by removing all τ2 -edges. We first show that Hn0 is a
spanning forest of Gn .
By the definition of the function IsFlipTree1 () from Section 2.5, the outdegree of every node
T ∈ Tn in Hn0 is at most 1. More precisely, if T has a thin leaf, then the outdegree is 1, as there is
∗ (let u denote a thin leaf, u0 the neighbor of
a way to root T such that it is contained in the set Tn,1
u and v the neighbor of u0 , then this rooted tree is given by Root(T, (v, u0 )), see the left hand side
of Figure 3), and only for the lexicographically smallest such tree the function returns true. On the
other hand, it T has no thin leaves, then the outdegree is 0, as there is no way to root T such that
∗ , so the function returns false for all these trees (see Figure 6). Moreover, as
it is contained in Tn,1
23
Pn00
G6
Pn0
τ1 -edges
τ2 -edges
H6
Pn
2 leaves
6 leaves
5 leaves
4 leaves
3 leaves
Figure 6. The graphs G6 and H6 with nodes are arranged in levels according to
the number of leaves of the corresponding plane trees. The figure shows for each
node, i.e., for each set of ordered rooted trees that differ by rotation, the plane tree
corresponding to these trees. In the figure, τ1 -edges are drawn as solid lines, τ2 -edges
as dashed lines, and the edges of H6 are highlighted in grey.
τ1 increases the number of leaves by one, the graph Hn0 is acyclic. An acyclic directed graph with
outdegree at most 1 at every node is a forest, proving that Hn0 is indeed a spanning forest of Gn .
In the following we argue that adding the τ2 -edges from Hn to Hn0 turns the spanning forest into a
spanning tree of Gn .
By what we said before the components (=trees) of the forest Hn0 can be characterized precisely:
Every component has a unique sink, i.e., a node with no outgoing edges, which corresponds to a
plane tree with no thin leaves. Also, the distance of a node T ∈ Tn from the sink of the component
it belongs to equals the number of thin leaves of T . Moreover, we obtain that the number of
components of Hn0 is given by the number of plane trees from Tn that have no thin leaves. In
Figure 6, Hn0 is obtained from Hn by removing two dashed τ2 -edges, resulting in three components
whose sinks appear as the leftmost node of each component.
We now consider a component of Hn0 that will play a special role in the following. Let Pn ∈ Tn
denote the path with n edges, and let (v1 , v2 , . . . , vn+1 ) be the sequence of vertices along this path.
We denote by Pn0 ∈ Tn the graph obtained from Pn by replacing the edge (vn , vn+1 ) by the edge
(vn−1 , vn+1 ). Furthermore, we denote by Pn00 ∈ Tn the graph obtained from Pn0 by replacing the edge
(v1 , v2 ) by the edge (v1 , v3 ). Clearly, Hn0 has a component consisting of the three nodes Pn , Pn0 and
Pn00 , with a τ1 -edge directed from Pn to Pn0 and a τ1 -edge directed from Pn0 to the sink Pn00 , see the
top part of Figure 6. We denote this component of Hn0 by Sn . Note that the definition of Sn is
meaningful only for n ≥ 4.
We proceed to show the following: (i) In every component of Hn0 except Sn there is exactly one node
that has an outgoing τ2 -edge in Hn , and the component Sn has no node with an outgoing τ2 -edge.
(ii) The τ2 -edges of Hn form an acyclic graph on the components of Hn0 (in particular, every τ2 -edge
of Hn starts and ends in a different component of Hn0 ). Combining our earlier observation that Hn0
is a spanning forest of Gn with claims (i) and (ii) clearly proves the lemma, so it remains to prove
(i) and (ii).
24
T0
d(T 0 ) ≥ 8 > d(T )
00
8
w
00
v v
v0
5
7
w
2
u v
v v0
T
6
u v v 0 v 00
Tb
v v 0 v 00
d(T ) = max{2, 5, 6, 7} = 7
u0
u
w
Tb0
u v v 0 v 00
w
Gn
w
τ1 -edges
τ2 -edges
Hn
Figure 7. Illustration of the notations used in the proof of Lemma 9.
Proof of (i): Consider a component of Hn0 with sink T ∈ Tn , and let N (T ) be the set of all nodes in
distance 1 from T in Hn0 . As T has no thin leaves, it can be rooted so that it is an image of τ1 : let u
and u0 be two leaves that have a common neighbor v such that u0 is the counterclockwise-next leaf
from u, then this rooted tree is given by Root(T, (v, u)). See the left hand side of Figure 3. The
preimage is a tree with exactly one thin leaf, which corresponds to a node T 0 ∈ Tn with a τ1 -edge
directed from T 0 to T that is present in the graph Hn0 . It follows that the set N (T ) is nonempty.
Note that the function IsFlipTree2 () returns true only if the given tree has a single thin leaf, so
the only outgoing τ2 -edges of Hn that start from a node in the same component as T start at a
node in N (T ). Moreover, as any tree in N (T ) can be rooted uniquely so that it is a preimage of τ2
(let v denote the thin leaf, u the counterclockwise-next leaf from v, u0 the neighbor of u and r the
neighbor of u0 next to u in the counterclockwise ordering of the neighbors of u0 , then this rooted tree
is given by Root(T, (r, u0 )), see the right hand side of Figure 3), every node in N (T ) has exactly
one outgoing τ2 -edge in Gn , see Figure 6. If the component containing T is different from Sn , then
the function IsFlipTree2 () returns true for exactly one of these rooted trees: The function first
computes a rooted version of T for each of them, thus consistently assigning an integer value to each
of the trees (the integers are derived from distances between certain vertices of T , but the origin of
these values is irrelevant for this part of the argument), and returns true only for the tree that was
assigned the largest value and that is lexicographically smallest among all trees with the same value.
However, if the component containing T is Sn (i.e., T = Pn ), then N (T ) = {Pn0 } and the unique
way to root Pn0 such that it is a preimage of τ2 is the tree 1n−1 ◦ 0n−2 ◦ 100, for which the function
IsFlipTree2 () exceptionally returns false. This exceptional rooted tree is not encountered in any
component other than Sn .
Proof of (ii): For the reader’s convenience, this part of the proof is illustrated in Figure 7. For any
plane tree T ∈ Tn , we define d(T ) as the maximum distance between any two leaves v and w that
satisfy the following conditions: w is the clockwise-next leaf from v and the counterclockwise-next
leaf v 0 from v has distance 2 from v. In other words, the leaves v and v 0 are adjacent to a common
vertex v 00 ; see the bottom left part of Figure 7. Claim (ii) is an immediate consequence of the
following observation: For any τ2 -edge from Hn directed from a component of Hn0 with sink T ∈ Tn
to a component of Hn0 with sink T 0 ∈ Tn , we have d(T ) < d(T 0 ). In particular, the edge starts and
ends in a different component of Hn0 . To see this let Tb ∈ Tn be the node in distance 1 of T in Hn0
that is the starting node of a τ2 -edge of Hn (this node exists and is unique by claim (i)), and let
Tb0 ∈ Tn be the end node of this τ2 -edge. As Tb has exactly one thin leaf, and as τ2 preserves one thin
leaf and possibly creates one additional thin leaf, Tb0 has distance 1 or 2 from T 0 . So in Hn there is a
25
τ1 -edge directed from Tb to T , a τ2 -edge directed from Tb to Tb0 and either a single τ1 -edge or a path
consisting of two τ1 -edges directed from Tb0 to T 0 , see Figure 7.
By the definition of IsFlipTree2 () and of the parameter d(T ), the tree T has two leaves v and w
in distance d(T ) such that w is the clockwise-next leaf from v and the counterclockwise-next leaf v 0
from v has distance 2 from v, implying that v and v 0 have a common adjacent vertex v 00 . Furthermore
Tb is obtained from T by replacing the edge (v, v 00 ) by (v, v 0 ). The function IsFlipTree2 () of course
computes a rooted version of T from a rooted version of Tb.
By the definition of τ2 , the tree Tb0 is obtained from Tb by considering the counterclockwise-next leaf
u from v, its neighbor u0 in Tb, and by replacing the edge (u, u0 ) by the edge (u, v). By the definition
of τ1 , T 0 is obtained from Tb0 by replacing the edge (u, v) by (u, v 0 ) plus possibly another application
of τ1 if Tb0 has two thin leaves. It follows that the distance between u and w in T 0 is d(T ) + 1,
implying that d(T 0 ) > d(T ), as claimed.
This completes the proof.
Remark 10. Here is how our algorithm can be modified to compute another Hamilton cycle in the
middle levels graph: The key is to modify the functions IsFlipTree1 () and IsFlipTree2 () such
that they define a spanning tree of Gn that is different from the ‘canonical’ spanning tree Hn . Each
different spanning tree of Gn clearly yields a different Hamilton cycle in the middle levels graph
Q2n+1 (n, n + 1) (cf. [Müt16, Lemma 6]). The only limitation an efficient algorithm must obey is
that the spanning tree must be locally computable, i.e., given information about the current node
T ∈ Tn of Gn , the algorithm must decide based on a polynomially sized neighborhood of T which
edges of Gn to include and which to exclude from a spanning tree. Note that the degree of the nodes
of Gn is O(n) and the total number of nodes is |Tn | = Θ(4n · n−5/2 ).
4. Running time and space requirements of the algorithm
In this section we prove that our Hamilton cycle algorithm can be implemented so that it satisfies
the runtime and space bounds claimed in the introduction. In the runtime analysis we apply some
results about the paths P2n (n, n + 1) that have been established in [MW12].
4.1. Running time. A naive implementation of the function Paths() takes time O(n2 ): To see
=0 (n − 1),
this observe that the membership tests whether x− is contained in one of the sets D2n−2
−
>0
(n − 1) or D2n−2
(n) in lines P12, P15, P16, P17, P20 and P21 and the application of the
D2n−2
mappings rev and π in line P22 take time O(n) (recall that rev−1 = rev and π −1 = π), and that the
value of n decreases by 1 with each recursive call. In the following we sketch how this can improved
so that each call of Paths() takes only time O(n). More details can be found in the comments of
our C++ implementation [www]. For this we maintain counters c0 and c1 for the number of zeros
and ones of a given bitstring x = (x1 , x2 , . . . , x2n ). Moreover, interpreting the bitstring x as a lattice
path (as described in Section 2.1), we maintain vectors c00 , c01 , c10 , c11 that count the number
of occurences of pairs of consecutive bits (x2i , x2i+1 ), i ∈ {1, 2, . . . , n − 1}, per height level of the
lattice path for each of the four possible value combinations of x2i and x2i+1 . E.g., for the bitstring
x = 1100001010 the vector c10 has a single nonzero entry 1 at height level (=index) 1 for the two bits
(x2 , x3 ) = 10, the vector c00 has a single nonzero entry 1 at height level (=index) 0 for the two bits
(x4 , x5 ) = 00, and the vector c01 has a single nonzero entry 2 at height level (=index) −1 for the pairs
of bits (x6 , x7 ) = (x8 , x9 ) = 01. Using these counters, the three membership tests mentioned before
=0 (n) if
can be performed in constant time. E.g., a bitstring x = (x1 , x2 , . . . , x2n ) is contained in D2n
and only if c0 = c1 and x1 = 1 and the entry of c00 at height level 0 equals 0 (i.e., the lattice path
never moves below the line y = 0). Moreover, these counters can be updated in constant time when
removing the last two bits of x and when applying the mappings rev and π: Note that rev simply
26
swaps the roles of c0 and c1 and the roles of c00 and c11 and possibly incurs an index shift, and that
π simply swaps the roles of c10 and c01 . To compute the applications of rev and π in line P22 in
constant time, we do not modify x at all, but rather count the number of applications of rev and π
and keep track of the middle range of bits of x that is still valid. When removing the last two bits of
x, this range shrinks by 2 on one of the sides. Taking into account that multiple applications of rev
and π cancel each other out, this allows us to compute the effect of applying those mappings lazily
when certain bits are queried later on (when testing the values of the last two bits of some substring
of x). E.g., by repeatedly removing the last two bits and applying the mappings rev and π, the
bitstring x = (x1 , x2 , . . . , x10 ) is transformed into (x8 , x6 , x7 , x4 , x5 , x2 , x3 , x1 ), (x2 , x4 , x5 , x6 , x7 , x8 ),
(x6 , x4 , x5 , x2 ), and finally into (x4 , x6 ).
The function IsFlipVertex() can be implemented to run in time O(n2 ). To see this observe that
the result of applying the function h−1 in line T1 can be computed in time O(n2 ), and that the
functions IsFlipTree1 () and IsFlipTree2 () called in lines T2–T5 also need time O(n2 ): For both
we need to rotate an ordered rooted tree with n edges (=bitstring of length 2n) for one full rotation,
and each rotation operation takes time O(n).
It was shown in [MW12, Lemma 9] that the length of any path P ∈ P2n (n, n + 1) with a first
=0 (n) is given by the following simple formula: Considering the unique
vertex F (P ) =: x ∈ D2n
=0 (k) for some k ≥ 0, the length of P is given by
decomposition x = 1 ◦ x` ◦ 0 ◦ xr with x` ∈ D2k
2|x` | + 2 ≤ 2(2n − 2) + 2 = 4n − 2. It follows that the while-loop in line H5 terminates after at most
O(n) iterations, i.e., the initialization phase of HamCycle() (lines H2–H7) takes time O(n2 ).
It was shown in [MW12, Theorem 10] that the distance between any two neighboring vertices of the
=0 (n) (x and x0 are first vertices of two paths P, P 0 ∈ P (n, n + 1))
form x ◦ 0, x0 ◦ 0 with x, x0 ∈ D2n
2n
on a cycle in (4) is exactly 4n + 2. Comparing the lengths of two paths from P2n (n, n + 1) that form
e2n (n, n + 1), we
a flippable pair with the lengths of the corresponding flipped paths from the set P
observe that either the length of one the paths decreases by 4 and the length of the other increases
by 4, or the lengths of the paths do not change. Specifically, the paths P, P 0 and R, R0 defined in (2)
and (3) have exactly the length differences −4 and +4, and these differences only propagate through
the first cases of the Paths() recursion, but not the last case in lines P19–P22. It follows that every
call of HamCycleNext() for which the condition in line N3 is satisfied and which therefore takes
time O(n2 ) due to the call of IsFlipVertex(), is followed by at least 4n − 3 calls in which the
condition is not satisfied, in which case HamCycleNext() terminates in time O(n). Consequently,
` consecutive calls of HamCycleNext() take time O(n2 + n`).
Summing up the time O(n2 ) spent for the initialization phase and O(n2 +n`) for the actual Hamilton
cycle computation, we conclude that the algorithm HamCycle(n, x, `) runs in time O(n2 + n`) =
O(n`(1 + n` )), as claimed.
4.2. Space requirements. The optimized variant of the algorithm Paths() discussed in the previous section requires space O(n), e.g., to store the counting vectors c00 , c01 , c10 , c11 . The functions
IsFlipTree1 () and IsFlipTree2 () require storing only a constant number of ordered rooted trees
with n edges (=bitstrings of length 2n) for lexicographic comparisons, so they also require space
O(n). Furthermore, it is readily checked that the additional space required by each of the calling
functions up to the top-level algorithm HamCycle() is only O(n). This proves that the total space
required by our algorithm is O(n).
5. Acknowledgements
We thank Günter Rote for persistently raising the question whether the proof of Theorem 1 could
be turned into an efficient algorithm. We also thank the referees of this paper and of the extended
abstract that appeared in the proceedings of the European Symposium on Algorithms (ESA) 2015
27
for numerous valuable comments that helped improving the presentation of this work and the C++
code.
References
[BER76]
J. Bitner, G. Ehrlich, and E. Reingold. Efficient generation of the binary reflected Gray code and its
applications. Comm. ACM, 19(9):517–521, 1976.
[BW84]
M. Buck and D. Wiedemann. Gray codes with restricted density. Discrete Math., 48(2-3):163–171, 1984.
[Che08]
Y. Chen. The Chung–Feller theorem revisited. Discrete Math., 308(7):1328–1329, 2008.
[Cum66] R. Cummins. Hamilton circuits in tree graphs. Circuit Theory, IEEE Transactions on, 13(1):82–90, Mar
1966.
[DG12]
P. Diaconis and R. Graham. Magical mathematics. Princeton University Press, Princeton, NJ, 2012. The
mathematical ideas that animate great magic tricks, With a foreword by Martin Gardner.
[DKS94] D. Duffus, H. Kierstead, and H. Snevily. An explicit 1-factorization in the middle of the Boolean lattice.
J. Combin. Theory Ser. A, 65(2):334–342, 1994.
[DSW88] D. Duffus, B. Sands, and R. Woodrow. Lexicographic matchings cannot form Hamiltonian cycles. Order,
5(2):149–161, 1988.
[Ehr73]
G. Ehrlich. Loopless algorithms for generating permutations, combinations, and other combinatorial configurations. J. Assoc. Comput. Mach., 20:500–513, 1973.
[EHR84] P. Eades, M. Hickey, and R. Read. Some Hamilton paths and a minimal change algorithm. J. ACM,
31(1):19–29, January 1984.
[EM84]
P. Eades and B. McKay. An algorithm for generating subsets of fixed size with a strong minimal change
property. Information Processing Letters, 19(3):131–133, 1984.
[FT95]
S. Felsner and W. Trotter. Colorings of diagrams of interval orders and α-sequences of sets. Discrete Math.,
144(1-3):23–31, 1995. Combinatorics of ordered sets (Oberwolfach, 1991).
[Gra53]
F. Gray. Pulse code communication, 1953. March 17, (filed Nov. 1947). U.S. Patent 2,632,058.
[GŠ10]
P. Gregor and R. Škrekovski. On generalized middle-level problem. Inform. Sci., 180(12):2448–2457, 2010.
[Hav83]
I. Havel. Semipaths in directed cubes. In Graphs and other combinatorial topics (Prague, 1982), volume 59
of Teubner-Texte Math., pages 101–108. Teubner, Leipzig, 1983.
[HH72]
C. Holzmann and F. Harary. On the tree graph of a matroid. SIAM Journal on Applied Mathematics,
22(2):187–193, 1972.
[HKRR05] P. Horák, T. Kaiser, M. Rosenfeld, and Z. Ryjácek. The prism over the middle-levels graph is Hamiltonian.
Order, 22(1):73–81, 2005.
[Joh63]
S. Johnson. Generation of permutations by adjacent transposition. Math. Comp., 17:282–285, 1963.
[Joh04]
R. Johnson. Long cycles in the middle two layers of the discrete cube. J. Combin. Theory Ser. A,
105(2):255–271, 2004.
[Knu11]
D. Knuth. The Art of Computer Programming, Volume 4A. Addison-Wesley, 2011.
[KT88]
H. Kierstead and W. Trotter. Explicit matchings in the middle levels of the Boolean lattice. Order,
5(2):163–171, 1988.
[Luc87]
J. Lucas. The rotation graph of binary trees is Hamiltonian. Journal of Algorithms, 8(4):503–535, 1987.
[LvBR93] J. Lucas, D. R. van Baronaigien, and F. Ruskey. On rotations and the generation of binary trees. Journal
of Algorithms, 15(3):343–366, 1993.
[MN17]
T. Mütze and J. Nummenpalo. A constant-time algorithm for middle levels Gray codes. In Proceedings
of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona,
Spain, Hotel Porta Fira, January 16-19, pages 2238–2253, 2017.
[Müt16]
T. Mütze. Proof of the middle levels conjecture. Proc. London Math. Soc., 112(4):677–713, 2016. All
references are to the unabridged version at arXiv:1404.4442.
[MW12]
T. Mütze and F. Weber. Construction of 2-factors in the middle layer of the discrete cube. J. Combin.
Theory Ser. A, 119(8):1832–1855, 2012.
[Rus88]
F. Ruskey. Adjacent interchange generation of combinations. Journal of Algorithms, 9(2):162–180, 1988.
[SA11]
M. Shimada and K. Amano. A note on the middle levels conjecture. arXiv:0912.4564, September 2011.
[Sav93]
C. Savage. Long cycles in the middle two levels of the Boolean lattice. Ars Combin., 35-A:97–108, 1993.
[Sav97]
C. Savage. A survey of combinatorial Gray codes. SIAM Rev., 39(4):605–629, 1997.
[Sed77]
R. Sedgewick. Permutation generation methods. ACM Comput. Surv., 9(2):137–164, June 1977.
[SS99]
I. Shields and C. Savage. A Hamilton path heuristic with applications to the middle two levels problem.
In Proceedings of the Thirtieth Southeastern International Conference on Combinatorics, Graph Theory,
and Computing (Boca Raton, Florida, 1999), volume 140, pages 161–178, 1999.
[SSS09]
I. Shields, B. Shields, and C. Savage. An update on the middle levels problem. Discrete Math.,
309(17):5271–5277, 2009.
28
[Sta99]
[SW95]
[Tro62]
[www]
R. Stanley. Enumerative combinatorics. Vol. 2, volume 62 of Cambridge Studies in Advanced Mathematics.
Cambridge University Press, Cambridge, 1999. With a foreword by Gian-Carlo Rota and appendix 1 by
Sergey Fomin.
C. Savage and P. Winkler. Monotone Gray codes and the middle levels problem. J. Combin. Theory Ser.
A, 70(2):230–248, 1995.
H. Trotter. Algorithm 115: Perm. Commun. ACM, 5(8):434–435, August 1962.
currently http://www.math.tu-berlin.de/~muetze and http://people.inf.ethz.ch/njerri.
| 8cs.DS
|
1
Collection and Dissemination of Data on
Time-Varying Digraphs
arXiv:1604.07478v3 [cs.SY] 29 Apr 2016
Kevin Topley
Email: [email protected]
Index Terms—data dissemination, broadcast algorithms, routing, random graph, data collection, consensus
Abstract—Given a network of fixed size n and an initial
distribution of data, we derive sufficient connectivity conditions
on a sequence of time-varying digraphs for (a) data collection
and (b) data dissemination, within at most (n−1) iterations. The
former is shown to enable distributed computation of the network
size n, while the latter does not. Knowledge of n subsequently
enables each node to acknowledge the earliest time point at
which they can cease communication, specifically we find the
number of redundant signals can be truncated at the finite time n.
Using a probabilistic approach, we obtain tight upper and lower
bounds for the expected time until the last node obtains the entire
collection of data, in other words complete data dissemination.
Similarly tight upper and lower bounds are also found for the
expected time until the first node obtains the entire collection of
data. Interestingly, these bounds are both Θ(log2 (n)) and in fact
differ by only two iterations. Numerical results are explored and
verify each result.
I. I NTRODUCTION
For dynamic communication networks, there may be a given
routing algorithm that is optimal for a specific dissemination
problem and particular mobility model of the network. Any
such routing algorithm is described first by its design, and then
is tested via simulations that employ the given assumptions
of the network dynamics. When mobility constraints are
“coarsely-grained”, that is, occur on a slow time-scale, then the
network is relatively stable and thus certain routing protocols
will allow for efficient search and usage of paths between the
source and destination [8]. Unfortunately, this process is likely
to lead only slowly to insights into the fundamental principles
underlying mobile computing [17]. We are thus interested to
what extent a high level of mobility can challenge the ability
of a network to collect and disseminate data.
Distributed algorithms for data collection and dissemination
on static graphs mainly focus on the issue of locality, in
other words they operate based on the direct neighborhood
of each node [16], [19]. In that context, a natural question
that may arise is what exactly can be computed locally
given limited node resources and/or number of rounds of
communication [14]. Similarly, if a network is highly mobile,
then the information being transmitted or received by any node
is also restricted to its local, current neighborhood, particularly
because the graph might change too quickly to gather any more
information. In this sense, we are not so much interested in
the message complexity of the task, but rather of its certainty
to be solvable in the first place.
We will consider directed graphs that are time-varying and
also, to varying degrees, unconstrained in regard to network
connectivity. In specific terms, we allow the network digraph
to change at each iteration, presuming the set of messages sent
at the previous iteration have been received. This in essence
is synonymous with synchronous communication; generalizations to analogous asynchronous models follow without any
greater insight of our results. To put it another way, we
make no assumptions about how long it takes for a message
to transfer or how long a path remains stable, only that
the network graph locally cannot change faster than it takes
for a single message to be delivered. This has been termed
“fine-grained” mobility [17], which implies that we let the
mobility of the nodes be fast enough such that there is no
guarantee for a node to send a query to its neighbor and
wait for an answer, and slow enough such that when a new
neighbor arrives, it will be able to receive a message from
its new neighborhood. We do not require that any node is
able to detect when its neighborhood changes. Despite this,
we are able to solve the iconic “two-army problem”, but only
when assuming the disconnectivity conditions of the network
are sufficiently constrained. In our deterministic approach we
consider varying degrees of network disconnectivity; in our
probabilistic approach we assume that the digraph is connected
at all times, which, due to our deterministic results, is sufficient
for the distributed algorithm to surely succeed after (n − 1)
iterations.
As a distributed algorithm, we focus mainly on flooding
and the specific case of flooding referred to as routing. We
consider routing to be the basic problem of a network where
information needs to be transported from all nodes to a unique
sink node, which concatenates the data. Conversely, routing
may require a unique source node to disseminate its local data
to all other nodes in the network. As the network changes,
any particular routing algorithm can become redundant and
ineffective as a method of data collection or dissemination,
therefore flooding is an effective remedy to the routing algorithm insofar as it is able to find a path to a set of possibly
numerous destinations. Flooding is also of interest when all
nodes wish to distribute their local information to the rest of
the entire network. This task results in a correct consensus of
all nodes regardless of the particular distributed problem that
is to be solved (i.e., [1], [15], [18], [24]). We obtain non-trivial
connectivity conditions on when this can be achieved if every
node simply broadcasts its set of collected messages, once
received, to its neighbors at all times. Furthermore, since each
node has limited resources at its disposal, we provide a time
point of termination at which all nodes can cease transmitting
data; in the deterministic case this time point is less than (n),
in the probabilistic case we can expect the termination point
2
to be earlier than 2 log2 (n) .
A. Related Work
A large number of specific protocols have been proposed
for efficient routing in mobile networks [9]. These algorithms
are based on the available topological knowledge of the
network communication graph, or sometimes on the relative
coordinates of each node. For the latter set of “geometric”
routing algorithms, it has been shown that greedy (i.e., local)
forwarding strategies may lead into dead-ends, while the optimal delivery strategy has only been guaranteed for the static
case [5] because it involves a preprocessing stage (see also
[2], [7], [10], [12], [23] for further developments in this area).
Considering only topology-based routing algorithms, reactive
protocols similar to DSR [6] and AODV [21] make sense in a
highly mobile environment. An alternative approach FRESH
[3] takes mobility into account, but, arguably for the worse, it
views mobility as a resource rather than a handicap. There also
exist hybrid protocols, such as IZR [22]. Nevertheless, none of
these protocols take into account both a potential topological
change at every single iteration, and simultaneously a worstcase perspective.
For a sequence of time-varying digraphs, the only thing
each node can know with certainty about the graph is the set
of incoming messages from their direct neighbors at each time
step. Thus we can claim that algorithms designed for dynamic
graphs are related to local graph algorithms with respect to
the complexity of broadcasts in the network and the limited
message sizes/knowledge sets at each node [20]. If nodes do
not know their neighborhoods, then any broadcast algorithm
needs messages in the order of the number of edges in the
graph (counting one message per edge). Various such other
lower bounds are given in [4]. However, these apply only to
static graphs; in dynamic graphs, the first question to address is
that of solvability. We will answer this problem via network
connectivity conditions which provably cannot be improved
upon.
B. Outline
The model of network mobility and the set of algorithmic
goals are introduced in Sec.II. In Sec.II-A we define the
“collection problem”, “dissemination problem”, “knowledge
problem”, and “termination problem”. The main results in
the deterministic case are presented in Sec.III, which includes our particular solutions for each of the aforementioned
problems. These solutions are the basic tools which we then
apply to our results concerning a probabilistic approach of
data collection and dissemination, which are presented in
Sec.IV . Numerical simulations illustrated in Sec.V verify our
probabilistic analysis and results. These simulations confirm
the empirical tightness of our results in terms of the expected
time until complete data dissemination and data collection.
The conclusions of our results are summarized in Sec.V I,
and some avenues of potential future work are discussed
in Sec.VI-A. The proofs of all results are contained in the
Appendix (Sec. V II).
II. M ODEL A SSUMPTIONS AND A LGORITHMIC G OALS
Consider a network V of n nodes, V = {1, 2, . . . , n},
where each node i ∈ V has initial data di ∈ R. Internode communication is assumed to be instantaneous and
occur at discrete times k ∈ N. At each time instant k, the
network communication is defined by a time-varying digraph
G(k) = {V, E(k)}, with vertex set V, and edge set E(k)
consisting of ordered pairs of vertices. Node i transmits a
message to node j at time k if and only if (i, j) ∈ E(k).
Define the “knowledge state” of node i as Ki (k) ⊇ {di } at
time k. At k = 0 initialize Ki (0) = {di } for all i ∈ V, and
assume di 6= dj ∀ i 6= j. If node i transmits a signal to node
j at time k, we assume that all data held at node i is then
contained at node j by time (k + 1),
[
(i, j) ∈ E(k) ⇔ Kj (k + 1) = Kj (k) Ki (k)
(1)
The update (1) is tantamount to the “flooding” algorithm
considered in [1], [11], [13], [24], or “broadcasting” as it is
referred to in, for example, [2], [17]. Implicitly (1) states
that the local processing time of any signal is negligible,
and generally speaking the graph cannot locally change faster
than it takes for a message to be transmitted. This type of
distributed protocol is not only ostensibly tangible, easily
extended to asynchronous communication models, but also has
been motivated in the past, for instance, by the shared medium
of wireless networks. Note that we do not require any node to
be able to learn their actual neighbors at any time, nor do they
even require to acknowledge a change in their neighborhood.
A. Algorithmic Goals
We say that node j has obtained a “full collection state”
(FCS) at time k if Kj (k) = {di : i ∈ V}. Conversely,
if V = {i ∈ V : Ki (k) ⊇ {dj }}, we say that node j
has obtained a “full disseminated state” (FDS) at time k. It
follows that if at time k all nodes j ∈ V have obtained either a
FCS or FDS, then the network has obtained a “full knowledge
state” (FKS) at time k. We refer the first two of these network
connectivity problems respectively as the “collection problem”
and “dissemination problem”.
Definition II.1. Collection Problem (CP): the data {di } of all
nodes i ∈ V must be obtained by a specific node q ∈ V.
Definition II.2. Dissemination Problem (DP): the data {dw }
of a specific node w ∈ V must be obtained by all nodes i ∈ V.
The third problem, which is clearly the composite of CP
and DP, is referred to as the “knowledge problem”.
Definition II.3. Knowledge Problem (KP): the data {dw } of
all nodes w ∈ V must be obtained by all nodes i ∈ V.
Note that, due to (1), the solution to the above 3 problems
can be parameterized entirely by network connectivity conditions. This is only one advantage of the flooding algorithm
(1). Another advantage is that, once a node i ∈ V reaches
FCS, it can compute any function of the initial set of data.
The FCS condition is the only knowledge state that allows
such a privilege. Conversely, if FDS is obtained with respect
3
to a node j ∈ V, than all nodes in the network V can
compute any function of the data dj , and thus the network
will have obtained a consensus on this function (presuming
all nodes know the common function that they are required
to solve). Lastly, if FKS is obtained then we have a network
consensus on any function of the initial data 1 . In terms of
distributed computation, there is no knowledge state that is
superior to FKS; however, we must assume that node storage
and transmission resources are of order (n), or perhaps even
larger, depending on what the initial data is and how efficiently
it be can be encoded [1], [18], [25].
In addition to the data dissemination and collection problems defined above, a fourth condition can be applied to
all 3. That is of achieving a “termination” state, wherein
every node within the network ceases to transmit signals,
thus eliminating the possibility of continuously increasing
redundant communication costs.
Definition II.4. Termination Problem (TP): once the problem
of interest (CP, DP, or KP) is solved, every node in the network
stops transmitting any further messages.
The TP is clearly tantamount to the “two-army problem”,
and our solution requires certain connectivity conditions,
without which the problem stands unresolved. We will show
that the solutions to CP and KP both allow for distributed
computation of the network size n. This in turn allows each
node to know when to cease transmitting signals and thus
solves TP when presuming the respective network connectivity
conditions hold. Conversely, the solution to DP does not permit
a distributed computation of the network size, and thus without
a priori knowledge of the value of n (or an upper bound for
n), the connectivity conditions that solve DP will not solve
TP, thus no node will know when DP is solved, and hence no
node will cease transmitting signals.
III. M AIN R ESULTS : D ETERMINISTIC C OMMUNICATION
In this section we present network connectivity conditions
that respectively solve CP, DP, KP, and TP (cf. Def.II.1−II.4)
within the finite time (n − 1). The connectivity conditions
cannot be weakened without increasing the upper bound of
(n − 1), thus they are not only valid solutions, but also
the least restrictive. In Sec.IV we consider sequences of
random digraphs, each of which is assumed to be connected,
thus satisfying all the conditions that solve each of the 4
aforementioned problems. In Sec.IV we prove that the upper
bound (n − 1) reduces in expectation to a function of n
that is less than 2 log2 (n) . Nonetheless, in the Appendix
it shown by example that the upper bound (n − 1) cannot
be improved upon, even for sequences of connected random
digraphs. We now proceed by introducing various types of
connectivity conditions that will be used to present our main
results.
1 Again we presume all nodes know the common function that they are
required to solve; however, if the initial data itself is able to convey the
desired common function, this problem does not require any common a priori
knowledge within the network, rather it is simply a “meta-problem” that can
solved in an identical fashion to the 3 problems already defined.
A. Network Connectivity Definitions: CP, KP, TP
Let V−i = V \ {i} for any node i ∈ V. At any time k,
define an “input-cord” to node i as an ordered set of nodes
I i (k) ⊆ V−i with the following properties,
(a) Iji (k) 6= Iri (k) , ∀ r 6= j
i
(k)) ∈ E(k) ,
(b) (Iji (k), Ij+1
∀ j ∈ {1, 2, . . . , |I i (k)| − 1}
i
(c) (I|I
i (k)| (k), i) ∈ E(k)
(2)
where Iji (k) denotes the j th entry in I i (k), and |I i (k)| is the
cardinality of I i (k). An example of an input-cord to node 6
is illustrated in Fig.1.
Figure 1.
An input-cord to node 6, I 6 (5) = {7, 9, 2}.
We say the input-cord I i (k) is “closed” if (i, I1i (k)) ∈ E(k).
The closure of I i (k), denoted Î i (k), is illustrated in Fig.2.
Figure 2.
A closed input-cord to node 6, Î 6 (2) = {7, 9, 2}.
Our first result provides sufficient connectivity conditions
to solve CP (cf. Def.II.1).
Lemma III.1. Given the update (1), if the sequence of inputcords to node q satisfies,
|I q (k)| ≥ (n − k − 1) , ∀ k ∈ {0, 1, . . . , n − 2}
(3)
then |Kq (k + 1)| ≥ (k + 2).
Given (3), the Lem.III.1 implies that after (n−1) iterations
the node q will know all data in the network, that is Kq (n −
1) = {dj : j ∈ V}, thus solving CP (cf. Def.II.1).
However, the Lem.III.1 states a stronger property of Kq (k)
than simply Kq (n − 1) = {dj : j ∈ V}. We will
utilize Lem.III.1 more completely to illustrate how the update
protocol (1) permits distributive computation of the network
size n. To do so, we require the following definition.
Definition III.2. χ(k)-cycle: if node i has a closed inputcord Î i (k) with cardinality greater than (χ − 2), then node i
is contained in a χ(k)-cycle. The graph G(k) contains χ(k)
if all nodes i ∈ V are contained in a χ(k)-cycle.
An illustration of a graph that contains 3(4) is illustrated in
Fig.3.
Given the above definition, we now apply Lem.III.1 to
all nodes i ∈ V, rather than just to node q, and obtain the
following solution to KP (cf. Def.II.3).
4
where Oji (k) denotes the j th entry in Oi (k), and |Oi (k)| is
the cardinality of Oi (k). An example of an output-cord from
node 6 is illustrated in Fig.4.
Figure 3. A graph G(4) that contains 3(4). Note that each node i ∈ V has
a closed input-cord with size greater than 1.
Theorem III.3. Define ψ(k) as follows,
(
n
if k ≤ dn/2e − 1
ψ(k) =
n − k if k ≥ dn/2e
An output-cord from node 6, O6 (3) = {7, 9, 2}.
Our next result provides sufficient connectivity conditions
to solve DP (cf. Def.II.2).
(4)
Lemma III.5. Given the update (1), if the sequence of ouputcords from node w satisfies,
|Ow (k)| ≥ (k + 1) , ∀ k ∈ {0, 1, . . . , n − 2}
Given the update (1), if ψ(k) ∈ G(k) (cf. Def.III.2) for k ∈
{0, 1, . . . , n − 2} then |Ki (k + 1)| ≥ (k + 2) for all i ∈ V.
The next corollary shows that Thm.III.3 allows for distributed computation of the network size n, and thus solves
TP (cf. Def.II.4).
Corollary III.4. Given the update (1), if ψ(k) ∈ G(k) (cf.
(4)) for k ∈ {0, 1, . . . , n − 2} then at time k = n all nodes
i ∈ V will know the size of the network |Ki (n)| = k = |V| =
n.
The connectivity condition defined in Thm.III.3 thus not
only guarantees FKS by time (n − 1), that is Ki (n − 1) =
{dj : j ∈ V} for all i ∈ V, but also permits distributed
computation of the network size 2 .
We now move on to the dissemination problem (cf.
Def.II.2). Although the connectivity conditions that solve
DP are time-symmetric to the conditions that solve CP,
when they are applied to the entire network the distributed
nature of the dissemination problem does not permit computation of the network size. However, when combining
the solution to CP (Lem.III.1) and the solution to DP
presented next (Lem.III.5), we obtain a solution to KP
that assumes weaker connectivity conditions than Thm.III.3.
However the Cor.III.4 cannot be similarly improved upon
(see Conj.V II.5).
B. Network Connectivity Definitions: DP, KP, TP
Analogous to the notion of an “input-cord”, we next define
an “output-cord” from node i as an ordered set of nodes
Oi (k) ⊆ V−i with the following properties,
(a) Oji (k) 6= Ori (k) , ∀ r 6= j
i
(b) (Oj+1
(k), Oji (k)) ∈ E(k) ,
∀ j ∈ {1, 2, . . . , |Oi (k)| − 1}
i
(c) (i, O|O
i (k)| (k)) ∈ E(k)
Figure 4.
(5)
2 Note that, although the connectivity condition in Thm.III.3 depends
functionally on the network size n, it is not required that any node specifically
knows the network size.
(6)
then |{Ki (k + 1) ⊇ dw : i ∈ V}| ≥ (k + 2).
Given (6), the Lem.III.5 implies that after (n−1) iterations
all nodes in the network will know the data dw , that is
|{Ki (n − 1) ⊇ dw : i ∈ V}| = n, thus solving DP (cf.
Def.II.2). Applying Lem.III.5 to all nodes i ∈ V, rather
than just node w, results in our next theorem, which is an
alternative to Thm.III.3 as a solution to KP (cf. Def.II.3).
Theorem III.6. Define ν(k) as follows,
(
k + 2 if k ≤ dn/2e − 1
ν(k) =
n
if k ≥ dn/2e
(7)
Given the update (1), if ν(k) ∈ G(k) (cf. Def.III.2) for k ∈
{0, 1, . . . , n − 2} then |{Ki (k + 1) ⊇ dw : i ∈ V}| ≥ (k + 2)
for all w ∈ V.
By time (n − 1), note that both Thm.III.3 and Thm.III.6
guarantee FKS, that is, the knowledge set of each node i ∈ V
contains all of the initial data {dj : j ∈ V}. This implies that
(a) the collection problem (cf. Def.II.1) is solved for all nodes
q ∈ V, and (b) the dissemination problem (cf. Def.II.2) is
solved for all nodes w ∈ V. By combining these two theorems,
we obtain a significantly weaker sufficient condition for FKS
by time (n − 1). The drawback of this result, similar to that of
Thm.III.6, is that it does not permit distributed computation
of the network size (in the Appendix this is conjectured, as
we foresee no way in which it can or cannot be proven).
Theorem III.7. Let η(k) = min{ψ(k), ν(k)}. Given the update (1),P
if η(k) ∈ G(k) (cf. Def.III.2) for k ∈ {0, 1, . . . , n−
n
2} then i=1 |Ki (n − 1)| = n2 .
The connectivity conditions for Thm.III.3, III.6 and
III.7 are plotted in Fig.5. Note that η(k) is significantly lower
than ψ(k) for k ∈ {0, 1, . . . , (n/2) − 2}, and significantly
lower than ν(k) for k ∈ {n/2, (n/2) + 1, . . . , n − 2}. At
k = (n/2)−1 Thm.III.7 requires n(k) ∈ G(k), thus all nodes
are contained in each other node’s input-cord. This condition,
known as a “connected” network graph, is illustrated in Fig.6.
5
γ − 1 + φ(n) k̂(n) to both hold without contradiction. The
following theorem places a tight Θ(log(n)) bound on k̂(n).
Theorem IV.1. Given the update (1) and a sequence of (n−1)
random graphs R(n − 1), the solution to (11) is γ̂ = 2.
Figure 5. The χ(k)-cycle conditions for Thm.III.3, III.6, III.7, plotted
respectively as ψ(k), ν(k), η(k).
Figure 6.
A connected digraph, n(2) = 6(2) ∈ G(2).
IV. M AIN R ESULTS : P ROBABILISTIC C OMMUNICATION
For a fixed network size n, define the set of permutations
of V = {1, 2, . . . , n} as p(V). The set p(V) has cardinality n!.
Denote the `th element of p(V) as p` (V), and j th element of
p` (V), as p`j (V). At time k let r(k) be a random integer from
the set {1, 2, . . . , n!}. We define a random connected graph
on n vertices as a digraph G̃(k) = {V, Ẽ(k)} with edge set
Ẽ(k),
r(k)
r(k)
(pj (V), pj+1 (V))
r(k)
r(k)
(pn (V), p1 (V))
∈ Ẽ(k) , ∀ j ∈ V−n
∈ Ẽ(k) .
(8)
Note that n(k) ∈ G̃(k) (cf. Def.III.2).
Let R(k) equal a sequence of k random graphs, R(k) =
{G̃(0), G̃(1), . . . , G̃(k − 1)}. If we define E(·) as the expectation operator, then the expected time k̂(n) at which the
knowledge set of all nodes equals {di : i ∈ V} is,
n
n
o
X
k̂(n) = min k ∈ R : E
|Ki (k)| = n2
(9)
i=1
This is the expected time at which a network with n nodes
will reach FKS.
Define φ(n) as,
log(n − 2)
φ(n) = −log 1 −
,
(10)
log(n)
assuming all logarithms are base 2. We will upper and lower
bound k̂(n) by a parameter γ̂, defined as,
n
o
γ̂ = min γ ∈ N : γ + φ(n) ≤ k̂(n) .
(11)
Note that the condition γ+φ(n) ≤ k̂(n) states that if k ≥ k̂(n)
then it is expected that at time k the network has reached
FKS. In this sense, it is possible for γ + φ(n) ≤ k̂(n) and
Note that the if the parameter γ̂ cannot be any smaller, than it
bounds k̂(n) tightly from both below and above (i.e., within
a single iteration). This is because network communication
can only occur at integer time instances k ∈ N. We show in
Lem.V II.1 that γ̂ cannot be reduced.
Thm.III.3, III.6, III.7
all
guarantee
that
PThe
n
2
|K
(n
−
1)|
=
n
for
any
sequence
of
(n
−
1)
random
i
i=1
graphs R(n − 1). The above result shows that, in expectation,
the network will reach FKS within the exponentially smaller
times log(n), log(n2 ) , see Prop.V II.3 in the Appendix.
The value k̂(n) defines the expected time at which the
last node in the network V reaches FCS, thus implying the
entire network has reached FKS. We next consider ǩ(n), which
will be defined as the expected time at which the first node
reaches FCS. Just as Thm.IV.1 can be viewed in regard to
Thm.III.3, III.6, III.7, the theorem below can be viewed in
regard to Lem.III.1, III.5.
Similar to (9), let us define ǩ(n) as follows,
n
o
ǩ(n) = mini∈V k ∈ R : E |Ki (k)| = n
(12)
The value ǩ(n) is thus the expected time at which the first
node in V will obtain FCS 3 . Similar to k̂(n), we will upper
and lower bound ǩ(n) by a parameter γ̌, defined as,
n
o
γ̌ = min γ ∈ N : γ + φ(n) ≤ ǩ(n) .
(13)
The following theorem places a tight Θ(log(n)) bound on
ǩ(n).
Theorem IV.2. Given the update (1) and a sequence of (n−1)
random graphs R(n − 1), the solution to (13) is γ̌ = 0.
Again, we emphasize that because network communication
occurs at integer time instances k ∈ N, if the parameter γ̌
can be proven as the absolute minimum such value, then it
bounds ǩ(n) tightly from both below and above (i.e., within
a single iteration). This is indeed proven within the proof of
Thm.IV.2, which is presented in the Appendix (Sec.V II).
Although Thm.IV.1 implies that the expected time of FKS
is upper bounded by 3 + φ(n) < log(2n2 ) for any random
sequence of connected graphs, this does not diminish the
significance of Thm.III.3, III.6, III.7. In the Appendix, by
way of Example V II.4, we show that there are sequences of
connected graphs that imply |Ki (k)| = k + 1 for all i ∈ V and
k ∈ {0, 1, . . . , n − 1}, thus yielding FKS only after (n − 1)
iterations, just as Thm.III.3, III.6, III.7 guarantee. Hence,
without changing the required connectivity conditions, the
deterministic bounds stated in Thm.III.3, III.6, III.7 cannot
be improved upon.
3 For convenience, we will, with abuse of notation, label the “first node to
obtain FCS”, as the “earliest time of FKS”. This is analogous to how the “last
node to obtain FCS” implies (with no abuse of notation) the “latest time of
FKS”.
6
Likewise, the Thm.IV.2 implies that the expected time of
full data dissemination (DP) (cf. Def.II.2), or equivalently the
time at which the first node reaches FCS, is upper bounded
by φ(n) < log(n2 /4), see Prop.V II.3 in the Appendix. In
comparison, the Lem.III.1, III.5 imply that for any random sequence of connected graphs, the maximum time of
complete data collection (resp. full data dissemination) is the
exponentially larger time (n − 1). However, this does not
necessarily diminish the results of Lem.III.1, III.5, since
the Appendix contains the Example V II.4 that proves that
there are sequences of connected graphs for which CP (cf.
Def.II.1) and DP (cf. Def.II.2) are solved only at the time
(n − 1). It follows then, that without changing the required
connectivity conditions, the deterministic upper bounds stated
in Lem.III.1, III.5 cannot be improved upon.
V. N UMERICAL S IMULATIONS
It is not hard to show that dlog(n)e is the earliest time for
FKS given any sequence of random graphs (see Thm.V II.2
in the Appendix). This implies that k̂(n)
≥ dlog(n)e. In
for network sizes
Fig.7 we plot the bounds φ(n) + 5±1
2
n ∈ {3, 4, . . . , 150}. On the same graph we plot the time at
which FKS occurs when averaged over 1000 simulations of
random graph sequences R(n − 1). It is clear that the lower
and upper bounds derived in Thm.IV.1 are a very close approximation to the empirical average time of FKS, particularly
they both remain with one iteration of the empirical average.
Analogously, we plot in Fig.7 the lower and upper bounds
φ(n) and (1+φ(n)) to the earliest time of FKS, which is also
plotted on the graph when averaged over 1000 simulations.
Again, it is clear that the empirical average time to earliest
FKS remains within one iteration of the bounds derived in
Thm.IV.2.
Figure 7. Empirical time until Earliest FKS and Latest FKS. Network size
∈ [3, 150]. Number of simulations = 1000.
for
In Fig.8 we plot the Thm.IV.1 bounds φ(n) + 5±1
2
network sizes n ∈ {5, 4, . . . , 500}. On the same graph we
plot the time at which FKS occurs when averaged over 20
simulations of random graph sequences R(n − 1). Similar to
Fig.7, we also plot the bounds of Thm.IV.2 and the empirical
average of the earliest time of FKS. The results are identical
to Fig.7, despite the fact that we average the times of FKS
over 50 times less than Fig.7, and increased the total network
size by a factor of 4/3.
Figure 8. Empirical time until Earliest FKS and Latest FKS. Network size
∈ [5, 500]. Number of simulations = 20.
An interesting phenomenon can be seen in both Fig.7 − 8,
which is the wave-like property of the empirical averaged FKS
times. The simulation conditions do not appear to account for
this unexpected result, and we can only leave its explanation
as potential future work.
VI. C ONCLUSION
Deterministic connectivity conditions were derived for
both data dissemination and data collection on time-varying
digraphs. The conditions assumed a “broadcast” protocol
wherein each node sent the entirety of its current data to all
nodes within its communication range, which was allowed to
vary among nodes and thus requiring the use of digraphs,
rather than the far stricter condition of symmetric communication channels. The deterministic conditions were proven
using a matrix representation of each communication graph,
and non-trivial manipulations thereof. Conditions were given
for the ubiquitous termination of all network communication
once the particular distribution problem of interest had been
solved. The termination problem is tantamount to the “2-army
problem”, and, to our knowledge, the conditions given are a
unique solution to it.
A probabilistic approach was introduced wherein at each
iteration an identically uniform and independently chosen random connected digraph was used to model the communication
links for that iteration. Lower and upper bounds were obtained
in regard to the expected times of termination for each of
the various data distribution problems that were addressed in
the deterministic setting. The probabilistic bounds were shown
to be exponentially smaller than the upper bounds given by
the deterministic results. Nonetheless, examples were given to
demonstrate how the deterministic results cannot be improved
upon without strengthening the connectivity constraints.
Empirical results confirmed the tightness of the probabilistic
bounds, which remained both analytically and empirically
within one iteration of the average time to termination, and
thus cannot be improved upon since we have assumed communication occurs at discrete time iterations. A curious wavelike property was observed in the empirical average time until
termination; an explanation of this remains an open question.
A. Future Work
Our first concern is regarding the two solutions to KP
which do not also solve TP. We predict there may be various
7
connectivity conditions that when applied to the iterations after
the time (n − 1), would lead a termination point within at
most the next (n−1) iterations. These conditions would likely
depend on the two different sets of conditions assumed in
Thm.III.6 and Thm.III.7. Furthermore, these hypothetical
conditions would not improve upon the time until FKS, which
is already guaranteed by these theorems to occur by time
(n − 1), but would only yield the benefit of solving TP. In
regard to issue of connectivity conditions beyond the time
(n − 1), we may consider the case when the initial set of data
changes with time, in which case none of our results could
be directly applied, but rather they would need modifications
contingent on the particular dynamics of the local data.
Another source of insight towards efficient data dissemination is that of ordering or classifying different types of
network dynamics in terms of the associated redundancy
of communications, expediency of data proliferation, and
constraints on connectivity. For instance, the sequence of
communication graphs presented in Thm.V II.2 yield FKS in
minimal time and with minimal redundancy, at the cost of
assuming very strict connectivity conditions. Conversely, the
Example V II.4 presents a simple, perhaps overly structured
but yet easily implemented sequence of graphs (all of which
are identical), that yield FKS in maximum time and with very
large (perhaps maximum) communication redundancy. Our
probabilistic model can be seen as a mid-point between these
two deterministic extremes, as it allows for an independent
and identically distributed type of random connected graph
at each iteration. The speed of convergence to FKS using
this probabilistic model has been quantified and numerically
verified in the present work, however the (average) redundancy
of communication associated with this model has not been
addressed. Quantifying how connectivity conditions relate to
communication redundancy would seem to be of worthy
consequence, since node storage and communication resources
are limited, if not in actuality than at least intuitively, and in
some cases expediency to FKS may be trumped by network
resources. Even defining an appropriate metric for communication redundancy, and more so for connectivity constraints,
remain debatable.
Lastly, it is conceivable that the sequence of connected
graphs may not be reducable to a uniformly random element
of the set of permutations p(V). Given a different distribution
and sample space from which the randomly connected graphs
are chosen, the expected times of FKS (resp. FCS, FDS) will
certainly vary. It appears to be a considerable task to obtain
expectation bounds for FKS when assuming a more general
model for the sequence of random communication graphs.
VII. A PPENDIX
Lemma III.1 Proof. The proof of Lem.III.1 is most easily
illustrated by appealing to the proof of Lem.III.5. The
communication conditions assumed in Lem.III.5 and therein
proven to solve DP can be stated as follows in terms of a
sequence of edge sets:
E(0) ⊇ {(1, 20 )}
E(1) ⊇ {(1, 21 ), (21 , 31 )}
..
.
E(k) ⊇ {(1, 2k ), . . . ,
(k + 1)k , (k + 2)k }
..
.
(14)
E(n − 2) ⊇ {(1, 2n−2 ), . . . ,
(n − 1)n−2 , nn−2 }
The set of nodes {1, 2, . . . , n} that comprise the network V
is assumed to be invariant with time, thus the sequence of
edge sets E(k) : k ∈ {0, 1, . . . , n − 2} fully
defines
the sequence of network communication graphs G(k) :
k ∈ {0, 1, . . . , n − 2} . Further note that each communication graph G(k) can be defined by a non-negative (n × n)
matrix M (k) = [Mij (k)] with positive entries for each pair
(i, j) ∈ E(k),
Mij (k) > 0 ⇔ (i, j) ∈ E(k) .
(15)
The update (1) implies that the cardinality of any knowledge
set Ki (k) will never decrease, thus we let Mii (k) > 0 for all
i ∈ V.
. Qk
Define M(k) = r=0 M (r). In accordance with the update
(1), the knowledge set of node i will contain data dj at time
(k + 1) if and only if Mji (k) > 0,
Ki (k + 1) ⊇ {dj } ⇔ Mji (k) > 0 .
(16)
The Lem.III.5 implies that if (14) holds, then M1j (n − 2) >
0 for all j ∈ V (note that in Lem.III.5 we define without loss
of generality (WLG) w = 1).
To summarize the
above: the Lem.III.5 places conditions
on the sequence E(k) : k ∈ {0, 1, . . . , n − 2} that in
turn define the matrices M (k) : k ∈ {0, 1, . . . , n − 2} ,
which, when multiplied together, yield M1j (n−2) > 0 for all
j ∈ V. It thus follows that by transposing
the product M(k)
we obtain a sequence of matrices M 0 (k) = M (n − k − 2) :
k ∈ {0, 1, . . . , n − 2} that in turn define a sequence of edge
sets and thus communication graphs.
The CP requires that a single node q ∈ V obtain the data that
is initially held at every other node i ∈ V−q . Let q = n WLG.
In terms of the matrix M(k), the CP is solved when Mjn > 0
for all j ∈ V. This condition is simply the transpose of the
matrix M(n − 2) once DP is solved, that is M1j (n − 2) > 0
for
j ∈ V. Thus to solve CP we need only the condition
all
M 0 (k) = M (n − k − 2) : k ∈ {0, 1, . . . , n − 2} , which,
when put in terms of the sequence of edge sets defined in
(14), is:
8
E(0) ⊇ {(10 , 20 ), (20 , 30 ), . . . ,
((n − 1)0 , n)}
E(1) ⊇ {(21 , 31 ), (31 , 41 ), . . . ,
((n − 1)1 , n)}
..
.
E(k) ⊇ { (k + 1)k , (k + 2)k , . . . ,
((n − 1)k , n)}
..
.
(17)
E(n − 2) ⊇ {((n − 1)n−2 , n)}
The proof of Lem.III.1 is complete by noting that the set of
conditions (17) is identical to the set of conditions (3) defined
in Lem.III.1.
Theorem III.3 Proof. If k ∈ {0, 1, . . . , dn/2e − 1} then
n(k) ∈ G(k) and thus all nodes have input-cords of length
(n − 1). It follows that (3) is satisfied, and thus by Lem.III.1
we have |Ki (k+1)| ≥ (k+2) for all k ∈ {0, 1, . . . , dn/2e−1}
and i ∈ V. If k ∈ {dn/2e, dn/2e + 1, . . . , n − 2} then
(n − k)(k) ∈ G(k) which implies each node has an input-cord
of at least length (n−k−1). This is precisely the condition (3)
required in Lem.III.1, thus implying |Ki (k + 1)| ≥ (k + 2)
for all k ∈ {dn/2e, dn/2e + 1, . . . , n − 2} and i ∈ V.
Corollary III.4 Proof. Given the update protocol (1)
and communication condition ψ(k) ∈ G(k) (cf. (4)), the
Thm.III.3 guarantees that |Ki (k)| > k + 1 for all i ∈ V
and k ∈ {0, 1, . . . , n − 1}. Each node i ∈ V initially holds
the unique data {di }, thus we have |Ki (k)| ≤ n. It then
follows that once |Ki (n)| = n, the node i knows that the
condition |Ki (k)| > k + 1 no longer holds, and thus n must
be the network size. All nodes i ∈ V can thus distributively
compute the network size n by time k = n under the assumed
communication conditions.
Lemma III.5 Proof. Let w = 1 WLG. We will assume
|O1 (k)| = (k + 1) , ∀ k ∈ {0, 1, . . . , n − 2}, since adding
more communication links can only increase the quantity of
interest |{Ki (k) ⊇ d1 : i ∈ V}|, which we intend to show is
lower bounded by (k + 1).
At time k = 0, only node 1 contains the data d1 . At k = 0
we have |O1 (0)| = 1, so let us label the single node contained
in O1 (0) as 20 ∈ V−1 . Now at time k = 1, both nodes 1 and 20
contain the data d1 . At time k = 1 we have |O1 (1)| = 2, so let
us label the two nodes contained in O1 (1) as 21 , 31 ∈ V−1 . If
21 = 20 , then 31 (6= 20 ) will obtain the data d1 from node 21 ,
and otherwise the node 21 (6= 20 ) will obtain the data d1 from
node 1; in both cases we end up at time k = 2 with 3 nodes
that contain d1 . In summary, if |{Ki (k) ⊇ d1 : i ∈ V}| =
(k + 1), all that is required for |{Ki (k + 1) ⊇ d1 : i ∈ V}| to
stay above (k + 2) is for some node i(k) that does not contain
data d1 to take a place in O1 (k). Since |O1 (k)| = (k + 1),
the latter condition necessarily must hold. On the other hand,
if |{Ki (k) ⊇ d1 : i ∈ V}| > (k + 1), then the condition to
be proven already holds, and so, if necessary, we can apply
the previous argument at the subsequent time (k + 1).
Theorem III.6 Proof. If k ∈ {0, 1, . . . , dn/2e − 1} then
(k + 2)(k) ∈ G(k) and thus all nodes have output-cords of
at least length (k + 1). This is precisely the condition (6)
in Lem.III.5, and thus |{Ki (k + 1) ⊇ dw : i ∈ V}| ≥
(k + 2) for all k ∈ {0, 1, . . . , dn/2e − 1} and w ∈ V. If
k ∈ {dn/2e, . . . , n − 2} then (n)(k) ∈ G(k) which implies
each node has an output-cord of length (n − 1). It follows that
(6) is satisfied and thus by Lem.6 we have |{Ki (k + 1) ⊇
dw : i ∈ V}| ≥ (k + 2) for all k ∈ {dn/2e, . . . , n − 2} and
w ∈ V.
Theorem III.7 Proof. For k ∈ {0, 1, . . . , d n2 e − 1} we have
η(k) = k + 2 and thus Lem.III.5 implies |{Ki (k + 1) ⊇ dw :
i ∈ V}| ≥ (k+2) for all k ∈ {0, 1, . . . , dn/2e−1} and
w ∈ V.
By (15) we can define the set of matrices A η(k) ∈ Rn×n
for which η(k) ∈ G(k). By (16) we then have M(d n2 e) =
Qd n2 e−1
A η(r) , which by Lem.III.5 will satisfy,
r=0
n
X
I M dn/2e ij ≥ dn/2e + 1 ,
(18)
j=1
where I(·) is the indicator function defined as,
(
1
if Aij > 0
I(Aij ) =
0
if Aij = 0
for a matrix A = [Aij ].
For k ∈ {d n2 e, . . . , n − 2} we have η(k)
= n − k. By (15)
this defines the set of matrices A η(k) ∈ Rn×n for which
η(k) ∈ G(k), η(k) = n − k and k ∈ {d n2 e, . . . , n − 2}. We
now let,
n−2
Y
M̃(n − 1) =
A η(r) ,
r=d n
2e
and seek to prove,
10n I M(dn/2e) · M̃(n − 1) 1n = n2 ,
(19)
where 1n ∈ Rn×1 is a vector of unit values. Note that (19) is
equivalent to,
(20)
10n I M̃(n − 1)0 · M(dn/2e)0 1n = n2 .
Consider the transpose of M̃(n − 1),
Qd n2 c
M̃(n − 1)0 = r=n−2
A0 η(r)
n
Qd 2 e−1 0
= r=0
A (r + 2)
Qd n2 e−1
= r=0 A(r + 2)
= M(dn/2e) .
(21)
From (21) we obtain,
I M̃(n − 1)0 · M(dn/2e)0 ij =
Pn
.
`=1 I M(dn/2e)i` M(dn/2e)j`
(22)
The condition (18) implies that, for any (i, j) ∈ V 2 , there must
exist at least one value of ` ∈ V for which both M(dn/2e)i`
and M(dn/2e)j` are non-zero, thus implying (20).
Theorem IV.1 Proof. Define P (k) ∈ Rn×n as follows
P (k)ij = P[Ki (k) ⊇ dj ], where P[A] is the probability that
condition A is true. Due to symmetry, the expected value of
any sequence of random graphs will yield at time k,
1 xk · · · xk
xk 1 · · · xk
P (k) = .
..
..
..
..
.
.
.
xk
xk
···
1
9
where xk is the probability that at time k the knowledge set
of node i contains the data dj for any j ∈ V−i ,
To lower bound (25) we upper bound
which yields,
xk = P[Ki (k) ⊇ dj : i ∈ V , j ∈ V−i ] .
Ek > (n − 1) 1 −
Next we let,
xk0 = P Ki (k) + dj : i ∈ V , j ∈ V−i
= 1 − xk ,
thus xk0 equals the probability that at time k node i does not
contain the data dj .
We will denote Ek = E(|Ki (k) \ di |) for any i ∈ V, and
similarly denote Ek0 = E(|Ki (k)|) = Ek + 1. Accordingly, we
initially have E0 = 0 and x00 = 1. The probability xk0 is
recursively defined as,
xk0 = PKi (k) + dj
= P Ki (k − 1) + dj ·
P (`, i) ∈ E(k − 1), K` (k − 1) + dj |
Ki (k −1) + dj
n−1−Ek0
k
= x(k−1)0 n−2−E
,
= x(k−1)0
n−1
n−1
wherefrom the expectation Ek can be defined recursively as,
Ek
= (n − 1)xk
= n − 1 − (n − 1)xk0
=n−1−
Qk−1
r=0 (n−2−Er )
(n−1)k−1
(23)
.
Rearranging (23) and taking logarithms yields,
Ek
log 1 − n−1
= k · log n−2
n−1 +
Pk−1
Er
r=1 log 1 − n−2
from which we obtain,
Ek
Ek
< log 1 − n−1
=
log 1 − n−2
P
k−1
Er
k · log n−2
+
log
1
−
r=1
n−1
n−2
Pk−2
< k + r=1 2r−1 (k − r) · log n−2
n−1
n−3
k−2
+2
log n−2 .
which we now show is lower bounded by,
!
n − 2 k−1
n 1−
−1 .
n
n−2
n−1
χk
<
(n − 3)(n − 2) 2k−2
(n − 1)2
n−2
n−1
2k−1
!
.
(27)
Canceling the equivalent terms in (27) and (26) then rearranging it will suffice to show,
2k−2
(n−1)3
(n−1)2
<
.
(n−2)n
n2 (n−3)
Assuming k ≥ 2 we can proceed,
1
1 + n2 −2n
< 1 + n3n−1
3 −3n2
3
2
∴ n − 3n < 3n3 − 7n2 + 2n
∴ 0 < (n − 1)2 ,
thus verifying the lower bound (26). P
n
0
0
Due again to symmetry, we have
i=1 Ek = nEk . We
0
2
nowPassume nEk > n − 1, from which we can then infer
n
2
E
i=1 |Ki (k)| = n . From (26) we have,
!
n − 2 2k−1
0
.
(28)
Ek > n 1 −
n
We now solve (28) for a sufficiently large k when supposing
Ek0 > n − (1/n),
!
2k−1
n 1 − n−2
≥ n − (1/n) ,
n
n−2
n
2k−1
≤
∴ k ≥ 2 + log
∴
(29)
1
n2 ,
log(n)
log n/(n−2)
.
Note that,
log(n)
= 2φ(n) ,
log n/(n − 2)
(24)
which can be easily verified by (10).
Note that (24) was obtained by iterating (23) backwards until
k = 2. At the point k = 2 we used the fact that E1 = 1, which
can be obtainedP
from (23) and the initial condition E0 = 0.
k−2
Let χk = k+ r=1 2r−1 (k−r). Note that χk = 3·2k−2 −1,
which we now prove by induction. Let k = 3, then χ3 = 5 =
3 · 23−2 − 1. Next assume χk = 3 · 2k−2 − 1, and consider
χk+1 ,
Pk−1
χk+1 = k + 1 + r=1 2r−1 (k − r + 1)
Pk−1
= χk + 1 + 2k−2
+ r=1 2r−1
= 3 · 2k−2 − 1 + 1 + 2k−2 + 2k−1 − 1
= 3 · 2k−1 − 1 .
Solving (24) for Ek yields,
!
n − 3 2k−2 n − 2 χk
Ek > (n − 1) 1 −
n−2
n−1
(25)
(26)
The following lemma will be used in the proof of Thm.IV.2.
Note that this lemma, together with Thm.IV.1, shows that
the bound k̂ cannot be reduced, and thus is tight. In other
words,Pone iteration below
k̂ and we do not expect FKS, that
n
2
is E
|K
(k)|
=
n
> k̂ − 1.
i
i=1
Lemma VII.1. For k̂ = 2 + φ(n) (cf.(10)), Ek̂−1 < (n2 −
n − 1)/n.
Proof:
For
convenience
denote
γφ(n)
=
2φ(n)
3·2φ(n) −1
n−3
n−2
. We denote Êk as the maximum
n−2
n−1
value of Ek . Applying Thm.IV.1 and (25) we have,
Ek̂−1 − Ek̂ < Êk̂−1 − (n − 1)(1 − γφ(n) )
2
< Êk̂−1 − n −n−1
n
∴ Ek̂−1 < Ek̂ + Êk̂−1 − n + 1 + n−1
< Êk̂−1 + n−1 ,
where we have used the fact Ek < n − 1 for all k ∈
{0, 1, . . . , n − 2}, which holds due to Example V II.4 (a fixed
connected digraph). Upper bounding the right-hand side (RHS)
10
of the above by (n − 1 − n−1 ) and solving for Êk̂−1 yields
the condition,
Êk̂−1 < n − 1 − 2n−1 .
−1
Notice that the upper bound (n − 1 − n ) was chosen
without any specific precondition. For this reason, along with
the fact that E` < E`+1 for all ` ∈ {0, 1, . . . , n − 2} (which
can be inferred simply by noting that more signals will have
occurred by time (` + 1) as compared to time `), we must
confirm the following,
Êk̂−1 − Ek̂ < Êk̂−1 − (n − 1)(1 − γφ(n) )
< n − 1 − 2n−1 − (n − 1)(1 − γφ(n) )
= −2n−1 + (n − 1)γφ(n) < 0 .
Rearranging the last inequality and taking logarithms yields,
2
log n(n−1)
2(n−2)
log(n)
>
.
(30)
(n−1)3
n
log n−2
log (n−2)
2 (n−3)
For notational
b = log(n − 2),
convenience let
a =2log(n),
(n−1)3
(n−1)
c = log n−3 , and d = log
. Note that a > b > 0,
2
c > 2b. Rearranging (30) yields,
a
a−b
> a−b+d
c−2b
∴ ac > a2 + b2 + 2 · log(n − 1) − 1 (a − b)
2
∴ c > a + ba + 2 · log(n − 1) − 1
−2(b/a) · log(n − 1) + (b/a) .
Upper bounding the last (b/a) term by 1, replacing a, b and c
with their definitions, and combining terms we obtain,
!
(n − 1)3
(n − 2)2 (n − 1)2
.
log
> log
log(n−2)
n−3
(n − 1)2 log(n)
Canceling the logarithms and rearranging yields,
(n − 1)1+2·
log(n−2)
log(n)
> (n − 3)(n − 2)2 .
Taking logarithms and rearranging then yields,
2 · log(n − 2)log(n − 1) + log(n − 1)log(n)
> log(n − 3)log(n) + 2 · log(n − 2)log(n) .
Concavity of the logarithm function implies log(n−2)/log(n−
3) > log(n)/log(n − 1), thus it remains to be shown,
log(n − 2)log(n − 1) + log(n − 1)log(n)
> 2 · log(n − 2)log(n).
Rearranging this inequality yields the sufficient condition,
log(n − 2) log(n − 1) − log(n)
+log(n) log(n − 1) − log(n − 2) > 0 .
Note that log(n − 1) − log(n) < 0 < log(n) − log(n − 1) <
log(n − 1) − log(n − 2), due to concavity of the logarithm
function. Furthermore, 0 < log(n−2) < log(n), thus verifying
(30).
Theorem IV.2 Proof. If Ek0 > (n − 1) then we can infer that
at least one node i ∈ V has reached FKS at time k, that is
|Ki (k)| = n. We thus seek to prove Ek̂−2 = Eφ(n) > (n − 2).
Applying (25) and Lem.V II.1 we have,
Ek̂−1 − Eφ(n) <
n2 −n−1
n
−
(n − 1) 1−
!
φ(n)−2
φ(n)−2 n−2 3·2
n−3 2
n−2
n−1
−1
.
Rearranging the above and upper bounding the RHS by (n-2)
yields the condition,
Eφ(n) > Ek̂−1 + n1 − (n − 1)
φ(n)−2
φ(n)−2
−1
n−2 3·2
n−1
φ(n)−2
n−3 2
− 2 + n1 − (n − 1) n−2
φ(n)−2 −1
n−2 3·2
>n−2 .
n−1
n−3 2
n−2
>n
(31)
Note that in the above derivation we assumed Ek̂−1 > n − 2.
This assumption can be proven by using (26) as follows,
!
2φ(n)+1
n−2
Ek̂−1 > n 1 − n
−1≥n−2
∴
∴
2φ(n)+1
≤ n−1
log(n)
φ(n) + 1 ≥ log log(n/(n−2))
,
n−2
n
where the last line holds due to (10).
Rearranging (31) and taking logarithms yields,
−2·log(n)
>
−8·log(n/(n−2))
2
(n−3)
n−2
log n(n−1)2 /log (n−2)
(n−1)3
.
The proof is complete by noting,
n−2 >
8
n−2
<
n
n−2
n(n−1)2 ,
(n−2)2 (n−3)
(n−1)3
,
the first line of which holds for all n ≥ 1 and the second holds
for all n ≥ 4.
Next we proof the tightness of the lower bound k̂ − 2 in
regard to the expected time until the first node reaches FKS.
This is done by showing that Ek̂−3 < n − 2, which follows
essentially the same logic of Lem.V II.1. For convenience
2φ(n)−1
3·2φ(n)−1 −1
n−3
n−2
denote χφ(n)−1 = n−2
. Applying
n−1
the previous result Ek̂−1 > n − 2, Lem.V II.1 and (25) we
have,
Ek̂−3 − Ek̂−1 < Êk̂−3 − (n − 1)(1 − γφ(n)−1 )
< Êk̂−3 − (n − 2)
∴ Ek̂−3 < Ek̂−1 + Êk̂−3 − n + 2
< n − 1 − n−1 + Êk̂−3 − n + 2
= Êk̂−3 + 1 − n−1 .
Upper bounding the RHS of the above by (n − 1 − 3n−1 ) and
solving for Êk̂−3 yields the condition,
Êk̂−3 < n − 2 − 2n−1 .
Notice that the upper bound (n − 1 − 3n−1 ) was chosen
without any specific precondition. For this reason, along with
11
the fact that Ek̂−3 < Ek̂−1 for all ` ∈ {0, 1, . . . , n − 2}, we
must confirm the following,
Êk̂−3 − Ek̂−1 < Êk̂−3 − (n − 1)(1 − γφ(n)−1 )
< n − 2 − 2n−1 − (n − 1)(1 − γφ(n) )
= −1 − 2n−1 + (n − 1)γφ(n) < 0 .
log(n) <
Rearranging the last inequality and taking logarithms yields,
n(n−1)2
log
2(n+1)(n−2)
log(n)
>
.
(32)
(n−1)3
n
2 · log (n−2)
log (n−2)
2 (n−3)
Note that
log(n)
log(n/(n−2)
2
as 2log(n) and canceling terms
Upper bounding log (n−1)
2
√
1/ 2
(n − 2). For this reason, we can utilize (30)
yields n < 2
and prove (32) by the following,
n(n−1)2
n(n−1)2
1
≥ log 2(n+1)(n−2)
2 log 2(n−2)
n4 (n−1)4
4(n+1)2 (n−2)2
2
≥
2(n − 2)(n + 1) ≥ n(n − 1)2
n3 + 2n2 + n − 4 ≥ 0 , ∀ n ≥ 1 .
Proof: At initial time k = 0 the knowledge set of each
node i ∈ V contains only {di }, that is Ki (0) = {di }. At time
k = 1 each node receives a signal from exactly one other
node in the network, thus |Ki (1)| = 2 for each node i ∈ V.
At time k = 2, each node receives a signal from exactly one
other node in the network, thus the maximum cardinality of
any knowledge set Ki (2) is 4. Proceeding in this way, we find
|Ki (k)| ≤ min{2k , n} for any sequence of (k) random graphs
R(k). If log(n) ∈ N, then all nodes reach FKS at k = log(n).
If log(n) ∈
/ N, then at time k = blog(n)c the cardinality of
each knowledge set is less than or equal to 2blog(n)c , from
whence it is clear that 2dlog(n)e > n, and hence the network
will reach FKS at a minimum time of k = dlog(n)e.
Proposition VII.3. For all n ≥ 3, log(n2 ) > 2 + φ(n) >
log(n).
(33)
Rearranging (33) yields,
−
log(1 − n2 )
4
>
.
n2
log(n)
for x ≥ 0 [26],
n−1
log(n) ≤ √
,
n
thus it remains to be shown,
n−1
√
n
<
2<
√
n
2
n √
n−(1/ n)
,
(34)
(35)
Rearranging (35) yields,
−
log(1 − n2 )
4
< ,
log(n)
n
which can be simplified as,
n − 2 > n1−(4/n) .
Theorem VII.2. For any set of (n−1) random graphs R(n−
1), the minimum time at which the entire network obtains FKS
is k = dlog(n)e.
Proof: Applying (10) we will first show,
log(n − 2)
log(n2 ) > 2 − log 1 −
.
log(n)
√x
1+x
which can be easily verified for all n ≥ 1.
Next we apply (10) to show,
log(n − 2)
2 − log 1 −
> log(n) .
log(n)
log(n/(n−2)
n(n−1)2
2(n−2)
n2
.
2n − 2
Applying the upper bound log(1 + x) ≤
we can let x = n − 1 and obtain,
2
since,
> log n(n−1)
2(n−2)
log(n)
n(n−1)2
>
log
log(n/(n−2)
2(n−2)
(n−1)2
= log
+ log n/(n − 2)
2
2
log(n)
2 > log (n−1)
∴
.
2
∴
∴
∴
2x
Using the upper bound log(1+x) ≤ 2+x
for x ∈ (−1, 0] [26],
we can let x = −2/n and obtain from (34),
(36)
It is clear that the left-hand side (LHS) of (36) increases with
n at a faster rate than the RHS of (36). It thus remains to be
shown that for n = 3 the inequality (36) holds. For n = 3 the
inequality (36) becomes 1 > 31−(4/3) , thus (35) holds for all
n ≥ 3.
Example VII.4. A Fixed Connected Digraph.
Consider a random graph G̃(0) (see (8) for a formal
definition, and see Fig.6 for an illustration). If G(k) = G̃(0)
for all k ∈ {0, 1, . . . , n−2}, then the digraph is connected and
also “fixed” (or “time invariant”). By rearranging the labels of
each node and utilizing (15), the communication pattern of a
fixed randomly chosen connected digraph can be defined by
the powers of the following matrix M̃ ∈ Rn×n ,
M̃n1 = M̃ii = M̃i(i+1) = 1 , ∀ i ∈ V−n .
(37)
It is trivial to show that each column of M(k) = M̃ k+1 has
all non-zero elements only when k ≥ n − 2, and thus by
(16) each node reaches FKS exactly at time k = n − 1. This
result validates the feasibility of the upper bounds stated in
Lem.III.1, III.5, and Thm.III.3, III.6, III.7. In summary,
the upper bound of each respective result cannot be improved
upon without further restricting the network communication
conditions.
Conjecture VII.5. Given the communication constraints defined in Lem.III.5, Thm.III.6, or Thm.III.7, there exists no
distributed algorithm that can solve TP (cf. Def.II.4) without
further conditions on the sequence of communication graphs
after G(n − 1).
12
R EFERENCES
[1] S. Boyd, L. Xiao, and S. Lall, “A Scheme for Robust Distributed Sensor
Fusion based on Average Consensus,” in Proc. of 4th IEEE Symposium
on Info. Proc. in Sensor Networks, Los Angeles, CA, 2005.
[2] A. Dimakis, A. Sarwate, and M. Wainwright, “Geographic gossip:
Efficient aggregation for sensor networks,” in Proc. ACM/IEEE Symp.
Inf. Process. Sens. Netw., pp.69-76, TN, Apr. 2006.
[3] H. Dubois-Ferriere, M. Grossglauser, and M. Vetterli, “Age Matters:
Efficient Route Discovery in Mobile Ad Hoc Networks Using Encounter
Ages,” in Proc. of the 4th ACM Intern. Symp. on Mobile Ad Hoc Netw.
and Computing, 2004.
[4] F. Fich and E. Ruppert, “Hundreds of impossibility results for distributed
computing,” Distributed Computing, Vol.16, 2003.
[5] F. Kuhn, R. Wattenhofer, Y. Zhang, and A. Zollinger, “Geometric ad-hoc
routing: Of theory and practice,” in Proc. of the 22nd ACM Symposium
on Principles of Distributed Computing, 2003.
[6] D. Johnson, D. Maltz, and J. Broch, “DSR: The Dynamic Source
Routing Protocol for Multi-Hop Wireless Ad Hoc Networks.” Ad Hoc
Networking. Addison-Wesley, 2001.
[7] E. Kokiopoulou and P. Frossard,“Polynomial ltering for fast convergence
in distributed consensus,” IEEE Trans. on Signal Process.,Vol.57, No. 1,
pp. 342-354, Jan. 2009.
[8] V. Krishnamurthy, K. Topley, and G. Yin, “Consensus formation in a
two-time-scale Markovian system.” SIAMJ. on Multi-Scale Model and
Simul., Vol. 7, No. 4, pp. 1898-1927, July 2009.
[9] D. Lang, “A comprehensive overview about selected ad hoc networking
routing protocols,” Technical report, Technical University of Munich,
Dept. of Computer Science, March 2003.
[10] W. Li and H. Dai, “Location-aided fast distributed consensus,” IEEE
Trans. on Inf. Theory, Vol.56, No.12, pp.6208-6227, Dec.2008.
[11] M. Mehyar, D. Spanos, J. Pongsajapan, S. Low, and R. Murray,
“Asynchronous Distributed Averaging on Communication Networks,”
IEEE Trans. on Networking, Vol.15, No.3, 2007.
[12] C. Moallemi and B. Van Roy, “Consensus Propagation”. IEEE Trans.
on Info. Theory, Vol.54, No.7, 2008.
[13] D. Mosk-Aoyama and D. Shah, “Fast distributed algorithms for computing separable functions,” IEEE Trans. on Inf. Theory,Vol. 54, No. 7,
pp. 2997-3007, July 2008.
[14] M. Naor and L. Stockmeyer, “What can be computed locally?”, in Proc.
of the 25th Annual ACM Symposium on Theory of Computing, 1993.
[15] A. Nedic, and A. Olshevsky, “Distributed optimization over time-varying
directed graphs,” arXiv:1303.2289 [math.OC], March 2014.
[16] A. Nedic, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, “On distributed averaging algorithms and quantization effects,” IEEE Trans. on
Autom.Control,Vol.54, No.11, pp.2506-2517, Nov. 2009.
[17] R. O’Dell, R. Wattenhofer, “Information Dissemination in Highly Dynamic Graphs”. DIALM-POMC05, Cologne, Germany, Sept. 2005.
[18] R. Olfati-Saber, E. Franco, E. Frazzoli, and J. Shamma, “Belief consensus and distributed hypothesis testing in sensor networks,” in Proc.
Workshop Netw. Embedded Sensor Control, IN, pp. 169-182, Oct. 2005.
[19] R. Olfati-Saber and R. Murray, “Consensus problems in networks
of agents with switching topology and time-delays,” IEEE Trans. on
Autom. Control, Vol. 49, No. 9, pp. 1520-1533, Sep. 2004.
[20] D. Peleg, “Distributed Computing: A Locality-Sensitive Approach,”
SIAM, Philadelphia. ISBN 0-89871-464-8 MATH, 2000.
[21] C. Perkins and E. Royer. “Ad hoc On-Demand Distance Vector Routing,”
In Proceedings of the 2nd IEEE Workshop on Mobile Comput. Sys. and
Applications, 1999.
[22] P. Samar, M. Pearlman, and Z. Haas, “Independent Zone Routing: An
Adaptive Hybrid Routing Framework for Ad Hoc Wireless Networks,”
IEEE/ACM Trans. on Networking, Vol.12, No.4, 2004.
[23] S. Sundaram and C. Hadjicostis,“Distributed consensus and linear functional calculation: An observability perspective,” in Proc. of 6th Int.
Conf. on Inf. Process. Sens. Netw., Cambridge, MA, pp. 99-108, Apr.
2007.
[24] K. Topley and V. Krishnamurthy, “Average-Consensus Algorithms in a
Deterministic Framework Part I: Strong Connectivity,” IEEE Trans. on
Signal Process., Vol. 60, No. 12, pp. 6590-6603, Dec. 2012.
[25] K. Topley and V. Krishnamurthy, “Average-consensus Algorithms in a
Deterministic Framework Part II: Central Connectivity,” IEEE Trans. on
Signal Process., Vol. 60, No. 12, pp. 6604-6616, Dec. 2012.
[26] F. Topse, “Some bounds for the logarithmic function,” University of
Copenhagen, 2006.
| 3cs.SY
|
arXiv:cs/0508102v1 [cs.CE] 23 Aug 2005
Investigations of Process Damping Forces in Metal Cutting
Emily Stone∗ & Suhail Ahmed
Department of Mathematics and Statistics
Utah State University
Logan, UT 84322-3900
and
Abe Askari & Hong Tat
Mathematics and Computing Technology
The Boeing Company
P.O. Box 3707 MS 7L-25
Seattle, WA 98124-2207
February 1, 2008
* Communicating author. Current address: Dept. of Mathematical Sciences, The University of Montana, Missoula, MT 59812
Running title: Process Damping Forces in Metal Cutting
1
Abstract
Using finite element software developed for metal cutting by Third Wave Systems we
investigate the forces involved in chatter, a self-sustained oscillation of the cutting tool.
The phenomena is decomposed into a vibrating tool cutting a flat surface work piece, and
motionless tool cutting a work piece with a wavy surface. While cutting the wavy surface,
the shearplane was seen to oscillate in advance of the oscillation of the depth of cut, as were
the cutting, thrust, and shear plane forces. The vibrating tool was used to investigate process
damping through the interaction of the relief face of the tool and the workpiece. Crushing
forces are isolated and compared to the contact length between the tool and workpiece. We
found that the wavelength dependence of the forces depended on the relative size of the
wavelength to the length of the relief face of the tool. The results indicate that the damping
force from crushing will be proportional to the cutting speed for short tools, and inversely
proportional for long tools.
keywords: metal cutting, chatter, process damping, shear plane dynamics.
AMS classification numbers: 74S05,74H45,74H15.
2
1
Introduction
In this paper we investigate the forces produced in dynamic metal cutting, where either the
chip load varies in time, or the tool itself oscillates. Chatter is a consequence of an instability in cutting that leads to self-sustained vibrations of the tool, hence it is composed of both
an oscillating tool and an undulating workpiece surface. We separate chatter into these two
pieces by using finite element software for metal cutting developed by Third Wave Systems,
Inc. called AdvantEdge, to simulate the processes. AdvantEdge is a validated software package that integrates advanced dynamic, thermo-mechanically coupled finite element numerics
and material modeling appropriate for machining processes. The simulation software provides
accurate estimates of thermo-mechanical properties of the machining process such as cutting
forces, chip morphology, machined surface residual stresses, and temperature behavior of the
tool and workpiece.
Recently Third Wave Systems has developed two new modules, one with custom work piece
geometry and another with a vibrating tool, that increase the range of cutting conditions that
can be simulated. The custom work piece geometry module allows for the design of arbitrary
cutting surfaces in addition to the traditional flat surface. We study the effect of a varying chip
load through a cut with this module, observing interesting shear plane dynamics and varying
thrust and cutting forces. With the vibrating tool module, the cutting tool can be horizontally
or vertically vibrated at a user specified frequency and amplitude. This allows us to investigate
the effect of contact between the material and the relief face of the cutter, which generates a
kind of process damping.
In section 2 we examine the forces involved in cutting a workpiece with a sinusoidally
varying surface and, as a consequense, a sinusoidally varying depth of cut. In section 3 we use
the vibrating tool module to investigate process damping on the relief face of the tool during
chatter. In section 4 we summarize our findings and suggest avenues for further research.
2
Wavy Surface Runs
The work piece geometry we use has a sinusoidal top surface which was generated by replacing
the horizontal flat surface of the standard geometry with a sine function. The wavelength of
the sinusoid was based on scaled parameters from experimental chatter measurements [8]. To
ensure the feed was always greater than zero, the feed was set to be larger than the peak to
valley amplitude of the sinusoid (3 mil). All these AdvantEdge runs were performed on AL
7050 with a carbide tool with a 10 degree rake angle, and cutting edge radius of 0.7874 mil and
cutting speed of 2680 SFM. Altogether three sets of runs were performed with three different
feeds: 6,9, and 12 mil.
From the animations of the Third Wave runs we saw that the shear plane angle during the
cut oscillates as the tool moves through sinusoidally varying depth of cut (see figure 1). The
3
shear plane, denoted by the band of higher heat rates extending from the tool tip, is seen to lift
up smoothly and shorten as the tool anticipates the thickest part of the chip. After reaching
its maximum slope, the band snaps down and elongates to connect the tool tip with the point
ahead on the surface that has the smallest chip thickness. The shear plane continues to extend
between the tool tip and this minimum until that point, now a part of the chip, is crushed into
the point on the surface with maximum chip thickness. The cycle repeats as the tool continues
to encounter the oscillating chip thickness. The experiments we performed were designed to
explore the effect of variation in depth of cut, shear plane angle, and shear plane length on the
instantaneous cutting forces.
2.1
Wavy Surface Force Predictions
Our first observation is that the measured forces in x (cutting) and y (thrust), while oscillating
at the same frequency as the chip thickness, lead it in phase. To illustrate this we plot the
thrust and cutting force through the cut, overlaid with a sine wave of the same frequency and
phase as the chip thickness, see figures 2 and 3. (The amplitude and DC offset of the sine wave
were scaled to match the force for illustration).
The phase lead makes sense after analyzing individual frames in the simulations. (See figure
1.) The chip begins to bunch up and hence offer more resistance, which creates larger cutting
forces prior to the maximum in the chip thickness. The forces drop before the tool tip encounters
the next minimum of the chip thickness, in reaction to the shearplane snapping down to meet
the minimum, which causes the chip to narrow down and offer less resistance.
This phase lead of the force over the chip thickness has been observed in experiments carried
out in Dr. Philip Bayly’s lab at Washington University (for representative publications from
Dr. Bayly’s group see [2, 3, 4]). A tube cutting experiment with a vibrating tool was set up on
a lathe, creating an essentially single degree of freedom system with one dominant mode. Tests
were performed to determine the frequency response (FRF) of the cutting system, and these
FRF were used to parameterize a single degree of freedom model for the cutting process. This
model was developed in part by Brian Whitehead, who describes in his thesis [1] an attempt to
include friction on the rake face of the tool by incorporating a weighting function that models
the variation of forces as the chip slides up the rake face of the tool, after Stépán [5]. One of the
parameters of the model is the contact length, L, between the tool relief face and chip. With
positive L the weighting functions, once incorporated into the frequency domain model, act as
low-pass filters, and induce a lag in the force behind the uncut chip thickness. Upon fitting
this model with the tube cutting experiment FRF, they determined that the best fit came with
using a negative L value, essentially making the forces on the tool anticipate the chip thickness.
This does not make sense if L is considered to be the contact length on the rake face, but with
L negative the force is being integrated over a region out in front of the tool, which induces the
phase lead of the force over the uncut chip thickness. We believe that this is a manifestation
4
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
1.5
Y (mm)
Y (mm)
1.5
1
1
2.5
3
3.5
4
2.5
X (mm)
3
3.5
4
X (mm)
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
1.5
Y (mm)
Y (mm)
1.5
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
1
1
2.5
3
3.5
4
2.5
X (mm)
3.5
4
X (mm)
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
Heat Rate (W/mm3)
50386
46787
43188
39589
35990
32391
28792
25193
21594
17995
14396
10797
7198
3599
0
1.5
Y (mm)
Y (mm)
1.5
1
2.5
3
1
3
3.5
4
2.5
X (mm)
3
3.5
4
X (mm)
Figure 1: Sequence of snapshots from a Third Wave run, time increases to the right and down.
5
Cutting force for 12 mil cut, overlaid with chip variation.
58
56
cutting force (pounds)
54
52
50
48
46
44
42
40
38
0
20
40
60
80
distance in cut (mils)
100
120
Figure 2: Cutting force vs. distance into cut, overlaid with chip thickness (dashed line).
Thrust force for 12 mil cut, overlaid with chip variation.
17
thrust force (pounds)
16
15
14
13
12
11
10
9
0
20
40
60
80
distance in cut (mils)
100
120
Figure 3: Thrust force vs. distance into cut, overlaid with chip thickness (dashed line).
6
α
λ
β
F_s
t_1
R
y
-R
θ
cutter
φ
x
workpiece
η
V
Figure 4: Diagram of angles and forces for orthogonal cutting.
of the shearplane oscillation effect we see in the Third Wave cutting simulations.
It is plausible that while the forces lead the chip thickness, they may be in phase with the
shearplane length, following a model popularized by Ernst and Merchant [6], that assumes that
the forces will be proportional to the shear plane length. The Merchant model predicts that
specifically the shear plane force (Fs ) is linearly dependent on the shear plane area, so Fs = τ wl,
where τ is the shear stress involved in cutting the material, w is the cutting width, and l is the
length of the shear plane, which is oriented at the angle φ (refer to figure 4). Also from the
figure note that Fs = ||R|| cos β, where R is the total force on the tool. To determine β we rely
on the traditional decomposition of R into vertical and horizontal components, Fy and Fx , the
radial or thrust, and the tangential or cutting, forces respectively. If the angle that R makes
F
with the horizontal is given by ψ = arctan( Fxy ), then β = φ + ψ. The AdvantEdge program
computes Fx and Fy , which also gives us a measure of the total cutting force as the tool moves
through the cut, since ||R|| =
q
Fx2 + Fy2 , as well as the angle ψ. We can measure the shear
plane length l, and the angle φ, for each step in the run, and since Fs = ||R|| cos(φ + ψ), we
can test the Merchant prediction:
Fs = ||R|| cos(φ + ψ) = τ wl.
(1)
The shearplane length was measured directly from the animation frames, where visualizations of the heat production are best for identifying the shearplane region. This is subject to
obvious error through the identification of a line that could be called the shearplane. We chose
to measure a line that started at the tip of the tool and ended at the surface of the workpiece,
parallel with the contours of heat production. See figure 5 for an example. The shear plane
length oscillates along with the undulations in the surface of the workpiece, though not neces7
1.4
3
Heat Rate (W/mm )
50386.9
46787.8
43188.8
39589.7
35990.6
32391.6
28792.5
25193.4
21594.4
17995.3
14396.2
10797.2
7198.12
3599.06
0
1.3
Y (mm)
1.2
1.1
1
0.9
0.8
3.25
3.5
3.75
4
X (mm)
Figure 5: How the shearplane angle and length is computed: the angle of the line shown here
with the horizontal and its length are measured directly from the frames of the Third Wave
animation.
sarily with the same phase. Following this observation, the shearplane length as the tool moves
through the workpiece was least squares fit to a sine wave with the same wavelength as the
wavy surface. The result of this exercise for the 12 mil feed run is plotted in figure 6, and the
fitted expression is
l(x) = 4.6 + 1.1(.151 cos(2πx/40) + 0.647 sin(2πx/40)).
The units on shearplane length are arbitrary, and vary with the three runs examined in this
section. The important feature is the phase of the oscillation.
The shearplane angle can be measured and fit in a similar way, the result is presented in
figure 7, and the fitted expression is
φ(x) = 0.49 + 0.1 cos(2πx/40) − 0.09 sin(2πx/40).
Similar fits were performed for the 6 and 9 mil cuts, and the units are radians. We also observe
that the shearplane length oscillation leads the chip thickness oscillation by a noticeable amount,
see figure 8.
We compute ||R|| =
q
F
F2x + F2y and β = φ + arctan( Fxy ) for each of the three runs as a
function of distance into the cut (x). Expression (1) implies that l(x) ∼ ||R(x)|| cos β(x), so
we conducted a least squares fit to ||R|| cos β(x) with the expression c1 + c2 l(x), the result is
presented in figure 9. The shear plane force is seen to lag the shear plane length, which fits
with the observation that while the shearplane force leads the chip thickness, the shearplane
length leads the chip thickness by an even greater amount, leading to the lag in shearplane force
with respect to the shearplane length. Also note that we fit the force Fs not with c1 l(x), as is
specified by Ernst and Merchant, instead it is a linear fit offset by a constant. This indicates
8
Shear Plane Length, 12 mil cut
5.4
shear plane length
5.2
5
4.8
4.6
4.4
4.2
4
3.8
3.6
0
20
40
60
80
distance into cut (mils)
100
120
Figure 6: Shear plane length measured from experimental snapshots, overlaid with fitted sine
wave.
Shear Plane Angle, 12 mil cut
0.75
shear plane angle (radians)
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0
20
40
60
80
distance into cut (mils)
100
120
Figure 7: Shear plane angle measured from experimental snapshots, overlaid with fitted sine
wave.
9
Shear plane length (−), chip thickness (−−)
Shear plane length compared to chip thickness, 12 mil cut
15
10
5
0
0
20
40
60
80
distance into cut (mils)
100
120
Figure 8: l(x) compared to chip thickness through the cut for 12 mil cut. The magnitude
of the chip thickness is in mils, the shearplane length has been scaled and shifted vertically
for comparison. The critical feature is the phase shift: the shearplane length leads the chip
thickness.
that there must be a nonlinear region near l = 0, a roll-off of the function so that Fs = 0 when
l = 0. This by itself does not contradict the Merchant model, which was meant to apply only
in steady cutting finite chip thickness regimes. However, the Merchant formula does fail to
completely capture the behavior of the forces in this dynamic situation.
This procedure suggests plotting the data ||R(x)|| cos β(x) parametrically against l(x) to
determine if this linear fit is appropriate, see figure 10. In this format it is clear that the
shearplane force depends on something in addition to l, since it is a multi-valued. The fitted
line is drawn for comparison for each of the three runs.
The direct measured forces, Fx and Fy can also be compared with the shearplane length
oscillation during the run. The result is presented in figure 11, for the 12 mil chip thickness run.
The cutting force also lags the shearplane length, similar to Fs , predictably since the cutting
force is the largest component of the total force. The thrust force, however, comes very close
to matching the oscillation of the shearplane length, a fact that could be explored further.
To conclude this section, we would like to comment on an idea presented in Wellbourne and
Smith [7]. They computed an expression for the cutting force involved in cutting a wavy surface
by assuming that the force was proportional to the shear plane length also. However, they were
limited to estimating the shearplane length by computing the length of a line, oriented at a
constant slope, as it intercepts a sine wave (see figure 12). The length can be found by solving
10
12 mil D.O.C.
45
Fs
40
35
30
25
20
30
40
50
60
70
80
90
100
110
120
9 mil D.O.C.
30
Fs
25
20
15
20
30
40
50
60
70
80
90
100
110
120
130
110
120
130
6 mil D.O.C.
20
Fs
15
10
5
0
20
30
40
50
60
70
80
90
distance into cut (mils)
100
Figure 9: Fs compared with fit to shearplane length for 12 mil, 9 mil and 6 mil depth of cut
(D.O.C.) runs, the length is the smooth curve.
11
6 mil feed
9 mil feed
20
30
15
25
Fs
35
Fs
25
10
20
5
15
0
0
2
4
6
shear plane length
8
10
10
1
1.5
2
2.5
shear plane length
12 mil feed
45
Fs
40
35
30
25
3
4
5
shear plane length
6
Figure 10: ||R(x)|| cos β(x) plotted parameterically with x against l, 6,9 and 12 mil feed. The
shear plane length is in different units in each figure, the measured shearplane force is in pounds.
12
3
Cutting and thrust forces compared with shearplane length, 12 mil D.O.C.
55
Fx
50
45
40
20
40
60
80
100
120
100
120
16
Fy
14
12
10
8
20
40
60
80
distance into cut (mils)
Figure 11: Fx and Fy plotted with l (the smooth curve), 12 mil feed.
15
workpiece surface
10
moving shearplane
5
moving tool tip
0
0
20
40
60
80
100
120
Figure 12: How the shear plane length might change through a wavy surface cut.
13
shearplane length through cut, varying shearplane angle
60
φ=0.25
length of shearplane
50
40
φ=0.35
30
φ=0.45
φ=0.55
20
10
workpiece surface
0
0
20
40
60
80
distance into cut (mils)
100
120
Figure 13: Plotting the shear plane length for varying shear plane angle (φ), through a wavy surface cut. The function plotted with a dashed line is the workpiece surface itself, for comparison
with the chip thickness.
a transcendental equation (the intersection of the moving line and a sine wave) numerically
for varying shear plane angles (figure 13). From this you can see that the length leads the
oscillation of the chip load, how much depending on the shear plane angle. The oscillation also
becomes more triangular as the angle is decreased, it will have a discontinuity when the multiple
roots to the transcendental equation appear, that is, for small enough shearplane angles.
These effects certainly play into the form of the forces determined with experimentally
measured shear plane lengths. But there are several confounding factors: the shear plane angle
is not a constant, and more importantly, the surface itself is deformed during a cut, and the
simple intersection of a line and a sine wave does not give the correct shear plane length, as we
saw in the animations, see figure 1.
3
Analyzing Crushing Forces
As described in the Introduction, the new Third Wave module that allows the tool to oscillate
either horizontally or vertically during a cut gave us an opportunity to explore the forces
involved when the relief face of the cutting tool impacts the workpiece. In our simulations to
investigate relief face crushing the cutting tool was vibrated sinusoidally in the vertical direction
14
Third Wave AdvantEdge
3
Heat Rate (hp/in )
1E+06
928571
857143
785714
714286
642857
571429
500000
428571
357143
285714
214286
142857
71428.6
0
0.14
0.13
0.12
0.11
0.1
Y (in)
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.25
0.3
0.35
0.4
0.45
X (in)
Figure 14: Third Wave run (AL7050) with relief angle of 6 degrees.
with a 3 mil amplitude and varying frequencies. Two simulations were performed to calculate
the crushing force for each set of cutting parameters, identical in all ways except for the relief
angle of the tool. In one simulation, the relief angle was small enough that the entire relief face
came into contact with the workpiece, causing the crushing of the workpiece at the relief face
of the tool. In the other simulation, the relief angle was made large enough so that there was
no contact between the relief face of the tool and the work piece.
To illustrate, a snapshot of a crushing run is shown in figure 14, where the tool has come
into contact with the workpiece during the downstroke of the tool and left a flattened region in
its otherwise sinusoidal path. This should be compared with a run with the relief angle drawn
up from 6 degrees to 25 degrees, figure 15, where the tool tip leaves a sinusoidal workpiece
surface in its wake. Assuming the forces involved in material removal and crushing combine
linearly, the crushing force can be resolved by subtracting the forces from two such runs. The
exercise is illustrated in figure 16, where the material being cut is AL7050, with a tool with rake
angle of 10 degrees and cutting edge radius of 0.5 mils. In the two runs the wavelength of the
oscillation is 80 mils, corresponding to a cutting speed of 2680 SFM and vibration frequency of
6.7 kHz. The nominal feed is 1 mil, and vibration amplitude is 3 mil. The spatial wavelength
λ of the undulating cut is equal to the ratio of the vibration frequency ν to the cutting speed
V , that is λ =
V
ν
. We focused our investigations on the vertical or thrust force in what follows.
15
Third Wave AdvantEdge
3
Heat Rate (hp/in )
500000
464286
428571
392857
357143
321429
285714
250000
214286
178571
142857
107143
71428.6
35714.3
0
0.14
0.13
0.12
0.11
0.1
Y (in)
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.25
0.3
0.35
0.4
0.45
X (in)
Figure 15: Third Wave run (AL7050) with relief angle of 25 degrees.
We found that the relief edge length, relative to the wavelength of the oscillation, was critical
in determining the behavior of the crushing force in these runs. Therefore we ran simulations
with tool relief lengths of 10 mil, 30 mil and 100 mil, which were much smaller, somewhat
smaller, and much larger than the range of wavelengths tested. The results are summarized
below.
3.1
Crushing force simulations
Once again the AdvantEdge runs were performed on AL 7050 with a carbide tool with a 10
degree rake angle and cutting edge radius of 0.7874 mils. The cutting speed was chosen to be
2680 SFM, and vibration frequencies near 13.4 KHz, and amplitudes of 3 mils, typical of an
axial-torsional vibration mode for a twist drill ([8]). In total we did 24 runs: 12 crushing, 12
no crushing. For each length of tool relief face (10, 30 and 100 mils) we ran four wavelengths:
40, 60, 80 and 100 mil. We found plotting the crushing force vs. the vertical amplitude of
the tool during a run a useful way of visualizing the data. In figures 17-19 these results are
presented for each tool relief length, with the varying wavelength runs plotted together for
comparison. The general form of each plot is the same, during the downstroke of the tool,
as expected, the crushing force increases up the angled portion of the loop, i.e. the loops are
traversed counterclockwise with increasing time. Near the end of the downstroke the force
16
Cutting Force
Thrust Force
Crushing Case
50
40
30
20
10
0
Forces for 10 mil DOC (lbs)
0
50
100
150
Non−Crushing Case
200
250
50
40
30
20
10
0
0
50
100
150
200
Difference Between Crushing & Non−Crushing Cases
250
50
40
30
20
10
0
0
50
100
150
200
250
cut position (mils)
Figure 16: Computing the crushing forces from two Third Wave runs.
17
Long tool relief length (100 mil)
60
40 mil λ
60 mil λ
80 mil λ
100 mil λ
50
Crushing forces
40
30
20
10
0
−10
−3
−2
−1
0
1
Vertical position of tool
2
3
Figure 17: Thrust crushing force vs. vertical position of tool, long tool relief length, multiple
wavelengths.
begins to decrease, and drops to zero abruptly as the tool begins its trip back up to the top
of the oscillation. During this phase the force is zero and the portion of the loop along the
displacement axis is traversed. Immediately after the top of the oscillation is reached and the
downstroke begins, the force starts to increase again, and another cycle is initiated.
The three sets show a qualitative difference in shape, and a quantitative difference in maximum force achieved relative to wavelength of the oscillation. The long relief length tool demonstrates the most easily understood force loop: the force increases approximately linearly as the
tool moves into the material, reaching a maximum close to the lowest point of the traverse.
In the short relief length tool runs the force increases approximately linearly until about one
third of the way down, when it flattens out. The short tool relief length is saturated at this
point: is has come into contact with as much material as it can, it cannot impact more length
as it moves down. The intermediate length tool runs have a shorter flat region that depends
on wavelength, with longer wavelengths showing some flattening.
The quantitative difference between the three sets of runs concerns the dependence of the
maximum force value vs. wavelength of oscillation. In Whitehead [1], a model for process
damping due to rubbing on the relief face is proposed that depends inversely on the wavelength
of the vibration, based on a suggestion by Tobias [9]. This translates to a term in his linear
model that is directly proportional vibration velocity (which depends directly on vibration
frequency) and inversely proportional to cutting speed. In figure 20 we plot the maximum
18
Intermediate tool relief length (30 mil)
40
40 mil λ
60 mil λ
80 mil λ
100 mil λ
35
Crushing forces
30
25
20
15
10
5
0
−5
−3
−2
−1
0
1
Vertical position of tool
2
3
Figure 18: Thrust crushing force vs. vertical position of tool, intermediate tool relief length,
multiple wavelengths.
Short tool relief length (10 mil)
20
40 mil λ
60 mil λ
80 mil λ
100 mil λ
18
16
Crushing forces
14
12
10
8
6
4
2
0
−2
−3
−2
−1
0
1
Vertical position of tool
2
3
Figure 19: Thrust crushing force vs. vertical position of tool, short tool relief length, multiple
wavelengths.
19
60
55
long relief length tool
maximum crushing force
50
45
40
35
intermediate relief length tool
30
25
short relief length tool
20
15
10
40
50
60
70
80
wavelength
90
100
Figure 20: Maximum crushing force vs. wavelength of oscillation.
force for each run vs. wavelength in the three cases. Though the data is too sparse to fit
any detailed model, it is clear that the force is inversely proportional to the wavelength only
in the short relief length runs, the long relief length runs demonstrate a direct dependence on
wavelength, and in the intermediate length runs the maximum force appears to be indifferent
to wavelength. The dependence of the damping on cutting speed and vibration frequency thus
will depend on the length of the relief face relative to the vibration wavelength, and given that
chatter vibrations are mainly high frequency, the long tool data will be the most relevant for
estimating chatter process damping.
3.2
Crushing force modeling
The magnitude of the crushing force will most certainly depend on the amount of contact
between the relief face of the tool and the material. We can analytically model the variation of
the contact during a crushing thrust of the tool by determining the contact length, as it depends
on frequency of the oscillation, height of oscillation and the cutting speed. The contact length
at each point along the cut can be calculated from the intersection of a tool edge with the
work piece (see figure 21). The sinusoid in the figure is the path that the tool takes during an
oscillation, the chip is not shown in the diagram. A small program in MATLAB calculated this
intersection length in the figures that follow.
The contact length during an oscillating cut is plotted vs. vertical position of the tool tip,
see figure 22. This representation can be directly compared to the force vs. vertical position
plot for the 3rd Wave run with a 100 mil relief length tool, figure 17. The contact length and the
20
cutter
tool tip
contact length
tool path
Figure 21: Illustration of contact length during crushing, long tool relief length.
force have similar dependence on wavelength, both increase with increasing wavelength at the
bottom of the downstroke of the tool (the highest point on the loop). There is also a “roll-over”
at the top of the downstroke, where the shorter wavelength force/contact length does not begin
to grow as soon as the longer wavelength case. This is a consequence of the relative slope of
the tool path to the relief angle of the tool. A similar diagram for the short relief length (10
mil) tool is shown in figure 24. The contact and the crushing forces in this case saturate when
the tool relief length is fully engaged with the workpiece, but an obvious difference between
the two is the invariance the maximum contact length shows with respect to wavelength. The
maximum forces, however, decrease with increasing wavelength for the short tool.
With this evidence in mind we can attempt to fit the crushing force data from the experiments with a model that has a linear dependence on contact length. If X is the contact length,
Y = mX + b is the crushing force, and m and b are fitted via least squares. The results depend
on the wavelength of the oscillation and the relief length of the tool, and are tabulated below.
Long relief length tool (100 mil):
wavelength (mils)
40
60
80
100
fitted line
Y = 1.102
Y = 1.132
Y = 1.184
Y = 1.305
Intermediate relief length tool (30 mil):
21
*
*
*
*
X
X
X
X
+ 0.397
- 1.539
- 9.698
- 20.493
error (lbs)
2.34
3.75
3.88
3.74
Contact Length vs. position of tool, long relief length
50
100 mil
80 mil
60 mil
40 mil
45
contact length (mils)
40
35
30
25
20
15
10
5
0
−3
−2
−1
0
1
vertical position of tool (mils)
2
3
Figure 22: Contact length during tool oscillation, long tool relief length (100 mils).
35
Contact length vs. position of tool, intermediate relief length (30 mil)
100 mil
80 mil
60 mil
40 mil
contact length (mils)
30
25
20
15
10
5
0
−3
−2
−1
0
1
2
vertical position of tool (mils)
3
Figure 23: Contact length during tool oscillation, long tool relief length (30 mils).
22
Contact length vs. position of tool, short relief length
contact length (mils)
20
15
100 mil
80 mil
60 mil
40 mil
10
5
0
−3
−2
−1
0
1
2
vertical position of the tool (mils)
3
Figure 24: Contact length during tool oscillation, short tool relief length (10 mils).
wavelength (mils)
40
60
80
100
fitted line
Y = 1.002
Y = 1.431
Y = 0.963
Y = 0.977
*
*
*
*
X
X
X
X
+ 3.628
- 6.472
- 7.128
- 7.877
error (lbs)
3.02
2.54
2.47
2.62
fitted line
Y = 0.548
Y = 0.234
Y = 0.207
Y = 0.167
*
*
*
*
X
X
X
X
+
+
+
+
error (lbs)
0.84
0.51
0.12
0.33
Short relief length tool (30 mil):
wavelength (mils)
40
60
80
100
1.501
1.350
1.484
0.957
The slope constant in the linear relationship gives the scale factor between force (in lbs)
and contact length (in mils) and thus has units lbs/mil of contact length. This conversion
factor varies somewhat within runs (mean and standard deviation are 1.18 ± 0.089, 1.093 ±
0.226, 0.289 ± 0.175, for the long, intermediate and short relief length runs, respectively) The
y intercept is determined by nonlinearity in the data at the top and bottom of the cycle in the
data, which causes the linear region to move up or down.
To illustrate these results we plot the scaled contact length vs. vertical position of the tool
along with the crushing force for a comparable run, for the three different relief lengths and
varying oscillation wavelengths (figures 22,23, and 24). The long tool force loop shows a close fit
with contact length, except near the top of the oscillation when the contact length grows more
quickly from zero. The intermediate tool shows a better fit with data at top of the oscillation,
23
Long Relief Tool (100 mils)− 80 mil wavelength
60
Crushing Force
Crushing force (lbs) and Scaled Contact Length
50
Contact Length
40
30
20
10
0
−10
−3
−2
−1
0
1
Vertical Position of tool (mils)
2
3
Figure 25: Crushing force and contact length during tool oscillation, long tool relief length, 80
mil wavelength.
though the force does not flatten out in the same way as the contact length. The short tool
forces are undercut the contact length, the contact length grows more quickly at the top of
the oscillation, again, and the forces begin to decrease from the plateau before the contact
length comes down. All these point to additional factors coming into play in the prediction of
crushing forces. The assumption of linear dependence on contact length holds primarily for the
mid-region of the downstroke of the tool, excluding the lift-off at the bottom and turn-around
point at the top.
To illustrate these results we plot the scaled contact length vs. vertical position of the
tool along with the crushing force for a comparable run, for the three different relief lengths
and varying oscillation wavelengths (figures 25, 26, and 27). The long tool force loop shows
a close fit with contact length, except near the top of the oscillation when the contact length
grows more quickly from zero. The intermediate tool shows a better fit with data at top of
the oscillation, though the force does not flatten out in the same way as the contact length.
The short tool forces are undercut the contact length, the contact length grows more quickly
at the top of the oscillation, again, and the forces begin to decrease from the plateau before
the contact length comes down. All these point to additional factors coming into play in the
prediction of crushing forces. The assumption of linear dependence on contact length holds
24
Intermediate relief tool − 30 mil, 80 mil wavelength
Contact Length (mils) and Scaled Crushing force
40
35
30
Contact Length
25
20
Crushing force
15
10
5
0
−5
−3
−2
−1
0
1
2
3
Vertical Position of tool (mils)
Figure 26: Crushing force and contact length during tool oscillation, intermediate tool relief
length.
Crushing Force (lbs) and Contact Length (mils)
Short Relief Tool (10 mils) − 40 mil wavelength
25
Crushing Force
Contact Length
20
15
10
5
0
−3
−2
−1
0
1
Vertical Position of tool (mils)
2
3
Figure 27: Crushing force and contact length during tool oscillation, short tool relief length.
25
primarily for the mid-region of the downstroke of the tool, excluding the lift-off at the bottom
and turn-around point at the top.
4
Discussion and Conclusions
Our investigations with Third Wave simulations were designed to isolate different effects present
during regenerative chatter vibrations in metal cutting: the vibration of the tool and the
oscillatory variation in the chip thickness. The tool vibration feature of the package was used
to investigate process damping due to crushing of the workpiece material on the relief face of
the tool, and the custom workpiece feature allowed us to set up a sinusoidally varying chipload. Simulations with the later demonstrated an oscillation of the shear plane as the tool
progressed, and we found that the cutting, thrust and shear plane forces all were in advance
of the chip thickness. The shear plane force, proposed to be directly dependent on shear plane
length by Ernst and Merchant [6], was found to lead the shear plane length as well, as the forces
anticipated the decrease in chip load. This anticipation is due primarily to the deformation of
the chip as it encounters the tool, not surprizingly the process is more complex than predicted
during dynamic cutting.
An important cause of process damping is the interaction of the relief face of the cutter
with the workpiece material. Here we studied the thrust forces involved in such damping by
running an oscillating tool in AdvantEdge for a variety of oscillation wavelengths and tool relief
lengths. We found that the wavelength dependence of the force was determined by the length
of the relief face of the cutter relative to the wavelength of the oscillation: for long relative relief
length the maximum force increased with increasing wavelength, for short relative relief length
it decreased, and for intermediate relative relief length the dependence was indeterminent.
This expands the intuition of Whitehead et al., [1] who used a process damping model that
depended inversely on wavelength, as it would for a short tool. Since the wavelength of the
oscillation equals the cutting speed divided by the oscillation frequency, this means the damping
is proportional to the cutting speed for short tools, and inversely proportional to the cutting
speed for long tools. This, however, assumes that the frequency of oscillation is independent of
cutting speed, which is most likely not the case, given the form of the linear stability diagrams
for regenerative chatter (for an example related to this study see [8]).
We also found that the force depended linearly on contact length between the relief face of
the tool and the workpiece material during the downward stroke of the tool. Other factors come
into play during the initial moments of the downward stroke, and through the “lift-off” of the
tool from the trough of the oscillation. This suggests that this form of process damping can be
explained purely geometrically, where the relevant factors are oscillation amplitude and relative
size of relief length to wavelength. Incorporating this into a model that would determine
stability of the steady cutting state to oscillations would not be straight-forward, however,
26
since the damping itself depends on the wavelength of the instability induced (typically the
wavelength is determined by the physical parameters of the system, not a priori).
In summary, this research highlights the importance of dynamic and nonlinear effects on
instabilities in metal cutting, even in the case of orthogonal cutting. Forces that act in anticipation of chipload increases and process damping that depends in a complicated way on vibration
frequency are examples of effects that could be studied further, both in the laboratory and with
large scale and predictive computer modeling.
Acknowledgements
This work was supported in part by the National Science Foundation (DMS-0104818) and the
Mathematics and Computing Technology Group, The Boeing Company, Bellevue, WA.
References
[1] Whitehead, B., The Effect of Process Damping on Stability and Hole Form in Drilling,
Master of Science Thesis, Sevier Institute of Technology, Washington University, St. Louis,
(2001).
[2] Bayly, P.V., S. A. Metzler, A.J. Schaut, and K.A. Young, Theory of torsional chatter in
twist drills: model, stability analysis and comparison to test, ASME J. Man. Sci. & Eng.,
123 (4), 552-561 (2001).
[3] Bayly, P.V., K.A. Young and J.E. Halley, Analysis of tool oscillation and hole roundness
error in a quasi-static model of reaming, ASME J. Man. Sci. & Eng., 123 (3), 387-396 (2001).
[4] Bayly, P.V., M.T. Lamar, and S.G. Calvert, Low frequency regenerative vibration and the
formation of lobed holes in drilling, ASME J. Man. Sci. & Eng., 124 (2), 275-285 (2001).
[5] Stépán, G., Delay differential equation models for machine tool chatter, Dynamics and
Chaos in Manufacturing Processes, F. Moon, ed., Wiley and Sons, 165-192 (1998).
[6] Ernst, H. and M.E. Merchant, Chip formation, friction and high quality machined surfaces,
Trans. Am. Soc. Met., 29, 299-378 (1941).
[7] Wellbourne, D.B, and Smith, J.D., Machine Tool Dynamics, an Introduction, Cambridge
University Press, Cambridge, U.K., p. 31 (1970).
[8] Stone, E., and A. Askari, Nonlinear models of chatter in drilling processes, Dynamical
Systems, 17 (1) 65-85 (2002).
[9] Tobias, S.A., Machine Tool Vibration, J. Wiley, New York, (1965).
27
| 5cs.CE
|
Towards the Interactive Effects of Demand Response
Participation on Electricity Spot Market Price
Saeed Mohajeryami, Milad Doostan, Seyedmahdi Moghadasi
Energy Production and Infrastructure Center (EPIC)
Electrical and Computer Engineering Department
University of North Carolina at Charlotte
Charlotte, NC, USA
Email:{smohajer,mdoostan,smoghada}@uncc.edu
Abstract—The electricity market is threatened by supply
scarcity, which may lead to very sharp price spikes in the spot
market. On the other hand, demand-side’s activities could
effectively mitigate the supply scarcity and absorb most of these
shocks and therefore smooth out the price volatility. In this paper,
the positive effects of employing demand response programs on
the spot market price are investigated. A demand-price elasticity
based model is used to simulate the customer reaction function in
the presence of a real time pricing. The demand achieve by DR
program is used to adjust the spot market price by using a price
regression model. SAS software is used to run the multiple linear
regression model and MATLAB is used to simulate the demand
response model. The approach is applied on one week data in
summer 2014 of Connecticut in New England ISO. It could be
concluded from the results of this study that applying DR program
smooths out most of the price spikes in the electricity spot market
and considerably reduces the customers’ electricity cost.
Keywords— demand response, electricity spot market price,
electricity market, real time pricing, price elasticity of demand
I. INTRODUCTION
The market for electricity are progressively developing over
numerous years of competition and reorganization.
Nevertheless, there are still several areas in the industry that are
kept shielded from the advancement in the market, which one
of them is demand-side. As a matter of fact, this area is
underdeveloped due to the detachment from market price
fluctuations as the regulatory bodies attempted to give
immunity to retail customers vulnerable to such fluctuations.
For instance, during 2000 and 2001, California experienced a
major power crisis under the restructured wholesale market.
Although numerous factors could be listed as reasons to create
this crisis, most people agree that the lack of demand response
exacerbated the situation[1].
However, studies over the recent years show that demand
response (DR) programs could create an environment in which
customers could engage in the process of optimizated decision
making. Consequently, it can change the customers’
consumption pattern in response to the price signals provided
by the wholesale market.
These programs can create numerous practical possibilities
for the power system operators and utilities to make an
improvement in both economic and technical indices of their
system. As a matter of fact, power system operators can
compensate lack of their supply during the peak time with DR
resources. It is estimated that the capacity to meet demand
during the top 100 peak hours accounts for 10-20% of
Peter Schwarz
Belk College of Business
University of North Carolina at Charlotte
Charlotte, NC, USA
Email: [email protected]
electricity cost annually [2]. On the other hand, utilities can also
benefit from DR by taking advantages of lower prices offered
by such resources compared to electricity spot market.
DR programs are generally could be separated into two
main categories: incentive-based programs (IBP) and timebased rate (TBR) programs. As shown in Fig. 1, each category
is composed of several programs. The authors in [3-4]
elaborated these programs in detail.
For many years, utilities have offered IBPs to to large
industrial and commercial customers. As an example, ERCOT
offers emergency interruptible load program to its large
customers. Moreover, Southern California Edison (SCE) has
offered a variety of DR programs such as automated demand
response (Auto-DR), permanent load shifting (PLS) and
scheduled load reduction programs (SLRP) to its large
customers [5]. On the other hand, TBR programs are typically
neglected by utilities due to the lack of proper infrastructure,
technical complication and highly capital intensive
infrastructure. However, over the last decade, the US
government and its energy sector, due to environmental
challenges of the traditional electricity generation, attempted to
adopte a supportive approach in order to provide the necessary
infrastructure for DR and energy efficiency programs. These
programs are mostly TBR programs. Another impact factor that
helps to launch TBR programs more conveniently is the
penetration of advanced smart meters. Advanced metering
penetration, based on 2010 FERC survey, has reached to a
considerable level of 8.7 percent in the US [6]. Advanced
metering is regarded by many as a cornerstone of the TBR
programs. Furthermore, many utilities have recently launched
pilot programs to evaluate the feasibility and technical
challenges of TBR programs in the new environment [7-11].
Although based on the previous discussion, the financial
support is more available in order to build the necessary
infrastructure for the TBR programs, these program face many
obstacles in the implementation stage. One of the most major
issues that utilities face with regard to the design of the
programs is finding the proper model to explain the customer’s
reaction to the incentives provided by each program. As a
matter of fact, the utilities cannot employ the proper profit
maximizing strategies in the absence of a reliable model.
Therefore, many of the programs might not even initiated or if
they do, they might be doomed to failure. Moreover, the wrong
models might leads to proposing the wrong incentive payments.
The improper incentive payment can discourage the
participation in the program. Therefore, to overcome this
problem, different models are proposed to explain the
customer’s reaction function. Authors in [12-14] employ
demand-price elasticity concept to model the effects of
implementing demand response programs on customer’s
reaction. In fact, demand-price elasticity is a concept borrowed
from the consumer theory in microeconomics which reflects the
relative change in the demand with respect to the relative
change in the price [15-16]. The approach utilized in [17]
models the customer’s reaction function with linear
optimization technique assuming the customers have access to
the real-time electricity prices. In this model, the objective
function is maximizing the utility and minimizing the cost of
electricity consumption.
A statistical method introduced in [18] uses the demandprice elasticity to explain the customer reaction function in the
direct load control (DLC) program. Moreover, the authors in
[19] applies self-organizing maps and statistical Ward’s linkage
to classify electricity market prices into different clusters. It
also uses a non-parametric curve estimation approach to
explore the underlying structure in different clusters which
leads to extraction of the proper customer’s reaction to the
different prices. Furthermore, the authors in [20] developed a
method based on consumers theory in microeconomics to
incorporate the customers’ willingness to shift consumption
cross-periods based on the pertinent rates.
In addition to the aforementioned models, several
forecasting based approaches are proposed to explain the
customer’s reaction function. These approaches which use expost and ex-ante data are being used to forecast the short-term
and long-term customer’s reaction function. Authors in [21]
report the current practice at Pacific Gas and Electric (PG&E)
company. ex-ante and ex-post reports are used to develop
individual customer regression models. In order to develop a
robust model, all the interdependencies of weather, calendar
days, etc are added to the regression model. The utilized models
need substantial amount of historical data and proper control
groups.
In this paper, the effect of demand response on the spot
market price is examined. The objective is to investigate the
impact of DR on the price volatility by proposing an algorithm
which feeds the outcome of DR model into the spot market and
explore the positive impact of demand response. It will be
shown how the full participation of the demand side under real
time pricing can decrease the wholesale price and its
fluctuations in the market. In order to carry out such task, the
customer’s reaction function model and day-ahead load
forecasting are required. In this paper, customer’s reaction
model is taken from [22]; also, for the forecasting part, multiple
linear regression model (MLR) is used.
Two main classes of load and price forecasting are
prevalent in the literature. One assumes that merely the
availability of the historical data of the desired variable is
sufficient for forecast purposes, while the other relies on
additional different parameters like weather, pressure,
humidity, seasonality, etc to do the forecasting. ARIMA models
which belong to the former class are popular in short term load
and price forecasting [23-24]. The latter class which uses MLR
Fig. 1. Categories of the demand response programs
models are suitable for short, medium and long term
forecasting. Both classes could be enhanced by different
intelligent techniques such as Artificial Neural Network
(ANN), Fuzzy logic and Wavelet [25-28]. [29] analyzes the
application of aforementioned load forecasting classes in the
presence of DR programs in detail.
Even though, numerous models are proposed for price
forecasting, there is not any single model that works for all the
situations; consequently, the utilities use multiple models in
parallel to create scenario based forecasts.
The organization of this paper is as follows. First, a full
description of the demand response model used in this paper is
presented in section II. Multiple linear regression model is
elaborated in Section III. Section IV explains the
implementation of the proposed algorithm. Then the results for
the case study as well as discussion of the results are provided
in section V. Section VI closes the paper with drawing
conclusion from the provided discussion and results.
II. DEMAND RESPONSE MODEL
In order to describe the employed demand response model in
this paper, it is necessary to understand the concept of demandprice elasticity. The demand for almost all goods and services
rises as the price decreases. Based on diminishing marginal
return law, this change in the demand is not linear [15]. In order
to quantify the aforesaid change in the demand, the concept of
demand-price elasticity has to be utilized. Indeed, the nonlinear
demand curve could be linearized around a given point. Then
the change in the demand relative to the change in price could
be measured which is known as demand-price elasticity. (1)
represents the demand-price elasticity function mathematically.
=
E
P0
d0
×
∂d
(1)
∂p
Where E is the price elasticity,
p
and d are price and
demand, P0 and d 0 are initial price and demand respectively.
Price elasticity has two components: self-elasticity and
cross-elasticity. In other words, between two competing
commodities, the percent change of demand with respect to the
percent change in its own price is self-elasticity, whereas the
percent change of demand with respect to the percent change in
the price of the other commodity is cross-elasticity.
U ( d=
(i )) U ( d 0 (i )) + p0 .∆d (i ) +
As it was mentioned, the demand response model in this
paper is defined based on the elasticity. To achieve this target
demand response model for 24 hours, first it is required that the
model for one hour to be extracted and then expanded to 24
hours. In what follows, this procedure is described.
(9) could be rewritten and expanded as follows:
A. Demand-price elasticity model for one hour
Suppose the customer’s benefit for the i-th hour is as follows:
B=
( d (i )) U ( d (i )) − d (i ). p (i )
U=
( d (i )) U ( d 0 (i )) +
1
∂ U ( d 0 (i ))
∂d (i )
× ∆d (i ) +
2
∂d (i )
2
× ( ∆d (i ))
= p (i )
(4)
(5)
Assuming that the initial demand before implementing the
DR program is in optimum point, (6) and (7) should hold.
∂d (i )
= p0
(6)
(7)
By using (5) and the definition of the price elasticity of
demand (1), (8) is obtained.
∂ U ( d (i )) ∂p 1 p0
=
=
×
(8)
2
∂d (i )
∂d E d 0
Plugging (7) and (8) into the Taylor series expansion of utility
function (3) gives,
2
∆d (i )
2 × E (i ) × d 0
]
(10)
Expanding ∆d (i ) = d (i ) − d 0 (i ) and then plugging (10)
into (5) yields (11) and (12),
p (i=
) p0 (i ) × [1 +
d (i ) − d 0 (i )
E (i ) × d 0 (i )
p (i ) = p0 (i ) + p0 (i ) ×
Hence, according to (5), in the optimum point, the marginal
utility is equal to the price of the electricity.
∂U ( d 0 (i ))
2
(i )) U ( d 0 (i )) + p0 .∆d (i )[1 +
U ( d=
]
(11)
d (i ) − d 0 (i )
(12)
E (i ) × d 0 (i )
E (i ) × ( p (i ) − p0 (i ))
p0 (i )
(13)
B. Demand-price elasticity model for 24 hours
To provide a model for 24 hours, both self- and crosselasticity have to be taken into the consideration. The crosselasticity between hours i and j is defined as:
P0 ( j ) ∂d (i )
, i≠ j
E (i , j ) =
×
d 0 (i ) ∂p ( j )
(14)
The demand response model for 24 hours of a day could be
obtained by combining self- and cross-elasticity of demand as
follows:
d (i )
d (i ) = d 0 (i ) + E (i ) × 0 × ( p (i ) − p0 (i )) +
(15)
p0 (i )
24
∂B0
∂U ( d 0 (i ))
=
=
− p0 0
∂d (i )
∂d (i )
.( ∆d (i ))
d (i )= d 0 (i ) × 1 +
Therefore,
∂d (i )
p0
Therefore, the customer’s demand can be represented as
follows:
Where ∆d (i ) is the customer demand change from d 0 (i )
(the initial demand) to d (i ) (optimum point).
The customer benefit can take different units; however, in
this paper, for the simplicity, it is assumed that this benefit is in
terms of dollar. Moreover, according to the classic economics,
it is assumed that every individual optimizes her benefit. To
obtain the optimum point, the derivative of the benefit function
with respect to the demand must be zero.
∂U ( d (i ))
.
2 E (i ) d 0
2
∂B ( d (i )) ∂U ( d (i ))
(i ) 0
=
− p=
∂d (i )
∂d (i )
1
(3)
2
×
.
(2)
Where U ( d (i )) is customer’s utility in i-th hour. This
function could be formulated with the Taylor serious expansion
accourding to [29].
∂U ( d 0 (i ))
1
(9)
d 0 (i )
1, 2,..., 24
∑ E (i, j ) × p (j) × ( p(j) − p (j)) , i =
0
j =1
j ≠i
0
The variation in the demand in (15) stems from two sources,
one source is the self-elasticity which is reflected by the first
and second terms and the other source is cross-elasticity which
is reflected by the third term. The obtained relation in (15) is
employed in this paper for demand response modeling part.
III. MULTIPLE LINEAR REGRESSION MODEL
As it was mentioned, to perform our proposed algorithm,
the day-ahead price forecasting is necessary. An MLR model is
used in this paper to carry out this task. This model is a linear
model of demand, weather, time of day, week and season. In
addition to the day-ahead price forecasting, this model could be
utilized to update the spot market prices which is a part of our
algorithm. Indeed, MLR function is able to model the
dependent variable (price) as a linear function of the
independent variables, independent dummy variables and
interaction variables. The abundance of sample data can make
the MLR model a very powerful tool.
Moreover, this model could be enhanced by adding different
lag orders of the variables or different functional forms of the
weather parameters; nevertheless, these extra variables are
avoided in this paper for simplicity. Indeed, the accuracy of
MLR model is enough for the purposes of this paper. The model
used for the forecasting purpose is as follows:
24
∑α k .hk (i ) + α 25.d (i ) + α 26.T (i ) + α 27 .W (i ) +
P (i ) =
α0 +
of day, week and season are loaded into the SAS software. Then
the algorithm continues with developing a simple basic
regression model. Since adding too many variables to the
regression model may lead to the reduction of efficiency and
accuracy, it is necessary to select the most efficient variables.
Therefore, a simple basic model is used at the beginning and the
START
(16)
k =1
α 28 .M ( i ) + α 29 .D H ( i ) + α 30 .DSat ( i ) + α 31.DSun ( i )
IV. IMPLEMENTATION
Before explaining the implementation of the proposed
algorithm, it is necessary to make several assumptions for this
study. These assumptions are listed as follows.
• The utility is an independent non-profit agent that
functions as an intermediary link between the customers
and the wholesale market and purchases the electricity
on behalf of the customers. However, in practice, the
utility and the customers are two separate entities and
have different profit functions.
• The spot market is the main market and the utilities
purchase their whole demand in this market. However,
in practice, the day-ahead market is the main market and
the spot market is the real time market where the
participants use it to meet their obligations in an
emergency case.
• Real time pricing is applied to DR model. Indeed, to
incent the customers to change their consumption
pattern, real time pricing is the best possible choice.
To justify the aforementioned assumptions, several reasons
could be provided. First, although considering the utility profit
is more realistic, it makes the problem extremely complicated
while does not provide any relevant outcome for this study.
Second, the price volatility mainly exists in spot market where
the limitation of supply leads to sharp price spikes which is the
main focus of this study. However, in the day-ahead market,
due to the more availability of supply, such sharp spikes are
nonexistent. Indeed, by considering the spot market as the main
market, it is expected to observe more pronounced results.
Finally, the customers typically pay flat rate for electricity.
However, in this paper, the real time pricing is selected to be
applied to DR model as it would be the best incentive for the
users to shift loads at different times.
By considering the aforementioned assumptions, the
proposed algorithm for the evaluation of the impact of DR on
spot market prices is illustrated in Fig. 2.
The procedure is as follows. First, the historical data
including the information about price, demand, weather, time
Develop a simple multivariable
regression model
Perform the in-sample analysis
Modify the model
Where αk is the coefficient of the independent variables, hk is
hours of the day, T(i) is the temperature at time i, W(i) is the
dew point at time i, M(i) is the month at time i, DH(i) is a binary
variable which indicates the holiday at time i, DSat(i) and DSun(i)
are a binary variables which indicate the Saturdar and Sunday,
respectively, at time i.
Load the historical data
Perform the out-of-sample analysis
NO
Is FERMS acceptable ?
YES
Use the real-time prices and compute the new
demand based on DR model
Use the regression model to compute the new
wholesale electricity prices
END
Fig. 2. flowchart illustrating the implementation process
TABLE I: SELF AND CROSS ELASTICITIES
Peak
Off-Peak
Low
Peak
-0.10
0.016
0.012
Off-Peak
0.016
-0.10
0.01
Low
0.012
0.01
-0.10
others variables are added later on one at a time to evaluate
whether or not it improves the out-of-sample forecasting error
root mean square (FERMS).
If the overall FERMS is acceptable, then the day-ahead
price forecasting could be performed. The resultant data could
be fed into the selected DR model in a software. In this paper
MATLAB is employed for this purpose. DR model which uses
the day-ahead forecasted prices as a real time pricing produces
a new demand. For implementing the DR model, the self- and
cross- elasticity values of table I have been used [30-31]. Then
by using the regression model and applying new demand, the
updated prices could be achieved.
V. CASE STUDY
TABLE II: MULTIVARIALBE REGRESSION MODEL
** 1% SIGNIFICANT LEVEL, * 5% SIGNIFICANT LEVEL, +
10% SIGNIFICANT LEVEL
Variable
Parameter
Estimate
t-value
Intercept
Demand
Temperature
Humidity
Month
holiday
Saturday
Sunday
hour1
hour2
hour3
hour4
hour5
hour6
hour7
hour8
hour9
hour10
hour11
hour12
hour13
hour14
hour15
hour16
hour17
hour18
hour19
hour20
hour21
hour22
hour23
-182.3786
0.0483
0.62756
-1.23893
9.66399
7.23457
9.30045
11.82146
15.86262
24.07332
27.62021
28.27496
28.42
24.2956
11.63418
0.09629
-12.40514
-23.12766
-29.87756
-32.4509
-34.80398
-29.86742
-35.05025
-28.77225
-23.27377
-35.94987
-45.16361
-39.38583
-35.50788
-29.14752
-15.38141
-10.81**
20.66**
2.63**
-7.22**
8.41**
1.17
3.53**
4.33**
2.65**
3.99**
4.55**
4.64**
4.67**
4.02**
1.93+
0.02
-2.05*
-3.79**
-4.82**
-5.14**
-5.43**
-4.57**
-5.32**
-4.34**
-3.5**
-5.45**
-6.97**
-6.21**
-5.67**
-4.73**
-2.56*
A. Description
The proposed approach is examined on the reported hourly
data of New England ISO [32]. This ISO provides the zonal
information for all its serving areas including the day-ahead
price, load forecasting, real locational marginal price (LMP),
load, Dry Bulb temperature and Dew Point. For this study, the
data of summer 2014 in Connecticut is used. The study has been
done for one week (August 18th to 24th).
Moreover, in this case study, it is assumed that the
customers had paid a flat rate of 30$/MWh for this week before
DR program. After DR program, the utility charges the
wholesale market LMP for each hour.
B. Results and discussion
After applying the proposed approach to the selected data
and several iterations, variables in table II are selected as the
most efficient variables for the forecasting purposes. Table II
lists the variables, their parameter estimates (coefficients) and
their t-value.
Fig. 3. forecasted price vs. real price
Fig. 4. Demand before and after DR program
Fig. 5. spot market price before and after DR program
TABLE III: SUMMARY OF THE EFFECT OF DR ON THE SPOT
MARKET PRICE
Change in Energy
-13058 MWh
-2.23 %
Change in Cost
-4,674,396 $
-26.23 %
The parameter estimate could be interpreted as the change
in the predicted value of the dependent variable (price) for one
unit increase in the independent variable. Also, t-value is
defined as a ratio of the departure of an estimated parameter
from its notional value and its standard error. In table II, with
31 degrees of freedom (number of variables and intercept), tvalues between 1.3 and 1.69 are significant with 10% error. tvalues between 1.69 and 2.45 are significant with 5% error and
t-values over 2.45 are significant with 1% error. Therefore,
from table II, it is understood that almost all of the variables are
significant with 1% error. However, the selection of some
variables with lower level of significance is necessary in terms
of improving the out-of-sample analysis and FERMS.
By using the coefficients of table II and comparing the
results with the real time LMP, the forecasting error root mean
square (FERMS) of electricity price for this case study is
computed as 11% which is an acceptable error for volatile
variable like electricity price. Fig. 3 illustrates the forecasting
prices and real time LMP.
The achieved forecasted prices are fed into the selected DR
model to produce the new demand. Fig. 4 shows the demand
before and after applying the demand response model. Finally,
the new demand is fed into the regression model to produce new
prices.
Fig. 5 shows the wholesale electricity price before and after
applying the DR model. As it is shown, the price spikes are
declined considerably.
Based on the demand and price before and after applying
DR model, the total pertinent cost of electricity could be
computed. According to the calculation, it is observed that the
total demand is reduced by 2.23% and the total cost of
electricity for this week is reduced by 26.23%. Table III
summarizes the change in the wholesale electricity cost.
26.23% change in the total cost stems from two major
sources. First, 2.23% change in the total demand; second, a
considerable shift of demand from expensive peak time to lessexpensive off-peak period of time due to customer’s exposure
to the real time pricing as discussed earlier.
VI. CONCLUSION
This paper studied the effect of demand response programs
on the electricity spot market price. Demand response model
was used to account for the customer reaction function facing
the real time pricing. MLR model was employed to perform the
price forecasting in day-ahead market. Then, an algorithm
comprised of MLR and DR model was introduced to feed the
outcome of DR model into the spot market and explore the
impact of demand response program. Finally, the proposed
approach was tested on a case study by using the real data of
New England ISO.
The results showed that the demand-side participation
through 2.23% reduction of demand as well as shifting load
from expensive peak time to less-expensive off-peak period of
time could reduce the total cost of electricity by approximately
26 percent.
Therefore, according to the results, it could be said that
keeping the demand side isolated from the market deprives the
market of an effective tool for smoothing out the price
volatility. However, with adopting the proper DR program, both
the supply- and demand-side of the market can benefit. The
customers can benefit from the reduced bill and the supply-side
can compensate shortage of its sources during the peak time.
REFERENCES
[1] A. Faruqui, S. George. "Quantifying customer response to dynamic
pricing." The Electricity Journal, Vol. 18, no. 4, pp. 53-63, 2005
[2] J. Wells, D. Haas. “Electricity markets: consumers could benefit from
demand programs, but challenges remain”. DIANE Publishing, 2004.
[3] M. H. Albadi, E. F. El-Saadany, "A summary of demand response in
electricity markets," Elec. Power Syst. Res., vol.78, no.11, pp.1989-1996, 2008
[4] P.T. Baboli, M.P. Moghaddam, M. Eghbal, "Present status and future
trends in enabling demand response programs," IEEE Power and Energy
Society General Meeting, pp.1-6, July 2011
[5] Q. Zhou, W. Guan, W. Sun, "Impact of demand response contracts on load
forecasting in a smart grid environment," IEEE Power and Energy Society
General Meeting, pp.1-4, July 2012
[6] Federal Energy regulatory Commission (FERC) (2011, February),
“Assessment of demand response and advanced metering”, Staff report,
[online], available: http://goo.gl/rqtcXM
[7] A. Faruqui, S. Sergici, "Household response to dynamic pricing of
electricity: a survey of 15 experiments." Journal of Regulatory Economics, Vol.
38, no. 2, pp.193-225, 2010
[8] A. Faruqui, S. Sergici, "Arcturus: international evidence on dynamic
pricing." The Electricity Journal, Vol. 26, no. 7, pp. 55-65, 2013
[9] H. Sæle, O. S. Grande. "Demand response from household customers:
experiences from a pilot study in Norway." IEEE Trans. on Smart Grid, Vol. 2,
no. 1, pp. 102-109. 2011
[10] V. Gungor, et al. "Smart grid and smart homes: key players and pilot
projects." IEEE Industrial Electronics Magazine, Vol. 6, no. 4, pp. 18-34, 2012
[11] J. Torriti, M. G. Hassan, M. Leach. "Demand response experience in
Europe: Policies, programmes and implementation." Energy, Vol. 35, no. 4, pp.
1575-1583, 2010
[12] M. P. Moghaddam, A. Abdollahi, M. Rashidinejad, “Flexible demand
response programs modeling in competitive electricity markets,” Applied
Energy, Vol. 88, no. 9, Pages 3257-3269, Sep. 2011.
[13] H.A. Aalami, M.P. Moghaddam, and G.R. Yousefi, “Modeling and
prioritizing demand response programs in power markets,” Elec. Power Syst.
Res. , vol. 80, no. 4, pp. 426-435, 2010.
[14] H.A. Aalami, M.P. Moghaddam, and G.R. Yousefi, “Demand response
modeling considering Interruptible/Curtailable loads and capacity market
programs,” Applied Energy, vol. 87, no. 1, pp. 243-250, 2010
[15] D. Kirschen, and S. Goran, “Fundamentals of Power System Economics”,
John Wiley & Sons, Ltd, 2004.
[16] M. Shahidehpour and M. Alomoush. “Restructured electrical power
systems: Operation: Trading, and volatility”, CRC Press, 2001.
[17] A.J. Conejo, J.M. Morales, and L. Baringo, “Real-time demand response
model,” IEEE Trans. on Smart Grid, vol. 1, no. 3, pp. 236-242, 2010.
[18] Y. Rongshan, Y. Wenxian, S. Rahardja, "A Statistical Demand-Price
Model with Its Application in Optimal Real-Time Price," IEEE Trans. on Smart
Grid, vol.3, no.4, pp.1734-1742, Dec. 2012
[19] A. Gabaldon, et al. "Development of a methodology for clustering
electricity-price series to improve customer response initiatives," IET Gen.,
Trans. & Dist., vol.4, no.6, pp.706-715, June 2010
[20] S. Mohajeryami, et al., “A novel economic model for price-based demand
response,” Electric Power Systems Research (EPSR) , Volume 135, June 2016
[21] H. Liang, M. Agustin, "Demand response forecasting in practice:
Challenges and opportunities," IEEE Power and Energy Society General
Meeting, pp.1-3, July 2012
[22] H. Aalami, G.R.Yousefi, M.P Moghaddam, "Demand Response model
considering EDRP and TOU programs," IEEE/PES Transmission and
Distribution Conference and Exposition, T&D., pp.1-6, April 2008
[23] G. Gross, F.D. Galiana, "Short-term load forecasting," Proc. IEEE , vol.75,
no.12, pp.1558,1573, Dec. 1987
[24] J. Contreras, R. Espinola, F.J. Nogales, A.J. Conejo, "ARIMA models to
predict next-day electricity prices," IEEE Trans. on Power Syst., vol.18, no.3,
pp.1014,1020, Aug. 2003
[25] N. Amjady, A. Daraeepour, F. Keynia, "Day-ahead electricity price
forecasting by modified relief algorithm and hybrid neural network," IET Gen.,
Trans. & Dist, vol.4, no.3, pp.432-444, March 2010
[26] N. Amjady, "Day-ahead price forecasting of electricity markets by a new
fuzzy neural network," IEEE Trans. on Power Syst., vol.21, no.2, pp.887,896,
May 2006
[27] C. P. Rodriguez, G.J. Anders, "Energy price forecasting in the Ontario
competitive power system market," IEEE Trans. on Power Syst., vol.19, no.1,
pp.366-374, Feb. 2004
[28] N.M. Pindoriya, S.N. Singh, S.K. Singh, "An Adaptive Wavelet Neural
Network-Based Energy Price Forecasting in Electricity Markets," IEEE Trans.
on Power Syst., vol.23, no.3, pp.1423,1432, Aug. 2008
[29] A. Garulli, S. Paoletti, A. Vicino, "Models and Techniques for Electric
Load Forecasting in the Presence of Demand Response," IEEE Trans. on Cont.
Syst. Tech. , vol.23, no.3, pp.1087-1097, May 2015
[30] P.T. Baboli, M. Eghbal, M.P Moghaddam, H. Aalami, "Customer behavior
based demand response model," IEEE Power and Energy Society General
Meeting, pp.1-7, July 2012
[31] D.S. Kirschen, G. Strbac, P. Cumperayot, D. Mendes, “Factoring the
elasticity of demand in electricity prices”, IEEE Trans. On power syst., Vol.15,
No. 2, pp. 612-617, May 2000
[32] New England ISO zonal information, 2013 and 2014 SMD hourly data
report, [online], available: http://goo.gl/7h390w
| 3cs.SY
|
Hierarchical Convolutional-Deconvolutional Neural
Networks for Automatic Liver and Tumor
Segmentation
arXiv:1710.04540v1 [cs.CV] 12 Oct 2017
Yading Yuan
Department of Radiation Oncology
Icahn School of Medicine at Mount Sinai
New York, NY, USA
Abstract—Automatic segmentation of liver and its tumors is an
essential step for extracting quantitative imaging biomarkers for
accurate tumor detection, diagnosis, prognosis and assessment
of tumor response to treatment. MICCAI 2017 Liver Tumor
Segmentation Challenge (LiTS)1 provides a common platform for
comparing different automatic algorithms on contrast-enhanced
abdominal CT images in tasks including 1) liver segmentation, 2)
liver tumor segmentation, and 3) tumor burden estimation. We
participate this challenge by developing a hierarchical framework based on deep fully convolutional-deconvolutional neural
networks (CDNN). A simple CDNN model is firstly trained to
provide a quick but coarse segmentation of the liver on the entire
CT volume, then another CDNN is applied to the liver region
for fine liver segmentation. At last, the segmented liver region,
which is enhanced by histogram equalization, is employed as
an additional input to the third CDNN for tumor segmentation.
Jaccard distance is used as loss function when training CDNN
models to eliminate the need of sample re-weighting. Our
framework is trained using the 130 challenge training cases
provided by LiTS. The evaluation on the 70 challenge testing
cases resulted in a mean Dice Similarity Coefficient (DSC) of
0.963 for liver segmentation, a mean DSC of 0.657 for tumor
segmentation, and a root mean square error (RMSE) of 0.017
for tumor burden estimation, which ranked our method in the
first, fifth and third place, respectively.
I. I NTRODUCTION
Liver cancer is the second leading cause of global cancer
mortality (after lung cancer), and is one of the most rapidly
increasing cancers in terms of incidence and mortality worldwide and in the United States [1], [2]. Although contrastenhanced computed tomography (CT) has been widely used
for liver cancer screening, diagnosis, prognosis, and the assessment of its response to treatment, proper interpretation of CT
images is normally time-consuming and prone to suffer from
inter- and intra-observer variabilities. Therefore, computerized
analysis methods have been developed to assist radiologists
and oncologists for better interpretation of liver CT images.
Automatically segmenting liver and viable tumors from
other tissue is an essential step in quantitative image analysis
of abdominal CT images. However, automatic liver segmentation is a challenging task due to the low contrast inside liver,
fuzzy boundaries to its adjacent organs and highly varying
shape. Meanwhile, automatic tumor segmentation on liver
1 http://lits-challenge.com
normally suffers from significant variety of appearance in
size, shape, location, intensity, textures, as well as the number
of occurrences. Although researchers have developed various
methods to conquer these challenges [3], [4], [5], interactive
approaches are still the only way to achieve acceptable tumor
segmentation.
In this paper, we present a fully automatic framework based
on deep fully convolutional-deconvolutional neural networks
(CDNN) [6], [7], [8] for liver and liver tumor segmentation
on contrast-enhanced abdominal CT images. Similar to [9],
our framework is hierarchical and includes three steps. In
the first step, a simple CDNN model is trained to obtain a
quick but coarse segmentation of the liver on the entire 3D
CT volume; then another CDNN is applied to the liver region
for fine liver segmentation; finally, the segmented liver region
is enhanced by histogram equalization and serves as an additional input to the third CDNN for tumor segmentation. Instead
of developing sophisticated pre- and post-processing methods
and hand-crafted features, we focus on designing appropriate
network architecture and efficient learning strategies such that
our framework can handle images under various acquisition
conditions.
II. DATASET AND P REPROCESSING
Only LiTS challenge datasets were used for model training
and validation. The LiTS datasets consist of 200 contrastenhanced abdominal CT scans provided by various clinical
sites around the world, in which 130 cases were used for training and the rest 70 for testing. The datasets have significant
variations in image quality, spatial resolution and field-of-view,
with in-plane resolution ranging from 0.6 × 0.6 to 1.0 × 1.0
mm and slice thickness from 0.45 to 6.0 mm. Each axial slice
has identical size of 512 × 512, but the number of slices in
each scan varies from 42 to 1026.
As for pre-processing, we simply truncated the voxel values
of all CT scans to the range of [-100, 400] HU to eliminate
the irrelevant image information. While a comprehensive 3D
contextual information could potentially improve the segmentation performance, due to the limited hardware resource, it
is infeasible to perform a fully 3D CDNN on the volumetric
CT scans in our experimental environment. Thus, our CDNN
model is based on 2D slice and the CT volume is processed
slice-by-slice, with the two most adjacent slices concatenated
as additional input channels to the CDNN model. Different
resampling strategies were applied at different hierarchical
levels and will be described below.
III. M ETHOD
A. CDNN model
Our CDNN model [8] belongs to the category of fully convolutional network (FCN) that extends the convolution process
across the entire image and predicts the segmentation mask
as a whole. This model performs a pixel-wise classification
and essentially serves as a filter that projects the 2D CT slice
to a map where each element represents the probability that
the corresponding input pixel belongs to liver (or tumor). This
model consists two pathways, in which contextual information
is aggregated via convolution and pooling in the convolutional
path and full image resolution is recovered via deconvolution
and up-sampling in the deconvolutional path. In this way, the
CDNN model can take both global information and fine details
into account for image segmentation.
We fix the stride as 1 and use Rectified Linear Units
(ReLUs) [10] as the activation function for each convolutional/deconvolutional layer. For output layer, we use sigmoid
as the activation function. Batch normalization is added to the
output of every convolutional/deconvolutional layer to reduce
the internal covariate shift [11].
We employ a loss function based on Jaccard distance
proposed in [8] in this study:
P
(tij pij )
i,j
P
P
P
,
(1)
LdJ = 1 −
t2ij + p2ij − (tij pij )
i,j
i,j
i,j
where tij and pij are target and the output of pixel (i, j), respectively. As compared to cross entropy used in the previous
work [9], [12], the proposed loss function is directly related to
image segmentation task because Jaccard index is a commonly
used metric to assess medical imaging segmentation. Meanwhile, this loss function is well adapted to the problems with
high imbalance between foreground and background classes
as it does not require any class re-balancing. We trained the
network using Adam optimization [13] to adjust the learning
rate based on the first and the second-order moments of the
gradient at each iteration. The initial learning rate was set as
0.003.
In order to reduce overfitting, we added two dropout layers
with p = 0.5 - one at the end of convolutional path and
the other right before the last deconvolutional layer. We
also employed two types of image augmentations to further
improve the robustness of the proposed model under a wide
variety of image acquisition conditions. One consists of a
series of geometric transformations, including randomly flipping, shifting, rotating and scaling. The other type focuses on
randomly normalizing the contrast of each input channels in
the training image slices.Note that these augmentations only
require little extra computation, so the transformed images are
generated from the original images for every mini-batch within
each iteration.
B. Liver localization
This step aims to locate the liver region by performing a fast
but coarse liver segmentation on the entire CT volume, thus
we designed a relatively simple CDNN model for this task.
This model, named CDNN-I, includes 19 layers with 230, 129
trainable parameters and its architectural details can be found
in [8]. For each CT volume, the axial slice size was firstly
reduced to 128 × 128 by down-sampling and then the entire
image volume was resampled with slice thickness of 3 mm.
We found that not all the slices in a CT volume were needed in
training this CDNN model, so only the slices with liver, as well
as the 5 slices superior and inferior to the liver were included
in the model training. For liver localization and segmentation,
the liver and tumor labels were merged as a single liver label
to provide the ground truth liver masks during model training.
During testing, the new CT images were pre-processed
following the same procedure as training data preparation, then
the trained CDNN-I was applied to each slice of the entire CT
volume. Once all slices were segmented, a threshold of 0.5 was
applied to the output of CDNN and a 3D connect-component
labeling was performed. The largest connected component was
selected as the initial liver region.
C. Liver segmentation
An accurate liver localization enables us to perform a
fine liver segmentation with more advanced CDNN model
while reducing computational time. Specifically, we firstly
resampled the original image with slice thickness of 2 mm,
then the bounding-box of liver was extracted and expanded
by 10 voxels in x, y and z directions to create a liver
volume of interest (VOI). The axial dimensions of VOI were
further adjusted to 256 × 256 either by down-sampling if any
dimension was greater than 256, or by expanding in x and/or y
direction otherwise. All slices in the VOI were used for model
training.
The CDNN model used in the liver segmentation
(named CDNN-II) includes 29 layers with about 5 M trainable parameters. As compared to CDNN-I, the size of
local receptive f ield (LRF), or filter size, is reduced in
CDNN-II such that the network can go deeper, i.e. more
number of layers, which allows applying more non-linearities
and being less prone to overfitting [14]. Meanwhile, the
number of feature channels is doubled in each layer. Please
refer to [15] for more details.
During testing, liver VOI was extracted based on the initial
liver mask obtained in the liver localization step, then the
trained CDNN-II was applied to each slice in the VOI to
yield a 3D probability map of liver. We used the same postprocessing as liver localization to determine the final liver
mask.
D. Tumor segmentation
The VOI extraction in tumor segmentation was similar
to that in liver segmentation, except that the original image
resolution was used to avoid potentially missing small lesions
due to image blurring from resampling. Instead of using all the
slices in the VOI, we only collected those slices with tumor as
training data so as to focus the training on the liver lesions and
reduce training time. Besides the original image intensity, a
3D regional histogram equalization was performed to enhance
the contrast between tumors and surrounding liver tissues, in
which only those voxels within the 3D liver mask were considered in constructing intensity histogram. The enhanced image
served as an additional input channel to another CDNN-II
model for tumor segmentation. We found this additional input
channel could further boost tumor segmentation performance.
During testing, liver VOI was extracted based on the liver
mask from the liver segmentation step. A threshold of 0.5 was
applied to the output of CDNN-II model and liver tumors were
determined as all tumor voxels within the liver mask.
E. Implementation
Our CDNN models were implemented with Python based
on Theano [16] and Lasagne2 packages. The experiments were
conducted using a single Nvidia GTX 1060 GPU with 1280
cores and 6GB memory.
We used five-fold cross validation to evaluate the performance of our models on the challenge training datasets. The
total number of epochs was set as 200 for each fold. When
applying the trained models on the challenge testing datasets, a
bagging-type ensemble strategy was implemented to combine
the outputs of six models to further improve the segmentation
performance [8].
An epoch in training CDNN-I model for liver localization
took about 70 seconds, but the average time per epoch became
610 seconds and 500 seconds when training CDNN-II models
for liver segmentation and tumor segmentation, respectively.
This increase was primarily due to larger slice size and more
complicated CDNN models. Applying the entire segmentation
framework on a new test case was, however, very efficient,
taking about 33 seconds on average (8, 8 and 17 s for
liver localization, liver segmentation and tumor segmentation,
respectively).
IV. R ESULTS
AND
TABLE I
L IVER SEGMENTATION RESULTS ( DEEP X) ON L I TS TESTING CASES
Dice / case
0.9630
Dice global
0.9670
VOE
0.071
RVD
−0.010
ASSD
1.104
MSSD
23.847
RMSD
2.303
TABLE II
T UMOR SEGMENTATION RESULTS ( DEEP X) ON L I TS TESTING CASES
Dice / case
0.6570
Dice global
0.8200
VOE
0.378
RVD
0.288
ASSD
1.151
MSD
6.269
RMSD
1.678
D ISCUSSION
We applied the trained models to the 70 LiTS challenge test
cases (team: deepX). Based on the results from the challenge
organizers, our method achieved an average dice similarity
coefficient (DSC) of 0.963 for liver segmentation, a DSC of
0.657 for tumor segmentation, and a root mean square error
(RMSE) of 0.017 for tumor burden estimation, which ranked
our method in the first, fifth and third place, respectively. The
complete evaluation results are shown in Table I-III.
To summarize our work, we develop a fully automatic
framework for liver and its tumor segmentation on contrastenhanced abdominal CT scans based on three steps: liver
localization by a simple CDNN model (CDNN-I), liver fine
segmentation by a deeper CDNN model with doubled feature
2 http://github.com/Lasagne/Lasagne
channels in each layer (CDNN-II), and tumor segmentation by
CDNN-II model with enhanced liver region as additional input
feature. Our CDNN models are fully trained in an end-to-end
fashion with minimum pre- and post-processing efforts.
While sharing some similarities with previous work such
as U-Net [12] and Cascaded-FCN [9], our CDNN model is
different from them in the following aspects: 1) The loss
function used in CDNN model is based on Jaccard distance
that is directly related to image segmentation task while
eliminating the need of sample re-weighting; 2) Instead of
recovering image details by long skip connections as in UNet, the CDNN model constructs a deconvolutional path where
deconvolution is employed to densify the coarse activation
map obtained from up-sampling. In this way, feature map
concatenation and cropping are not needed.
Due to the limited hardware resource, training a complex
CDNN model is very time consuming and we had to restrict
the total number of epochs to 200 in order to catch the deadline
of LiTS challenge submission. While upgrading hardware is
clearly a way to speed up the model training, we plan to
improve our network architectures and learning strategies in
our future work such that the models can be trained in a more
effective and efficient way. Other post-processing methods,
such as level sets [17] and conditional random field (CRF)
[18], can also be potentially integrated into our model to
further improve the segmentation performance.
TABLE III
T UMOR BURDEN RESULTS ( DEEP X) ON L I TS TESTING CASES
RMSE
0.0170
Max Error
0.0490
R EFERENCES
[1] J. Ferlay, I. Soerjomataram, R. Dikshit, S. Eser, C. Mathers, M. Robelo,
D. Parkin, D. Forman, and F. Bray, “Cancer incidence and mortality
worldwide: sources, methods and major patterns in GLOBCAN 2012,”
Int. J. Cancer, vol. 136, no. 5, pp. E359–86, 2015.
[2] J. Petrick, M. Braunlin, M. Laversanne, P. Valery, F. Bray, and K. McGlynn, “International trends in liver cancer incidence, overall and by
histologic subtype, 1978-2007,” Int. J. Cancer, vol. 137, no. 7, pp.
1534–45, 2016.
[3] T. Heimann, B. van Ginneken, M. A. Styner, Y. Arzhaeva, V. Aurich,
C. Bauer, A. Beck, C. Becker, R. Beichel, G. Bekes, F. Bello, G. Binnig,
H. Bischof, A. Bornik, P. M. M. Cashman, Y. Chi, A. Cordova, B. M.
Dawant, M. Fidrich, J. D. Furst, D. Furukawa, L. Grenacher, J. Hornegger, D. Kainmuller, R. I. Kitney, H. Kobatake, H. Lamecker, T. Lange,
J. Lee, B. Lennon, R. Li, S. Li, H.-P. Meinzer, G. Nemeth, A.-M. R.
D. S. Raicu, E. M. van Rikxoort, M. Rousson, L. Rusko, K. A. Saddi,
G. Schmidt, D. Seghers, A. Shimizu, P. Slagmolen, E. Sorantin, G. Soza,
R. Susomboon, J. M. Waite, A. Wimmer, and I. Wolf, “Comparison and
evaluation of methods for liver segmentation from CT datasets,” IEEE
Trans. Med. Imaging, vol. 28, pp. 1251–1265, 2009.
[4] Y. Yuan, M. Chao, R. Sheu, K. Rosenzweig, and Y. Lo, “Tracking fuzzy
borders using geodesic curves with application to liver segmentation on
planning CT,” Med. Phys., vol. 42, pp. 4015–26, 2015.
[5] M. G. Linguraru, W. J. Richbourg, J. Liu, J. M. Watt, V. Pamulapati,
S. Wang, and R. M. Summers, “Tumor burden analysis on computed
tomography by automated liver and tumor segmentation,” IEEE Trans.
Med. Imaging, vol. 31, pp. 1965–1976, 2012.
[6] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” in Proc. CVPR, 2015, pp. 3431–3440.
[7] H. Noh, S. Hong, and B. Han, “Learning deconvolution network for
semantic segmentation,” in Proc. ICCV 2015, 2015, pp. 1520–1528.
[8] Y. Yuan, M. Chao, and Y.-C. Lo, “Automatic skin lesion segmentation
using deep fully convolutional networks with Jaccard distance,” IEEE
Trans. Med. Imaging, vol. 36, pp. 1876–1886, 2017.
[9] Christ, Patrick Ferdinand and Elshaer, Mohamed Ezzeldin A and Ettlinger, Florian and Tatavarty, Sunil and Bickel, Marc and Bilic, Patrick
and Rempfler, Markus and Armbruster, Marco and Hofmann, Felix and
DAnastasi, Melvin and others, “Automatic liver and lesion segmentation
in CT using cascaded fully convolutional neural networks and 3D
conditional random fields,” in International Conference on Medical
Image Computing and Computer-Assisted Intervention. Springer, 2016,
pp. 415–423.
[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Adv. Neural Inf. Process.
Sys., 2012, pp. 1097–1105.
[11] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” arXiv preprint
arXiv:1502.03167, 2015.
[12] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
for biomedical image segmentation,” in Proc. MICCAI 2015. Springer,
2015, pp. 234–241.
[13] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
[14] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[15] Y. Yuan, and Y.-C. Lo, “Improving dermoscopic image segmentation
with enhanced convolutional-deconvolutional networks,” arXiv preprint
arXiv:1709.09780, 2017.
[16] The Theano Development Team, R. Al-Rfou, G. Alain, A. Almahairi,
C. Angermueller, D. Bahdanau, N. Ballas, F. Bastien, J. Bayer, A. Belikov et al., “Theano: A python framework for fast computation of
mathematical expressions,” arXiv preprint arXiv:1605.02688, 2016.
[17] K. H. Cha, L. Hadjiiski, R. K. Samala, H.-P. Chan, E. M. Caoili, and
R. H. Cohan, “Urinary bladder segmentation in CT urography using
deep-learning convolutional neural network and level sets,” Med. Phys.,
vol. 43, no. 4, pp. 1882–1896, 2016.
[18] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille,
“Semantic image segmentation with deep convolutional nets and fully
connected CRFs,” arXiv preprint arXiv:1412.7062, 2014.
| 1cs.CV
|
Productivity of Non-Orthogonal Term Rewrite Systems
Matthias Raffelsieper
Department of Computer Science, TU Eindhoven
P.O. Box 513, 5600 MB Eindhoven, The Netherlands
eMail: [email protected]
Productivity is the property that finite prefixes of an infinite constructor term can be computed using a
given term rewrite system. Hitherto, productivity has only been considered for orthogonal systems,
where non-determinism is not allowed. This paper presents techniques to also prove productivity
of non-orthogonal term rewrite systems. For such systems, it is desired that one does not have to
guess the reduction steps to perform, instead any outermost-fair reduction should compute an infinite
constructor term in the limit. As a main result, it is shown that for possibly non-orthogonal term
rewrite systems this kind of productivity can be concluded from context-sensitive termination. This
result can be applied to prove stabilization of digital circuits, as will be illustrated by means of an
example.
1
Introduction
Productivity is the property that a given set of computation rules computes a desired infinite object. This
has been studied mostly in the setting of streams, the simplest infinite objects. However, as already
observed in [14], productivity is also of interest for other infinite structures, for example infinite trees, or
mixtures of finite and infinite structures. A prominent example of the latter are lists in the programming
language Haskell [10], which can be finite (by ending with a sentinel “[]”) or which can go on forever.
Existing approaches for automatically checking productivity, e.g., [2, 3, 14], are restricted to orthogonal systems. The main reason for this restriction is that it disallows non-determinism. A complete
computer program (i.e., a program and all possible input sequences, neglecting sources of true randomness)
always behaves deterministically, as the steps of computation are precisely determined. However, often a
complete program is not available, too large to be studied, or its inputs are provided by the user or they
are not specified completely. In this case, non-determinism can be used to abstract from certain parts by
describing a number of possible behaviors. In such a setting, the restriction to orthogonal systems, which
is even far stronger than only disallowing non-determinism, should be removed. An example of such a
setting are hardware components, describing streams of output values which are depending on the streams
of input values. To analyze such components in isolation, all possible input streams have to be considered.
This paper presents an extension of the techniques in [14] to analyze productivity of specifications that
may contain non-determinism. As in that work, the main technique to prove productivity is by analyzing
termination of a corresponding context-sensitive term rewrite system [7]. Here however, overlapping
rules are allowed and the data TRS is only required to be terminating, but it need not be confluent nor
left-linear. This technique can be used to prove stabilization of hardware circuits, which have external
inputs whose exact sequence of values is unknown. Thus, stabilization should be proven for all possible
input sequences, which are therefore abstracted to be random Boolean streams, i.e., arbitrary streams
containing the data values 0 and 1.
Structure of the Paper. In Section 2 we introduce proper specifications, which are the forms of rewrite
systems studied in this paper. After that, in Section 3, the different notions of productivity are discussed.
S. Escobar (Ed.): Workshop on Reduction Strategies
in Rewriting and Programming (WRS2011)
EPTCS 82, 2012, pp. 53–67, doi:10.4204/EPTCS.82.4
54
Productivity of Non-Orthogonal Term Rewrite Systems
For non-orthogonal specifications as studied in this paper, there exist both weak and strong productivity.
We will motivate that strong productivity is the notion that we are interested in, as it does guarantee a
constructor term to be reached by any outermost-fair reduction. The theoretical basis is laid in Section 4,
proving our desired result that termination of a corresponding context-sensitive TRS implies strong
productivity of a proper specification. Section 5 then applies this theory to an example hardware circuit,
checking that for a given circuit the output values always stabilize, regardless of the sequence of input
values. Finally, Section 6 concludes the paper.
2
Specifications
A specification gives the symbols and rules that shall be used to compute an intended infinite object. This
section gives a brief introduction to term rewriting, mainly aimed at fixing notation. For an in-depth
description of term rewriting, see for example [1, 11]. All symbols are assumed to have one of two
possible sorts. The first sort d is for data. Terms of this sort represent the elements in an infinite structure,
but which are not infinite terms by themselves. An example for data are the Booleans false and true
(which are also written 0 and 1), or the natural numbers represented in Peano form by the two constructors
0 and succ. The set of all terms of sort d is denoted Td (Σd , Vd ), where Σd is a set of function symbols all
having types of the form d m → d and where Vd is a set of variables all having sort d. The second sort is
the sort s for structure. Terms of this sort are to represent the intended structure containing the data and
therefore are allowed to be infinite. The set of all well-typed structure terms is denoted Ts (Σd ∪ Σs , V),
where Σs is disjoint from Σd and contains function symbols having types of the form d m × sn → s and
where V = Vd ∪ Vs for a set Vs of variables all having sort s, which is disjoint from Vd . We define the set of
all well-typed terms as T (Σd ∪ Σs , V) = Td (Σd , Vd ) ∪ Ts (Σd ∪ Σs , V) and denote the set of all ground terms,
i.e., terms not containing any variables, by T (Σd ∪ Σs ) = T (Σd ∪ Σs , ∅). A term t ∈ T (Σd ∪ Σs , V) of sort
ς ∈ {d, s} is either a variable, i.e., t ∈ Vς , or t = f (u1 , . . . , um ,t1 , . . . ,tn ) with f ∈ Σς of type d m × sn → ς
(where n = 0 if ς = d), u1 , . . . , um ∈ Td (Σd , Vd ), and t1 , . . . ,tn ∈ Ts (Σd ∪ Σs , V). In the latter case, i.e.,
when t = f (u1 , . . . , um ,t1 , . . . ,tn ), we define the root of the term t as root(t) = f .
A Term Rewrite System (TRS) over a signature Σ is a collection of rules (`, r) ∈ T (Σ, V)2 such that
`∈
/ V and every variable contained in r is also contained in `. As usual, we write ` → r instead of (`, r). A
term t ∈ T (Σ, V) rewrites to a term t 0 ∈ T (Σ, V) with the rule ` → r ∈ R, denoted t →`→r,p t 0 at position
p ∈ Pos(t), if a substitution σ exists such that t| p = `σ and t 0 = t[rσ ] p . A position is as usual a sequence
of natural numbers that identifies a number of argument positions taken to reach a certain subterm. The
notation t[rσ ] p represents the term t in which the subterm at position p, that is denoted by t| p , has been
replaced by the term rσ . This is the term r in which all variables have been replaced according to the
substitution σ , which is a map from variables to terms. It is allowed to only indicate the term rewrite
system R instead of the specific rule ` → r or to leave out the subscripts in case they are irrelevant or clear
from the context. The set of all normal forms of a TRS R over a signature Σ is denoted NF(R) and is
defined as NF(R) = {t ∈ T (Σ, V) | ∀t 0 ∈ T (Σ, V) : t 6→R t 0 }. The set of ground normal forms NFgnd (R)
additionally requires that all contained terms are ground terms, i.e., NFgnd (R) = NF(R) ∩ T (Σ).
We still have to impose some restrictions on specifications to make our approach work. These
restrictions are given below in the definition of proper specifications, which are similar to those of [14].
Definition 1. A proper specification is a tuple S = (Σd , Σs , C, Rd , Rs ), where Σd is the signature of data
symbols, each of type d m → d (then the data arity of such a symbol g is defined to be ard (g) = m), Σs is the
signature of structure symbols f , which have types of the shape d m × sn → s (and data arity ard ( f ) = m,
structure arity ars ( f ) = n), C ⊆ Σs is a set of constructors, Rd is a terminating TRS over the signature Σd ,
Matthias Raffelsieper
55
and Rs is a TRS over the signature Σd ∪ Σs , containing rules f (u1 , . . . , um ,t1 , . . . ,tn ) → t that satisfy the
following properties:
• f ∈ Σs \ C with ard ( f ) = m, ars ( f ) = n,
• f (u1 , . . . , um ,t1 , . . . ,tn ) is a well-sorted linear term,
• t is a well-sorted term of sort s, and
• for all 1 ≤ i ≤ n and for all p ∈ Pos(ti ) such that ti | p is not a variable and root(ti | p ) ∈ Σs , it holds
that root(ti | p0 ) ∈
/ C for all p0 < p (i.e., no structure symbol is below a constructor).
Furthermore, Rs is required to be exhaustive, meaning that for every f ∈ Σs \ C with ard ( f ) = m,
ars ( f ) = n, ground normal forms u1 , . . . , um ∈ NFgnd (Rd ), and terms t1 , . . . ,tn ∈ T (Σd ∪ Σs ) such that for
every 1 ≤ i ≤ n, ti = ci (u01 , . . . , u0k ,t10 , . . . ,tl0 ) with u0j ∈ NFgnd (Rd ) for 1 ≤ j ≤ k = ard (ci ) and ci ∈ C, there
exists at least one rule ` → r ∈ Rs such that ` matches the term f (u1 , . . . , um ,t1 , . . . ,tn ).
A proper specification S is called orthogonal, if Rd ∪ Rs is orthogonal, otherwise it is called nonorthogonal.
The above definition coincides with the definition of proper specifications given in [14] for orthogonal
proper specifications.1 We will illustrate the restrictions in the above definition later in Section 4. In
the following, all examples except for Example 18 will be using the domain of Boolean streams, where
C = {:} and Σd ⊇ {0, 1} with ard (0) = ard (1) = 0 and ard (:) = ars (:) = 1. In these examples, only a data
TRS Rd and a structure TRS Rs are given from which the remaining symbols in Σd and Σs and their
arities can be derived. If the data TRS Rd is not provided it is assumed to be empty.
3
Productivity
For orthogonal proper specifications, productivity is the property that every ground term t of sort s can,
in the limit, be rewritten to a possibly infinite term consisting only of constructors. This is equivalent
to stating that for every prefix depth k ∈ N, the term t can be rewritten to another term t 0 having only
constructor symbols on positions of depth k or less.
Definition 2. An orthogonal proper specification S = (Σd , Σs , C, Rd , Rs ) is productive, iff for every
ground term t of sort s and every k ∈ N, there is a reduction t →∗Rd ∪Rs t 0 such that every symbol of sort s
in t 0 on depth less or equal to k is a constructor.
Productivity of an orthogonal proper specification is equivalent to the following property, as was
shown in [14].
Proposition 3. An orthogonal proper specification S = (Σd , Σs , C, Rd , Rs ) is productive, iff for every
ground term t of sort s there is a reduction t →∗Rd ∪Rs t 0 such that root(t 0 ) ∈ C.
It was already observed in [4, 2] that productivity of orthogonal specifications is equivalent to the
existence of an outermost-fair reduction computing a constructor prefix for any given depth. Below, we
give a general definition of outermost-fair reductions, as they will also be used in the non-orthogonal
setting.
1
To see this, one should observe that a defined symbol cannot occur on a non-root position of a left-hand side. This holds
since otherwise the innermost such symbol would have variables and constructors as structure arguments and data arguments that
do not unify with any of the data rules (due to orthogonality), which therefore are normal forms and can be instantiated to ground
normal forms. Thus, exhaustiveness would require a left-hand side to match this term when instantiating all structure variables
with some terms having a constructor root, which would give a contradiction to non-overlappingness.
56
Productivity of Non-Orthogonal Term Rewrite Systems
Definition 4.
• A redex is a subterm t| p of a term t at position p ∈ Pos(t) such that a rule ` → r and a substitution
σ exist with t| p = `σ . The redex t| p is said to be matched by the rule ` → r.
• A redex is called outermost iff it is not a strict subterm of another redex.
• A redex t| p = `σ is said to survive a reduction step t →`0 →r0 ,q t 0 if p k q, or if p < q and t 0 = t[`σ 0 ] p
for some substitution σ 0 (i.e., the same rule can still be applied at p).
• A rewrite sequence (reduction) is called outermost-fair, iff there is no outermost redex that survives
as an outermost redex infinitely long.
• A rewrite sequence (reduction) is called maximal, iff it is infinite or ends in a normal form (a term
that cannot be rewritten further).
For non-orthogonal proper specifications, requiring just the existence of a reduction to a normal form
(or to a constructor prefix of arbitrary depth) does not guarantee the computation to reach it, due to the
possible non-deterministic choices. This can be observed for the term maybe in the following example.
Example 5. Consider a proper specification with the TRS Rs consisting of the following rules:
maybe → 0 : maybe
maybe →
maybe
random → 0 : random
random → 1 : random
This specification is not orthogonal, since the rules for maybe as well as those for random overlap.
We do not want to call this specification productive, since it admits the infinite outermost-fair reduction
maybe → maybe → . . . that never produces any constructors. However, there exists an infinite reduction
producing infinitely many constructors starting in the term maybe, namely maybe → 0 : maybe → 0 : 0 :
maybe → . . .. When only considering the rules for random then we want to call the resulting specification
productive, since no matter what rule of random we choose, an element of the stream is created.
Requiring just the existence of a constructor normal form is called weak productivity in [4, 2]. We
already stated above that this is not the notion of productivity we are interested in. The one we are
interested in is strong productivity, which is also defined in [4, 2], since it requires all reductions that
make progress on outermost positions to reach constructor normal forms.
Definition 6. A proper specification S is called strongly productive iff for every ground term t of sort s
all maximal outermost-fair rewrite sequences starting in t end in (i.e., have as limit for infinite sequences)
a constructor normal form.
It was observed in [4, 2] that weak and strong productivity coincide for orthogonal (proper) specifications. However, for non-orthogonal (proper) specifications this is not the case anymore. The rules for
maybe in Example 5 are not strongly productive, since they allow the infinite outermost-fair reduction
maybe → maybe → . . .. However, these rules are weakly productive, since any ground term can be
rewritten to an infinite stream containing only 0 elements after some finite prefix. For example, the ground
term 1 : maybe can be rewritten to the infinite stream 1 : 0 : 0 : . . ..
An example of a non-orthogonal proper specification that is both strongly and weakly productive
are the rules for random in Example 5, which always produce an infinite stream. In this case, the
restriction to outermost-fair reductions is not needed. However, if we add the rule id(xs) → xs and
replace the rule random → 1 : random by the rule random → id(1 : random), then the infinite reduction
random → id(1 : random) → id(1 : id(1 : random)) → . . . exists. This reduction is not outermost-fair
since the outermost redex id(. . .) survives infinitely often. When restricting to outermost-fair reductions,
then indeed an infinite stream of Boolean values is obtained for every such reduction, so this is a strongly
productive proper specification, too. Note that strong productivity implies weak productivity, so the
example is also weakly productive.
Matthias Raffelsieper
4
57
Criteria for Strong Productivity
For orthogonal proper specifications, it is sufficient to just consider reductions that create a constructor
at the top, as stated in Proposition 3. We will show next that this is also the case for non-orthogonal
proper specifications. However, in contrast to [14], here we have to consider all maximal outermost-fair
reductions, instead of just requiring the existence of such a reduction.
Proposition 7. A proper specification S = (Σd , Σs , C, Rd , Rs ) is strongly productive iff for every maximal
outermost-fair reduction t0 →Rd ∪Rs t1 →Rd ∪Rs . . . with t0 being of sort s there exists k ∈ N such that
root(tk ) ∈ C.
Proof. The “only if”-direction is trivial. For the “if”-direction, we show inductively that for every depth
z ∈ N and every maximal outermost-fair reduction ρ ≡ t0 → p0 t1 → p1 . . . there exists an index j ∈ N such
that for all positions p ∈ Pos(t j ) of sort s with |p| < z, root(t j | p ) ∈ C.
For z = 0, the index j can be set to 0, thus here the claim trivially holds. Otherwise, we get that
an index k ∈ N exists such that root(tk ) ∈ C. Let tk = c(u01 , . . . , u0m ,t10 , . . . ,tn0 ) with c ∈ C. Because c is
a constructor, we know that pl > ε for all l ≥ k. Define Pr = {p0i | pi = (m+r).p0i } for 1 ≤ r ≤ n (i.e.,
the positions in the maximal outermost-fair reduction that are occurring in structure argument r). Then,
for 1 ≤ r ≤ n and Pr = {pr0 , pr1 , . . . } the reduction tr0 = tr,0 → pr0 tr,1 → pr1 . . . is also a maximal outermostfair reduction, otherwise an infinitely long surviving outermost redex would also be an infinitely long
surviving outermost redex of the reduction ρ. By the induction hypothesis for z − 1 we get that indices
jr for 1 ≤ r ≤ n exist such that root(tr, jr | p ) ∈ C for all positions p ∈ Pos(tr, jr ) with |p| < z − 1. Since all
these reductions were taken from the original reduction, we define j = k + #d-red + ∑ni=1 ji , where #d-red
denotes the number of reductions performed in the data arguments of the constructor c such that p j = prjr
for the last r. This shows that the initial reduction ρ has the form t0 →∗ tk = c(u01 , . . . , u0m ,t10 , . . . ,tn0 ) →∗
c(u001 , . . . , u00m ,t100 , . . . ,tn00 ) = t j+1 , where tr, jr →∗ tr00 for every 1 ≤ r ≤ n. Since there are only constructors in
tr, jr for depths 0, . . . , z − 2, these constructors are still present in tr00 . This proves the proposition, since
c ∈ C and thus for all positions p ∈ Pos(t j ) of sort s with |p| < z we have root(t j | p ) ∈ C.
This characterization of strong productivity will be used in the remainder of the paper. Note that
it is similar to the requirements for infinitary strong normalization SN∞ observed in [12], where it is
found that for left-linear and finite term rewrite systems, SN∞ holds if and only if every infinite reduction
only contains a finite number of root steps. Thus, it could seem possible to define strong productivity of
proper specifications by requiring that every reduction starting in a finite ground term is infinitary strongly
normalizing, i.e., SN∞ holds for the relation →Rd ∪Rs ∩ T (Σd ∪ Σs )2 . However, this is not the case, as the
following example shows.
Example 8. Consider the proper specification containing the following TRS Rs :
a → f(a)
f(x : xs) → x : xs
This TRS has the property SN∞ , intuitively because either the symbol f remains at the root position
and can never be rewritten again (in case the first rule is applied), or the constructor : is created at the root.
Formally, this can for example be proven by the technique presented in [12]: Let Σ# = Σ ] {g# | g ∈ Σ},
where Σ = {0, 1, :, a, f} is the signature of the specification. Then we choose the finite weakly monotone Σ#
algebra ({0, 1, 2}, [·], ⊥, >, ≥), where ⊥ = 0, [0] = 0, [1] = 0, [a] = 1, [f](n) = n, [:](m, n) = min{m+n, 2},
[a# ] = 2, [ f# ](n) = 1, and [:# ](m, n) = 0 for m, n ∈ {0, 1, 2} and > and ≥ are the natural comparison
operators on the numbers {0, 1, 2}. It is easy to check that this algebra is indeed weakly monotone
58
Productivity of Non-Orthogonal Term Rewrite Systems
(i.e., that > is well-founded, > · ≥ ⊆ > ⊆ ≥, and for every g ∈ Σ# , the operation [g] is monotone with
respect to ≥). Additionally, the requirements of the combination of [12, Theorem 5 and Theorem 6]
are satisfied, i.e., {0, 1, 2} is finite, ≥ is transitive, a ≥ b implies a > b or a = b, a ≥ ⊥ = 0 for all
a, b ∈ {0, 1, 2}, and [` σ ] ≥ [r σ ] and [`# σ ] > [r# σ ] for all ` → r ∈ Rs and all substitutions σ , where
g(t1 , . . . ,tk )# = g# (t1 , . . . ,tk ). This proves SN∞ of →Rs , which especially entails SN∞ of the relation
→Rs ∩ T (Σd ∪ Σs )2 .
However, the above proper specification is not strongly productive, since the infinite outermostfair reduction a →Rs f(a) →Rs f(f(a)) →Rs . . ., continued by repeatedly reducing the symbol a, never
produces any constructors.
The above example shows that even though we require exhaustiveness of proper specifications, this
exhaustiveness only refers to constructor terms, i.e., the objects we are interested in, and not to arbitrary
terms. A similar observation, namely that top termination is not equivalent to productivity, was already
made in [15].
A first technique to prove strong productivity of proper specifications is given next. It is a simple
syntactic check that determines whether every right-hand side of sort s starts with a constructor. For
orthogonal proper specifications, this was already observed in [14]. It has to be proven again since here we
consider strong productivity, which requires all possible outermost-fair reductions to reach a constructor
normal form, instead of weak productivity as in [14], for which only a single reduction to a constructor
normal form needs to be constructed.
Theorem 9. Let S = (Σd , Σs , C, Rd , Rs ) be a proper specification. If for all rules ` → r ∈ Rs we have
root(r) ∈ C, then S is strongly productive.
Proof. Let ρ ≡ t0 → p0 t1 → p1 . . . be a maximal outermost-fair reduction and let t0 = f (u01 , . . . , u0m ,t10 , . . . ,tn0 ).
If f ∈ C we are done, so we assume f ∈ Σs \ C and perform structural induction on t0 to prove that
root(tk ) ∈ C for some k ∈ N.
From the induction hypothesis we get that for every 1 ≤ i ≤ n and every maximal outermost-fair
reduction ti0 = ti,0 → ti,1 → . . . there exists an index ki ∈ N such that root(ti,ki ) ∈ C.
Assume that for all j ∈ N, p j 6= ε. As in the proof of Proposition 7, we therefore again obtain maximal
outermost-fair reductions ti0 → . . . , thus we get indices ki ∈ N such that root(ti,ki ) ∈ C, as explained above.
This makes our reduction ρ have the shape t0 = f (u01 , . . . , u0m ,t10 , . . . ,tn0 ) →∗ f (u001 , . . . , u00m ,t100 , . . . ,tn00 ) = t j
for some j ∈ N, where u001 , . . . , u00m ∈ NFgnd (Rd ) (since the reduction ρ is maximal outermost-fair and Rd
is terminating) and ti,ki →∗ ti00 , thus also root(ti00 ) ∈ C. Because Rs is exhaustive, we get that t j contains a
redex at the root position ε, which of course is outermost. This gives rise to a contradiction to ρ being
outermost fair, as this outermost redex survives infinitely often, because p j 6= ε for all j ∈ N. Therefore,
p j = ε for some j ∈ N and the reduction has the shape t0 →∗ t j →ε rσ , where the last step is with respect
to some rule ` → r ∈ Rs . By the assumption on the shape of the rules in Rs , we have root(r) ∈ C, hence
also root(rσ ) ∈ C, which proves productivity according to Proposition 7.
This technique is sufficient to prove strong productivity of the proper specification consisting of the
two rules for random in Example 5, since both have right-hand sides with the constructor : at the root.
However, it is easy to create examples which are strongly productive, but do not satisfy the syntactic
requirements of Theorem 9.
Example 10. Consider the proper specification with the following TRS Rs :
ones → 1 : ones
finZeroes → 0 : 0 : ones
f(0 : xs) → f(xs)
finZeroes → 0 : ones
finZeroes → 0 : 0 : 0 : ones
f(1 : xs) → 1 : f(xs)
Matthias Raffelsieper
59
The constant finZeroes produces non-deterministically a stream that starts with one, two, or three
zeroes followed by an infinite stream of ones. Function f takes a binary stream as argument and filters out
all occurrences of zeroes. Thus, productivity of this example proves that only a finite number of zeroes
can be produced. This however cannot be proven with the technique of Theorem 9, since the right-hand
side of the rule f(0 : xs) → f(xs) does not start with the constructor :.
Another technique presented in [14] to show productivity of orthogonal proper specifications is
based on context-sensitive termination [7]. The idea is to disallow rewriting in structure arguments of
constructors, thus context-sensitive termination implies that for every ground term of sort s, a term starting
with a constructor can be reached (due to the exhaustiveness requirement). As was observed by Endrullis
and Hendriks recently in [5], this set of blocked positions can be enlarged, making the approach even
stronger.
Below, the technique for proving productivity by showing termination of a corresponding contextsensitive TRS is extended to also be applicable in the case of our more general proper specifications. This
version already includes an adaption of the improvement mentioned above.
Definition 11. Let S = (Σd , Σs , C, Rd , Rs ) be a proper specification. The replacement map µS : Σd ∪ Σs →
2N is defined as follows: 2
• µS ( f ) = {1, . . . , ard ( f )}, if f ∈ Σd ∪ C
• µS ( f ) = {1, . . . , ard ( f ) + ars ( f )} \ {1 ≤ i ≤ ard ( f ) + ars ( f ) | t|i is a variable for all ` → r ∈ Rs and
all non-variable subterms t of ` with root(t) = f },3 otherwise
In the remainder, we leave out the subscript S if the specification is clear from the context. The
replacement map µ is used to define the set of allowed positions of a non-variable term t as Posµ (t) = {ε}∪
{i.p | i ∈ µ(root(t)), p ∈ Posµ (t|i )} and the set of blocked positions of t as blockedµ (t) = Pos(t) \ Posµ (t).
Context-sensitive rewriting [7] then is the restriction of the rewrite relation to those redexes on positions
µ
from Posµ . Formally, we have t →`→r,p t 0 iff t →`→r,p t 0 and p ∈ Posµ (t) and we say a TRS R is
µ
µ-terminating iff no infinite →R -chain exists.
The replacement map µS is canonical [8] for the left-linear TRS Rs , guaranteeing through the second
condition of the above Definition 11 that non-variable positions of left-hand sides are allowed. In that
definition, the replacement map µS is extended to the possibly non-left-linear TRS Rd ∪ Rs by allowing
all arguments of symbols from Σd .
Our main result of this paper is that also for possibly non-orthogonal proper specifications, µtermination implies productivity.
Theorem 12. A proper specification S = (Σd , Σs , C, Rd , Rs ) is strongly productive, if Rd ∪ Rs is µS terminating.
Before proving the above theorem, we will show first that it subsumes Theorem 9. Intuitively, this
holds because structure arguments of constructors are blocked, and if every right-hand side of Rs starts
with a constructor then the number of allowed redexes of sort s in a term steadily decreases.
2
Note that in [5], Endrullis and Hendriks consider orthogonal TRSs and also block arguments of symbols in Σd which only
contain variables. This however is problematic when allowing data rules that are not left-linear. Example:
Rs :
Rd :
f(1)
d(x, x)
→
→
f(d(0, d(1, 0)))
1
f(0)
d(0, x)
→
→
0 : f(0)
0
d(1, x)
→
0
Here, the term f(d(0, d(1, 0))) can only be µ-rewritten to the term f(0) (which then in turn has to be rewritten to 0 : f(0)) if
defining µ(d) = {1}, since the subterm d(1, 0) can never be rewritten to 0. However, the example is not strongly productive,
as reducing in this way gives rise to an infinite outermost-fair reduction f(d(0, d(1, 0))) → f(d(0, 0)) → f(1) → . . . . Blocking
arguments of data symbols can only be done when Rd is left-linear.
3
The requirement of t not being a variable ensures that root(t) is defined.
60
Productivity of Non-Orthogonal Term Rewrite Systems
Proposition 13. Let S = (Σd , Σs , C, Rd , Rs ) be a proper specification. If for all rules ` → r ∈ Rs we have
root(r) ∈ C, then Rd ∪ Rs is µS -terminating.
Proof. Let t ∈ T (Σd ∪ Σs , V) be well-typed. If t has sort d, then all subterms must also be of sort d, as
symbols from Σd only have arguments of that sort. Hence, rewriting can only be done with rules from Rd ,
which is assumed to be terminating.
µ
Otherwise, let t be of sort s and assume that t starts an infinite µ-reduction t = t0 →`0 →r0 ,p0
µ
µ
0
0
s 0
t1 →`1 →r1 ,p1 t2 →`2 →r2 ,p2 . . . . We define Posred
µ (t ) = {p ∈ Posµ (t ) |µ t | p is a redex of sort s} for
any term t 0 ∈ T (Σd ∪ Σs , V). It will be proven that in every step ti →`i →ri ,pi ti+1 of the infinite rereds
reds
s
duction, |Posred
µ (ti+1 )| ≤ |Posµ (ti )| and that for steps with `i → ri ∈ Rs , we even have |Posµ (ti+1 )| <
s
|Posred
µ (ti )|. To this end, case analysis of the rule `i → ri is performed. If `i → ri ∈ Rd , then ti = ti [`i σi ] pi
reds
s
and ti+1 = ti [ri σi ] pi for some substitution σi . Because `i , ri ∈ T (Σd , V), |Posred
µ (`i σi )| = |Posµ (ri σi )| = 0
reds
s
since all symbols in Σd have arguments of sort d. Thus, Posred
µ (ti+1 ) = Posµ (ti ). In the second
s
case, `i → ri ∈ Rs . Let ti = ti [`i σi ] pi and ti+1 = ti [ri σi ] pi for some substitution σi . Then, Posred
µ (ti ) =
reds
reds
Posµ (ti [z] pi ) ] {pi .p | p ∈ Posµ (ti | pi )} for any variable z ∈ V of sort s. For ti+1 we observe that
reds
reds
reds
s
Posred
µ (ti+1 ) = Posµ (ti [ri σi ] pi ) = Posµ (ti [z] pi ) ] {pi .p | p ∈ Posµ (ti [ri σi ] pi | pi )} for any variable z ∈
reds
reds
s
V of sort s. Here, it holds that Posred
µ (ti | pi ) = Posµ (`i σi ) 3 ε, therefore pi ∈ Posµ (ti ). Furtherreds
reds
more, Posµ (ti [ri σi ] pi | pi ) = Posµ (ri σi ) = ∅, since root(ri ) ∈ C by assumption, hence µ(root(ri )) =
s
{1, . . . , ard (root(ri ))} and because symbols from Σd only have arguments of sort d. Thus, Posred
µ (ti+1 ) (
s
Posred
µ (ti ).
Combining these observations, we therefore only have finitely many reductions with rules from Rs in
the infinite reduction. Thus, an infinite tail of steps with rules from Rd exists. This however contradicts
the assumption that Rd is terminating, hence no infinite µ-reduction can exist which proves µ-termination
of Rd ∪ Rs .
Hence, we could restrict ourselves to analyzing context-sensitive termination only. However, the
syntactic check of Theorem 9 can be done very fast and should therefore be the first method to try.
In order to prove Theorem 12 we will show that a maximal outermost-fair reduction that never reaches
a constructor entails an infinite µ-reduction. For this purpose we need the following lemma, which shows
that in every ground term not starting with a constructor there exists a redex that is not blocked by the
replacement map µ.
Lemma 14. Let S = (Σd , Σs , C, Rd , Rs ) be a proper specification. For all ground terms t of sort s with
root(t) ∈
/ C there exists a position p ∈ Posµ (t) such that t → p .
Proof. Let t = f (u1 , . . . , um ,t1 , . . . ,tn ). We perform structural induction on t. If ui → p0 for some 1 ≤ i ≤ m
with i ∈ µ( f ), then t →i.p0 and i.p0 ∈ Posµ (t) since arguments of data symbols are never blocked. Thus,
we assume in the remainder that ui ∈ NFgnd (Rd ) for all 1 ≤ i ≤ m with i ∈ µ( f ). If root(ti ) ∈ C for
all 1 ≤ i ≤ n, i ∈ µ( f ), then t →ε by the exhaustiveness requirement (and because all arguments u j , t j
with j ∈
/ µ( f ) are being matched by pairwise different variables, due to left-linearity). Otherwise, there
exists 1 ≤ i ≤ n, i ∈ µ( f ) such that root(ti ) ∈
/ C. By the induction hypothesis we get that ti → p0 for some
0
0
p ∈ Posµ (ti ). Therefore, we also have i.p ∈ Posµ (t) and t →i.p0 .
A second lemma that is required for the proof of Theorem 12 states that a specialized version of
the Parallel Moves Lemma [1, Lemma 6.4.4] holds for our restricted format of term rewrite systems.
It allows us to swap the order of reductions blocked by µ with reductions not blocked by µ. To
q
formulate the lemma, we need the notion of a parallel reduction step t →P t 0 , which is defined for
Matthias Raffelsieper
61
a set P = {p1 , . . . , pn } ⊆ Pos(t) such that for every pair 1 ≤ i < j ≤ n we have pi k p j and a term
t = t[`1 σ1 ] p1 . . . [`n σn ] pn as t 0 = t[r1 σ1 ] p1 . . . [rn σn ] pn for rules `i → ri ∈ Rd ∪ Rs and substitutions σi ,
1 ≤ i ≤ n.
Lemma 15. Let S = (Σd , Σs , C, Rd , Rs ) be a proper specification. For all ground terms t,t 0 ,t 00 and
q
positions p ∈ Posµ (t 0 ), P ⊆ blockedµ (t) with t →P t 0 →`→r,p t 00 , a term tˆ and a set P0 ⊆ Pos(tˆ) exist such
q
that t →`→r,p tˆ →P0 t 00 .
Proof. Let P = {p1 , . . . , pk } ⊆ blockedµ (t). Then t = t[`1 σ1 ] p1 . . . [`k σk ] pk →P t[r1 σ1 ] p1 . . . [rk σk ] pk = t 0 =
t 0 [`σ ] p for some rules `1 → r1 , . . . , `k → rk , ` → r ∈ Rd ∪ Rs and substitutions σ1 , . . . , σk , σ . W.l.o.g.,
let 0 ≤ j ≤ k be such that pi 6k p for all 1 ≤ i ≤ j and pi k p for all j < i ≤ k. Since p ∈ Posµ (t 0 ) and
0
0
pi ∈ blocked
µ (t ), it must hold that
p < pi for all 1 ≤ i ≤ j. Therefore, the term t must have the shape
0
t = t `σ [r1 σ1 ] p1 −p . . . [r j σ j ] p j −p p [r j+1 σ j+1 ] p j+1 . . . [rk σk ] pk .
If ` → r ∈ Rd , then it must hold that j = 0, since arguments of data symbols are never blocked. Hence,
the lemma trivially holds in this case, as all reductions are on independent positions.
Otherwise, ` → r ∈ Rs . Because the positions pi for 1 ≤ i ≤ j are blocked, it must be the case that they
are either below a variable in all rules containing a certain symbol f (hence, they are also below a variable
in `), or they are below a structure argument of a constructor c ∈ C. By requirement of specifications, if a
constructor is present on a left-hand side of a rule, all its structure arguments must be variables. Thus, we
conclude that all positions pi , and thereby all terms ri σi , are below some variable of ` in t 0 . Additionally,
the left-hand side ` is required to be linear, therefore there exist pairwise different variables x1 , . . . , x j ,
contexts C1 , . . . ,C j , and a substitution σ 0 being like σ except that σ 0 (xi ) = xi for 1 ≤ i ≤ j such that:
q
t 0 = t [`σ 0 {x1 :=C1 [r1 σ1 ], . . . , x j :=C j [r j σ j ]}] p [r j+1 σ j+1 ] p j+1 . . . [rk σk ] pk
→ p t [rσ 0 {x1 :=C1 [r1 σ1 ], . . . , x j :=C j [r j σ j ]}] p [r j+1 σ j+1 ] p j+1 . . . [rk σk ] pk = t 00
We conclude that p ∈ Posµ (t), as all reduction steps in t →P t 0 are either below or independent of p.
Thus:
t = t [`σ 0 {x1 :=C1 [`1 σ1 ], . . . , x j :=C j [` j σ j ]}] p [` j+1 σ j+1 ] p j+1 . . . [`k σk ] pk
→ p t [rσ 0 {x1 :=C1 [`1 σ1 ], . . . , x j :=C j [` j σ j ]}] p [` j+1 σ j+1 ] p j+1 . . . [`k σk ] pk = tˆ
q
→P0 t [rσ 0 {x1 :=C1 [r1 σ1 ], . . . , x j :=C j [r j σ j ]}] p [r j+1 σ j+1 ] p j+1 . . . [rk σk ] pk = t 00
q
In the second reduction step, the positions of the terms `i σi in tˆ constitute the set P0 ⊆ Pos(tˆ).
We are now able to prove our main theorem, showing that context-sensitive termination implies
productivity of the considered proper specification.
Proof of Theorem 12. Assume S = (Σd , Σs , C, Rd , Rs ) is not strongly productive. Then, a maximal
outermost-fair reduction sequence ρ ≡ t0 → t1 → . . . exists where for all k ∈ N, root(tk ) ∈
/ C.
This reduction sequence is infinite, since otherwise it would end in a term tm for some m ∈ N with
root(tm ) ∈
/ C. Then however, according to Lemma 14, the term tm would contain a redex, giving a
contradiction to the sequence being maximal.
The sequence might however perform reductions that are below a variable argument of a constructor
or below a variable in all left-hand sides of a defined symbol. These reduction steps are not allowed when
considering context-sensitive rewriting with respect to µ. Such reductions however can be reordered.
First, we observe that there is always a redex which is not blocked, due to Lemma 14, thus there is also an
outermost such one. Because the reduction is outermost-fair, and because reductions below a variable
cannot change the matching of a rule, as shown in Lemma 15, such redexes must be contracted an infinite
62
Productivity of Non-Orthogonal Term Rewrite Systems
number of times in the infinite reduction sequence ρ. Thus, we can reorder the reduction steps in ρ: If
there is a (parallel) reduction below a variable before performing a step that is allowed by µ, then we
swap these two steps using Lemma 15. Repeating this, we get an infinite reduction sequence ρ 0 consisting
of steps which are not blocked by µ. Thus, this is an infinite µ-reduction sequence, showing that Rd ∪ Rs
is not µ-terminating, which proves the theorem.
The technique of Theorem 12, i.e., proving µ-termination of the corresponding context-sensitive TRS,
is able to prove strong productivity of Example 10. By Definition 11, the corresponding replacement
map µ is defined as µ(0) = µ(1) = µ(ones) = µ(finZeroes) = ∅ and µ(f) = µ(:) = {1}, i.e., rewriting
is allowed on all positions except those that are inside a second argument of the constructor :. Contextsensitive termination of the TRS together with the above replacement map µ can for example be shown
by the tool AProVE [6]. Thus, productivity of that example has been shown according to Theorem 12.
Also, strong productivity of the proper specification consisting of the rules random → 0 : random,
random → id(1 : random), and id(xs) → xs can be proven using Theorem 12 and the tool AProVE [6],
where µ(0) = µ(1) = µ(random) = µ(id) = ∅ and µ(:) = {1} according to Definition 11. Note that for
this example, one could also have used µ(id) = {1}, i.e., here the removal of argument positions in the
second item of Definition 11 is irrelevant.
This is not the case in the next example, showing that this improvement, which was inspired by [5]
and blocks more argument positions, allows to prove productivity of specifications where this would
otherwise not be possible.
Example 16. Consider the following proper specification, given by the TRS Rs :
a → f(1 : a, a)
f(x : xs, ys) → x : ys
f(f(xs, ys), zs) → f(xs, f(ys, zs))
When defining µ(1) = µ(a) = ∅ and µ(:) = {1} by the first case of Definition 11, and defining
µ(f) = {1, 2} (i.e., not removing any argument positions, as was done in the orthogonal case in [14]),
µ
µ
µ
then an infinite µ-reduction exists: a → f(1 : a, a) → f(1 : a, f(1 : a, a)) → . . .
This reduction can be continued in the above style by reducing the underlined redex further, which will
always create the term a on an allowed position of the form 2n . However, such positions are not required
for any of the f-rules to be applicable; for both rules it holds that all subterms of left-hand sides that start
with the symbol f, which are the terms f(x : xs, ys), f(f(xs, ys), zs), and f(xs, ys), have a variable as second
argument. Thus, according to Definition 11, the replacement map µ 0 can be defined to be like µ, except
that µ 0 (f) = {1}. With this improved replacement map, µ 0 -termination of the above TRS can for example
be proven by the tool AProVE [6], which implies productivity by Theorem 12.
Checking productivity in this way, i.e., by checking context-sensitive termination, can only prove
productivity but not disprove it. This is illustrated in the next example.
Example 17. Consider the proper specification with the following rules in Rs :
a → f(a)
f(x : xs) → x : f(xs)
f(f(xs)) → 1 : xs
Starting in the term a, we observe that an infinite µ-reduction starting with a → f(a) exists, which can
be continued by reducing the underlined redex repeatedly, since µ(f) = {1}. Thus, the example is not
µ-terminating. However, the specification is productive, as can be shown by case analysis based on the
root symbol of some arbitrary ground term t. In case root(t) = :, then nothing has to be done, according
to Proposition 7. Otherwise, if root(t) = a, then any maximal outermost-fair reduction must start with
Matthias Raffelsieper
63
t = a → f(a), thus we can reduce our analysis to the final case, where root(t) = f. In this last case, t = f(t 0 ).
Due to the rules for the symbol f, we have to perform a further case analysis based on the root symbol of
t 0 . If root(t 0 ) = :, i.e., t 0 = u : t 00 for some terms u and t 00 , then this constructor cannot be reduced further.
Also, t = f(u : t 00 ) is a redex, due to the second rule. Hence, in any maximal outermost-fair reduction
sequence this redex must eventually be reduced using the second rule, which results in a term with the
constructor : at the root. For root(t 0 ) = a we again must reduce t = f(a) → f(f(a)). Finally, in case
root(t 0 ) = f, we have two possibilities. The first one occurs when the term t 0 is eventually reduced at the
root. Since root(t 0 ) = f, this has to happen with either of the f-rules, creating a constructor : which, as we
already observed, must eventually result in the term t also being reduced to a term with the constructor :
at the root. Otherwise, in the second possible scenario, the term t 0 is never reduced at the root. Then
however, an outermost redex of the shape f(f(t 00 )) exists in all terms that t can be rewritten to in this way,
thus it has to be reduced eventually with the third rule. This again creates a term with constructor : at the
root. Combining all these observations, we see that in every maximal outermost-fair reduction there exists
a term with the constructor : as root symbol, which proves productivity due to Proposition 7.
In the remainder of this section we want to illustrate the requirements of proper specifications in
Definition 1, namely that the TRS Rs should be left-linear and that structure arguments of constructors in
left-hand sides must not be structure symbols, i.e., they must be variables. We begin with an example
specification that is not left-linear and not productive, but µ-terminating.
Example 18. We consider the non-proper specification S = (Σd , Σs , C, Rd , Rs ) with Σd = Rd = ∅, C =
{a, c} ⊆ Σs = {a, b, c, f}, and the following rules in Rs which also imply the arities of the symbols:
b → a
f(a) → a
f(c(x, x)) → f(c(a, b))
f(c(x, y)) → c(x, y)
The example specification is not productive, as it admits the infinite outermost-fair reduction sequence
f(c(a, a)) → f(c(a, b)) → f(c(a, a)) → . . . . However, the TRS is µ-terminating, as shown by the tool
AProVE [6], where µ(f) = {1} and µ(a) = µ(b) = µ(c) = ∅. This is the case because rewriting below
the constructor c is not allowed, thus the second step of the above reduction sequence is blocked. The
reason why Theorem 12 fails is the reordering of reductions, since in this example a reduction of the form
q
q
t →P t 0 →`→r,p t 00 (here: f(c(a, b)) →{1.1} f(c(a, a)) →f(c(x,x))→f(c(a,b)),ε f(c(a, b))) does not imply that
t →`→r,p (in the example, f(c(a, b)) 6→f(c(x,x))→f(c(a,b)),ε ), i.e., Lemma 15 does not hold.
The next example illustrates why non-variable structure arguments of constructors are not allowed in
left-hand sides.
Example 19. Let Rs contain the following rules:
ones → 1 : ones
f(x : y : xs) → f(y : xs)
f(x : xs) → x : xs
Here, we have non-productivity of the corresponding non-proper specification due to the infinite
outermost-fair reduction sequence f(ones) →1 f(1 : ones) →1.2 f(1 : 1 : ones) →ε f(1 : ones) → . . .,
however the second step is not allowed when performing context-sensitive rewriting, since µ(:) = {1}.
Using the tool AProVE [6], context-sensitive termination of the above TRS together with the replacement
map µ can be shown.
We can however unfold this example (cf. [5, 13]), which makes the resulting specification proper, by
introducing a fresh symbol g and replacing the two rules for f with the following three rules:
f(x : xs) → g(x, xs)
g(x, y : xs) → f(y : xs)
g(x, xs) → x : xs
Then, in the corresponding context-sensitive TRS, we have µ(f) = µ(:) = {1}, µ(g) = {2}, and
µ(ones) = µ(0) = µ(1) = ∅. This context-sensitive TRS is not µ-terminating, since it admits the infinite
µ
µ
µ
µ
µ
reduction f(ones) →1 f(1 : ones) →ε g(1, ones) →2 g(1, 1 : ones) →ε f(1 : ones) → . . . .
64
Productivity of Non-Orthogonal Term Rewrite Systems
QN
n2
D
Q
SI
next
n1
SE
CK
Figure 1: Example hardware circuit
It should be noted that the restriction for left-hand sides to only contain variables in constructor
arguments was already made in [14]. This is the case because matching constructors nested within
constructors would otherwise invalidate the approach of disallowing rewriting inside structure arguments
of constructors.
5
Application to Hardware Circuits
Proving productivity can be used to verify stabilization of hardware circuits. In such a circuit, the inputs
can be seen as an infinite stream of zeroes and ones, which in general can occur in any arbitrary sequence.
Furthermore, a circuit contains a number of internal signals, which also carry different Boolean values
over time.
To store a value over time, feedback loops are used. In such a loop, a value that is computed from
some logic function is also used as an input to that function. Thus, it is desired that such values stabilize,
instead of oscillating infinitely.
To check this, productivity analysis can be used. We will illustrate this by means of an example, that
will be considered throughout the rest of this section.
Consider the circuit shown in Figure 1, which was constructed from the transistor netlist of the cell
SDFF X1 in the Nangate Open Cell Library [9] and which implements a scanable D flip-flop. This circuit
first selects, based on the value of the input SE (scan enable), either the negation of the data input D (in
case SE=0) or the negation of the scan data input SI (in case SE=1). This value, called next in Figure 1,
is then fed into another multiplexer (mux), for which a feedback loop exists. This mux is controlled by the
negation of the clock input CK. If the clock is 0 then the negated value of next is forwarded to the output
n1, otherwise the stored value of n1 is kept. Similarly, n2 implements such a latch structure, however this
time the latch forwards the negation of the n1 input in case CK is 1, and it keeps its value when CK is 0.
The outputs Q and QN are computed from this stored value n2.
Note that a lot of the negations are only contained to refresh the signals, otherwise a high voltage
value might decay and not be detected properly anymore.
From the example circuit, we create a proper specification, where the data symbols consist of the two
Boolean values 0 and 1 and the symbol not used for negating:
not(0) → 1
not(1) → 0
Matthias Raffelsieper
65
rand → 0 : rand
rand → 1 : rand
next(0 : ses, d : ds, si : sis) → not(d) : next(ses, ds, sis)
next(1 : ses, d : ds, si : sis) → not(si) : next(ses, ds, sis)
n1(0 : cks, nextv : nexts, n1l) → not(nextv) : n1(cks, nexts, not(nextv))
n1(1 : cks, nextv : nexts, n1l) → n10 (cks, nexts, n1l, not(not(n1l)))
n10 (cks, nexts, 0, 0) → 0 : n1(cks, nexts, 0)
n10 (cks, nexts, 1, 1) → 1 : n1(cks, nexts, 1)
n10 (cks, nexts, 0, 1) → n10 (cks, nexts, 1, not(not(1)))
n10 (cks, nexts, 1, 0) → n10 (cks, nexts, 0, not(not(0)))
n2(0 : cks, n1v : n1s, n2l) → n20 (cks, n1s, n2l, not(not(n2l)))
n2(1 : cks, n1v : n1s, n2l) → not(n1v) : n2(cks, n1s, not(n1v)))
n20 (cks, n1s, 0, 0) → 0 : n2(cks, n1s, 0)
n20 (cks, n1s, 1, 1) → 1 : n2(cks, n1s, 1)
n20 (cks, n1s, 0, 1) → n20 (cks, n1s, 1, not(not(1)))
n20 (cks, n1s, 1, 0) → n20 (cks, n1s, 0, not(not(0)))
q(n2v : n2s) → not(n2v) : q(n2s)
qn(qv : qs) → not(qv) : qn(qs)
Figure 2: Structure TRS Rs for the circuit shown in Figure 1
The structures we are interested in are infinite streams containing Boolean values, thus the set of
constructors is C = {:}. The structure TRS Rs is shown in Figure 2.
It should be remarked that in the shown rules, some simplifications regarding the clock input CK have
been made. The inverters for the clock have been removed, and the two muxes that output the signals n1
and n2 are provided with decoupled clock values.
The defined function symbols next, n1, n2, q, and qn reflect the wires and output signals with the
corresponding name in Figure 1. The constant rand is added to abstract the values of the inputs. It
provides a random stream of Boolean values, thus it is able to represent any sequence of input values
provided to the circuit. The rules of the symbol next implement the mux selecting either the next data
input value d in case the next scan enable input value se is 0, or the next scan input value si in case se is 1.
The output of n1 is also computed by a mux, however, here the previous output value has to be
considered due to the feedback loop. We break the cycle by introducing a new parameter n1l that stores
the previously output value. Then, the next value of the stream at n1 is computed from the next value of
the clock ck, the input stream nextv : nexts coming from the previously described multiplexer, and from
66
Productivity of Non-Orthogonal Term Rewrite Systems
the previous output value n1l. If the clock ck is 0, then the latch simply outputs the negated value of nextv
and continues on the remaining streams, setting the parameter n1l to this value to remember it. Otherwise,
if ck is 1, then the feedback loop is active and has to be evaluated until it stabilizes. This is done by the
function n10 . It has as arguments the remaining input stream of the clock, the remaining input stream
of the scan multiplexer, and the previous output value and the newly computed output value. If both of
these values are the same, then the value of the wire n1 has stabilized and hence can be output. The tail of
the output stream is computed by again calling the function n1 with the remaining streams for the clock
and the scan multiplexer. Otherwise, the new output value (the last argument of n10 ) differs from the old
output value (the penultimate argument of n10 ). In that case, the new output value becomes the old output
value and the new output is recomputed. This is repeated until eventually the output value stabilizes, or it
will oscillate and never produce a stable output.
Similar to the function n1, the function n2 computes stable values for the corresponding wire in
Figure 1. Again, the parameter n2l is added to store a previously output value, and the auxiliary function n20
is used to compute a stable value for the feedback loop. The only difference to the function n1 is that the
cases of the clock are inverted, due to the additional inverter in Figure 1 that feeds the select input of the
multiplexer that computes n2. Finally, the functions q and qn implement the two inverters that feed the
corresponding output signals in Figure 1.
The above specification is productive, since the TRS Rd ∪ Rs can be proven context-sensitive terminating, for example by the tool AProVE [6]. Hence, according to Theorem 12, the specification is productive,
meaning that every ground term of sort s rewrites to a constructor term. This especially holds for the ground
terms tq = q(tn2 ) and tqn = qn(tq ), where tn2 = n2(rand, n1(rand, nexts(rand, rand, rand), n1l), n2l) and
the variables n1l and n2l are instantiated with all possible combinations of 0 and 1. Thus, the circuit
produces an infinite stream of stable output values, regardless of its initial state and input streams, and
does not oscillate infinitely long. This illustrates that productivity analysis can be used to prove stabilization of digital circuits with arbitrary input sequences, when encoding them as non-orthogonal proper
specifications.
6
Conclusions and Future Work
We have presented a generalization of the productivity checking techniques in [14] (including the improvements of [5]) to non-orthogonal specifications, which are able to represent non-deterministic systems.
These naturally arise for example when abstracting away certain details of an implementation, such as the
concrete sequence of input values. This was used to verify stabilization of hardware descriptions whose
environment is left unspecified, as was demonstrated in Section 5.
Our setting still imposes certain restrictions on the specifications that can be treated. The most severe
restriction is the requirement of left-linear rules in the structure TRS Rs . Dropping this requirement
however would make Theorem 12 unsound. Similarly, also the requirement that structure arguments of
constructors must be variables cannot be dropped without losing soundness of Theorem 12. This requirement however is not that severe in practice, since many specifications can be unfolded by introducing
fresh symbols, as was presented in [5, 13].
In the future, it would be interesting to investigate whether transformations of non-orthogonal proper
specifications, similar to those in [14], can be defined. It is clear that rewriting of right-hand sides for
example is not productivity-preserving for non-orthogonal specifications, since it only considers one
possible reduction. However, it would be interesting to investigate whether for example narrowing of
right-hand sides is productivity preserving, as it considers all possible reductions.
Matthias Raffelsieper
67
Acknowledgment. The author would like to thank the anonymous reviewers for their valuable comments
and suggestions that helped to improve the paper.
References
[1] F. Baader & T. Nipkow (1998): Term Rewriting and All That. Cambridge University Press. Available at
http://dx.doi.org/10.2277/0521779200.
[2] J. Endrullis (2010): Termination and Productivity. Ph.D. thesis, Vrije Universiteit Amsterdam. Available at
http://hdl.handle.net/1871/15879.
[3] J. Endrullis, C. Grabmayer & D. Hendriks (2008): Data-oblivious Stream Productivity. In: Proceedings of the
11th International Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR’08),
Lecture Notes in Computer Science 5330, Springer-Verlag, pp. 79–96. Available at http://dx.doi.org/10.
1007/978-3-540-89439-1_6. Web interface tool: http://infinity.few.vu.nl/productivity/.
[4] J. Endrullis, C. Grabmayer & D. Hendriks (2009): Complexity of Fractran and Productivity. In: Proceedings
of the 22nd Conference on Automated Deduction (CADE’09), Lecture Notes in Computer Science 5663,
Springer-Verlag, pp. 371–387. Available at http://dx.doi.org/10.1007/978-3-642-02959-2_28.
[5] J. Endrullis & D. Hendriks (2011): Lazy Productivity via Termination. Theoretical Computer Science 412(28),
pp. 3203–3225. Available at http://dx.doi.org/10.1016/j.tcs.2011.03.024.
[6] J. Giesl, P. Schneider-Kamp & R. Thiemann (2006): AProVE 1.2: Automatic Termination Proofs in the Dependency Pair Framework. In: Proceedings of the 3rd International Joint Conference on Automatic Reasoning
(IJCAR’06), Lecture Notes in Computer Science 4130, Springer-Verlag, pp. 281–286. Available at http://dx.
doi.org/10.1007/11814771_24. Web interface tool: http://aprove.informatik.rwth-aachen.de.
[7] S. Lucas (1998): Context-sensitive Computations in Functional and Functional Logic Programs. Journal
of Functional and Logic Programming 1998(1). Available at http://mitpress.mit.edu/e-journals/
JFLP/articles/1998/A98-01/A98-01.html.
[8] S. Lucas (2002): Context-Sensitive Rewrite Strategies. Information and Computation 178(1), pp. 294–343.
Available at http://dx.doi.org/10.1006/inco.2002.3176.
[9] Nangate Inc. (2008): Open Cell Library v2008 10 SP1.
openlibrary/.
Available at http://www.nangate.com/
[10] S. Peyton Jones (2003): Haskell 98 Language and Libraries: The Revised Report. Cambridge University
Press. Available at http://www.haskell.org/definition.
[11] Terese (2003): Term Rewriting Systems. Cambridge Tracts in Theoretical Computer Science 55, Cambridge
University Press, Cambridge, UK. Available at http://dx.doi.org/10.2277/0521391156.
[12] H. Zantema (2008): Normalization of Infinite Terms. In: Proceedings of the 19th International Conference on
Rewriting Techniques and Applications (RTA’08), Lecture Notes in Computer Science 5117, Springer-Verlag,
pp. 441–455. Available at http://dx.doi.org/10.1007/978-3-540-70590-1_30.
[13] H. Zantema (2009): Well-definedness of Streams by Termination. In: Proceedings of the 20th International
Conference on Rewriting Techniques and Applications (RTA’09), Lecture Notes in Computer Science 5595,
Springer-Verlag, pp. 164–178. Available at http://dx.doi.org/10.1007/978-3-642-02348-4_12.
[14] H. Zantema & M. Raffelsieper (2010): Proving Productivity in Infinite Data Structures. In: Proceedings of
the 21st International Conference on Rewriting Techniques and Applications (RTA’10), Leibniz International
Proceedings in Informatics (LIPIcs) 6, Schloss Dagstuhl–Leibniz-Zentrum für Informatik, pp. 401–416.
Available at http://dx.doi.org/10.4230/LIPIcs.RTA.2010.401.
[15] H. Zantema & M. Raffelsieper (2010): Stream Productivity by Outermost Termination. In: Proceedings of the
9th International Workshop in Reduction Strategies in Rewriting and Programming (WRS’09), Electronic
Proceedings in Theoretical Computer Science 15. Available at http://dx.doi.org/10.4204/EPTCS.15.7.
| 6cs.PL
|
Improved Neural Text Attribute Transfer with
Non-parallel Data
arXiv:1711.09395v2 [cs.CL] 4 Dec 2017
Igor Melnyk∗
Cicero Nogueira dos Santos
Kahini Wadhawan
Inkit Padhi
Abhishek Kumar
IBM Research AI
T. J. Watson Research Center
Yorktown Heights, NY
Abstract
Text attribute transfer using non-parallel data requires methods that can perform
disentanglement of content and linguistic attributes. In this work, we propose multiple improvements over the existing approaches that enable the encoder-decoder
framework to cope with the text attribute transfer from non-parallel data. We
perform experiments on the sentiment transfer task using two datasets. For both
datasets, our proposed method outperforms a strong baseline in two of the three
employed evaluation metrics.
1
Introduction
The goal of the text attribute transfer task is to change an input text such that the value of a particular
linguistic attribute of interest (e.g. language = English, sentiment = Positive) is transferred to a
different desired value (e.g. language = French, sentiment = Negative). This task needs approaches
that can disentangle the content from other linguistic attributes of the text. The success of neural
encoder-decoder methods to perform text attribute transfer for the tasks of machine translation and
text summarization rely on the use of large parallel datasets that are expensive to be produced. The
effective use of non-parallel data to perform this family of problems is still an open problem.
In text attribute transfer from non-parallel data, given two large sets of non-parallel texts X0 and
X1 , which contain different attribute values s0 and s1 , respectively, the task consists in using the
data to train models that can rewrite a text from X0 such that the resulting text has attribute value s1 ,
and vice-versa. The overall message contained in the rewritten text must be relatively the same of
the original one, only the chosen attribute value should change. Two of the main challenges when
using non-parallel data to perform such task are: (a) there is no straightforward way to train the
encoder-decoder because we can not use maximum likelihood estimation on the transferred text due
to lack of ground truth; (b) it is difficult to preserve content while transferring the input to the new
style. Recent work from Shen et al. [9] showed promising results on style-transfer from non-parallel
text by tackling challenging (a).
In this work, we propose a new method to perform text attribute transfer that tackles both challenges
(a) and (b). We cope with (a) by using a single collaborative classifier, as an alternative to commonly
used adversarial discriminators, e.g., as in [9]. Note that a potential extension to a problem of
multiple attributes transfer would still use a single classifier, while in [9] this may require as many
discriminators as the number of attributes. We approach (b) with a set of constraints, including the
attention mechanism combined with cyclical loss and a novel noun preservation loss to ensure proper
∗
Corresponding author. Email: [email protected]
Workshop on Learning Disentangled Representations: from Perception to Control (NIPS 2017), Long Beach,
CA, USA.
si
xik
Classification
Loss
Reconstruction
Loss
Classifier
(CNN)
Forward Transfer
Decoder
x̂i!i
k
Decoder
x̂i!j
k
(RNN)
si
Encoder
xik
Backward Transfer
si
(RNN)
Hki
sj
Attention
sj
si
Encoder
(RNN)
(RNN)
Classifier
(CNN)
i!j
nouns( x̂k
Classification
Loss
Content
Loss
sj
nouns(xk)
)
i
Hki!j
Attention
x̂i!j!i
k
Decoder
(RNN)
Classifier
(CNN)
Classification
Loss
Reconstruction
Loss
si
xik
Figure 1: Proposed framework of a Neural Text Attribute Transfer algorithm using non-parallel data.
content transfer. We compared our algorithm with Shen et al. [9] on the sentiment transfer task on
two datasets using three evaluation metrics (sentiment transfer accuracy, a novel content preservation
metric and a perplexity), outperforming the baseline in terms of the first two.
2
Proposed Method
We assume access to a text dataset consisting of two non-parallel corpora X = X0 ∪X1 with different
attribute values s0 and s1 of a total of N = m + n sentences, where |X0 | = m and |X1 | = n. We
denote a randomly sampled sentence k of attribute si from X as xik , for k ∈ 1, . . . , N and i ∈ {0, 1}.
A natural approach to perform text attribute transfer is to use a regular encoder-decoder network,
however, the training of such network requires parallel data. Since in this work we consider a
problem of attribute transfer on non-parallel data, we propose to extend the basic encoder-decoder by
introducing a collaborative classifier and a set of specialized loss functions that enable the training
on such data. Figure 1 shows an overview of the proposed attribute transfer approach. Note that for
clarity in the Figure 1 we have used multiple boxes to show encoder, decoder and classifier, the actual
model contains a single encoder and decoder, and one classifier.
The encoder (in the form of RNN), E(xik , si ) = Hki , takes as input a sentence xik together with its
attribute label si , and outputs Hki , a sequence of hidden states. The decoder/generator (also in the
form of RNN), G(Hki , sj ) = x̂i→j
for i, j ∈ 0, 1, takes as input the previously computed Hki and a
k
desired attribute label sj and outputs a sentence x̂i→j
, which is the original sentence but transferred
k
from attribute value i to attribute value j. The hidden states Hki are used by the decoder in the
attention mechanism [7, 2], and in general can improve the quality of the decoded sentence. For
i = j, the decoded sentence x̂i→i
is in its original attribute si (top part of Figure 1); for i 6= j, the
k
decoded/transferred sentence x̂i→j
is in a different attribute sj (bottom part of Figure1). Denote all
k
transferred sentences as X̂ = {x̂i→j
| i 6= j, k = 1, . . . , N }. The classifier (in the form of CNN),
k
then takes as input the decoded sentences and outputs a probability distribution over the attribute
labels, i.e., C(x̂i→j
) = pC (sj |x̂i→j
) (see Eq. (3) for more details). By using the collaborative
k
k
classifier our goal is to produce a training signal that indicates the effectiveness of the current decoder
on transferring a sentence to a given attribute value.
2
Note that the top branch of Figure 1 can be considered as an auto-encoder and therefore we can
enforce the closeness between x̂i→i
and xik by using a standard cross-entropy loss (see (1) below).
k
However, for the bottom branch, due to lack of parallel data, we cannot use the same approach, and
for this purpose we proposed a novel content preservation loss (see Eq. (2)). Finally, note that once
we transferred X to X̂ (forward-transfer step), we can now transfer X̂ back to X (back-transfer step)
by using the bottom branch in Figure 1 (see Eq. (5) and Eq. (6) below).
In what follows, we present the details of the loss functions employed in training of our model.
2.1
Forward Transfer
Reconstruction Loss. Given the encoded input sentence xik and the decoded sentence x̂i→i
k , the
reconstruction loss measures how well the decoder G is able to reconstruct it:
Lrec = Exik ∼X − log pG (xik |E(xik , si ), si ) .
(1)
Content Preservation Loss. To enforce closeness between xik and x̂i→j
for i 6= j, we utilize the
k
attention mechanism. Recall, that this mechanism enables to establish an approximate correspondence
between the words in the original (encoded) and transferred (decoded) sentences. For example, denote
i→j
i
the words in sentence xik as xik = {wkr
| r = 1, . . . , |xik |} and similarly x̂i→j
= {wkr
| r0 =
0
k
i→j
1, . . . , |xk |}. Utilizing the attention mechanism, we can establish the correspondence (r, r0 )
i
between the words. Among different pairings of such words we select only the ones where wkr
is a
i→j
noun (e.g., as detected by a POS tagger),and enforce that the corresponding transferred word wkr
matches that noun, i.e.,
h
i
Lcnt_rec = Exj ={...,wj ...}∼X − log pG xjk = {. . . , wkrj . . .}|E(xik , si ), sj ,
(2)
k
kr
0
i
for indices r and r such that wkr
is a noun and (r, r0 ) is a pair established by attention mechanism.
We note that although not always applicable, the above heuristic is very effective for attributes where
the sentences can share the nouns (e.g., for sentiment transfer considered in Section 4).
Classification Loss. The loss is formulated as follows:
i
h
Lclass_td = Ex̂i→j ∼X̂ − log pC (sj |x̂i→j
)
.
k
(3)
k
For the encoder-decoder this loss gives a feedback on the current generator’s effectiveness on
transferring sentences to a new attribute. For the classifier, it provides an additional training signal
from generated data, enabling the classifier to be trained in a semi-supervised regime.
Classification Loss - Original Data. In order to enforce a high classification accuracy, the classifier also uses a supervised classification loss, measuring the classifier predictions on the original
(supervised) instances xik ∈ X:
Lclass_od = Exik ∼X − log pC (si |xik ) .
(4)
2.2
Backward Transfer
Reconstruction Loss. The back-transfer (or cycle) loss [10, 5] is motivated by the difficulty of
imposing constraints on the transferred sentences. Back-transfer transforms the transferred sentences
x̂i→j
back to the original attribute si , i.e., x̂i→j→i
and compares them to xik . This also implicitly
k
k
imposes the constraints on the generated sentences and improves the content preservation (in addition
to (2)). The loss is formulated as follows:
h
i
Lback_rec = Exik ∼X − log pG (xik |E(x̂i→j
, sj ), si ) ,
(5)
k
which can be thought to be similar to an auto-encoder loss in (1) but in the attribute domain.
Classification Loss. Finally, we ensure that the back-transferred sentences x̂i→j→i
have the correct
k
attribute label si :
h
i
Lclass_btd = Ex̂i→j ∼X̂ − log pC (si |G(E(x̂ki→j , sj ), si ))
(6)
k
3
In summary, the training of the components of our architecture consists in optimizing the following
loss function using stochastic gradient descent with back-propagation for some weights λi > 0:
L(θE , θG , θC ) =
= min λ1 Lrec + λ2 Lcnt_rec + λ3 Lback_rec + λ4 Lclass_od + λ5 Lclass_td + λ6 Lclass_btd . (7)
E,G,C
The Algorithm 1 summarizes the above discussion and shows the main steps of the training of the
proposed approach.
Algorithm 1 Training of the Neural Text Attribute Transfer Algorithm using Non-parallel Data.
Require: Two non-parallel corpora X = X0 ∪ X1 with different attribute values s0 and s1 .
Initialize θE , θG , θC
repeat
— Sample a mini-batch of l original sentences A = {xik }lk=1 from X, with i ∈ {0, 1}
— Sample a mini-batch of l transferred sentences B = {x̂i→j
}lk=1 from the generator’s
k
i→j
i
distribution pG , where x̂k = G(E(xk , si ), sj ) with i, j ∈ {0, 1}
— Sample a mini-batch of l back-transferred sentences C = {x̂i→j→i
}lk=1 from the generator’s
k
i→j→i
i→j
distribution pG , where x̂k
= G(E(xk , sj ), si ) with i, j ∈ {0, 1}
— Compute Lrec (1), Lcnt_rec (2), Lclass_td (3), Lclass_od (4), Lback_rec (5), and Lclass_btd (6)
— Update {θE , θG θC } by gradient descent on loss L(θE , θG , θC ) in Eq. (7)
until convergence
3
Related Work
Attribute transfer has been studied more extensively in the context of images than in the text domain,
with several works studying the style transfer task under the setting of non-parallel data [4, 11].
However, style/attribute transfer in text is fundamentally different as textual data is sequential and of
potentially varying lengths, versus constant-sized images. In the image domain, one of the similar
works is CycleGAN [11], which also employs a cycle consistency loss (similar to our back-transfer
loss) that ensures that composition of a transfer and its reverse is close to the identity map. However,
there are several key differences between CycleGAN and our work: (i) we use of a single generator
for generating both styles which makes it easier to scale to multiple style transfer, (ii) we use a
collaborative classifier for measuring the style instead of a adversarial discriminator, which imparts
stability to the training, (iii) additional syntactic regularizers for better content preservation.
Controlled text generation and style transfer without parallel data has also received attention from the
language community recently [8, 6, 3, 9]. Ficler and Goldberg [3] consider the problem of attribute
conditioned generation of text in a conditioned language modeling setting using LSTM. Mueller et al.
[8] allows modifying the hidden representations to generate sentences with desired attributes which
is measured by a classifier, however their model does not explicitly encourage content preservation.
Our proposed model has some similarities with the approach taken by Hu et al. [6] and Shen et al. [9],
with the main differences being that instead of VAE and adversarial discriminators we use a simple
encode-decoder framework with a collaborative classifier augmented with the attention mechanism
and a set of specially designed content preservation losses.
4
Experiments and Results
In this Section we present experimental results of applying the proposed approach for sentiment
transfer as one example of text attribute transfer. We compared the algorithm with the approach of [9]
on two datasets. One is the dataset from [9], which is based on Yelp restaurant reviews and contains
(179K, 25K, 51K) sentences for (training, validation, testing) based on negative reviews and similarly
(268K, 38K, 76K) positive sentences. The sentences had a maximum length of 17 words. The second
dataset is based on general customer reviews on Amazon [1], from which we selected (265K, 33K,
33K) positive and the same number of negative sentences, each having up to 7 tokens per sentence.
We used three evaluation metrics: (i) sentiment accuracy, which is computed based on pre-trained
classifier (estimated on the training part of each dataset) and measures the percentage of sentences in
4
Table 1: Evaluation results on Yelp and Amazon datasets. For Yelp, the pre-trained classifier had a
default accuracy of 97.4% and the pre-trained language model had a default perplexity of 23.5. For
Amazon, these values were 82.02% for classification and 25.5 for perplexity.
Yelp
Amazon
Sentiment Content Perplexity Sentiment Content Perplexity
Shen et. al [9] 86.5
38.3
27.0
32. 8
71.6
27.3
Our Method
94.4
77.1
80.1
59.5
77.5
43.7
Table 2: Examples of sentences transferred from positive to negative sentiment on Yelp dataset
Original
their food was definitely delicious
love the southwestern burger
Shen et. al [9] there was so not spectacular
avoid the pizza sucks
Our Method
their food was never disgusting
avoid the grease burger
Original
restaurant is romantic and quiet
the facilities are amazing
Shen et. al [9] the pizza is like we were disappointed the drinks are gone
Our Method
restaurant is shame and unprofessional the facilities are ridiculous
Table 3: Examples of sentences transferred from negative to positive sentiment on Yelp dataset
Original sorry they closed so many stores
these people will try to screw you over
[9]
thanks and also are wonderful
these guys will go to work
these people will try to thank you special
Ours
amazing they had so many stores
Original i wish i could give them zero stars
seriously , that ’s just rude
[9]
i wish i love this place
clean , and delicious ...
Ours
i wish i ’ll give them recommended stars seriously , that ’s always friendly
the test set with correct sentiment label; (ii) content preservation accuracy, a new evaluation metric
proposed in this work, which is computed as the percentage of the transferred sentences where each
of them has at least one of the nouns present in the original sentence; (iii) perplexity score, which is
computed based on pre-trained language model (estimated on the training part of each dataset) and
measures the quality of the generated text.
The results are presented in Table 1. As compared to the algorithm of Shen et. al [9], the proposed
method although not able to get better perplexity scores, it can achieve more accurate sentiment
transfer and better content preservation. A possible explanation of having higher perplexity is that
since the algorithm of [9] does not explicitly enforce content similarity, it has an easier job of
achieving high sentiment accuracy and low perplexity of the transferred sentences. Our algorithm,
on the other hand, is penalized if the content changes, which forces it to sacrifice the perplexity.
Achieving better results across all the metrics still remains a challenge.
In Table 3 we also show some of the sentences generated by both algorithms on Yelp dataset. The
algorithm of [9], although able to create well structured sentences with correct sentiment labels, in
many cases it cannot accurately preserve the content. On the other hand, our approach may generate
text with somewhat higher perplexity but ensures a better sentiment and content transfer.
5
Conclusion
In this work we proposed a novel algorithm for text attribute transfer with non-parallel corpora based
on the encoder-decoder architecture with attention, augmented with the collaborative classifier and a
set of content preservation losses. Although the experimental evaluations showed promising results, a
number of challenges remain: (i) achieve better results across all the three metrics and propose new
evaluation metrics to better capture the quality of transfer; (ii) improve the architecture to enable
transfer for more challenging text attributes (e.g., such as professional-colloquial) where the text
goes under more significant transformation then in a simpler sentiment transfer tasks; (iii) extend the
architecture to work in a multi-attribute transfer, a more challenging problem.
5
References
[1] Amazon reviews dataset. https://www.kaggle.com/bittlingmayer/amazonreviews.
Accessed: 2017-11-26.
[2] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473, 2014.
[3] J. Ficler and Y. Goldberg. Controlling linguistic style aspects in neural language generation.
arXiv preprint arXiv:1707.02633, 2017.
[4] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural
networks. In CVPR, pages 2414–2423, 2016.
[5] D. He, Y. Xia, T. Qin, L. Wang, N. Yu, T. Liu, and W. Ma. Dual learning for machine translation.
In Advances in Neural Information Processing Systems, pages 820–828, 2016.
[6] Z. Hu, Z. Yang, X. Liang, R. Salakhutdinov, and E. P. Xing. Towards controllable generation of
text. In International Conference on Machine Learning, 2017.
[7] T. Luong, H. Pham, and C. D. Manning. Effective approaches to attention-based neural machine
translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing, pages 1412–1421, September 2015.
[8] J. Mueller, D. Gifford, and T. Jaakkola. Sequence to better sequence: continuous revision of
combinatorial structures. In ICML, pages 2536–2544, 2017.
[9] T. Shen, T. Lei, R. Barzilay, and T. Jaakkola. Style transfer from non-parallel text by crossalignment. In Advances in Neural Information Processing Systems, 2017.
[10] J. Zhu, T. Park, P. Isola, and A. Efros. Unpaired image-to-image translation using cycleconsistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
[11] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using
cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
6
| 2cs.AI
|
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
arXiv:1803.08033v1 [math.GN] 21 Mar 2018
W. XI, D. DIKRANJAN, M. SHLOSSBERG, AND D. TOLLER
Abstract. We study locally compact groups having all subgroups minimal. We call such groups hereditarily
minimal. In 1972 Prodanov proved that the infinite hereditarily minimal compact abelian groups are precisely
the groups Zp of p-adic integers. We extend Prodanov’s theorem to the non-abelian case at several levels. For
infinite hypercentral (in particular, nilpotent) locally compact groups we show that the hereditarily minimal ones
remain the same as in the abelian case. On the other hand, we classify completely the locally compact solvable
hereditarily minimal groups, showing that in particular they are always compact and metabelian.
The proofs involve the (hereditarily) locally minimal groups, introduced similarly. In particular, we prove a
conjecture by He, Xiao and the first two authors, showing that the group Qp ⋊ Q∗p is hereditarily locally minimal,
where Q∗p is the multiplicative group of non-zero p-adic numbers acting on the first component by multiplication.
Furthermore, it turns out that the locally compact solvable hereditarily minimal groups are closely related to this
group.
March 22, 2018
1. Introduction
A Hausdorff topological group (G, τ ) is called minimal if there exists no Hausdorff group topology on G which
is strictly coarser than τ (see [11, 31]). This class of groups, containing all compact ones, was largely studied in
the last five decades, (see the papers [2, 8, 9, 12, 24, 28], the surveys [4, 32] and the book [7]). Since it is not
stable under taking quotients, the following stronger notion was introduced in [6]: a minimal group G is totally
minimal, if the quotient group G/N is minimal for every closed normal subgroup N of G. This is precisely a
group G satisfying the open mapping theorem, i.e., every continuous surjective homomorphism with domain G
and codomain any Hausdorff topological group is open. Clearly, every compact group is totally minimal.
Examples of locally compact non-compact minimal groups can be found in [29], they are all non-abelian.
Indeed, Stephenson [31] noticed much earlier that local compactness and minimality jointly imply compactness,
for abelian groups:
Fact 1.1. [31, Theorem 1] A minimal locally compact abelian group is compact.
In particular, R is not minimal. This failure of minimality to embrace also local compactness was repaired by
Morris and Pestov [26]. They called locally minimal a topological group (G, τ ), having a neighborhood V of the
identity of G, such that for every coarser Hausdorff group topology σ ⊆ τ with V ∈ σ, one has σ = τ . Clearly,
every minimal group is locally minimal. Moreover, every locally compact group is locally minimal.
Neither minimality nor local minimality are inherited by all subgroups (although they are inherited by closed
central subgroups). This justifies the following definition, crucial for this paper (these properties are abbreviated
sometimes to HM and HLM in the sequel):
Definition 1.2. A topological group G is said to be hereditarily (locally) minimal, if every subgroup of G is
(locally) minimal.
The following theorem of Prodanov provided an interesting characterization of the group of p-adic integers Zp
in terms of hereditary minimality.
Fact 1.3. [28] An infinite compact abelian group K is isomorphic to Zp for some prime p if and only if K is
hereditarily minimal.
Dikranjan and Stoyanov classified all hereditarily minimal abelian groups.
2010 Mathematics Subject Classification. 20F16, 20F19, 20F50, 22A05, 22B05, 22D05.
Key words and phrases. locally minimal group, Lie group, p-adic number, p-adic integer, hereditarily non-topologizable group,
categorically compact group.
1
2
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
Fact 1.4. [8] Let G be a topological abelian group. Then the following conditions are equivalent:
(1) each subgroup of G is totally minimal;
(2) G is hereditarily minimal;
(3) G is topologically isomorphic to one of the following groups:
(a) a subgroup of Zp for some prime p,
(b) a direct sum ⊕ Fp , where for each prime p, the group Fp is a finite abelian p-group,
(c) X × Fp , where X is a rank-one subgroup of Zp , and Fp is a finite abelian p-group.
Note that only the groups from item (a) can be infinite locally compact. Indeed, using Fact 1.1 one can extend
Fact 1.3 to locally compact abelian groups as follows:
Corollary 1.5. An infinite hereditarily minimal locally compact abelian group is isomorphic to Zp .
The main aim of this paper is to extend this result to non-abelian groups at various levels of non-commutativity.
In particular, we obtain an extension to hypercentral (e.g., nilpotent) groups (Corollary 4.3), yet some nonabelian groups may appear beyond the class of hypercentral groups (see Theorem C or Theorem D for a complete
description in the case of solvable groups). Let us mention here that without any restraint on commutativity one
can find examples of very exotic hereditarily minimal groups even in the discrete case (see §3.1, entirely dedicated
to discrete hereditarily minimal groups and their connection to categorically compact groups).
As far as hereditarily locally minimal groups are concerned, the following question was raised in [4, Problem
7.49] in these terms: if a connected locally compact group G is hereditarily locally minimal, is G necessarily a Lie
group? Inspired by this question and the above results, hereditarily locally minimal groups were characterized in
[3] among the locally compact groups which are either abelian or connected as follows:
Fact 1.6. [3, Corollary 1.11] For a locally compact group K that is either abelian or connected, the following
conditions are equivalent:
(a) K is a hereditarily locally minimal group;
(b) K is either a Lie group or has an open subgroup isomorphic to Zp for some prime p.
This provides, among others, a characterization of the connected Lie groups as connected locally compact
hereditarily locally minimal groups.
1.1. Main Results. It was mentioned in [3], that item (a) in Fact 1.6 might not be equivalent to item (b) in the
non-abelian case, and it was conjectured (see [3, Conjecture 5.1]) that a possible counter-example could be the
group (Qp , +) ⋊ Q∗p , where Q∗p = (Qp ∖ {0}, ⋅) (here the natural action of Q∗p on Qp by multiplication is intended).
We prove that this conjecture holds true.
Theorem A. Let p be a prime. Then (Qp , +) ⋊ Q∗p is hereditarily locally minimal.
We report in Theorem 2.10 a result of Megrelishvili, ensuring that the group in Theorem A is minimal. Although
it is not hereditarily minimal (e.g., its subgroup (Qp , +) is not minimal by Fact 1.1), we show in the classification
Theorem D that it contains (up to isomorphism) most locally compact solvable HM groups.
Clearly, a hereditarily minimal group is hereditarily locally minimal. In order to ensure the converse implication,
we give the next definition.
Definition 1.7. For a topological group G we consider the following properties:
(Nf n ) G contains no finite normal non-trivial subgroups;
(Cf n ) Every infinite compact subgroup of G satisfies (Nf n ).
Remark 1.8. Obviously, a torsionfree group G satisfies both properties. Moreover, if G is abelian, then G is
torsionfree if and only if it satisfies (Nf n ). It is also clear that (Cf n ) implies (Nf n ) when G is infinite compact.
Theorem B. For a compact group G, the following conditions are equivalent:
(a) G is a hereditarily minimal group;
(b) G is a hereditarily locally minimal group satisfying (Cf n ).
Remark 1.9. One cannot replace compact by locally compact as the groups Qp and R ⋊ Z(2) show (easy counterexamples are provided also by arbitrary infinite discrete abelian groups).
Our next result extends Corollary 1.5.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
3
Theorem C. Let G be an infinite hereditarily minimal locally compact group that is either compact or locally
solvable. Then G is either center-free or isomorphic to Zp , for some prime p.
In order to introduce our main result we need to first recall some folklore facts and fix the relevant notation.
Notation 1.10. For a prime p, let Z∗p = (Zp ∖ pZp , ⋅). Its torsion subgroup Fp consists only of the (p − 1)-th roots
of unity if p > 2, and Fp = {1, −1} (the square roots of unity) if p = 2. Moreover, it is cyclic and
⎧
⎪
if p = 2,
⎪Z(2)
Fp ≅ ⎨
⎪
Z(p
−
1)
otherwise.
⎪
⎩
∗
We denote by Cp the subgroup of Zp defined by
⎧
⎪
⎪1 + 4Z2 if p = 2,
Cp = ⎨
⎪
⎪
⎩1 + pZp otherwise.
n
n
Finally, for n ∈ N we let Cpp = {xp ∶ x ∈ Cp } ≤ Cp , so
n
⎧
⎪
n
⎪(1 + 4Z2 )2 = 1 + 2n+2 Z2 if p = 2,
(1.1)
Cpp = ⎨
n
⎪
(1 + pZp )p = 1 + pn+1 Zp otherwise.
⎪
⎩
It is well known that Cp ≅ (Zp , +) and Z∗p = Cp Fp ≅ Cp × Fp ≅ (Zp , +) × Fp as topological groups.
Now we provide two series of examples of hereditarily minimal metabelian locally compact groups that play a
prominent role in our Theorem D:
Example 1.11. Consider the natural action of Z∗p on (Zp , +) by multiplication, and the semidirect product
K = (Zp , +) ⋊ Z∗p .
Then K is a subgroup of the group considered in Theorem A, so it is hereditarily locally minimal. Moreover,
K ≅ (Zp , +) ⋊ ((Zp , +) × Fp ), so K is compact and all of its non-abelian subgroups are minimal by Corollary 2.9.
However, K is not hereditarily minimal, as for example its compact abelian subgroup {0} ⋊ Z∗p ≅ Z∗p ≅ (Zp , +) × Fp
is not hereditarily minimal by Fact 1.3.
(i) For a subgroup F of Fp and an integer n ∈ N, consider the following subgroups of K:
Kp,F = (Zp , +) ⋊ F ≤ Kp,Fp
n
Mp,n = (Zp , +) ⋊ Cpp ≤ Mp,0 .
For example, Kp,{1} = (Zp , +)⋊{1} is isomorphic to Zp . In Example 6.5(a) we use a criterion for hereditary
minimality of a compact solvable group (Theorem 6.3) in order to prove that Kp,F is hereditarily minimal,
while we use Theorem A and Theorem B to prove that Mp,n is hereditarily minimal in Example 3.12.
n
(ii) In case p = 2 and n ∈ N, we consider also the group Tn = (Z2 , +) ⋊βn C22 with the faithful action
n
βn ∶ C22 × Z2 → Z2 defined by
n+1
⎧
⎪
if y ∈ C22 ,
⎪yx
βn (y, x) = ⎨
2n
2n+1
⎪
⎪
⎩−yx if y ∈ C2 ∖ C2 .
(when no confusion is possible, we denote βn simply by β). Note that the restriction β ′ = βn ↾C 2n+1 ×Z2 ∶
2
n+1
n+1
C22 × Z2 → Z2 is the natural action by multiplication in the ring Z2 , so Tn ≥ (Z2 , +) ⋊β ′ C22
= M2,n+1 .
Obviously [Tn ∶ M2,n+1 ] = 2, so M2,n+1 is normal in Tn . Another application of the criterion for hereditary
minimality (Theorem 6.3) shows in Example 6.5(b) that also the groups Tn are hereditarily minimal.
The following theorem classifies the locally compact solvable hereditarily minimal groups by showing that these
are precisely the groups described in Example 1.11. Note that the only abelian ones among them are the groups
Kp,{1} = (Zp , +) ⋊ {1} ≅ Zp , for prime p.
Theorem D. Let G be an infinite locally compact solvable group, then the following conditions are equivalent:
(1) G is hereditarily minimal;
(2) G is topologically isomorphic to one of the following groups:
(a) Kp,F = Zp ⋊ F , where F ≤ Fp for some prime p;
4
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
n
(b) Mp,n = Zp ⋊ Cpp , for some prime p and n ∈ N;
n
(c) Tn = (Z2 , +) ⋊β C22 , for some n ∈ N.
A locally compact hereditarily minimal group need not be compact in general, as witnessed by the large supply
of infinite discrete hereditarily minimal groups. Nonetheless, as the groups in Theorem D are compact and
metabelian, one has the following:
Corollary 1.12. If G is an infinite hereditarily minimal locally compact solvable group, then G is compact
metabelian.
Another nice consequence of Theorem D is the following: every closed non-abelian subgroup of Mp,n is isomorphic to one of the groups in (b) or (c), while the closed abelian subgroups of Mp,n are isomorphic to Zp ≅ Kp,{1} .
The proof of Theorem D is covered by Theorem 6.11 and Theorem 6.13, where we consider the torsionfree case
and the non-torsionfree case, respectively.
Another application of Theorem D is Theorem 6.21 in which we classify the infinite locally compact, solvable,
hereditarily totally minimal groups (see Definition 6.15).
Remark 1.13. Call a topological group G compactly hereditarily (locally) minimal, if every compact subgroup of G
is hereditarily (resp., locally) minimal. We abbreviate it to CHM and CHLM in the sequel. For compact groups,
being hereditarily minimal is equivalent to being compactly hereditarily minimal. Clearly, every discrete group
is compactly hereditarily minimal. Applying Theorem B to compact subgroups one can prove that a topological
group G is compactly hereditarily minimal if and only if G is a compactly hereditarily locally minimal group
satisfying (Cf n ).
The next diagram summarizes some of the interrelations between the properties considered so far. The double
arrows denote implications that always hold. The single arrows denote implications valid under some additional
assumptions.
ks
CHLM
KS
5 CHM
KS
(1)
compact
(2)
HLM ks
(3)
)
HM
(1): This implication holds true for groups satisfying (Cf n ) (Remark 1.13).
(2): This implication holds true for totally disconnected locally compact groups (Proposition 3.13(1)).
(3): This implication holds true for compact groups satisfying (Cf n ) (Theorem B).
The group Qp witnesses that the implication HM Ô⇒ CHM cannot be inverted if compact is replaced by
locally compact (and totally disconnected). Indeed, Qp is not minimal (so not hereditarily minimal), yet every
compact subgroup of Qp is isomorphic to Zp , which means that Qp is CHM. On the other hand, every non-trivial
compact subgroup of the locally compact group G = R × Zp is topologically isomorphic to Zp , so G is CHM by
Prodanov’s theorem. Yet G is not hereditarily locally minimal by Fact 1.6. This shows that none of the vertical
arrows in the diagram can be inverted in general.
The paper is organized as follows. The proof Theorem A, given in §2, is articulated in several steps. We first
recall the crucial criteria for (local) minimality of dense subgroups (Fact 2.2). Using these criteria, we show in
§2.1 that every non-abelian subgroup H of L is locally minimal, while §2.2 takes care of the abelian subgroups of
L.
Section §3 contains some general results on locally compact HM groups and the proof of Theorem B. In §3.1
we provide a brief review on the relevant connection between discrete categorically compact groups and the
discrete HM groups. In §3.2 we give some general results on non-discrete locally compact HM groups, proving
in Proposition 3.9 that they are totally disconnected, and contain a copy of Zp (in particular, an infinite locally
compact HM group is torsion if and only if it is discrete, and in this case it is not locally finite). Furthermore,
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
5
such a group G satisfies (Nf n ), so that either Z(G) = {e}, or Z(G) ≅ Zp for a prime p. In the latter case, G is
also torsionfree by Corollary 4.2. So a non-discrete locally compact HM group is either center-free or torsionfree.
Finally, we prove Theorem B (making use of Proposition 3.11) and apply Theorem B to see in Example 3.12 that
the groups Mp,n are HM.
In §4 we explore infinite non-discrete locally compact HM groups with non-trivial center, proving in Corollary
4.3 that the hypercentral ones are isomorphic to Zp and we give a proof of Theorem C. In Theorem 4.6 we collect
some necessary conditions a non-discrete locally compact HM group with non-trivial center must satisfy.
In §5 we prepare the tools for the proof of Theorem D, by proving that the groups introduced in Example
1.11 are pairwise non-isomorphic (Corollary 5.4 and propositions 5.6 and 5.7) and by classifying the semidirect
products of Zp with some compact subgroups of Z∗p (Proposition 5.8 and Lemma 5.9).
Theorem D is proved in §6. To this end we first provide a criterion Theorem 6.3, used in Example 6.5 to check
that the groups Kp,F are HM. Another consequence of Theorem 6.3 is Lemma 6.4, that we apply to show that the
groups Tn are HM in Example 6.5. We start the proof of Theorem D in §6.1, by proving some reduction results
(Proposition 6.7), and some general results (Propositions 6.8, 6.9 and 6.10). In §6.2 we consider the torsionfree
case of Theorem D in Theorem 6.11, while the non-torsionfree case Theorem 6.13 is proved in §6.3, based on the
technical result Proposition 6.12 dealing with the case p = 2.
We dedicate §6.4 to hereditarily totally minimal groups (HTM for short, see Definition 6.15). The only locally
compact solvable ones to consider are the groups classified in Theorem D, and we first prove in Proposition 6.18
that the groups Kp,F are HTM. Then we see in Proposition 6.20 that the HM groups Mp,n and Tn are not HTM,
leading us to Theorem 6.21, which describes the groups Kp,F as the only infinite locally compact solvable HTM
groups.
The last §7 collects some open questions, a partial converse to Theorem 4.6, and some final remarks.
1.2. Notation and terminology. We denote by Z the group of integers, by R the real numbers, and by N and
N+ the non-negative integers and positive natural numbers, respectively. For n ∈ N+ , we denote by Z(n) the finite
cyclic group with n elements. If p is a prime number, Qp stands for the field of p-adic numbers, and Zp is its
subring of p-adic integers.
Let G be a group. We denote by e the identity element. If A is a non-empty subset of G, we denote by ⟨A⟩
the subgroup of G generated by A. In particular, if x is an element of G, then ⟨x⟩ is a cyclic subgroup. If F = ⟨x⟩
is finite, then x is called a torsion element, and o(x) = ∣F ∣ is the order of x. We denote by t(G) the torsion part
of the group G and G is called torsionfree if t(G) is trivial. The centralizer of x is CG (x). If the center Z(G) is
trivial, then we say that G is center-free. A group G is called n-divisible if nG = G for n ∈ N+ .
Let P be an algebraic (or set-theoretic) property. A group is called locally P if every finitely generated subgroup
has the property P. For example, in a locally finite group every finitely generated subgroup is finite.
The n-th center Zn (G) is defined as follows for n ∈ N. Let Z0 (G) = {e}, Z1 (G) = Z(G), and assume that
n > 1 and Zn−1 (G) is already defined. Consider the canonical projection π∶ G → G/Zn−1 (G) and let Zn (G) =
π −1 Z(G/Zn−1 (G)). Note that Zn (G) = {x ∈ G ∶ [x, y] ∈ Zn−1 (G) for every y ∈ G}. This produces an ascending
chain of subgroups Zn (G) called the upper central series of G, and a group is nilpotent if Zn (G) = G for some
n ∈ N. In this case, its nilpotency class is the minimum of such n. For example, the groups with nilpotency
class at most 1 are the abelian groups. One can continue the upper central series to infinite ordinal numbers
via transfinite recursion: for a limit ordinal λ, define Zλ (G) = ⋃α<λ Zα (G). A group is called hypercentral if it
coincides with Zα (G) for some ordinal α. Nilpotent groups are obviously hypercentral, while hypercentral groups
are locally nilpotent.
We denote by G′ = G(1) the derived subgroup of G, namely the subgroup of G generated by all commutators
[a, b] = aba−1 b−1 , where a, b ∈ G. For n ≥ 1, define G(n) = (G(n−1) )′ and also G(0) = G. We say that G is solvable
of class n for some n ∈ N, if G(n) = {e} and G(m) ≠ {e} for 0 ≤ m < n. If G is solvable of class n, then G(n−1) is
abelian. In particular, G is metabelian, if G is solvable of class at most 2.
For an integral domain (in particular, a field) A we denote by A∗ the multiplicative group of all invertible
elements of A (resp., the group (A ∖ {0}, ⋅)).
All the topological groups in this paper are assumed to be Hausdorff. For a topological group G, the connected
component of G is denoted by c(G). For a subgroup H ≤ G, the closure of H is denoted by H. A topological
group is precompact if it is isomorphic to a subgroup of a compact group. Let S and T be topological groups and
6
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
α ∶ S × T → T be a continuous action by automorphisms. We say that the action α is faithful if ker α = {s ∈ S ∶
∀t ∈ T α(s, t) = t} is trivial.
All unexplained terms related to general topology can be found in [13]. For background on abelian groups, see
[14].
2. Proof of Theorem A
There exist useful criteria for establishing the minimality (local minimality) of a dense subgroup of a minimal
(respectively, locally minimal) group. These criteria are based on the following definitions.
Definition 2.1. Let H be a subgroup of a topological group G.
(1) [28, 31] H is said to be essential in G if H ∩ N ≠ {e} for every non-trivial closed normal subgroup N of
G.
(2) [1] H is locally essential in G if there exists a neighborhood V of e in G such that H ∩ N ≠ {e} for every
non-trivial closed normal subgroup N of G which is contained in V.
Fact 2.2. Let H be a dense subgroup of a topological group G.
(1) [2, Minimality Criterion] H is minimal if and only if G is minimal and H is essential in G (for compact
G see also [28, 31]).
(2) [1, Local Minimality Criterion] H is locally minimal if and only if G is locally minimal and H is locally
essential in G.
In this section, L denotes the group (Qp , +) ⋊ Q∗p .
In §2.1, we show that every non-abelian subgroup H of L is essential (in particular, locally essential) in its
closure H. Since the latter group is locally compact (and thus locally minimal), we conclude by the above criterion
that H is also locally minimal. At the same time, we deduce by the above Minimality Criterion that H is minimal
if and only if H is minimal in Corollary 2.8. In particular, every precompact non-abelian subgroup of L is minimal.
Finally, using Fact 1.6, we prove in §2.2 that also the abelian subgroups of L are locally minimal.
2.1. Non-abelian subgroups of L. We begin this section with some easy lemmas of independent interest.
Lemma 2.3. If G is a group, then G′ non-trivially meets every normal subgroup N of G not contained in Z(G).
In particular, if G is a topological group with trivial center, then G′ is essential in G.
Proof. Let a ∈ N ∖ Z(G), and let b ∈ G be such that e ≠ [a, b] = aba−1 b−1 . Then [a, b] ∈ N ∩ G′ , as N is normal.
Lemma 2.4. If H is a non-abelian subgroup of a group G, then H ∩ G′ is non-trivial.
Proof. If a, b are non-commuting elements of H, then {e} ≠ [a, b] ∈ H ∩ G′ .
Lemma 2.5. Every non-trivial subgroup H of Qp is essential.
Proof. First note that an element e ≠ x ∈ Qp has the form x = pn a for a ∈ Z∗p , and n ∈ Z, so its generated subgroup
is ⟨x⟩ = xZ = pn aZ (note that here Z carries the p-adic topology, induced by Qp ).
Let N be a non-trivial closed subgroup of Qp , and let H ≥ pn aZ, and N ≥ pm bZ for some a, b ∈ Z∗p and n, m ∈ Z.
Being closed, N also contains the subgroup pm bZp , which coincides with pm Zp . Then e ≠ apmax{n,m} ∈ N ∩ H.
Lemma 2.6. If H is a non-abelian subgroup of L, then Z(H) is trivial.
Proof. Let (m, n) ∈ Z(H), and we show that m = 0 and n = 1.
Let (a, b), (c, d) ∈ H be non-commuting elements. Then (m, n) commutes with both (a, b) and (c, d) while the
latter two elements do not commute with each other. This implies the following:
(1) a(1 − n) = m(1 − b),
(2) c(1 − n) = m(1 − d),
(3) a(1 − d) ≠ c(1 − b).
Multiply (1) by c to obtain ac(1 − n) = mc(1 − b). This together with (2) imply that am(1 − d) = mc(1 − b). In
view of (3), the latter equality is possible only if m = 0. Using (1) − (2) we now obtain a(1 − n) = c(1 − n) = 0. By
(3), either a ≠ 0 or c ≠ 0, and thus n = 1.
Proposition 2.7. If H is a non-abelian subgroup of L, then H is essential in H.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
7
Proof. We first consider the subgroup H1 = H ∩ L′ , which is non-trivial by Lemma 2.4. As L′ = Qp ⋊ {1} is
isomorphic to Qp , Lemma 2.5 implies that H1 is essential in L′ .
Now let N be a non-trivial closed normal subgroup of H, and we have to prove that N ∩ H is non-trivial.
′
′
Since H is non-abelian, its center is trivial by Lemma 2.6, so H is essential in H by Lemma 2.3. Then N ∩ H
is non-trivial, and, in particular, N1 = N ∩L′ is non-trivial. Obviously, N is closed in L, so N1 is a closed subgroup
of L′ . The essentiality of H1 in L′ gives that N1 ∩ H1 is non-trivial, so also N ∩ H is non-trivial.
Corollary 2.8. Let H be a non-abelian subgroup of L. Then:
(1) H is locally minimal;
(2) H is minimal if and only if H is minimal.
Proof. (1): Since H is locally compact, and thus locally minimal, we can apply Proposition 2.7 and the Local
Minimality Criterion.
(2): Apply the Minimality Criterion.
Since L is complete, and compact groups are minimal, we immediately obtain the following consequence of
Corollary 2.8(2).
Corollary 2.9. If H is a precompact non-abelian subgroup of L, then H is minimal.
Recall that a group G is perfectly minimal if G × M is minimal for every minimal group M . The above result
should be compared with the following Megrelishvili’s theorem, ensuring that L is perfectly minimal.
Let K be a topological division ring. A subset B of K is called bounded if for every neighborhood X of 0
there is a neighborhood Y of 0 such that Y B ⊆ X and BY ⊆ X. A subset V of K is retrobounded if 0 ∈ V and
(K ∖ V )−1 is bounded. Then K is called locally retrobounded if it has a local base at 0 consisting of retrobounded
neighborhoods. For example, K is locally retrobounded if it is: locally compact, topologized by an absolute value,
or a linearly ordered field.
Fact 2.10. [24, Theorem 4.7(a)] Let K be a non-discrete locally retrobounded division ring. Then the group
G = K ⋊ K ∗ is perfectly minimal, where the natural action of K ∗ on K = (K, +) by multiplication is considered.
As K = Qp is locally retrobounded, Fact 2.10 entails that L is perfectly minimal. By Lemma 2.6, the non-abelian
subgroups of L are center-free, so we can apply the above results to obtain the following corollary.
Corollary 2.11. For every i ∈ I, let Hi be a non-abelian subgroup of L that is either dense or precompact. Then,
the product ∏i∈I Hi is perfectly minimal.
Proof. If Hi is precompact, then it is minimal by Corollary 2.9. If Hi is dense in L, then it is minimal by Fact
2.10 and Corollary 2.8(2). Thus, Hi is minimal for every i ∈ I. These subgroups are center-free according to
Lemma 2.6. As the arbitrary product of center-free minimal groups is perfectly minimal by [23, Theorem 1.15],
we conclude that ∏i∈I Hi is perfectly minimal.
2.2. Abelian subgroups of L.
Lemma 2.12. An infinite compact subgroup C of G = Q∗p contains an open subgroup of G isomorphic to Zp .
Proof. Note that G ≅ Z × Z∗p , where Z is equipped with the discrete topology, so G ≅ Z × Fp × Zp , and we identify
these two groups.
Being infinite compact, C non-trivially meets the open subgroup U = {0}×{0}×Zp of G, otherwise the projection
of C in the discrete quotient group G/U would be compact infinite. So consider the non-trivial subgroup O = C ∩U
of C. As O is a closed subgroup of U ≅ Zp , it is isomorphic to Zp itself and it has finite index in U , so it is also
open in U . As U is open in G, we conclude that O is open in G.
Now we are in position to prove Theorem A.
Proof of Theorem A. If H is a non-abelian subgroup of L, then Corollary 2.8(1) applies. Let H be an abelian
subgroup of L, and consider the following two possibilities:
Case 1: there is e ≠ h ∈ H ∩ L′ , then obviously H ≤ CL (h). Since L′ = (Qp , +) ⋊ {1}, we have CL (h) = L′ . It
follows that H ≤ L′ ≅ Qp . Then H is locally minimal by Fact 1.6.
Case 2: the subgroup H ∩ L′ is trivial, so the projection L → L/L′ restricted to H gives a continuous group
isomorphism q ∶ H → q(H).
8
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
It is not restrictive to assume H to be non-discrete, and if we prove that the closure of H in L has an open
subgroup isomorphic to Zp , it will follow by Fact 1.6 that H is locally minimal; so we can assume H to be closed
in L.
Then H is a non-discrete locally compact and totally disconnected group, so H contains an infinite compact
open subgroup K by van Dantzig’s theorem [33], and the restriction q ↾K ∶ K → q(K) is a closed map, hence
a topological group isomorphism. The infinite compact subgroup q(K) of q(H) ≤ L/L′ ≅ Q∗p contains an open
subgroup O isomorphic to Zp by Lemma 2.12. Obviously, also q ↾K −1 (O) is isomorphic to Zp , and it is open in
K, hence in H.
3. Locally compact HM groups
In this section we consider hereditarily minimal locally compact groups G. We start with the following immediate consequence of Corollary 1.5.
Lemma 3.1. Let G be a hereditarily minimal locally compact group, and let A be a closed abelian subgroup of G.
Then either A is finite, or A ≅ Zp for some prime p.
In particular, if G is not a torsion group, then it contains a copy of some Zp , so G is not discrete.
Now we give some results on the discrete hereditarily minimal groups.
3.1. Discrete HM groups. Clearly, every locally finite group is torsion. In the following result we recall that
the converse holds true for locally solvable groups.
Fact 3.2. [10, Proposition 1.1.5] Every torsion locally solvable group is locally finite.
The next fact guarantees the existence of an infinite abelian group inside every infinite locally finite group.
Fact 3.3. [15] Every infinite locally finite group contains an infinite abelian subgroup.
From Fact 3.2 and Fact 3.3 we obtain the following corollary.
Corollary 3.4. An infinite locally solvable group contains an infinite abelian subgroup.
The next result will be used in the sequel to conclude that an infinite locally solvable hereditarily minimal
group is not discrete.
Lemma 3.5. If G is an infinite discrete hereditarily minimal group, then the abelian subgroups of G are finite.
In particular, the center of G is finite, G is torsion but it is neither locally finite nor locally solvable.
Proof. By Lemma 3.1, G has no infinite abelian subgroups, so the center of G is finite.
In particular G is torsion, but it is not locally finite by Fact 3.3. Finally, G is not locally solvable by Fact
3.2.
So a discrete HM group is torsion by Lemma 3.5. This result should be compared with Proposition 3.9(2),
where we prove that non-discrete locally compact HM groups are not torsion, as they contain a copy of the p-adic
integers Zp for some p.
The minimality of infinite discrete groups has been a long standing open question of Markov. Since minimal discrete groups admit no non-discrete Hausdorff group topologies at all, such groups are also called non-topologizable.
The first example of a non-topologizable group was provided by Shelah [30] under the assumption of the Continuum Hypothesis CH. His example is simple and torsionfree, so that discrete group is also totally minimal, yet
not hereditarily minimal. A countable example of a non-topologizable group was built a bit later by Ol′ shanskij
[27] (it was an appropriate quotient of Adjan’s group A(m, n) built for the solution of Burnside problem).
There exist infinite hereditarily minimal discrete groups and this fact, recently established by [19] is related to
another interesting topic, namely categorically compact groups. According to [9] a Hausdorff topological group G
is categorically compact (briefly, C-compact) if for every topological group H the projection p ∶ G × H → H sends
closed subgroup of G × H to closed subgroups of H. Compact groups are obviously C-compact by Kuratowski
closed projection theorem, while solvable C-compact groups can be shown to be compact [9, Corollary 3.5]. The
C-compact groups are two-sided complete and the class of C-compact groups has a number of nice properties
typical for the compact groups: stability under taking closed subgroups, finite products and Hausdorff quotient
groups. Moreover, the ω-narrow (in particular, separable) C-compact groups are totally minimal [9, Corollary
3.5]. The question of whether all C-compact groups are compact, raised in [9], remained open for some time
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
9
even in the case of locally compact groups (the connected C-compact locally compact groups are compact [9,
Proposition 5.1]). It was proved in [9, Theorem 5.5], that a countable discrete group G is C-compact if and only
if every subgroup of G is totally minimal (i.e., non-topologizable along with all its quotients). This gives rise to
the following notion (the specific term was proposed later in [20]).
Definition 3.6. A group G is hereditarily non-topologizable when all subgroups of G are non-topologizable along
with all their quotients (i.e., totally minimal in the discrete topology).
It was shown in [9, Corollary 5.4] that every discrete hereditarily non-topologizable group is C-compact, yet
the existence of infinite hereditarily non-topologizable groups remained open until the recent paper [19] which
provided many examples of such groups (hence, discrete C-compact that obviously fail to be compact). These
examples have various additional properties displaying various levels of being infinite (infinite exponent, nonfinitely generated, uncountable, etc.).
3.2. Proof of Theorem B. The first item of the next result is taken from [3, Lemma 2.8].
Lemma 3.7. Let G be a topological group, and N be an open subgroup of G.
(1) If N is locally minimal, then G is locally minimal.
(2) If N is hereditarily locally minimal, then G is hereditarily locally minimal.
Proof. (2): Let H be a subgroup of G. By our assumption on N , the locally minimal subgroup H ∩ N is an open
in H. It follows from item (1) that H is locally minimal.
The following two deep results are due to Zelmanov and Kaplansky, respectively. Note that the second part of
(a) makes use of Fact 3.3.
Fact 3.8.
(a) [34, Theorem 2] Every compact torsion group is locally finite. In particular, every infinite
compact group contains an infinite compact abelian subgroup.
(b) [18, Theorem 6] Every non-discrete Lie group contains a non-trivial continuous homomorphic image of
R.
The next result is the starting point of our exploration of hereditarily minimal locally compact groups. Recall
that a topological group G is called compactly covered if each element of G is contained in some compact subgroup.
Proposition 3.9. Let G be a hereditarily minimal locally compact group.
(1) Then G is totally disconnected and compactly covered.
(2) If G is non-discrete, then it contains a copy of Zp for some prime p and satisfies (Nf n ). In particular,
Z(G) is trivial or Z(G) ≅ Zp for some prime p.
Proof. (1) : Assume that the connected component c(G) is non-trivial. As c(G) is a hereditarily minimal
connected locally compact group, it is a Lie group by Fact 1.6, and of course it is non-discrete. Using Fact 3.8(b),
c(G) contains an infinite abelian Lie group K, which is isomorphic to Zp by Corollary 1.5, contradicting the fact
that K is a Lie group. The second assertion follows from Lemma 3.1.
(2): By (1), G is a non-discrete totally disconnected locally compact group. By [33], it contains an infinite
compact open subgroup H. According to Fact 3.8(a), H contains an infinite hereditarily minimal compact abelian
group A. By Fact 1.3, A is isomorphic to Zp for some prime p.
To prove the second assertion we need to check that G contains no finite normal non-trivial subgroups. Assume
for a contradiction that G has a finite non-trivial normal subgroup F . As G ≥ A ≅ Zp , we deduce that A ∩ F is
trivial, and the natural action by conjugations α ∶ A × F → F has an infinite kernel M = ker α. Being a non-trivial
closed subgroup of A ≅ Zp , M is also isomorphic to Zp . It follows that F ⋊α M = F × M is hereditarily minimal.
Let C be a cyclic non-trivial subgroup of F . Then C × Zp is a hereditarily minimal compact abelian group,
contradicting Fact 1.3. For the final assertion, apply Lemma 3.1 to the closed normal subgroup Z(G).
The following example shows that the hypotheses of the above proposition cannot be relaxed.
Example 3.10. Our first item shows that hereditary minimality cannot be relaxed to hereditary local minimality,
while in item (b) we show that an infinite discrete HM group may have finite non-trivial center, so need not satisfy
(Nf n ). So “non-discrete” cannot be removed in Proposition 3.9(2).
(a) The Lie group R × Z(2) is certainly hereditarily locally minimal by Fact 1.6, but not totally-disconnected.
10
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
(b) Let T be a countable discrete hereditarily non-topologizable group (see §3.1). By [9, Theorem 5.5], T is Ccompact. As this class is stable under taking finite products, also the direct product G = Z(2)×T is C-compact.
Since G is also countable and discrete, [9, Theorem 5.5] applies again, and G is hereditarily non-topologizable.
Thus, G is a discrete hereditarily minimal group with finite non-trivial center.
Proposition 3.9(2) has the following easy consequence: as Zp × Zq is not HM for any pair p, q of primes, direct
products of non-discrete locally compact hereditarily minimal groups are never hereditarily minimal.
Recall that a profinite group is a totally disconnected compact group. The following proposition can be applied
for example to profinite torsionfree groups, as a torsionfree group obviously satisfies (Nf n ).
Proposition 3.11. Let G be an infinite profinite group satisfying (Nf n ). If H is a locally minimal dense subgroup
of G, then H is minimal.
Proof. We prove first that if H is a locally essential subgroup of G, then it is essential in G. Since G is profinite,
there exists a local base at the identity B consisting of compact open normal subgroups. Suppose that H is locally
essential in G, and let K ∈ B be such that M ∩ H is not trivial for every non-trivial closed normal subgroup M
of G contained in K.
Let N be a non-trivial closed normal subgroup of G, and we will show that N ∩ H is not trivial. As N is
compact, and M = N ∩ K is an open subgroup of N , the index [N ∶ M ] is finite. This yields that M is infinite
since N is infinite by our assumption on G. Hence, M ⊆ K is a non-trivial closed normal subgroup of G. This
implies that M ∩ H is not trivial. Since M ∩ H ⊆ N ∩ H the latter group is also non-trivial.
Let H be a locally minimal dense subgroup of the compact group G. By the Local Minimality Criterion, H is
locally essential in G, hence essential by the above argument. Now apply the Minimality Criterion to conclude
that H is minimal.
The compact group T has plenty of dense non-minimal subgroups and they are all locally minimal by Fact 1.6,
so “profinite” cannot be relaxed to “compact” in the above proposition.
By Proposition 3.9, the compact HM groups satisfy the assumptions of Proposition 3.11. Using this fact, now
we prove Theorem B: a compact group is hereditarily minimal if and only if it is hereditarily locally minimal and
satisfies (Cf n ).
Proof of Theorem B. If G is a hereditarily minimal group, then it is hereditarily locally minimal. To prove
G satisfies (Cf n ), pick an infinite compact subgroup H of G. Clearly, H is also hereditarily minimal, so satisfies
(Nf n ) by Proposition 3.9(2). This means that G satisfies (Cf n ).
For the converse implication, suppose that G is a hereditarily locally minimal compact group satisfying (Cf n ).
We first show that G is totally disconnected. Assuming the contrary, the connected component c(G) is a
non-trivial hereditarily locally minimal, connected, compact group. By Fact 1.6, c(G) is a non-discrete Lie group.
According to Fact 3.8(b), c(G) contains an infinite compact abelian Lie group C. In particular, C contains a
copy of T, hence contains torsion elements. So C does not satisfy (Nf n ), contradicting our assumption that G
satisfies (Cf n ).
Let H be a subgroup of the profinite group G. Without loss of generality we may assume that H is infinite.
By our assumption, H is a profinite hereditarily locally minimal group satisfying (Nf n ). By Proposition 3.11,
applied to H and its subgroup H, we deduce that H is minimal.
Our first application of Theorem B shows that the groups Mp,n are HM for every prime p and every n ∈ N.
Example 3.12. Being a subgroup of (Zp , +) ⋊ Z∗p , the group Mp,n = (Zp , +) ⋊ Cpp is hereditarily locally minimal
by Theorem A. As Mp,n is torsionfree, it is also hereditarily minimal by Theorem B.
n
The following proposition proves the implication (2) in the diagram in the introduction under an assumption
much weaker than total disconnectedness (equivalent to the property of having a compact connected component1).
Proposition 3.13. Let G be a locally compact group having a compact open subgroup H.
(1) If G is compactly hereditarily locally minimal, then it is hereditarily locally minimal.
(2) If G is compactly hereditarily minimal, then it is either discrete or contains a copy of Zp for some prime
p.
1or containing no lines, i.e., copies of R.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
11
Proof. (1): Since G is compactly hereditarily locally minimal, H is hereditarily locally minimal, and Lemma
3.7(2) applies.
(2): If G is non-discrete and compactly hereditarily minimal, then also H is non-discrete. Being also hereditarily
minimal and compact, Proposition 3.9(2) applies to H.
4. Proof of Theorem C
We start this section with the following apparently folklore property of central extensions. Recall that a group
is p-torsionfree if it has no element of order p.
Lemma 4.1. Let G be a group with non-trivial center. If Z(G) is p-torsionfree and G/Z(G) is a p-group, then
Z2 (G) = Z(G). If in addition G/Z(G) is finite, then G is abelian.
Proof. Let Z1 = Z(G) be p-torsionfree and G/Z1 be a p-group. Assume for a contradiction that Z1 ≨ Z2 , where
Z2 = Z2 (G). We can choose x ∈ Z2 ∖ Z1 such that xp ∈ Z1 . Since x ∉ Z1 , there exists y ∈ G that does not
commute with x. Then a = [x, y] = xyx−1 y −1 ≠ e, and a ∈ Z1 , by the definition of Z2 . Let φ ∶ G → G be the inner
automorphism induced by x and observe that φ(y) = xyx−1 = ay. Since xp ∈ Z1 is a central element, φp = IdG .
On the other hand, as a ∈ Z1 , we also have φp (y) = xp yx−p = ap y. So y = ap y, and it follows that ap = e. Since
a ∈ Z1 , which is p-torsionfree, this yields a = e, a contradiction.
For the last assertion, recall that finite p-groups are nilpotent, so if G is not abelian, then Z(G/Z1 (G)) is
non-trivial, i.e., Z1 (G) ≨ Z2 (G).
In the following result, we prove some more properties a non-discrete hereditarily minimal, locally compact
group with non-trivial center satisfies.
Corollary 4.2. Let G be a non-discrete hereditarily minimal locally compact group with non-trivial center. Then
there is a prime number p such that:
(1) for every non-central element x ∈ G, we have Zp ≅ Z(G) ≤ ⟨x⟩ ≅ Zp . In particular, for every closed
subgroup H of G, either H ≤ Z(G), or Z(G) ≤ H;
(2) G is torsionfree, and G/Z(G) is a center-free p-group.
Proof. (1): The center Z(G) is isomorphic to Zp for some prime p by Proposition 3.9(2).
Let x ∈ G be a non-central element, and consider the abelian subgroup Nx = ⟨Z(G), x⟩ of G. As Nx contains
Z(G) ≅ Zp , it is isomorphic to Zp by Lemma 3.1. As the closed subgroups of Zp are totally ordered by inclusion,
we obtain that Z(G) ≤ ⟨x⟩ = Nx ≅ Zp . It follows that if H is a closed non-central subgroup of G, then Z(G) ≤ H.
Nx .
n
(2): If x ∉ Z(G), then by item (1), the index [⟨x⟩ ∶ Z(G)] = pn for some n ∈ N, so xp ∈ Z(G). This proves
that G/Z(G) is a p-group and that G is torsionfree, so Z2 (G) = Z(G) by Lemma 4.1.
In the next result, we use Corollary 4.2 to extend Corollary 1.5.
Corollary 4.3. An infinite hereditarily minimal locally compact hypercentral group G is isomorphic to Zp for
some prime p.
Proof. By a well-known theorem of Mal’cev (see [10, page 8]), all hypercentral groups are locally nilpotent, so in
particular G is locally solvable. Then G is not discrete by Lemma 3.5, and clearly it has non-trivial center. Hence
Z2 (G) = Z(G) ≅ Zp by Corollary 4.2. Since G is hypercentral, we deduce that G = Z(G) ≅ Zp .
The next results will be used in the subsequent proof of Theorem C.
Proposition 4.4. If G is a hereditarily minimal locally compact group with non-trivial center, then every abelian
subgroup of the quotient group G/Z(G) is finite.
Proof. Assume that G/Z(G) has an abelian subgroup, so also its closure A is abelian. Let H = q −1 (A), where
q ∶ G → G/Z(G) is the canonical homomorphism. So, H is a locally compact subgroup of G with H ≥ Z(G).
Hence, Z(H) ≥ Z(G) ∩ H = Z(G), and by the third isomorphism theorem we deduce that H/Z(H) is a quotient
of the abelian group H/Z(G) ≅ A, so H is nilpotent (of class ≤ 2).
If G is discrete, then H is finite by Lemma 3.5. As H = q −1 (A), we deduce that also A is finite. Now we assume
that G is non-discrete. By Proposition 3.9(2), Z(G) ≅ Zp for some prime p. As H contains Z(G), Corollary 4.3
implies that also H is isomorphic to Zp . So A ≅ H/Z(G) is finite.
12
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
The quotient group Q = G/Z(G) is torsion, by Lemma 3.5 and Corollary 4.2. According to Fact 3.2 and
Fact 3.3, the property of Q described in Proposition 4.4 is equivalent also to having no infinite locally solvable
subgroup, or having no infinite locally finite subgroup.
Corollary 4.5. Let G be a non-discrete hereditarily minimal locally compact group with non-trivial center. If
G/Z(G) is a locally finite group, then G ≅ Zp for some prime p.
Proof. By Corollary 4.2, we obtain that G/Z(G) is a p-group and Z(G) ≅ Zp .
If G/Z(G) is infinite, then it has an infinite abelian subgroup by Fact 3.3, contradicting Proposition 4.4. So
G/Z(G) is finite, and G is abelian by Lemma 4.1.
As a consequence of Corollary 4.5 we now prove Theorem C.
Proof of Theorem C. Let G be an infinite hereditarily minimal locally compact group with non-trivial center
that is either compact or locally solvable. We have to prove that G ≅ Zp , for some prime p.
First note that G is non-discrete by Lemma 3.5. Applying Corollary 4.2, we obtain that G/Z(G) is a p-group
and Z(G) ≅ Zp . In view of Corollary 4.5, it suffices to prove that G/Z(G) is locally finite. If G is locally solvable,
then its quotient G/Z(G) is locally finite by Fact 3.2. If G is compact, then G/Z(G) is compact torsion, so locally
finite by Fact 3.8(a).
Let us see that the assumption “compact or locally solvable” cannot be removed in Theorem C. Recall that the
countable discrete group G = Z(2) × T from Example 3.10 is hereditarily minimal with non-trivial center. Clearly,
this group is non-abelian, and it is neither locally solvable nor compact.
Applying Theorem C to the groups studied in Corollary 4.2 (see also Proposition 4.4), we can deduce additional
properties they share.
Theorem 4.6. Let G be a non-discrete hereditarily minimal locally compact group, and assume {e} ≠ Z(G) ≠ G.
Then there exists a prime p such that:
(1) every non-trivial closed subgroup H of G (e.g., Z(G)) is open;
(2) every non-trivial compact subgroup H of G is isomorphic to Zp ;
(3) every finite subgroup of G/Z(G) is cyclic and G/Z(G) satisfies (Nf n ).
Proof. By Corollary 4.2, G is torsionfree, Z(G) ≅ Zp and G/Z(G) is a p-group, for some prime p.
(1): First we prove that Z(G) is open in G. By Proposition 3.9(1), G is totally disconnected, so it has a local
base at the identity consisting of compact open subgroups, and let K be one of these subgroups. If K is central,
then Z(G) is open. Otherwise, we have K ≥ Z(G) by Corollary 4.2(1), so K ≅ Zp by Theorem C. Moreover, the
index [K ∶ Z(G)] is finite and Z(G) is open in K, so Z(G) is open in G.
Now let H be a non-trivial closed subgroup of G. If it is contained in Z(G) then, being closed in Z(G) ≅ Zp ,
it is also open in Z(G), hence open in G. Otherwise, H contains Z(G) by Corollary 4.2(1), so H is open.
(2): If H is a non-trivial compact subgroup of G, then H is infinite as G is torsionfree. This implies that
H ∩ Z(G) ≠ {e}, since Z(G) is open in G. So we deduce that H ≅ Zp by Theorem C.
(3): Let π ∶ G → G/Z(G) be the canonical map, and let F be a finite subgroup of G/Z(G). Then π −1 (F ) is a
closed subgroup of G, containing ker π = Z(G) ≅ Zp , and such that [π −1 (F ) ∶ Z(G)] = ∣F ∣ is finite. Then π −1 (F )
is compact, hence isomorphic to Zp by item (2). If g ∈ G is such that Zp ≅ ⟨g⟩ = π −1 (F ), then F = ⟨π(g)⟩.
To check that G/Z(G) satisfies (Nf n ) pick a finite non-trivial normal subgroup N of G/Z(G) and let ∣N ∣ = pn .
By what we have just proved, H = π −1 (N ) is isomorphic to Zp , and indeed Zp ≅ Z(G) = pn H. Let x ∈ H ∖ Z(G),
and y ∈ G be an element non-commuting with x. As H is normal in G, we can consider the conjugation by y as a
map φ ∶ H → H. Since pn x ∈ Z(G), we have φ(pn x) = pn x; on the other hand, φ(pn x) = pn φ(x), so we conclude
pn φ(x) = pn x. As H is abelian and torsionfree, we deduce φ(x) = x, a contradiction.
Note that G as in Theorem 4.6 is neither compact nor locally solvable, by Theorem C. See Questions 7.2 and
7.3 for further comments and Proposition 7.4 for a partial converse.
We conclude this section by listing the three possibilities (trichotomy) for an infinite locally compact nonabelian HM group G:
(1) G is discrete, if and only if G is torsion (by Lemma 3.5 and Proposition 3.9(2)). In this case G is not
locally finite but may have finite non-trivial center (see Example 3.10).
If G is not discrete, we apply Proposition 3.9, and we obtain the following two cases:
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
13
(2) Z(G) = {e}, G contains a copy of Zp , and satisfies (Nf n ).
(3) Z(G) = Z2 (G) ≅ Zp is open and proper in G, and G has the properties listed in Corollary 4.2 and Theorem
4.6.
To case (1) was dedicated §3.1, while case (2) (for solvable groups) will be the subject of the rest of the paper.
Note that the groups in (1) and in (3) are neither compact, nor locally solvable, by Lemma 3.5 and Theorem
C.
5. Semidirect products of p-adic integers
We begin this section with a general result on semidirect products, and their quotients. We then apply it in
the subsequent Lemmas 5.2 and 5.5, where we consider some of the semidirect products introduced in Example
1.11.
Lemma 5.1. Let α ∶ Y ×X → X be a continuous action of an abelian group Y on an abelian group X, and consider
the topological semidirect product G = X⋊α Y . Consider the subgroup of X defined by A = ⟨x−α(y, x) ∶ x ∈ X, y ∈ Y ⟩.
Then the derived group G′ coincides with A ⋊α {eY }, and the quotient group G/G′ is topologically isomorphic
to (X/A) × Y .
Proof. For y ∈ Y , let Ay = {x − α(y, x) ∶ x ∈ X}, then A = ⟨Ay ∶ y ∈ Y ⟩.
[(x, eX ), (eY , y)] = (x − α(y, x), eY ) for every x ∈ X and y ∈ Y , so A ⋊ {eY } ≤ G′ .
On the other hand, every y, y ′ ∈ Y commute, thus
Note that the commutator
α(y ′ , Ay ) = {α(y ′ , x) − α(y ′ , α(y, x)) ∶ x ∈ X} = {α(y ′ , x) − α(y, α(y ′ , x)) ∶ x ∈ X} ≤ Ay ≤ A,
which implies that A ⋊α {eY } is normal in G. Let χ ∶ G → (X/A) × Y be defined by χ(x, y) = (φ(x), y), where
φ ∶ X → X/A is the canonical map. Since φ is a continuous open surjection it follows that χ is a continuous
open surjection. Moreover, the definitions of φ and A imply that χ is also homomorphism. It is easy to see that
ker χ = A ⋊α {eY }, so G/(A ⋊α {eY }) ≅ (X/A) × Y is abelian and G′ ≤ A ⋊α {eY }.
Recall that the topological group (Cp , ⋅) is isomorphic to the group (Zp , +) (essentially, via the p-adic logarithm),
n
so the closed subgroups of (Cp , ⋅) are totally ordered, have the form Cpp , for n ∈ N, and have been described in
(1.1).
n
Now we apply Lemma 5.1 to the case when (Cpp , ⋅), viewed as a subgroup of the automorphism group Aut (Zp ),
n
acts on Zp via the natural action by multiplication. Recall that we denote by Mp,n = Zp ⋊ Cpp the semidirect
product arising this way.
n
Lemma 5.2. For the group Mp,n = Zp ⋊ Cpp , the following hold:
′
′
● if p > 2, then Mp,n
= pn+1 Zp ⋊ {1} and Mp,n /Mp,n
≅ Z(pn+1 ) × Cpp ;
n
′
n+2
′
● if p = 2, then M2,n = 2 Z2 ⋊ {1} and M2,n /M2,n ≅ Z(2n+2 ) × C22 .
n
Proof. In the notation of Lemma 5.1, let A = ⟨(1 − y)x ∶ x ∈ Zp , y ∈ Cpp ⟩.
By (1.1), one immediately obtains that if p > 2, then A = pn+1 Zp , while A = 2n+2 Z2 otherwise.
n
In the following remark we give the explicit isomorphisms stated in the above Lemma 5.2.
Remark 5.3. Using the proof of Lemma 5.1 we can write explicitly the isomorphisms in Lemma 5.2. Let p be a
′
̃ satisfies the equality ψ((x,
̃
prime and n ∈ N. The isomorphism ψ
y)Mp,n
) = ψ(x, y) for every (x, y) ∈ Mp,n , where:
● if p > 2, ψ ∶ Mp,n → Z(pn+1 ) × Cpp is defined by ψ(x, y) = (x mod pn+1 , y);
n
● if p = 2, ψ ∶ M2,n → Z(2n+2 ) × C22 is defined by ψ(x, y) = (x mod 2n+2 , y).
̃ ○ q, where q ∶ Mp,n → Mp,n /M ′ is the canonical map.
In other words, we have ψ = ψ
p,n
n
Obviously, if two groups Mp,n , Mp′ ,n′ are isomorphic, then p = p′ . Under this assumption, we now prove that
also n = n′ .
n
Corollary 5.4. For n ∈ N, the subgroups Mp,n = Zp ⋊ Cpp of Mp,0 = Mp = Zp ⋊ Cp are pairwise non-isomorphic.
′
′
Proof. Let n, m ∈ N, and assume that ψ ∶ Mp,n → Mp,m is an isomorphism. Then ψ(Mp,n
) = Mp,m
, and ψ induces
′
′
an isomorphism ψ̄ ∶ A → B, where A = Mp,n /Mp,n , and B = Mp,m /Mp,m .
By Lemma 5.2, comparing the torsion subgroups of A and B, we obtain Z(pn+1 ) ≅ Z(pm+1 ) when p > 2, or
Z(2n+2 ) ≅ Z(2m+2 ) when p = 2. In any case, n = m.
14
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
The following result is the counterpart of Lemma 5.2 for the groups Tn .
Lemma 5.5. For every n ∈ N, the commutator subgroup of the group Tn is Tn′ = 2Z2 ⋊ {1}, and the quotient group
n
Tn /Tn′ is isomorphic to Z(2) × C22 .
Proof. Let A = ⟨x − β(y, x) ∶ x ∈ Z2 , y ∈ C22 ⟩ ≤ Z2 . In view of Lemma 5.1, we have to prove that A = 2Z2 . By
n+1
the definition of β we have A = ⟨V, W ⟩, where V = ⟨x − yx ∶ x ∈ Z2 , y ∈ C22 ⟩ ≤ Z2 and W = ⟨x + yx ∶ x ∈ Z2 , y ∈
n
n+1
C22 ∖ C22 ⟩ ≤ Z2 .
n+1
n+1
Note that if y ∈ C22 , then y ∈ 1 + 2Z2 , so 1 − y ∈ 2Z2 . In particular, V = ⟨(1 − y)x ∶ x ∈ Z2 , y ∈ C22 ⟩ ≤ 2Z2 .
n
n+1
To study W , first observe that C22 ∖ C22
= (1 + 2n+2 Z2 ) ∖ (1 + 2n+3 Z2 ) = 1 + 2n+2 + 2n+3 Z2 . Hence,
n
W = ⟨x(1 + y) ∶ x ∈ Z2 , y ∈ 1 + 2n+2 + 2n+3 Z2 ⟩ = ⟨xt ∶ x ∈ Z2 , t ∈ 2 + 2n+2 + 2n+3 Z2 ⟩ ≤ 2Z2 .
On the other hand, 2Z2 = (2 + 2n+2 )Z2 ≤ W , since t = 2 + 2n+2 ∈ 2 + 2n+2 + 2n+3 Z2 . It is now clear that
A = ⟨V, W ⟩ = W = 2Z2 , which completes the proof.
Proposition 5.6. For n ∈ N, the groups Tn are pairwise non-isomorphic.
m
Proof. Assume that there exists a topological isomorphism ψ ∶ Tn → Tm . Let π2 ∶ Tm → C22 be the projection on
the second coordinate. We first show that π2 (ψ(Z2 ⋊ {1})) = 1. If a ∈ Z2 and π2 (ψ(a, 1)) = c, then π2 (ψ(2a, 1)) =
′
c2 . By Lemma 5.5, Tn′ = Tm
= 2Z2 ⋊ {1}, which implies that ψ(2Z2 ⋊ {1}) = 2Z2 ⋊ {1} and π2 (ψ(2Z2 ⋊ {1})) = 1.
m
As (2a, 1) ∈ 2Z2 ⋊ {1}, we deduce that c2 = 1. It follows that c = 1, since C22 is torsionfree.
n+1
m+1
Consider the subgroups M2,n+1 = Z2 ⋊βn C22
and M2,m+1 = Z2 ⋊βm C22
of Tn and Tm , respectively. We
will prove that ψ(M2,n+1 ) = M2,m+1 . Since ψ −1 is also a topological isomorphism, it suffices to show that
ψ(M2,n+1 ) ≤ M2,m+1 .
m+1
For this aim we prove that π2 (ψ(M2,n+1 )) ≤ C22 . Note that an element of M2,n+1 has the form (a, b2 ), where
n
m
a ∈ Z2 and b ∈ C22 . Clearly, π2 (ψ(0, b)) ∈ C22 , so
m+1
π2 (ψ(a, b2 )) = π2 (ψ(a, 1))π2 (ψ(0, b))2 = 1 ⋅ π2 (ψ(0, b))2 ∈ C22
Hence, π2 (ψ(M2,n+1 )) ≤
m = n.
m+1
C22 ,
.
which means that ψ(M2,n+1 ) ≤ M2,m+1 . By Corollary 5.4, we deduce that
It is not difficult to see that if Tn ≅ Mp,m for some n, m, p, then p = 2. In the next result we apply Lemma 5.5
to prove also that a group Tn is not isomorphic to any of the groups M2,m .
Proposition 5.7. For every n, m ∈ N, and every prime number p, the groups Tn and Mp,m are not isomorphic.
Proof. By contradiction, assume that ψ ∶ Tn → Mp,m is an isomorphism. Similar to the proof of Corollary 5.4, ψ
′
induces an isomorphism between the torsion groups t(Tn /Tn′ ) and t(Mp,n /Mp,n
). But t(Tn /Tn′ ) ≅ Z(2) by Lemma
′
n+2
′
n+1
5.5, while t(M2,n /M2,n ) ≅ Z(2 ) and t(Mp,n /Mp,n ) ≅ Z(p ) when p > 2 by Lemma 5.2.
In the sequel, we consider a faithful action α ∶ Zp × Zp → Zp of Zp on Zp , and the semidirect product Mp,α =
(Zp , +) ⋊α (Zp , +) arising this way. Recall that Aut (Zp ) ≅ Z∗p , as every φ ∈ Aut (Zp ), has the form φ(x) = m ⋅ x
for m = φ(1) ∈ Z∗p ; identifying φ with φ(1), the action α gives a group monomorphism f ∶ (Zp , +) → Z∗p such that
α(y, x) = f (y) ⋅ x.
Proposition 5.8. For a prime p, consider the semidirect product Mp,α = (Zp , +) ⋊α (Zp , +), where α is a faithful
action.
● If p > 2, then Mp,α ≅ Mp,n for some n ∈ N.
● If p = 2, then either Mp,α ≅ Mp,n , or Mp,α ≅ Tn , for some n ∈ N.
Proof. Since α is faithful, there is a group monomorphism f ∶ (Zp , +) → Z∗p = Cp Fp such that α(y, x) = f (y) ⋅ x.
Now we consider two cases, depending on whether the image of f is contained in Cp or not.
If it is, then f ∶ (Zp , +) → Cp is continuous when we equip these two copies of (Zp , +) with the p-adic topology.
n
n
So f (Zp ) = Cpp for some n ∈ N, and f ∶ (Zp , +) → Cpp is a topological isomorphism.
n
We define φ ∶ (Zp , +) ⋊α (Zp , +) → (Zp , +) ⋊ (Cpp , ⋅) by (x, y) → (x, f (y)). To prove φ to be a topological
isomorphism, it remains only to check it is a homomorphism, as follows:
φ((x1 , y1 )(x2 , y2 )) = φ(x1 + α(y1 , x2 ), y1 y2 ) = φ(x1 + f (y1 ) ⋅ x2 , y1 y2 ) = (x1 + f (y1 ) ⋅ x2 , f (y1 )f (y2 )),
(5.1)
φ(x1 , y1 )φ(x2 , y2 ) = (x1 , f (y1 ))(x2 , f (y2 )) = (x1 + f (y1 ) ⋅ x2 , f (y1 )f (y2 )).
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
15
Now we consider the case when f (Zp ) ⊈ Cp . First we see that this happens only if p = 2; indeed this follows
from the fact that if p > 2 then (Zp , +) is (p − 1)-divisible, while Fp has cardinality p − 1. So we have p = 2 and
f ∶ (Z2 , +) → Z∗2 = C2 F2 such that α(y, x) = f (y) ⋅ x and f (Z2 ) ⊈ C2 . Recall that C2 = 1 + 4Z2 and F2 = {1, −1}.
Equipping C2 ≅ Z2 with the 2-adic topology, F2 with the discrete topology, and the codomain of f with the
product topology, it is easy to see that f ∶ (Z2 , +) → C2 ⋅ F2 is continuous, so f (1) ∈ (−1) ⋅ C2 .
Consider the (continuous) canonical projection π ∶ C2 ⋅ F2 → C2 , and call f˜ the composition map f˜ = π ○ f ∶
Z2 → C2 . Then f˜ is a continuous homomorphism, f˜(1) = −f (1), and it is easy to see that
⎧
⎪
if y ∈ 2Z2 ,
⎪f (y)
f˜(y) = ⎨
⎪
⎪
⎩−f (y) if y ∈ Z2 ∖ 2Z2 .
n
n
Since Z2 is compact, there is n ∈ N such that f˜(Z2 ) = C22 = (1+4Z2 )2 , and we now prove that M2,α is isomorphic
to Tn .
n
Let φ ∶ M2,α → Tn be defined by φ(x, y) = (x, f˜(y)). As f˜ ∶ (Z2 , +) → C22 is a topological group isomorphism
we deduce that φ is homeomorphism. Let us show that φ is also a homomorphism. If (x1 , y1 ), (x2 , y2 ) ∈ M2,α ,
then
φ(x1 , y1 )φ(x2 , y2 ) = (x1 , f˜(y1 ))(x2 , f˜(y2 )) = (x1 + βn (f˜(y1 ), x2 ), f˜(y1 )f˜(y2 )).
On the other hand, since f˜ is a homomorphism we have
φ((x1 , y1 )(x2 , y2 )) = φ(x1 + f (y1 ) ⋅ x2 , y1 + y2 ) = (x1 + f (y1 ) ⋅ x2 , f˜(y1 + y2 )) = (x1 + f (y1 ) ⋅ x2 , f˜(y1 )f˜(y2 )).
To finish the proof, we now check that βn (f˜(y1 ), x2 ) = f (y1 ) ⋅ x2 .
n+1
If y1 ∈ 2Z2 , then f˜(y1 ) = f (y1 ) and also f˜(y1 ) ∈ C22 , so βn (f˜(y1 ), x2 ) = f˜(y1 ) ⋅ x2 = f (y1 ) ⋅ x2 .
n
n+1
If y1 ∈ Z2 ∖ 2Z2 , then f˜(y1 ) = −f (y1 ), and moreover f˜(y1 ) ∈ C22 ∖ C22 , so βn (f˜(y1 ), x2 ) = −f˜(y1 ) ⋅ x2 =
f (y1 ) ⋅ x2 .
In the following lemma we describe the faithful actions of a finite group on the p-adic integers.
Lemma 5.9. Let α ∶ H × Zp → Zp be a faithful action of a finite group H on Zp . Then H is isomorphic to a
subgroup F of Fp , and G = (Zp , +) ⋊α H is topologically isomorphic to Kp,F .
Proof. Since α is faithful, there is a group monomorphism f ∶ H → Z∗p such that α(h, x) = f (h) ⋅ x. As H is finite,
its image F is contained in the torsion subgroup Fp of Z∗p .
Consider the map φ ∶ G → Kp,F defined by φ(a, b) = (a, f (b)) for every (a, b) ∈ Kp,F . Following the argument
in (5.1), one can verify that φ is a topological isomorphism.
6. Proof of Theorem D
In this section we prove that infinite locally compact solvable HM groups are metabelian (see Proposition 6.7).
Then we use it to classify all locally compact solvable HM groups.
We start this section with two general results we use in the sequel. Especially Lemma 6.2 will be used in
Theorem 6.3, Proposition 6.8, Proposition 6.10 and Theorem 6.13.
Lemma 6.1. Let G be an infinite hereditarily minimal locally compact solvable group of class n > 1. Then, G(n−1)
is isomorphic to Zp for some prime p.
Proof. Since G is solvable of class n > 1, G(n−1) is a non-trivial normal abelian subgroup of G. By Lemma 3.5, G
is non-discrete. Hence, G(n−1) is infinite, by Proposition 3.9. Being infinite hereditarily minimal locally compact
abelian group, G(n−1) is isomorphic to Zp for some prime p, by Corollary 1.5.
Let p be a prime. Recall that an abelian group G is p-divisible if pG = G.
Lemma 6.2. Let p and q be distinct primes and α ∶ Zq × Zp → Zp be a continuous action by automorphisms.
Then Zp ⋊α Zq is not hereditarily locally minimal.
Proof. In case K = ker α is trivial, then Zq is algebraically isomorphic to a subgroup of Aut (Zp ) ≅ Z∗p . This
is impossible since Zq is p-divisible, while Z∗p = Cp Fp contains no infinite p-divisible subgroups. Hence, K is a
non-trivial closed subgroup of Zq , so isomorphic to Zq itself. By Fact 1.6, the group Zp ⋊α K ≅ Zp × Zq is not
hereditarily locally minimal.
16
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
The following theorem provides a criterion for hereditary minimality in terms of properties of the closed
subgroups of a compact solvable group. We are going to use it in Example 6.5 to check that the groups Kp,F are
hereditarily minimal.
Theorem 6.3. Let G be a compact solvable group. The following conditions are equivalent:
(1) G is hereditarily minimal;
(2) there exists a prime p, such that for every infinite closed subgroup H of G either Z(H) = {e} or H ≅ Zp ;
(3) there exists a prime p, such that for every infinite closed subgroup H of G either Z(H) = {e} or Z(H) ≅ Zp .
Proof. Since the assertion of the theorem is trivially true for finite or abelian groups, we can assume that G is
infinite and solvable of class n > 1.
(1) ⇒ (2): By Lemma 6.1, B = G(n−1) ≅ Zp for some prime p. Let H be an infinite closed subgroup of G such
that Z(H) is non-trivial. By Theorem C, H ≅ Zq for some prime q. If q ≠ p, then H ∩ B is trivial. Since B is
normal subgroup of G we deduce that B ⋊ H ≅ Zp ⋊ Zq is hereditarily minimal, contradicting Lemma 6.2.
(2) ⇒ (3): Trivial.
(3) ⇒ (1): Let H be an infinite subgroup of G. We will prove that H is minimal. If H is abelian, then
Z(H) = H ≤ Z(H) and by (3) we deduce that Z(H) is isomorphic to Zp , so, in particular, its subgroup H is
minimal by Prodanov’s theorem. Hence, we can assume that H is non-abelian, so solvable of class m where
n ≥ m > 1, and we have to show that it is essential in H.
Consider the closed non-trivial abelian subgroup M = H (m−1) of H. We prove that M ≅ Zp . It suffices to show,
by our assumption (3) that M is infinite. Assuming the contrary, let M be finite. By Corollary 3.4, H contains
an infinite closed abelian subgroup A. By (3), A ≅ Zp and thus A ∩ M is trivial. As M is a normal subgroup
of H, the topological semidirect product M ⋊α A is well defined, where α is the natural action by conjugations
α ∶ A × M → M. Being an infinite group that acts on a finite group, A must have a non-trivial kernel K = ker α.
Hence, K ≅ Zp as a closed non-trivial subgroup of A ≅ Zp . It follows that H contains a subgroup isomorphic to
M ⋊α K ≅ M × K ≅ M × Zp . Let C be a non-trivial cyclic subgroup of M , then G contains an infinite closed
abelian subgroup L ≅ C × Zp that is not isomorphic to Zp , contradicting (3).
Coming back to the proof of the essentiality of H in H, let N be a non-trivial closed normal subgroup of H.
If N1 = N ∩ M is trivial, then N M ≅ N × M ≅ N × Zp . Following the same ideas of the first part of the proof, one
can find an infinite abelian subgroup of G not isomorphic to Zp , contradicting (3). Therefore N1 and H1 = H ∩ M
are non-trivial subgroups of M ≅ Zp , with N1 closed in M . By Lemma 2.5, N1 ∩ H1 is non-trivial.
Since N ∩ H ≥ N1 ∩ H1 , we conclude that N ∩ H ≠ {e}, proving the essentiality of H in H.
Now we give an application of Theorem 6.3 that we use in Example 6.5.
Lemma 6.4. Let G be a compact solvable torsionfree group containing a closed hereditarily minimal solvable
(k−1)
subgroup G1 of class k. If G1
minimal.
≅ Zp , and [G ∶ G1 ] = pn , for some n ∈ N and a prime p, then G is hereditarily
Proof. According to Theorem 6.3, it suffices to check that for every infinite closed subgroup H of G, either
Z(H) = {e} or H ≅ Zp . Assume that H is an infinite closed subgroup of G, and let H1 = G1 ∩ H. Then there
exists r ≤ n such that
(6.1)
[H ∶ H1 ] = pr .
So, H1 is an infinite closed subgroup of G1 .
m
Now assume that Z(H) is non-trivial and pick any e ≠ z ∈ Z(H). Then e ≠ z p ∈ H1 for some integer 0 < m ≤ r,
pm
by (6.1). As z ∈ Z(H1 ), this proves that Z(H1 ) ≠ {e}, so H1 ≅ Zp by Theorem 6.3.
m
Let A = Z(H) ∩ H1 . As z p ∈ A, it is a non-trivial closed subgroup of H1 ≅ Zp , so A is isomorphic to Zp and
[H1 ∶ A] = ps for some s ∈ N. Then pr+s = [H ∶ H1 ][H1 ∶ A] = [H ∶ A] = [H ∶ Z(H)][Z(H) ∶ A], so H/Z(H) is a
finite p-group. Thus H is abelian by Lemma 4.1.
Since the compact abelian torsionfree group H contains a subgroup isomorphic to Zp of finite index, we deduce
that H itself is isomorphic to Zp .
Example 6.5. Now we use Theorem 6.3 and Lemma 6.4 to see that the the infinite compact metabelian groups
n
Kp,F and Tn = (Z2 , +) ⋊β C22 are hereditarily minimal.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
17
(a) To show that Kp,F = Zp ⋊ F is hereditarily minimal, pick an infinite closed subgroup H of Kp,F . If H is
non-abelian, then Z(H) is trivial by Lemma 2.6. If H is abelian, then there exists e ≠ x ∈ H ∩ (Zp ⋊ {1})
since F is finite. As H ≤ CKp,F (x) ≤ Zp ⋊ {1}, we obtain that H ≅ Zp . By Theorem 6.3, we conclude that
Kp,F is hereditarily minimal.
n+1
(b) The group Tn is torsionfree, and its compact subgroup (Z2 , +) ⋊β C22
≅ M2,n+1 has index 2 in Tn .
′
Moreover, M2,n+1 is hereditarily minimal by Example 3.12, while M2,n+1
≅ Z2 by Lemma 5.2. So we
deduce by Lemma 6.4 that Tn is hereditarily minimal.
Alternatively, since its open subgroup M2,n+1 is hereditarily locally minimal, the group Tn is also hereditarily locally minimal by Lemma 3.7(2). As Tn is torsionfree, Theorem B implies that Tn is hereditarily
minimal.
The next result, in which we study when some semidirect products are hereditarily minimal, should be compared
with Proposition 5.8 and Lemma 5.9.
Lemma 6.6. Let G = (Zp , +) ⋊α T , where T is either finite or (topologically) isomorphic to (Zp , +). Then G is
hereditarily minimal if and only if α is faithful.
Proof. If K = ker α is not trivial, then (Zp , +) × K ≤ G is not hereditarily minimal by Prodanov’s theorem, so G
is also not hereditarily minimal.
Now assume that α is faithful. If T is finite, then T is isomorphic to a subgroup F of Fp , and G ≅ Kp,F by
Lemma 5.9. By Example 6.5, G is hereditarily minimal. In case T is isomorphic to Zp , then by Proposition 5.8
either G ≅ Mp,n or G ≅ Tn for some n ∈ N. We use Example 3.12 and Example 6.5, respectively, to conclude that
G is hereditarily minimal.
6.1. The general case. The next proposition offers a reduction from the general case to a specific situation that
will be repeatedly used in the sequel, very often without explicitly giving/recalling all details.
Proposition 6.7. Let G be an infinite hereditarily minimal locally compact group, which is either compact or
locally solvable.
If G has a non-trivial normal solvable subgroup, then G is metabelian. In particular, it has a normal subgroup
N ≅ Zp , such that N = CG (N ), and there exists a monomorphism
j ∶ G/N ↪ Aut (N ) ≅ Aut (Zp ) ≅ Zp × Fp .
(†)
Proof. Let A be a non-trivial normal solvable subgroup of G, and assume that n > 0 is the solvability class of A.
Then A(n−1) is non-trivial and abelian, it is characteristic in A, so normal in G. Hence, we can assume A to be
abelian.
Now note that G is non-discrete by Lemma 3.5, so A is infinite by Proposition 3.9.
Let N1 = CG (A), and note that also N1 is normal in G. As A is abelian, we have A ≤ N1 , and indeed A ≤ Z(N1 ),
so Z(N1 ) is infinite. Moreover, N1 is closed in G, so it is hereditarily minimal, and locally compact. Since N1 is
also either compact or locally solvable, we have N1 ≅ Zp for some prime p according to Theorem C.
By induction, define Nn+1 = CG (Nn ) for n ≥ 1, and similarly prove that Nn ≅ Zp is normal in G for every
n. Then H = ⋃ Nn is abelian, locally compact, hereditarily minimal, so also H ≅ Zp by Corollary 1.5. Then
n
[H ∶ N1 ] is finite, so the ascending chain of subgroups {Nn }n stabilizes and there exists n0 ∈ N+ such that
H = Nn0 = CG (Nn0 ) ≅ Zp is normal in G.
Let N = Nn0 . Then G acts on N by conjugation, and the kernel of this action is N = CG (N ). Hence
G/CG (N ) = G/N is isomorphic to a subgroup of Aut (N ) via a monomorphism j as in (†). Since N ≅ Zp , we
have Aut (N ) ≅ Aut (Zp ) ≅ Zp × Fp . This implies that G/N is abelian, so G is metabelian.
Let us see that the assumption ‘compact or locally solvable’ cannot be removed in Proposition 6.7. Consider
the countable discrete group G = Z(2) × T from Example 3.10. The hereditarily minimal group G has a normal
abelian subgroup, but it is not compact, nor even locally solvable.
In the rest of this section G is a hereditarily minimal locally compact, solvable group, and N and j are as in
Proposition 6.7.
Proposition 6.8. If x ∈ G∖N , then Hx = ⟨x⟩ trivially meets N . Moreover, if x is non-torsion, then G1 = N ⋅Hx ≅
N ⋊ Hx is isomorphic to either Mp,n or Tn , for some prime p and n ∈ N.
18
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
Proof. Clearly, we can consider only the case when x is non-torsion, as N is torsionfree. Hence, Hx ≅ Zq for some
prime q by Lemma 3.1. If q ≠ p, then Hx ∩ N = {e} and G ≥ N ⋅ Hx ≅ N ⋊ Hx , where the action is the conjugation
in G. Moreover, this semidirect product is also isomorphic to Zp ⋊α Zq for some action α, contradicting Lemma
6.2. So we deduce that Hx ≅ Zp .
As G is hereditarily minimal so is its subgroup C. In particular, C is essential in its closure Hx . Since N is
normal in G, it follows that Hx ∩ N is trivial if and only if C ∩ N is trivial.
Assume by contradiction that C ∩ N is non-trivial. Let ϕx ∶ G → G be the conjugation by x, and note that
ϕx ↾N ≠ IdN , as x ∉ N = CG (N ). Obviously, ϕx ↾C = IdC , so ϕx ↾Hx = IdHx , and in particular ϕx ↾N ∩Hx = IdN ∩Hx .
As N ∩ Hx is a non-trivial closed subgroup of N ≅ Zp , we have N ∩ Hx = pk N for an integer k. Take an arbitrary
element t ∈ N , so pk t ∈ N ∩ Hx and ϕx (pk t) = pk t. On the other hand, ϕx (pk t) = pk ϕx (t), which means ϕx (t) = t,
as N is torsionfree. So ϕx ↾N = IdN , a contradiction.
Thus, G1 = N ⋅ Hx ≅ N ⋊ Hx ≅ Zp ⋊α Zp , for a faithful action α by Lemma 6.6. Finally, apply Proposition
5.8.
In the following result, we consider the canonical projection q ∶ G → G/N , and we study how the torsion
elements of G are related to those of G/N . To this end, recall that G/N is (algebraically isomorphic to) a
subgroup of Aut (N ) ≅ Aut (Zp ) ≅ Zp × Fp , so the torsion subgroup of G/N is isomorphic to a subgroup of Fp .
Proposition 6.9. (a) Let x ∈ G ∖ N . Then C = ⟨x⟩ ≅ ⟨q(x)⟩, so x is torsion if and only if q(x) is torsion. In
the latter case, C is isomorphic to a subgroup of Fp .
(b) G/N is torsionfree if and only if G is torsionfree.
(c) Assume G has torsion, and fix a torsion element x0 of G of maximum order m. Then the subgroup Sp = N ⋅⟨x0 ⟩
contains all torsion elements of G, and indeed t(G) = t(Sp ) = Sp ∖ N . Furthermore, Sp ≅ Kp,F , where F is
the subgroup of Fp isomorphic to ⟨x0 ⟩.
Proof. (a) By Proposition 6.8, the map q restricted to C induces an isomorphism C ≅ ⟨q(x)⟩. For the last part,
use the fact that the torsion subgroup of G/N is isomorphic to a subgroup of Fp .
(b) If G has a non-trivial torsion element g, then g ∉ N ≅ Zp . So, q(g) is a non-trivial torsion element of G/N .
By item (a), if q(g) is a non-trivial torsion element of G/N , then g is a non-trivial torsion element of G.
(c) Using item (a) and the fact that t(G/N ) is cyclic we deduce that t(G/N ) = t(Sp /N ) = ⟨q(x0 )⟩. Moreover,
we have
t(G) = q −1 (t(G/N )) ∖ N = q −1 (t(Sp /N )) ∖ N = Sp ∖ N.
As Sp is hereditarily minimal (being a subgroup of G), Lemma 6.6 implies that Sp ≅ N ⋊α ⟨x0 ⟩, for some faithful
action α. The isomorphism Sp ≅ Kp,F now follows From Lemma 5.9.
By Proposition 6.7 we may identify G/N with a subgroup of Zp × Fp via the monomorphism j. For p > 2 the
following dichotomy holds.
Proposition 6.10. If p > 2, then either j(G/N ) ≤ Zp or j(G/N ) ≤ Fp .
Proof. By contradiction, let (x, t) ∈ j(G/N ) be such that x ≠ 0Zp and t ≠ eFp . Let q ∶ G → G/N be the canonical
map and pick z ∈ G such that q(z) = (x, t). Then C = ⟨z⟩ misses N , as (x, t) is non-torsion. Furthermore,
C1 = C ≅ Zq for some prime q. As N1 = N ∩ C1 misses C and as C is minimal (since G is HM), we deduce that
the closed subgroup N1 of C1 must be trivial by the Minimality Criterion. Hence, the subgroup N ⋊ C1 ≅ Zp ⋊ Zq
is hereditarily minimal. This is possible only if q = p by Lemma 6.2. Thus, C1 ≅ Zp , so also H = q(C1 ) is (even
topologically) isomorphic to Zp . Since Zp is (p−1)-divisible we have H = (p−1)H. Now let π2 ∶ Zp ×Fp → Fp be the
projection on the second coordinate. We have π2 (H) = π2 ((p − 1)H) = {eFp }. On the other hand, eFp ≠ t ∈ π2 (H),
a contradiction.
6.2. Torsionfree case. The next theorem classifies the locally compact solvable HM groups, that are also torsionfree. The non-torsionfree groups are considered in Theorem 6.13.
Theorem 6.11. Let G be an infinite locally compact solvable torsionfree group, then the following conditions are
equivalent:
(1) G is hereditarily minimal;
(2) G is topologically isomorphic to one of the following groups:
(a) Zp for some prime p;
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
19
n
(b) Mp,n = Zp ⋊ Cpp , for some prime p and n ∈ N;
n
(c) Tn = (Z2 , +) ⋊β C22 , for some n ∈ N.
Proof. (2) ⇒ (1) ∶ If G ≅ Zp , then it is hereditarily minimal by Prodanov’s theorem. If G ≅ Mp,n , then it is
hereditarily minimal by Example 3.12(a). If G ≅ Tn for some n, then G is hereditarily minimal by Example
3.12(b).
(1) ⇒ (2): If G is abelian, then G ≅ Zp for some prime p, by Prodanov’s theorem.
Assume from now on that G is non-abelian, so G ≠ N and let x1 ∈ G ∖ N . By Proposition 6.8, N trivially meets
⟨x1 ⟩, and the subgroup G1 = N ⋅ ⟨x1 ⟩ is isomorphic to Zp ⋊α Zp , for some faithful action α. Clearly, G′ ≤ N ≤ G1 ,
so G1 is normal. Moreover, G1 /N is topologically isomorphic to ⟨x1 ⟩, hence to Zp by Lemma 3.1.
Now we claim that there is a monomorphism f ∶ G/N → Zp . Since G is torsionfree, G/N is also torsionfree by
Proposition 6.9. Moreover, according to Proposition 6.7, there exists a group monomorphism j ∶ G/N → Zp × Fp ,
so also W = j(G/N ) is torsionfree. If p > 2, then W ≤ Zp by Proposition 6.10, so we simply take f = j. If p = 2,
the subgroup 2W of W is isomorphic to a subgroup of Z2 × {eF2 }. We define φ ∶ Z2 × F2 → Z2 by φ(x) = 2x.
Then φ ↾W is an isomorphism between W and 2W since W is torsionfree. Hence, f = φ ○ j is a monomorphism
G/N → Z2 .
Equipping the codomain of f with the p-adic topology, we now prove that f is a topological isomorphism onto its
range. As G1 /N is topologically isomorphic to Zp , the restriction f ↾G1 /N ∶ G1 /N → Zp is continuous. In particular,
f (G1 /N ) ≅ G1 /N is a non-trivial compact subgroup of Zp . Hence, f (G1 /N ) is open and G1 /N ≅ f (G1 /N ) = pk Zp
for some k ∈ N. We deduce that f (G/N ) is also open, so f (G/N ) = pt Zp for some t ∈ N. Since N is a common
closed normal subgroup of both G and G1 , we get
G/N /G1 /N ≅ f (G/N )/f (G1 /N ) ≅ pt Zp /pk Zp ≅ Z(pt−k ).
As G1 /N is open in G/N which is compact, we obtain that f ∶ G/N → f (G/N ) is a topological isomorphism. In
particular, G/N is topologically isomorphic to Zp .
For every x ∈ G ∖ N , let Gx be the closed subgroup of G generated by N and x. Consider the families
F1 = {Gx }x∉N and F2 = {q(Gx )}x∉N , where q ∶ G → G/N is the canonical map. As G/N ≅ Zp , the family F2 is
totally ordered by inclusion and every member of F2 has finitely many successors in F2 . Since all the subgroups
Gx contain N , it follows that F1 is also totally ordered and has the same property. In particular, the members
in F1 that contain G1 are finitely many, so there is a maximal member among them and it coincides with G as
⋃ Gx = G. Finally, apply Proposition 6.8.
x∉N
6.3. Non-torsionfree case. Given an action α ∶ (Z2 , +) × (Z2 , +) → (Z2 , +) and (b, c) ∈ Z2 × Z2 we write shortly
b(c) in place of α(b, c). A map ε ∶ (Z2 , +) → (Z2 , +) is called a crossed homomorphism, if
ε(b1 + b2 ) = ε(b1 ) + b1 (ε(b2 )), ∀b1 , b2 ∈ Z2 .
The following technical result is used in the proof of Theorem 6.13, in the special case of p = 2.
Proposition 6.12. Let α ∶ (Z2 , +) × (Z2 , +) → (Z2 , +) be a continuous faithful action by automorphisms and let
′
′
f be an automorphism of M2,α = (Z2 , +) ⋊α (Z2 , +) of order 2. If the automorphism f¯ ∶ M2,α /M2,α
→ M2,α /M2,α
induced by f is identical, then the group G = ((Z2 , +) ⋊α (Z2 , +)) ⋊ F , where F = {f, Id}, is not hereditarily
minimal.
Proof. Since the group F is finite we deduce that the topological semidirect product G = M2,α ⋊ F is well defined.
According to Theorem B, it suffices to show that G does not satisfy (Cf n ). Consider the element g0 ∶= (e, f ) of G,
note that it has order 2 in G, and let P = ⟨g0 ⟩. It is sufficient to find an infinite compact abelian subgroup H of G
containing P . To this end, we shall provide a non-torsion element g ∈ G, commuting with g0 , and let H = ⟨g, g0 ⟩.
In view of the semidirect product structure of G and since M2,α is torsion free, this means to find a fixed point
of f , i.e., an element e ≠ x ∈ M2,α with f (x) = x, since in this case g = (x, f ) ∈ G works, as g commutes with g0 .
First of all, we need to find an explicit form of f . For the sake of brevity put N = Z2 × {0} and T = {0} × Z2 .
Our hypothesis that f¯ is identical allows us to claim that also the automorphism T → T , induced by f¯, of the
quotient T ≅ M2,α /N is identical, so we can write f (0, b) = (ε(b), b) for every b ∈ Z2 , where ε ∶ Z2 → Z2 is a
continuous crossed homomorphism, as f is a continuous homomorphism. Moreover, since f is an automorphism
20
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
′
of M2,α and N ≥ M2,α
it follows that f ↾N is an automorphism of N . In fact, as N ≅ Z2 , f ↾N is the multiplication
∗
by some m ∈ Z2 , and m ∈ F2 = {±1} as o(x0 ) = 2. Hence, we can write f as follows:
(6.2)
f (a, b) = f (a, 0)f (0, b) = (ma, 0)(ε(b), b) = (ma + ε(b), b).
Assume that m = 1, then (6.2) and f 2 = Id give 2ε(b) = 0, hence ε(b) = 0, for every b = 0. Take any b ≠ 0 in Z2
and put x = (0, b). Then obviously f (x) = x and we are done.
In case m = −1. Take any b ≠ 0 in Z2 , so 2b ≠ 0. Recall that Aut (Z2 ) ≅ Z∗2 = Z2 ∖ 2Z2 = 1 + 2Z2 , so the action α
gives mb ∈ Z∗2 = 1 + 2Z2 such that b(ε(b)) = mb ε(b). Hence,
ε(2b) = ε(b) + mb ε(b) = (1 + mb )ε(b) ∈ 2Z2 ,
which means there exists s ∈ Z2 such that ε(2b) = 2s. Consider the element x = (s, 2b) ∈ M2,α , and observe that it is
not torsion as 2b ≠ 0. Finally, x is a fixed point of f . Indeed, f (x) = (−s+ε(2b), 2b) = (−s+2s, 2b) = (s, 2b) = x.
Here follows the counterpart of Theorem 6.11 for non-torsionfree groups, which concludes the classification of
the locally compact solvable HM groups given in Theorem D.
Theorem 6.13. Let G be an infinite locally compact solvable non-torsionfree group. Then the following conditions
are equivalent:
(a) G is hereditarily minimal;
(b) G is topologically isomorphic to Kp,F for some prime p and a non-trivial subgroup F of Fp .
Proof. (b) ⇒ (a): If G ≅ Kp,F , then G is hereditarily minimal by Example 6.5.
(a) ⇒ (b): In the notation of Proposition 6.9, let Sp = N ⋅ ⟨x0 ⟩, where x0 is a torsion element of G of maximal
order. Moreover, t(G) ⊆ Sp and Sp ≅ Kp,F for some non-trivial F ≤ Fp . So, it suffices to show that G = Sp .
By contradiction, pick x ∈ G ∖ Sp and observe that x ∉ N and x is non-torsion. The subgroup T = ⟨x⟩ is
isomorphic to Zq for some prime q by Prodanov’s theorem. Moreover,
(6.3)
T ∩ N = {e},
by Proposition 6.8 If p > 2, then G/N is isomorphic to a subgroup of Fp by Proposition 6.10 and Proposition
6.9(b). Hence [G ∶ N ] < ∞, and this contradicts (6.3), as T is infinite.
Now assume that p = 2. In view of the normality of N ≅ Zp in G and (6.3), we can apply Lemma 6.2 to deduce
that q = 2 as well. So L = N T is isomorphic to N ⋊ T ≅ M2,α , where α is a faithful action, by Lemma 6.6. In the
sequel we identify L = N T ≅ N ⋊ T and M2,α . Since G′ ≤ N ≤ L by Proposition 6.7, it follows that L is normal in
G. As x0 has order two and L is torsionfree, we obtain
G ≥ L ⋊ ⟨x0 ⟩ ≅ (Z2 ⋊α Z2 ) ⋊β F2 ,
where β is a faithful action. Indeed, if ker β is non-trivial, then ker β = F2 and we deduce that
(Z2 ⋊α Z2 ) ⋊β F2 ≅ (Z2 ⋊α Z2 ) × F2 ,
contradicting the fact that G is hereditarily minimal.
The action β coincides with the action φ ∶ L → L by conjugation by x0 , and o(φ) = 2 as o(x0 ) = 2 and
x0 ∉ CG (N ) = N ≤ L. Clearly φ(L′ ) = L′ , as L′ is a normal subgroup of G, so φ induces an automorphism
φ̄ ∶ L/L′ → L/L′. Since φ̄ is still an internal automorphism of L/L′ and this quotient is abelian, we deduce that
φ̄ = IdL/L′ . Now we can apply Proposition 6.12 to f = φ to deduce that (Z2 ⋊α Z2 )⋊F2 is not hereditarily minimal,
contradicting the fact that G is hereditarily minimal.
Corollary 6.14. Let G be an infinite hereditarily minimal locally compact group, which is either compact or
locally solvable.
Then G contains a subgroup K ≅ Zp for some prime p, such that its normalizer NG (K) is isomorphic to one
of the groups in (a), (b) or (c), in Theorem D.
Proof. As G is non-discrete by Lemma 3.5, the existence of K is guaranteed by Proposition 3.9(2). The subgroup
N = NG (K) is closed in G and contains K, so N satisfies all the properties of G listed in the statement. As K
is normal in N , we can apply Proposition 6.7 to conclude that N is metabelian. Finally, Theorem D applies to
N.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
21
6.4. Hereditarily totally minimal topological groups. In this subsection we classify the infinite locally
compact solvable hereditarily totally minimal groups. We also provide an equivalent condition for a non-discrete
locally compact group with non-trivial center to be HTM.
Definition 6.15. Call a topological group G hereditarily totally minimal, if every subgroup of G is totally minimal.
The following concept has a key role in the Total Minimality Criterion (see Fact 6.17).
Definition 6.16. A subgroup H of a topological group G is totally dense if for every closed normal subgroup N
of G the intersection N ∩ H is dense in N .
Fact 6.17. [6, Total Minimality Criterion] Let H be a dense subgroup of a topological group G. Then H is totally
minimal if and only if G is totally minimal and H is totally dense in G.
As hereditarily totally minimal groups are HM, the groups Kp,F , Mp,n and Tn are the only locally compact
solvable groups we need to consider, according to Theorem D. In the next proposition we prove that the groups
Kp,F are hereditarily totally minimal.
Proposition 6.18. Let p be a prime and F ≤ Fp . Then Kp,F is hereditarily totally minimal.
Proof. Let H be an infinite subgroup of Kp,F and we have to prove that H is totally minimal. As Kp,F is compact,
it suffices to show that H is totally dense in H, by Fact 6.17. To this aim, let N be a non-trivial closed normal
subgroup of H. As H is an infinite compact HM group, its subgroup N is infinite by Proposition 3.9, so it intersects
non-trivially the subgroup Zp ⋊ {1}. Moreover, there exists n ∈ N such that N ≥ N ∩ (Zp ⋊ {1}) = pn Zp ⋊ {1}. This
implies that N is open in Kp,F , so in H, hence H ∩ N = N .
The following Lemma will be used in the proof of Proposition 6.20.
Lemma 6.19. If G is a hereditarily totally minimal group, then all quotients of G are hereditarily totally minimal.
Proof. Let N be a closed normal subgroup of G and let q ∶ G → G/N be the quotient map. Take a subgroup D of
G/N and let D1 = q −1 (D). Now we prove that D is totally minimal. Consider the restriction q ′ ∶ D1 → D, clearly,
q ′ is a continuous surjective homomorphism. Since D1 is totally minimal by our hypothesis on G, we obtain that
q ′ is open. Moreover, being a quotient group of D1 , we deduce that D itself is totally minimal.
Now we show that the HM groups Mp,n and Tn are not hereditarily totally minimal.
Proposition 6.20. Let p be a prime and n ∈ N, then:
(1) Mp,n is not hereditarily totally minimal;
(2) Tn is not hereditarily totally minimal.
Proof. (1): Assume for a contradiction that Mp,n is hereditarily totally minimal for some prime p and n ∈ N.
′
Then the quotient group Mp,n /Mp,n
is hereditarily totally minimal by Lemma 6.19.
n
′
If p > 2, then Mp,n /Mp,n ≅ Z(pn+1 ) × Cpp by Lemma 5.2, but this group is not even hereditarily minimal by
n
′
Fact 1.3. In case p = 2, as M2,n /M2,n
≅ Z(2n+2 ) × C22 , we get a similar contradiction.
n
(2): By Lemma 5.5 the quotient group Tn /Tn′ of Tn is isomorphic to Z(2) × C22 , which is not hereditarily
totally minimal. Alternatively, one can note that Tn contains M2,n+1 , which is not hereditarily totally minimal
by (1).
Theorem 6.21. Let G be an infinite locally compact solvable group, then the following conditions are equivalent:
(1) G is hereditarily totally minimal;
(2) G is topologically isomorphic to Kp,F = Zp ⋊ F , where F ≤ Fp for some prime p.
Proof. (1) ⇒ (2): Clearly, a hereditarily totally minimal group is hereditarily minimal. So, by Theorem D, G is
topologically isomorphic to one of the three types of groups: Kp,F , Mp,n or Tn . In addition, neither Mp,n nor
Tn are hereditarily totally minimal by Proposition 6.20. Hence, G is topologically isomorphic to Kp,F for some
prime p and F ≤ Fp .
(2) ⇒ (1): Use Proposition 6.18.
The next fact was originally proved in [12], and we use it in the subsequent Theorem 6.23.
22
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
Fact 6.22. [7, Theorem 7.3.1] If a topological group G contains a compact normal subgroup N such that G/N is
(totally) minimal, then G is (resp., totally) minimal.
The groups Kp,F , where F is a non-trivial subgroup of Fp , are center-free. The next theorem deals with the
case of non-trivial center.
Theorem 6.23. Let G be a locally compact non-discrete group with non-trivial center. The following conditions
are equivalent:
(1) G is hereditarily totally minimal;
(2) every non-trivial closed subgroup of G is open, Z(G) ≅ Zp for some prime p and G/Z(G) is hereditarily
totally minimal.
Proof. By Fact 1.4, the assertion holds true in the abelian case, so we may assume that Z(G) ≠ G.
(1) ⇒ (2): If G is hereditarily totally minimal, then G/Z(G) is hereditarily totally minimal, by Lemma 6.19.
Applying Theorem 4.6 we conclude that every closed non-trivial subgroup of G is open and Z(G) ≅ Zp for some
prime p.
(2) ⇒ (1): Note that G/Z(G) is a discrete hereditarily totally minimal group.
We first prove that if H is a closed subgroup of G, then H is totally minimal. As Z(G) ≅ Zp , the subgroup
H ∩Z(G) is compact and normal in H. We also have H/(H ∩Z(G)) ≅ HZ(G)/Z(G) ≤ G/Z(G), so H/(H ∩Z(G))
is totally minimal since G/Z(G) is hereditarily totally minimal. By Fact 6.22, H is totally minimal.
Now let H be a subgroup of G. By Fact 6.17, it remains to show that H is totally dense in H. To this aim,
take a non-trivial closed normal subgroup N of H. Clearly, N is also closed in G and by our assumption it is
open in G. In particular, N is open in H so H ∩ N = N .
Now we recall of special case of a theorem, known as Countable Layer Theorem.
Theorem 6.24. [16, Theorem 9.91] Any profinite group G has a canonical countable descending sequence
(6.4)
G = Ω0 (G) ⊇ . . . ⊇ Ωn (G) ⊇ . . . . . .
of closed characteristic subgroups of G with the following two properties:
(1) ⋂∞
n=1 Ωn (G) = {e};
(2) for each n ∈ N+ , the quotient Ωn−1 (G)/Ωn (G) is isomorphic to a cartesian product of simple finite groups.
Theorem 6.25. Every locally compact hereditarily totally minimal group is metrizable.
Proof. Let G be a locally compact HTM group. By Proposition 3.9(1), G is totally disconnected. Let K be a
compact open subgroup of G. It is enough to show that K is metrizable. According to the Countable Layer
Theorem, one has a decreasing chain of closed normal subgroups (6.4) with ⋂∞
n=1 Ωn (K) = {e}. By Lemma 6.19,
the quotient group G/Ω1 (K) is HM. Since this is a direct product of finite groups, this is possible only if G/Ω1 (K)
is finite. Similarly, each quotient group Ωn (K)/Ωn+1 (K) is finite. Hence, also each quotient group G/Ωn (K) is
finite. This implies that each subgroup Ωn (K) is open. Now ⋂∞
n=1 Ωn (K) = {e} and the compactness of K imply
that the subgroups {Ωn (K) ∶ n ∈ N} form a local base at e. Therefore, K is metrizable.
7. Open questions and concluding remarks
Recall that a group is hereditarily non-topologizable when it is hereditarily totally minimal in the discrete
topology.
Question 7.1. Are discrete hereditarily minimal groups also hereditarily non-topologizable ?
In case of affirmative answer, one can deduce from the results quoted in §3.1 that the class of countable discrete
hereditarily minimal groups is stable under taking finite products. On the other hand, if G and H are discrete
hereditarily minimal groups such that their product is not hereditarily minimal, then one of these groups is not
hereditarily non-topologizable, so provides a counter-example to Question 7.1.
In view of Theorem 4.6, the following natural question arises:
Question 7.2. Does there exist a non-discrete hereditarily minimal locally compact group G with {e} ≠ Z(G) ≠ G?
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
23
By Theorem 4.6, a positive answer to Question 7.2 provides a group G such that the center is open, and has
infinite index in G, so G is not compact (this follows also from Theorem C). So by Corollary 1.5, Question 7.2
has this equivalent formulation: does there exist a hereditarily minimal locally compact group G which is neither
discrete nor compact, and has non-trivial center? In other words:
Question 7.3.
(1) Does there exist a hereditarily minimal locally compact group which is neither discrete
nor compact?
(2) Can such a group have non-trivial center?
As already noted, a positive answer to Question 7.2 gives a positive answer to both items in Question 7.3. On
the other hand, a group G providing a positive answer to Question 7.3, if not center-free, gives also a positive
answer to Question 7.2.
One can focus his search for an answer to Questions 7.2 and 7.3 using Proposition 7.4 below.
Theorem 4.6 provides some necessary conditions for a non-discrete locally compact group G satisfying {e} ≠
Z(G) ≠ G to be also hereditarily minimal. Among others: being torsionfree, having the center open and isomorphic
to Zp , having the quotient G/Z(G) a (discrete) p-group. This justifies the hypotheses in the following proposition,
which proves a partial converse to Theorem 4.6.
Proposition 7.4. Let p be a prime number, and G be a torsionfree group such that Z(G) ≅ Zp is open in G. If
G/Z(G) is hereditarily minimal, then G is hereditarily minimal.
Proof. Note that G/Z(G) is a discrete hereditarily minimal group by our assumption. By Fact 6.22 and using
similar arguments as in the proof of Theorem 6.23 (see the implication (2) ⇒ (1)), one can prove that every
closed subgroup of G is minimal.
Now let H be a subgroup of G, and we show that H is essential in H. Since G is torsionfree and G/Z(G) is
torsion by Lemma 3.5, every non-trivial subgroup of G meets the center non-trivially. Let N be a closed normal
non-trivial subgroup of H. So, H ∩ Z(G) and N ∩ Z(G) are non-trivial subgroups of Z(G) and N ∩ Z(G) is also
closed. As Z(G) ≅ Zp , Lemma 2.5 implies that N ∩H ∩Z(G) is non-trivial. In particular, N ∩H is non-trivial.
If G is a torsionfree group such that Z(G) ≅ Zp and G/Z(G) is an infinite discrete hereditarily minimal p-group,
then G is hereditarily minimal by Proposition 7.4, and clearly G is neither discrete nor compact. So such a group
provides a positive answer to Question 7.2 (so in particular also to Question 7.3). We conjecture that such a
group exists on the base of a remarkable example of a locally compact group M built in [17, Theorem 4.5], having
the following properties:
(1) Z(M ) is open and Z(M ) ≅ Zp ;
(2) M /Z(M ) is a Tarski monster of exponent p having p conjugacy classes;
(3) every element of M is contained in a subgroup, which contains Z(M ) and is isomorphic to Zp (in
particular, M is torsionfree);
(4) all normal subgroups of M are central;
(5) for every proper closed subgroups H of M , H ≅ Zp and either H ≤ Z(M ) or Z(M ) ≤ H.
In particular, we see that M has all of the properties listed in Theorem 4.6, but it is not clear if such an M
must be HM. This will be ensured by Proposition 7.4 if one can ensure the Tarski monster M /Z(M ) to be HM
(i.e., hereditarily non-topologizable). Tarski monsters T with this property were built in [19], so it remains to
check if for such a T one can build an extension M as above.
By Proposition 6.7, if G is an infinite locally compact HM group, which is either compact or locally solvable,
with a non-trivial normal solvable subgroup, then G is metabelian. So in particular G is solvable, and Theorem
D applies.
Question 7.5. Can Theorem D be extended to locally solvable groups?
Fact 1.4 shows that the abelian HM groups are second countable, while Theorem D shows that the conclusion
remains true if “abelian” is replaced by “solvable and locally compact”. On the other hand, uncountable discrete
HTM groups were built in [19], showing that a locally compact HTM non-solvable group need not be second
countable even in the discrete case. Yet this leaves open the following:
Question 7.6. Are locally compact HM groups metrizable?
24
XI, DIKRANJAN, SHLOSSBERG, AND TOLLER
Theorem 6.25 shows that the answer is affirmative for locally compact HTM groups. On the other hand, Fact
1.6 suggests that the answer can be affirmative even for locally compact HLM groups.
Proposition 6.20 provides examples of locally compact hereditarily minimal groups with trivial center that
are not HTM. We are not aware if a locally compact hereditarily minimal group with non-trivial center can be
non-HTM.
Acknowledgments. The first-named author takes this opportunity to thank Professor Dikranjan for his generous
hospitality and support. The second-named author is partially supported by grant PSD-2015-2017-DIMA-PRID2017-DIKRANJAN PSD-2015-2017-DIMA - progetto PRID TokaDyMA of Udine University. The third and
fourth-named authors are supported by Programma SIR 2014 by MIUR, project GADYGR, number RBSI14V2LI,
cup G22I15000160008 and by INdAM - Istituto Nazionale di Alta Matematica.
References
[1] L. Außenhofer, M.J. Chasco, D. Dikranjan and X. Dominguez, Locally Minimal Topological Groups 2, J.
Math. Anal. Appl. 380 (2011), 552–570.
[2] B. Banaschewski, Minimal topological algebras, Math. Ann. 211 (1974), 107–114.
[3] D. Dikranjan, W. He, Z. Xiao and W. Xi, Locally Compact Groups and Locally Minimal Group Topologies,
Fundam. Math. 2017, to appear.
[4] D. Dikranjan and M. Megrelishvili, Minimality conditions in topological groups, Recent progress in general
topology III, Atlantis Press, Paris, (2014), 229–327.
[5] D. Dikranjan and S. Morris, Subgroups of products of locally compact groups, Topol. Proc. 26 (2001/2002),
533–544.
[6] D. Dikranjan and Iv. Prodanov, Totally minimal groups, Annuaire Univ. Sofia Fat. Math. Méc. 69 (1974/75)
5–11.
[7] D. Dikranjan, Iv. Prodanov and L. Stoyanov, Topological Groups: Character, Dualities and Minimal Group
Topologies, Pure and Applied Mathematics, Vol. 130, Marcel Dekker Inc., New York-Basel, 1989.
[8] D. Dikranjan and L. Stoyanov, Criterion for minimality of all subgroups of a topological abelian group, C. R.
Acad. Bulgare Sci. 34 (1981), 635–638.
[9] D. Dikranjan and V. Uspenskij, c-compact topological groups, J. Pure Appl. Algebra 126 (1998), no. 1-3,
149–168.
[10] M.R. Dixon, Sylow theory, formations and Fitting classes in locally finite groups, World Scientific Publishing,
River Edge, NJ, 1994.
[11] D. Doïtchinov, Produits de groupes topologiques minimaux, Bull. Sci. Math. 97 (2) (1972), 59–64.
[12] V. Eberhardt, S. Dierolf and U. Schwanengel, On products of two (totally) minimal topological groups and
the three-space-problem, Math. Ann. 251 (1980) 123–128.
[13] R. Engelking, General Topology, Sigma Series in Pure Math, Vol. 6, Heldermann, Berlin, 1989.
[14] L. Fuchs, Infinite Abelian groups, Vol. I, Academic Press, New York, 1970.
[15] P. Hall and C.R. Kulatilaka, A property of locally finite groups, J. London Math. Soc. 39 (1964), 235–239.
[16] K. Hofmann and S. Morris, The structure of compact groups. A primer for the student—a handbook for the
expert, Gruyter Studies in Mathematics 25, Walter de Gruyter & Co., Berlin, 1998.
[17] K. Hofmann, S. Morris, S. Oates-Williams and V. Obraztsov, Locally compact groups with closed subgroups
open and p-adic, Mathematical Proceedings of the Cambridge Philosophical Society, 118 (2) (1995), 303–313.
[18] I. Kaplansky, Lie Algebras and Locally Compact Groups, Chicago Lectures in Mathematics, Univ. of Chicago
Press, 1971.
[19] A.A. Klyachko, A.Yu. Ol′ shanskij and D.V. Osin, On topologizable and non-topologizable groups, Topol. Appl.
160 (16) (2013) 2104–2120.
[20] G. Lukács, Hereditarily h-complete groups, Topol. Appl. 145 (2004), no. 1–3, 177–189.
[21] G. Lukács, Hereditarily non-topologizable groups, Topol. Proc. 33 (2009), 269–275.
[22] A.A. Markov, Three papers on topological groups: I. On the existence of periodic connected topological groups;
II. On free topological groups; III. On unconditionally closed sets, Amer. Math. Soc. Transl. (1950), no. 30.
[23] M. Megrelishvili, Group representations and construction of minimal topological groups, Topol. Appl. 62
(1995), 1–19.
HEREDITARILY MINIMAL TOPOLOGICAL GROUPS
25
[24] M. Megrelishvili, G-minimal topological groups, Proc. of Orsatti Conference ‘Abelian groups, module theory,
and topology’ (Padua, 1997), Lecture Notes in Pure and Appl. Math., Dekker, New York, 201 (1998), 289–
299.
[25] S. Morris and V. Obraztsov, Embedding free amalgams of discrete groups in non-discrete topological groups,
Geometric group theory down under (Canberra, 1996), 203–223, de Gruyter, Berlin, 1999.
[26] S. Morris and V. Pestov, On Lie groups in varieties of topological groups, Colloq. Math. 78 (1998), 39–47.
[27] A. Ol′ shanskij, A note on countable non-topologizable groups, Vestnik Mosk. Gos. Univ. Mat. Mekh., 3 (1980),
103.
[28] Iv. Prodanov, Precompact minimal group topologies and p-adic numbers, Annuaire Univ. Sofia Fac. Math.
Méc. 66 (1971/72), 249–266.
[29] D. Remus and L. Stojanov, Complete minimal and totally minimal groups, Topol. Appl. 42 (1991), no. 1,
57–69.
[30] S. Shelah, On a problem of Kurosh, Jónsson groups, and applications, In: S. I. Adian, W. W. Boone and G.
Higman, Eds., Word Problems II, North-Holland, Amsterdam, (1980) 373–394.
[31] R.M. Stephenson, Jr., Minimal topological groups, Math. Ann. 192 (1971), 193–195.
[32] R.M. Stephenson, Jr., Some unsolved problems concerning P-minimal and P-closed spaces, Proc. Charlotte
Topological Conf., Academic Press, New York, (1974), 249–257.
[33] D. van Dantzig, Studien over Topologische Algebra, Dissertation, Amsterdam 1931. Math. Soc. 55 (1), (1997),
143–146.
[34] E. I. Zelmanov, On periodic compact groups, Israel J. Math. 77 (1992), 83–95.
(W. Xi)
School of Mathematical Sciences
Nanjing Normal University
Wenyuan Road No. 1, 210046 Nanjing
China
E-mail address: [email protected]
(D. Dikranjan)
Dipartimento di Matematica e Informatica
Università di Udine
Via delle Scienze 206, 33100 Udine
Italy
E-mail address: [email protected]
(M. Shlossberg)
Dipartimento di Matematica e Informatica
Università di Udine
Via delle Scienze 206, 33100 Udine
Italy
E-mail address: [email protected]
(D. Toller)
Dipartimento di Matematica e Informatica
Università di Udine
Via delle Scienze 206, 33100 Udine
Italy
E-mail address: [email protected]
| 4math.GR
|
Learning Dexterous Manipulation
Policies from Experience and Imitation
Journal Title
XX(X):1–17
c The Author(s) 2016
Reprints and permission:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/ToBeAssigned
www.sagepub.com/
arXiv:1611.05095v1 [cs.LG] 15 Nov 2016
Vikash Kumar1 , Abhishek Gupta2 , Emanuel Todorov1 and Sergey Levine2
Abstract
We explore learning-based approaches for feedback control of a dexterous five-finger hand performing non-prehensile
manipulation. First, we learn local controllers that are able to perform the task starting at a predefined initial state.
These controllers are constructed using trajectory optimization with respect to locally-linear time-varying models
learned directly from sensor data. In some cases, we initialize the optimizer with human demonstrations collected
via teleoperation in a virtual environment. We demonstrate that such controllers can perform the task robustly, both
in simulation and on the physical platform, for a limited range of initial conditions around the trained starting state.
We then consider two interpolation methods for generalizing to a wider range of initial conditions: deep learning, and
nearest neighbors. We find that nearest neighbors achieve higher performance. Nevertheless, the neural network has its
advantages: it uses only tactile and proprioceptive feedback but no visual feedback about the object (i.e. it performs the
task blind) and learns a time-invariant policy. In contrast, the nearest neighbors method switches between time-varying
local controllers based on the proximity of initial object states sensed via motion capture. While both generalization
methods leave room for improvement, our work shows that (i) local trajectory-based controllers for complex nonprehensile manipulation tasks can be constructed from surprisingly small amounts of training data, and (ii) collections
of such controllers can be interpolated to form more global controllers. Results are summarized in the supplementary
video: https://youtu.be/E0wmO6deqjo
Keywords
Reinforcement Learning, Dexterous Manipulation, Trajectory Optimization
1 Introduction
Dexterous manipulation is among the most challenging
control problems in robotics, and remains largely unsolved.
This is due to a combination of factors including high
dimensionality, intermittent contact dynamics, and underactuation in the case of dynamic object manipulation.
Here we describe our efforts to tackle this problem in
a principled way. We do not rely on manually designed
controllers. Instead we synthesize controllers automatically,
by optimizing high-level cost functions, as well as by
building off of human-provided expert demonstrations.
The resulting controllers are able to manipulate freelymoving objects, as shown in Figure 1. Such non-prehensile
manipulation is challenging, since the system must reason
about both the kinematics and the dynamics of the
interaction Lynch and Mason (1999). We present results
for learning both local models and control policies that
can succeed from a single initial state, as well as more
generalizable global policies that can use limited onboard
sensing to perform a complex grasping behavior. The small
amount of data needed for learning each controller (around
60 trials on the physical hardware) indicate that the approach
can practically be used to learn large repertoires of dexterous
manipulation skills.
We use our ADROIT platform Kumar, Xu and Todorov
(2013), which is a ShadowHand skeleton augmented with
high-performance pneumatic actuators. This system has
a 100-dimensional continuous state space, including the
Prepared using sagej.cls [Version: 2015/06/09 v1.01]
positions and velocities of 24 joints, the pressures in 40
pneumatic cylinders, and the position and velocity of the
object being manipulated.
Pneumatics have non-negligible time constants (around
20 ms in our system), which is why the cylinder pressures
represent additional state variables, making it difficult to
apply torque-control techniques. The system also has a
40-dimensional continuous control space – namely the
commands to the proportional valves regulating the flow
of compressed air to the cylinders. The cylinders act on
the joints through tendons. The tendons do not introduce
additional state variables (since we avoid slack via pretensioning) but nevertheless complicate the dynamics.
Overall this is a daunting system to model, let alone control.
Depending on one’s preference of terminology, our
method can be classified as model-based Reinforcement
Learning (RL), or as adaptive optimal control (Bellman
and Kalaba 1959). While RL aims to solve the same
general problem as optimal control, its uniqueness comes
from the emphasis on model-free learning in stochastic
1 Department of Computer Science & Engineering, University of
Washington, WA, USA
{vikash, todorov}@cs.washington.edu
2 Department
of Electrical Engineering & Computer Sciences, University
of California at Berkeley, CA, USA
{abhigupta, svlevine}@berkeley.edu
Corresponding author:
Vikash Kumar, add postal address
2
Journal Title XX(X)
Figure 1. Learned hand manipulation behavior involving clockwise rotation of the object
domains (Sutton and Barto 1998). The idea of learning
policies without having models still dominates RL, and
forms the basis of the most remarkable success stories, both
old (Tesauro 1994) and new (Mnih, Kavukcuoglu, Silver,
Rusu, Veness, Bellemare, Graves, Riedmiller, Fidjeland and
Ostrovski 2015). However RL with learned models has
also been considered. Adaptive control on the other hand
mostly focuses on learning the parameters of a model
with predefined structure, essentially interleaving system
identification with control (Åström and Wittenmark 2013).
Our approach here lies somewhere in between (to fix
terminology, we call it RL in subsequent sections). We
rely on a model, but that model does not have any
informative predefined structure. Instead, it is a timevarying linear model learned from data, using a generic
prior for regularization. Related ideas have been pursued
previously (Mitrovic, Klanke and Vijayakumar 2010; Levine
and Abbeel 2014; Levine, Wagener and Abbeel 2015b).
Nevertheless, as with most approaches to automatic control
and computational intelligence in general, the challenge is
not only in formulating ideas but also in getting them to
scale to hard problems – which is our main contribution here.
In particular, we demonstrate scaling from a 14-dimensional
state space in (Levine, Wagener and Abbeel 2015b) to
a 100-dimensional state space here. This is important in
light of the curse of dimensionality. Indeed RL has been
successfully applied to a range of robotic tasks (Tedrake,
Zhang and Seung 2004; Kober, Oztop and Peters 2010;
Pastor, Hoffmann, Asfour and Schaal 2009; Deisenroth,
Rasmussen and Fox 2011), however dimensionality and
sample complexity have presented major challenges (Kober,
Bagnell and Peters 2013; Deisenroth, Neumann and Peters
2013).
The manipulation skills we learn are initially represented as time-varying linear-Gaussian controllers. These
controllers are fundamentally trajectory-centric, but otherwise are extremely flexible, since they can represent any
trajectory with any linear stabilization strategy. Since the
controllers are time-varying, the overall learned control
law is nonlinear, but is locally linear at each time step.
These types of controllers have been employed previously
for controlling lower-dimensional robotic arms (Lioutikov,
Paraschos, Neumann and Peters 2014; Levine, Wagener and
Abbeel 2015b).
For learning more complex manipulation skills, we also
explore the use of human demonstrations to initialize the
controllers. Complex tasks with delayed rewards, such as
grasping and in-hand repositioning of a heavy object, are
difficult to learn from scratch. We show that a teleoperation
system can be used to provide example demonstrations from
Prepared using sagej.cls
a human operator using a glove-based interface, and that
these demonstrations can be used to initialize learning for
complex skills.
Finally, to move beyond local policies that can succeed
from only a narrow range of initial states, we explore
generalization through two distinct approaches. In the
approach method, we train a collection of local policies, each
initialized with a different initial demonstration, and then
use a nearest neighbor query to select the local policy for
which the initial conditions best match the current pose of
the manipulated object. In the second method, we use a deep
neural network to learn to mimic all of the local policies,
thus removing the need for nearest neighbor queries. Our
experimental results show that the nearest neighbor approach
achieves the best success rate, but at the cost of requiring
the variables for the nearest neighbor queries (the pose of
the object) to be provided by hand. We also show that
the deep neural network can learn a time-invariant policy
for performing the task without requiring knowledge of the
object pose at all, using only onboard sensing on the fivefinger hand to perform the task.
The work on local trajectory-based control (Sections 5
and 6) was previously described in conference proceedings
(Kumar, Todorov and Levine 2016) while the work on
generalization (Sections 7 and 8) is new and is described here
for the first time.
2
Related Work
Although robotic reinforcement learning has experienced
considerable progress in recent years, with successful results
in domains ranging from flight (Abbeel, Coates, Quigley
and Ng 2006) to locomotion (Tedrake, Zhang and Seung
2004) to manipulation (Peters, Mülling and Altun 2010a;
Theodorou, Buchli and Schaal 2010; Peters and Schaal
2008), comparatively few methods have been applied to
control dexterous hands. (van Hoof, Hermans, Neumann and
Peters 2015) report results for simple in-hand manipulation
with a 3-finger hand, and our work reports learning of simple
in-hand manipulation skills, such as rotating a cylinder, using
time-varying linear-Gaussian controllers (Kumar, Todorov
and Levine 2016). However, neither of these prior methods
demonstrate generalization to conditions not seen during
training. Our experiments demonstrate that our approach can
learn policies for a complex precision grasping task, and can
generalize to variation in the initial position of the target
object. In contrast to in-hand manipulation, this task exhibits
complex discontinuities at the point of contact. We overcome
this challenge by combining learning from experience with
Kumar, Gupta, Todorov and Levine
imitation learning from human demonstrations, provided
through a data glove teleoperation interface.
Initialization of controllers from demonstration is a widely
employed technique in robotic reinforcement learning
(Peters, Mülling and Altun 2010a; Theodorou, Buchli and
Schaal 2010). However, most prior robotic reinforcement
learning methods still use a hand-specified reward or
cost function to provide the goal of the task during
learning. Specifying suitable cost functions for complex
dexterous manipulation can be exceedingly challenging,
since simple costs can lead to poor local optima, while
complex shaped costs require extensive intuition about the
task. In our work, we define the cost in terms of the
example demonstrations. This approach resembles the work
of (Gupta, Eppner, Levine and Abbeel 2016), which used
an EM-style algorithm to associate demonstrations with
initial states in a reinforcement learning scenario. However,
this prior work showed results on a simple inflatable hand
with limited actuation, and did not demonstrate dexterous
manipulation for complex tasks.
3 Overview
The ADROIT platform, which serves as the experimental
platform for all of our dexterous manipulation experiments,
is described in detail in Section 4. This system is used in three
sets of experiments: the first set of experiments examines
learning dexterous manipulation skills from scratch using
trajectory-centric reinforcement learning, the second set of
experiments is focused on learning more complex skills with
a combination of trajectory-centric reinforcement learning
and learning from demonstration, and the third set of
experiments examines how various approximation methods,
including nearest neighbor and deep neural networks, can
be used to learn from multiple learned behaviors to acquire
a single generalizable skill that succeeds under various
circumstances.
The trajectory-centric reinforcement learning algorithm
that we use combines the linear-quadratic regulator (LQR)
algorithm with learned time-varying local linear models.
This algorithm, which follows previous work (Levine and
Abbeel 2014), is described in Section 5. We then present
results on both a real-world and simulated version of the
ADROIT platform using the algorithm, in Section 6. This
first set of experiments focuses primarily on the capability
of the trajectory-centric reinforcement learning method to
efficiently learn viable and robust manipulation skills.
The second set of experiments, presented in Section 7,
is aimed at evaluating how human demonstrations can
be used to aid learning for more complex skills. In this
section, we examine a complex grasping scenario, where
trajectory-centric reinforcement learning on its own does
not produce sufficiently successful behaviors, while human
demonstrations alone also do not achieve a sufficient
degree of robustness in the face of perturbations. We
demonstrate that combining demonstrations with trajectorycentric reinforcement learning produces effective skills with
a high degree of robustness to variation in the initial
placement of objects in the world.
Our final set of experiments, presented in Section 8,
address the question of generalization: can we use multiple
Prepared using sagej.cls
3
skills, learned with a combination of imitation and trajectorycentric reinforcement learning, to acquire a single robust
and generalizable dexterous manipulation policy? To that
end, we explore the use of deep neural networks to achieve
generalization by combining the behaviors of multiple
skills from the previous section. We also compare to
nearest neighbor as a baseline method. We demonstrate that
deep neural networks can learn time-invariant manipulation
policies that acquire the strategies represented by the
time-varying controllers learned with trajectory-centric
reinforcement learning, and furthermore can perform those
skills using onboard sensing in a simulated experiment,
without knowledge of the true position of the manipulated
object.
4
System
Modularity and the ease of switching robotic platforms
formed the overarching philosophy of our system design.
The learning algorithm (algorithm 1) has no dependency on
the selected robotic platform except for step 3, where the
policies are shipped to the robotic platform for evaluation
and the resulting execution trajectories are collected. This
allows the training to happen either locally (on the machine
controlling the robot) or remotely (if more computational
power is needed).
Manipulation strategies were studied for two different
platforms detailed below.
4.1
Hardware Platform
The ADROIT platform is described in detail in (Kumar, Xu
and Todorov 2013). Here we summarize the features relevant
to the present context. ADROIT manipulation platform is an
anthropomorphic arm-hand system actuated using a custom
build high-performance pneumatic actuation. It consists of
a 24 dof hand and a 4 dof arm. As our motivation here is
to understand in-hard dexterous manipulation, we mounted
the 24 dof hand on a fixed base to promote finger centric
behaviors. Fixed base severely limits the workspace of the
overall system but renders the system amenable to only
finger-based and wrist-based manipulation strategies. For
this work, we use the term ‘ADROIT’ to refer to the fixed
base 24 dof hand setup. 20 out of the 24 hands joints are
independently actuated using 40 antagonistic tendons. The
DIP joints are coupled with the respective PIP joint for
the 4 fingers. Finger-tendons can exert up to 42 Newton,
while the wrist-tendon can exert up to 120 Newton of
force. Each cylinder is supplied with compressed air via a
high-performance Festo valve. The cylinders are fitted with
solid-state pressure sensors. The pressures together with the
joint positions and velocities (sensed by potentiometers in
each joint) are provided as state variables to our controller.
ADROIT’s low-level driver runs on a 12 core 3.47GHz
Intel(R) Xeon(R) processor with 12GB memory running
Windows x64.
The manipulation task also involves an object – which is a
long tube filled with coffee beans, inspired by earlier work on
grasping (Amend Jr, Brown, Rodenberg, Jaeger and Lipson
2012). The object is fitted with PhaseSpace active infrared
markers on each end. The markers are used to estimate the
object position and velocity (both linear and angular) which
4
Journal Title XX(X)
Pressure
Tendon
Joint
Valve
(NI) (35 KHz)
(NI) (9 KHz)
µU (500 Hz)
(NI)(200 Hz)
Filte r
Driver
Hardware
are also provided as state variables. Since all our sensors have
relatively low noise, we apply a minimal amount of filtering
before sending the sensor data to the controller.
and
subsample
ADROIT Control
ADROIT State
Controller
Socket (200hz)
State =
𝑗𝑜𝑖𝑛𝑡
𝑗𝑜𝑖𝑛𝑡 𝑣𝑒𝑙
𝑝𝑟𝑒𝑠𝑠𝑢𝑟𝑒
Socket (200hz)
(a) End pose, learned
(b) End pose, human
(c) End pose, learned
(d) End pose, human
Controls =
Controller
(100,1)
𝑉𝑎𝑙𝑣𝑒 𝑣𝑜𝑙𝑡𝑎𝑔𝑒
(40,1)
Figure 2. System overview
4.2
Simulation Platform
We model the ADROIT hand, including the antagonistic
tendon transmission, joint coupling and pressure dynamics,
using the MuJoCo simulator we have developed (Todorov,
Erez and Tassa 2012). Pressure dynamics are implemented
by extending the default actuation model with user callbacks.
Simulating a 5 s trajectory at 2 ms timestep takes around 0.7
s of CPU time or around 0.3 ms of CPU time per simulation
step. This includes evaluating the feedback control law
(which needs to be interpolated because the trajectory
optimizer uses 50 ms time steps) and advancing the physics
simulation.
Having a fast simulator enables us to prototype and
quickly evaluate candidate learning algorithms and cost
function designs, before testing them on the hardware.
Apart from being able to run much faster than real-time,
the simulator can automatically reset itself to a specified
initial state (which needs to be done manually on the
hardware platform). Note that the actual computation time
(involving GMM fitting, policy update, and network training)
is practically the same for both system as the system is
oblivious to the source of the observations (i.e if they were
generated by the hardware or the simulation platform).
Ideally, the results on the simulation platform should be
leveraged to either transfer behaviors or seed the learning
on the hardware platform. This, however, is hard in practice
(and still an active area of research in the field) due to (a)
the difficulty in aligning the high dimensional state space of
the two platforms, (b) the non-deterministic nature of the
real world. State space alignment requires precise system
identification and sensor calibration which otherwise are not
necessary, as our algorithm can learn the local state space
information directly from the raw sensor values.
5
Reinforcement Learning with Local
Linear Models
In this section, we describe the reinforcement learning
algorithm (summarised in algorithm 1) that we use to
control our pneumatically-driven five finger hand. The
derivation in this section follows previous work (Levine
and Abbeel 2014), but we describe the algorithm in this
Prepared using sagej.cls
Figure 3. Object rotation task: end poses for the two rotation
directions (a-b: clockwise, c-d: anticlockwise), comparing the
learned controller to the movement of a human who has not
seen the robot perform the task.
Algorithm 1 RL with linear-Gaussian controllers
initialize p(ut |xt )
for iteration k = 1 to K do
run p(ut |xt ) to collect trajectory samples {τi }
fit dynamics p(xt+1 |xt , ut ) to {τj } using linear
regression with GMM prior
5:
fit p = arg minp Ep(τ ) [`(τ )] s.t. DKL (p(τ )kp̂(τ )) ≤
6: end for
1:
2:
3:
4:
section for completeness. The aim of the method is to
learn a time-varying linear-Gaussian controller of the form
p(ut |xt ) = N (Kt xt + kt , Ct ), where xt and ut are the
state and action at time step t. The actions in our system correspond to the pneumatic valve’s input voltage,
while the state space is described in the preceding section. The aim of the algorithm is to minimize the expectation Ep(τ ) [`(τ )] over trajectories τ = {x1 , u1 , . . . , xT , uT },
PT
where `(τ ) = t=1 `(xt , ut ) is the total cost, and the expecQT
tation is under p(τ ) = p(x1 ) t=1 p(xt+1 |xt , ut )p(ut |xt ),
where p(xt+1 |xt , ut ) is the dynamics distribution.
5.1
Optimizing Linear-Gaussian Controllers
The simple structure of time-varying linear-Gaussian
controllers admits a very efficient optimization procedure
that works well even under unknown dynamics. The method
is summarized in Algorithm 1. At each iteration, we run the
current controller p(ut |xt ) on the robot to gather N samples
(N = 5 in all of our experiments), then use these samples
to fit time-varying linear-Gaussian dynamics of the form
p(xt+1 |xt , ut ) = N (fxt xt + fut ut + fct , Ft ). This is done
by using linear regression with a Gaussian mixture model
prior, which makes it feasible to fit the dynamics even when
Kumar, Gupta, Todorov and Levine
5
at each iteration. The standard iterative LQR method can
change the controller drastically at each iteration, which can
cause it to visit parts of the state space where the fitted
dynamics are arbitrarily incorrect, leading to divergence.
Furthermore, due to the non-deterministic nature of the real
world domains, line search based methods can get misguided
1
`(xt , ut ) ≈ [xt ; ut ]T `xu,xut [xt ; ut ] + [xt ; ut ]T `xut + const. leading to unreliable progress.
2
To address these issues, we solve the following
where subscripts denote derivatives, e.g. `xut is the gradient optimization problem at each iteration:
of ` with respect to [xt ; ut ], while `xu,xut is the Hessian.
min Ep(τ ) [`(τ )] s.t. DKL (p(τ )kp̂(τ )) ≤ ,
The particular cost functions used in our experiments are
p(ut |xt )
described in the next section. When the cost function
is quadratic and the dynamics are linear-Gaussian, the where p̂(τ ) is the trajectory distribution induced by the
optimal time-varying linear-Gaussian controller of the form previous controller. Using KL-divergence constraints for
p(ut |xt ) = N (Kt xt + kt , Ct ) can be obtained by using the controller optimization has been proposed in a number of
LQR method. This type of iterative approach can be thought prior works (Bagnell and Schneider 2003; Peters and Schaal
of as a variant of iterative LQR (Li and Todorov 2004), 2008; Peters, Mülling and Altün 2010b). In the case of
where the dynamics are fitted to data. Under this model linear-Gaussian controllers, a simple modification to the
of the dynamics and cost function, the optimal policy can LQR algorithm described above can be used to solve this
be computed by recursively computing the quadratic Q- constrained problem. Recall that the trajectory distributions
function and value function, starting with the last time step. are given by p(τ ) = p(x ) QT p(x |x , u )p(u |x ).
1
t+1 t
t
t t
t=1
These functions are given by
Since the dynamics of the new and old trajectory
the number of samples is much lower than the dimensionality
of the system (Levine and Abbeel 2014). We also compute a
second order expansion of the cost function around each of
the samples, and average the expansions together to obtain a
local approximate cost function of the form
distributions are assumed to be the same, the KL-divergence
1 T
xt Vx,xt xt + xT
t Vxt + const
is given by
2
1
T
Q(xt , ut ) = [xt ;ut ]T Qxu,xut [xt ;ut ]+[xt ;ut ]T Qxut +const
X
2
Ep(xt ,ut ) [log p̂(ut |xt )] − H(p),
DKL (p(τ )kp̂(τ )) =
We can express them with the following recurrence:
t=1
V (xt ) =
T
Qxu,xut = `xu,xut + fxut
Vx,xt+1 fxut
Qxut =
Vx,xt =
Vxt =
T
`xut + fxut
Vxt+1
T
Qx,xt − Qu,xt Q−1
u,ut Qu,xt
−1
T
Qxt − Qu,xt Qu,ut Qut ,
which allows us to compute the optimal control
law
as
g(xt ) = ût + kt + Kt (xt − x̂t ),
where
−1
Kt = −Q−1
Q
and
k
=
−Q
Q
.
If
we
consider
t
u,ut u,xt
u,ut ut
p(τ ) to be the trajectory distribution formed by the
deterministic control law g(xt ) and the stochastic dynamics
p(xt+1 |xt , ut ), LQR can be shown to optimize the standard
objective
T
X
min
Ep(xt ,ut ) [`(xt , ut )].
(1)
g(xt )
t=1
However, we can also form a time-varying linear-Gaussian
controller p(ut |xt ), and optimize the following objective:
min
p(ut |xt )
T
X
t=1
Ep(xt ,ut ) [`(xt , ut )] − H(p(ut |xt )).
As shown in previous work (Levine and Koltun 2013),
this objective is in fact optimized by setting p(ut |xt ) =
N (Kt xt + kt , Ct ), where Ct = Q−1
u,ut . While we ultimately aim to minimize the standard controller objective in
Equation (1), this maximum entropy formulation will be a
useful intermediate step for a practical learning algorithm
trained with fitted time-varying linear dynamics.
5.2 KL-Constrained Optimization
In order for this learning method to produce good results, it
is important to bound the change in the controller p(ut |xt )
Prepared using sagej.cls
and the Lagrangian of the constrained optimization problem
is given by
Ltraj (p, η) = Ep [`(τ )] + η[DKL (p(τ )kp̂(τ )) − ] =
"
#
X
Ep(xt ,ut ) [`(xt , ut )−η log p̂(ut |xt )] −ηH(p(τ ))−η.
t
The constrained optimization can be solved with dual
gradient descent (Boyd and Vandenberghe 2004), where we
alternate between minimizing the Lagrangian with respect
to the primal variables, which are the parameters of p, and
taking a subgradient step on the Lagrange multiplier η. The
optimization with respect to p can be performed efficiently
using the LQG algorithm, by observing that the Lagrangian
is simply the expectation of a quantity that does not depend
on p and an entropy term. As described above, LQR can be
used to solve maximum entropy control problems where the
objective consists of a term that does not depend on p, and
another term that encourages high entropy. We can convert
the Lagrangian primal minimization into a problem of the
form
min
p(ut |xt )
T
X
t=1
˜ t , ut )] − H(p(ut |xt ))
Ep(xt ,ut ) [`(x
˜ t , ut ) = 1 `(xt , ut ) − log p̂(ut |xt ).
by using the cost `(x
η
This objective is simply obtained by dividing the Lagrangian
by η. Since there is only one dual variable, dual gradient
descent typically converges very quickly, usually in under
10 iterations, and because LQR is a very efficient
trajectory optimization method, the entire procedure can be
implemented to run very quickly.
6
Journal Title XX(X)
Table 1. Different hand positioning task variations learned
Platform
Hardware
Simulated
# different task variations learned
5 (move assisted by gravity)
2 (move against gravity)
5 (move assisted by gravity)
3 (move against gravity)
We initialize p(ut |xt ) with a fixed covariance Ct and zero
mean, to produce random actuation on the first iteration. The
Gaussian noise used to sample from p(ut |xt ) is generated
in advance and smoothed with a Gaussian kernel with a
standard deviation of two time steps, in order to produce
more temporally coherent noise.
(a) Start pose, movement
against gravity
(b) End pose, movement
against gravity
Table 2. Different object manipulation task variations learned
Platform
Hardware +
an elongated
object (Fig:3)
Simulated +
4 object
variations
# different task variations learned
2 ({clockwise & anti-clockwise}
object rotations along vertical)
13 ({clockwise, anti-clockwise,
clockwise then anti-clockwise}
object rotation along vertical
8 ({clockwise, anti-clockwise}
object rotation along horizontal)
q denotes the vector of hand joint angles, q̇ is the vector of
joint angular velocities, a the vector of cylinder pressures,
and the actions ut correspond to the valve command signals,
which are real-valued and correspond to the degree to which
each valve is opened at each time step. The tasks in this
section require moving the hand to a specified pose from a
given initial pose. We arranged the pair of poses such that in
one task-set the finger motions were helped by gravity, and
in another task-set they had to overcome gravity, as shown
in Figure 4. Note that for a system of this complexity, even
achieving a desired pose can be challenging, especially since
the tendon actuators are in agonist-antagonist pairs and the
forces have to balance to maintain posture. The cost function
at each time step is given by
`(xt , ut ) = ||qt − q ∗ ||2 + 0.001||ut ||2 ,
and the cost at the final time step T emphasizes the target
pose to ensure that it is reached successfully:
`(xT , uT ) = 10||qt − q ∗ ||2 .
6.2
(c) Start pose, movement
assisted by gravity
(d) End pose, movement
assisted by gravity
Figure 4. Positioning task
6
Learning Policies from Experience
In this section, we will describe our first set of experiments,
which uses the trajectory-centric reinforcement learning
algorithm in the previous section to learn dexterous
manipulation skills from scratch on both the physical and
simulated ADROIT platform. The experiments in this section
are aimed to ascertain whether we can learn complex
manipulation behaviors entirely from scratch, using only
high-level task definitions provided in terms of a cost
function, with the controller learned at the level of valve
opening and closing commands. The particular tasks are
detailed in Table 1 and 2 and shown in the accompanying
video, and include both hand posing behaviors and object
manipulation skills.
6.1
Hand Behaviors
In the first set of tasks, we examine how well trajectorycentric reinforcement learning can control the hand to reach
target poses. The state space is given by x = (q, q̇, a). Here,
Prepared using sagej.cls
Object Manipulation Behaviors
The manipulation tasks we focused on require in-hand
rotation of elongated objects. We chose this task because
it involves intermittent contacts with multiple fingers and
is quite dynamic, while at the same time having a certain
amount of intrinsic stability. We studied different variations
(table 2) of this task with different objects: rotation
clockwise (figure 1), rotation counter-clockwise, rotation
clockwise followed by rotation counter-clockwise, and
rotation clockwise without using the wrist joint (to encourage
finger oriented maneuvers) – which was physically locked in
that condition. Figure 3 illustrates the start and end poses
and object configurations in the task learned on the ADROIT
hardware platform. The running cost was
`(xt , ut ) =0.01||qt − q ∗ ||2 + 0.001||ut ||2 +
||qtpos − q pos∗ ||2 + 10||qtrot − q rot∗x ||2
where x = (q, q pos , q rot , q̇, q̇ pos , q̇ rot , a). Here q denotes the
vector of hand joint angles, q pos the object positions, q rot the
object rotations, a the vector of cylinder pressures, and ut the
vector of valve command signals. At the final time we used
`(xt , ut )t=T =2[0.01||qt − q ∗ ||2 + ||qtpos − q pos∗ ||2
+ 10||qtrot − q rot∗x ||2 ].
Here, the cost function included an extra term for desired
object position and orientation. The final cost was scaled by
a factor of 2 relative to the running cost.
Kumar, Gupta, Todorov and Levine
7
Hand posing progress
11000
Effects of injected noise
6800
curling with gravity(hardware)
curling with gravity(sim)
curling against gravity(hardware)
curling against gravity(sim)
10000
9000
Sigma=1
Sigma=2
Sigma=3
6600
6400
8000
6200
6000
6000
Cost
Cost
7000
5000
4000
5600
3000
5400
2000
5200
1000
1
2
3
4
5
6
7
8
9
5000
10
# Iteration
Object twirling progress
7000
4800
0
5500
5000
4500
4000
3500
3000
2
4
6
8
10
12
14
16
18
# Iteration
Figure 5. Learning curves for the positioning (top) and
manipulation (bottom) tasks. ∗
Results
Besides designing the cost functions, minimal parameter
tuning was required to learn each skill. Training consisted
of around 15 iterations. In each iteration we performed 5
trials with different instantiations of the exploration noise
in the controls. The progress of training as well as the final
performance is illustrated in the video accompanying the
submission, and in the figure at the beginning of the paper.
Here we quantify the performance and the robustness to
noise. Figure 5 shows how the total cost for the movement
(as measured by the cost functions defined above) decreased
over iterations of the algorithm. The solid curves are data
from the physical system. Note that in all tasks and task
variations we observe very rapid convergence. Surprisingly,
the manipulation task which is much harder from a control
viewpoint takes about the same number of iterations to learn.
In the positioning task we also performed a systematic
comparison between learning in the physical system and
learning in simulation. Performance early in training was
comparable, but eventually the algorithm was able to find
better policies in simulation. Although it is not shown in
the figure, training on simulation platform happens a lot
faster, because the robot can only run in real-time while the
simulated platform runs faster than real-time, and because
resetting between repetitions needs to be done manually on
the robot.
Prepared using sagej.cls
6
8
10
12
14
We further investigated the effects of exploration noise
magnitude injected during training. Figure 6 shows that for
a relatively small amount of noise performance decreases
monotonically. As we increase the noise magnitude,
sometimes we see faster improvement early on but the
behavior of the algorithm is no longer monotonic. These are
data on the ADROIT hardware platform.
6.4
6.3
4
Figure 6. Effect of noise smoothing on learning. Sigma=1
(width of the Gaussian kernel used for smoothing noise), takes
a slow start but maintains a constant progress. Higher sigma
favors steep decent but it fails to maintain the progress as it is
unable to successfully maintain the stability of the object being
manipulated and ends up dropping it. The algorithm incurs a
huge cost penalty and restarts its decent from there. ∗
6000
0
2
# Iteration
Rotate clockwise
Rotate anticlockwise
Rotate clockwise(WR)
6500
Cost
5800
Delayed Robustification
Finally, we used the simulation platform to investigate
robustness to perturbations more quantitatively, in the
manipulation task. We wanted to quantify how robust our
controllers are to changes in initial state (recall that the
controllers are local). Furthermore, we wanted to see if
training with noisy initial states, in addition to exploration
noise injected in the controls, will result in more robust
controllers. Naı̈vely adding initial state noise at each iteration
of the algorithm (Algorithm 1) severely hindered the overall
progress. However, adding initial state noise after the policy
was partially learned (iteration ≥ 10 in our case) resulted
in much more robust controllers. We term this strategy as
delated robustification.
The results of these simulations are shown in Figure 7.
We plot the orientation of the object around the vertical axis
as a function of time. The black curve is the unperturbed
trajectory. As expected, noise injected in the initial state
makes the movements more variable, especially for the
controller that was trained without such noise. Adding initial
state noise during training substantially improved the ability
of the controller to suppress perturbations in initial state.
Overall, we were surprised at how much noise we could add
(up to 20 % of the range of each state variable) without the
∗
At each iteration, the current controller p(ut |xt ) is deployed on the robot
to gather N samples (N = 5 in all of our experiments).
8
Journal Title XX(X)
5% noise
10% noise
20% noise
1.8
1.6
1.4
radian
1.2
1
0.8
0.6
0.4
0.2
0
-0.2
0
5
0
sec
5% noise
5
0
sec
10% noise
5
sec
20% noise
2
radian
1.5
1
7.1
0.5
0
0
5
sec
0
5
sec
0
5
sec
Figure 7. Robustness to noise in initial state. Each column
corresponds to a different noise level: 5, 10, 20 % of the range
of each state variable. The top row is a controller trained with
noise in the initial state. The bottom row is a controller trained
with the same initial state (no noise) for all trials.
hand dropping the object, in the case of the controller trained
with noise. The controller trained without noise dropped the
object in 4 out of 20 test trials. Thus injecting some noise
in the initial state (around 2.5 %) helps improve robustness.
Of course on the real robot we cannot avoid injecting such
noise, because exact repositioning is very difficult.
7
robots are involved. Furthermore, many useful task goals
are non-convex, and while the method can optimize nonconvex cost functions, like all local optimization methods,
it does not necessarily converge to a global optimum.
Random exploration can help mitigate some of these issues.
To encourage exploration, we add random noise while
collecting trajectory samples(step 3 of the algorithm 1). For
tasks where the reward is delayed and there are multiple local
minima, randomly exploring around and hoping to get lucky
takes a big toll on sample efficiency. A well know strategy to
overcome some of these challenges is to imitate an expert,
in order to effectively steer the learned controller towards
an effective solution. However, simply following an expertprovided behavior does not necessarily produce behavior
that is robust to perturbations. In this section, we describe
how we can combine learning from expert teleoperation with
trajectory-centric reinforcement learning to acquire more
complex manipulation skills that overcome local optima
by following expert demonstrations while still retaining the
robustness benefits of learning from experience.
Learning Policies from Experience and
Imitation
The previous section highlighted the strengths of our
reinforcement learning algorithm, outlined in Section 5,
in synthesizing the details of dexterous manipulation
strategies while still preserving sample efficiency. However,
as our algorithm makes progress by optimizing a quadratic
approximation of the cost over a local approximation of
the dynamics, it can become stuck in local minima if the
approximation of the learned dynamics isn’t sufficiently
accurate, or when the cost is not convex. In principle,
arbitrary precision can be achieved by increasing the number
of Gaussian kernels used for the dynamics prior, and by
increasing the number of trajectory samples N used for
the fitting the dynamics. In practice, collecting an arbitrary
number of samples might not be feasible due to time
and computational limitations, especially when physical
Prepared using sagej.cls
Task Details
The task in this second set of experiments consists of picking
up an elongated tube from the table. This task is challenging
because the cost depends on the final configuration of the
tube, which depends in a discontinuous manner on the
positions of the fingers. Furthermore, grasping the tube from
various initial poses involves the use of multiple different
strategies. Note that the hand is rigidly mounted, so the
picking must be done entirely using motion of the wrist
and fingers. The shape of the tube further complicates the
task. Naı̈ve strategies like aligning the hand to the object’s
principal axis and force-closure will not perform well, since
only the wrist is available for positioning the hand. Instead,
the hand must use complex finger gaits to maneuver the
object into position. The significant weight of the tube (in
comparison to the lifting capability of the hand) constantly
tilts the object out of the grasp. Successful strategies need to
discover ways to use the thumb and fingers to act as ‘support’
and ‘pivot’ for each other, in order to reposition and reorient
the object leading to a successful grasp and pick up. Failure
to establish either the support or pivot results in the tube
flying out of the workspace due to a net unbalanced force
on the object.
The hand is mounted approximately thirty degrees to the
horizontal. The task starts with the hand in zero position
(Figure 8). The goal is to lift the object from a known initial
condition (the object is being tracked using the PhaseSpace
motion capture system). The tube is considered lifted if it is
in complete control of the hand (i.e. not falling out of the
hand or resting on the table) and all points on the object
are above the ground by a certain height. Note that for
the resemblance between the hardware and the simulated
platform, the contact between the fingers and the ground
plane is disabled for the simulated platform in order to allow
the fingers to curl from below the object.
7.2
Mujoco Haptix
While leveraging an expert is quite desirable, deploying and
exploiting an expert is exceptionally difficult for dexterous
Kumar, Gupta, Todorov and Levine
9
(a) Overall setup
(a) Hardware platform
(b) Simulation platform
Figure 8. Initial pose for the pickup task
manipulation. This is due to two principal factors. First,
dexterous manipulation strategies are extremely sensitive to
minor variations in the contact forces, contact locations, and
object positions. Thus, minor deviations from the expert
demonstrations render the demonstration useless. Second,
technology to capture the details of hand manipulation
is often unreliable. Unlike full body movements, hand
manipulation behaviors unfold in a compact region of space
co-inhabited by the objects being manipulated. This makes
motion capture difficult, due to occlusions and, in the
case of passive systems, marker confusion. Manipulation
also involves large numbers of contacts, including dynamic
phenomena such as rolling, sliding, stick-slip, deformations,
and soft contacts. The human hand takes advantage of these
rich dynamics, but recording the data and interpreting it with
regard to well-defined physics models is challenging.
To address these challenges, we exploit the adaptation
capabilities of the brain in order to ship the data collection
from the real world to a physically realistic simulation.
The Mujoco Haptix system (Kumar and Todorov 2015)
was developed to facilitate physically-consistent recording
of rich hand-object interactions. This was done in the
context of the DARPA HAPTIX program and was adapted
for our purposes here. The simulation is based on the
MuJoCo physics engine. The Haptix framework augments
the simulator with real-time motion capture of arm and hand
movements, and stereoscopic visualization using OpenGL
projection from the viewpoint of the users head (which
is also tracked via motion capture.) The resulting system
has empirically-validated end-to-end latency of 42 msec. It
creates a sense of realism which is sufficient for human
users to interact with virtual objects in a natural way. Since
the interaction happens in simulation, we can record every
aspect of it including joint kinematics and dynamics, contact
interactions, simulated sensor readings etc. There are no
sensor technologies available today that could record such
rich dataset from hand-object interactions in the physical
world. Furthermore, since the interaction is based on our
simulation model of multi-joint and contact dynamics, the
dataset is by definition physically-consistent.
Prepared using sagej.cls
(b) Expert interacting with the simulation using Cyber glove
system
Figure 9. Mujoco Haptix system. NVIDIA 3D Vision 2 glasses
are used for stereoscopic visualization, together with a BenQ
GTG XL2720Z stereo monitor. OptiTrack motion capture system
is used for 3D glasses, head, and forearm tracking. A
CyberGlove, calibrated for the ADROIT model, is used for
tracking the finger joints.
7.3
Expert Demonstrations
We use Mujoco Hapix to capture expert demonstrations.
The expert first goes through a regression based calibration
process, that maps the Cyberglove sensors to the ADROIT
joint space. To enable hand manipulation, we map the joint
angle q c reported by the Cyberglove to actuator commands
using the Equation (2):
ut = kJ tendon (qt − qct ),
(2)
where qt is the current joint configuration of the hand,
J tendom is the tendon jacobian that maps the joint space
to the tendon space, and k is the gain vector. ut is applied
as controls to the pneumatic actuators and we let the
simulation evolve by stepping the physics of the world
forward in time. Hand-object interactions evolve as the
simulation steps forward. The expert is in a tight feedback
loop with the simulation via the stereoscopic rendering. As
the user interacts with the system by changing its strategy,
hand manipulation behaviors emerge. We record the state
xt and the control trajectory ut over time as demonstration
trajectory. where x = (q, qpos , qrot , q̇, q̇pos , q̇rot , a). Here
q denotes the vector of hand joint angles, qpos the object
positions, qrot the object rotations, a the vector of cylinder
pressures, and u the vector of valve command signals.
10
Journal Title XX(X)
7.4 Learning Imitation Policies
We first attempted to learn the pickup task from scratch using
the following cost function:
`(xt , ut ) =α1 ||qt − q ∗ ||2 + α2 ||ut ||2 +
α3 ||qtpos − q pos∗ ||2 + α4 ||qtrot − q rot∗x ||2 ,
where q pos∗ and q rot∗x denote the goal pose of the object,
and q ∗ is a target joint angle configuration that corresponds
to a grasp. Simple learning by experience fails to come
up with a strategy that can pick up the rod even after
significant cost parameter tuning. The observed behavior is a
combination of the following: (a) Hand randomly explores
for a while and fails to find the object; (b) The fingers
eventually find the object and knock it away from the
manipulate-able workspace; (c) The fingers keep tapping on
the top of the tube without any significant improvement.
The difficulty in finding a successful pickup behavior stems
from a combination of factors. First, the cost function only
produces a meaningful signal toward the end of the episode,
once the object is already grasped. Second, almost every
strategy that can succeed has to first clear the object and
only then dig the fingers into the space below the tube, with
fingers on both sides to restrict the object’s motion. Lastly,
the high dimensionality of the ADROIT platform means that
the space of successful solutions is very narrow, and a huge
variety of movements are available that all fail at the task.
Note that, for cases where the object isn’t aligned well with
the palm, additional reorientation maneuvers are required in
order to pick up the object, due to the hand’s restricted lateral
mobility.
Our learning process for these skills combines learning
from demonstrations with learning from experience, with
the expert demonstration used to bootstrap the learning
process. During the learning from experience phase, we use
an additional shaping cost, which is a weak cost term that
prevents the learning process from deviating too far away
from the expert demonstration. The assumption here is that
the expert demonstration is already quite good and we don’t
need to deviate too far to improve the solution. The strength
of the shaping cost controls the amount of deviation from the
expert demonstration.
For bootstrapping the learning using expert demonstrations, instead of using a random policy in step 3 of the Algorithm 1 at the initial iteration, small random noise is injected
into the control trajectory of the expert demonstration to
collect trajectory samples {τi }. The running cost we use is
`(xt , ut ) =||qt − q̂t ||2 + 0.1||ut ||2 + 50||qtposZ − 0.12||2 ,
where q̂t is the hand configuration of the expert at time t
and qtposZ is the vertical height of the object at time t. The
first term is the shaping cost that restricts the learning from
deviating too far from the demonstration, the second term
is the control cost and the final term encourages picking up
the object to a height of 12 cm above the ground. The final
cost is the same as the running cost. Figure 10 presents a
representative pickup behavior achieved using this method.
8 Policy Generalization
Section 6 and Section 7 explores the capabilities of
trajectory-centric reinforcement learning to learn robust
Prepared using sagej.cls
“local” policies for dexterous manipulation of freely-moving
objects. However, the resulting local policies succeed from
specific initial states, and are not designed to handle variation
in the initial conditions, such as different initial placement of
the manipulated object. In this section, we will discuss how
we can use the local policies to learn a single global policy
that generalizes effectively across different initial conditions.
Generalizable global policies can typically be learned
using reinforcement learning with more expressive function
approximators. Deep neural networks represent a particularly
general and expressive class of function approximators,
and have been recently used to learn skills that range
from playing Atari games (Mnih, Kavukcuoglu, Silver,
Rusu, Veness, Bellemare, Graves, Riedmiller, Fidjeland
and Ostrovski 2015) to vision-based robotic manipulation
(Levine, Finn, Darrell and Abbeel 2015a). In this section,
we evaluate how well deep neural networks can capture
generalizable policies for the pickup skill discussed in the
previous section, where variation in the task corresponds to
changes in the initial pose of the rod. Our goal is to explore
the generalization capabilities of deep neural networks along
two axes: the ability to handle variability in the initial
conditions, and the ability to handle partial observability
and limited sensing. To that end, we will evaluate deep
neural network policies that perform the task either with full
state observation, or with only the onboard sensing available
on the ADROIT platform, without external motion capture
markers to provide the pose of the rod.
8.1
Task Details
We evaluate generalization on the same pickup task as
outlined in Section 7.1. To analyze generalization, we vary
the orientation of the rod at the initial state. The goal
is to learn a strategy for this task that succeeds for any
initial rod orientation. This is particularly challenging since
the robot cannot translate or reorient the wrist (since the
hand is stationary), and therefore must utilize substantially
different grasping strategies for different rod orientations,
including the use of auxiliary finger motions to reposition
the rod into the desired pose. Besides this challenge, the
task also inherits all of the difficulties detailed in Section 7,
including the high dimensionality of the system and the
challenge of delayed rewards. To mitigate the challenge
of local optima, we again use expert demonstrations. We
collected a set of 10 demonstrations across 180 degrees of
variation in the rod orientation. Figure 11 shows the 10
initial configurations from which an expert was engaged for
providing demonstrations.
8.2
Local Policies
Before evaluating generalization, we first analyze the
performance of the individual local policies trained with
imitation and learning from experience. For each task
execution, we evaluate the success or failure of trial
according to the following criteria: a successful picking trial
must result in the object being stationary, both extremely of
the rod being above the ground by a certain height, and the
entire rod aligned with the x-axis, so as to ensure a successful
grasp into the desired goal position. Note that partially
successful executions, where the tube’s center of mass is
Kumar, Gupta, Todorov and Levine
11
Figure 10. Pick up strategy learned using learning via imitation. The movement starts with aligning the palm with the object, then
curling the fingers under the object followed by an object-reorientation movement that further aligns the object with the palm. The
final movement engages the thumb is squeeze hard again the palm before lifting the wrist upward.
As shown in Figure 12b, we can improve the robustness of
the local policies corresponding to each condition by further
training each of those policies from experience, following
the method described in Section 7. Each demonstration
was subjected to 10 iterations of learning from experience,
and the resulting controllers were evaluated using the same
procedure as in the previous paragraph. The controllers
succeed in a wider neighborhood around their default rod
pose, with the brittle strategy in condition 6 becoming
much more robust, and the region of success for conditions
9, 3, and 0 expanding substantially on both sides. The
overall costs and success rates for individual conditions are
summarized in Figure 13. While we observe that the overall
cost remained similar, the success rates for varying rod poses
increased substantially, particularly for conditions 6, 8, 9,
and 10. Figure 14 evaluates the effectiveness of the local
policies for the entire range of task variation. We observe
that the policies are less effective as they move away from
the zero point (marked with dark vertical line).
Figure 11. The initial tube configuration of the expert
demonstration set.
lifted up but one of the end points remains on the ground,
are still marked as failures. This allows us to determine not
only whether the tube was picked up, but whether it was also
positioned in the desired pose – an important condition for
any subsequent manipulations that might be applied to it.
In Figure 12a, we analyze the performance obtained by
following the demonstrated behaviors directly, for variation
in the rod angle within the local neighborhood of the rod pose
for which the demonstration was generated. The numbers
correspond to the conditions illustrated in Figure 11. The xaxis corresponds to variation in the orientation of the rod.
Successful trials are marked as green circles, while failures
are marked as red crosses. Overall, we observe that most
demonstrations are somewhat successful for the particular
rod angle for which they were created, but the success
rate decreases sharply under variation in the rod pose, with
discontinuous boundaries in the success region (particularly
for conditions 9, 3, and 0). Some conditions, such as
condition 6, are typically successful, but exhibit a brittle
strategy that sometimes fails right at the demonstration angle,
while others, such as conditions 3 and 1, fail sporadically at
various rod poses.
Prepared using sagej.cls
8.3
Nearest Neighbor
We can observe from the results in Figure 12b that, after
training, each local policy succeeds in a neighborhood that
extends to the boundary of the next local policy. This
suggests that a relatively simple nearest-neighbor technique
could in principle allow for a nonparametric strategy to
expand the success region to the entire range of rod
orientations. In Figure 15, we evaluate the performance of
this nearest neighbor strategy, which simply deploys the
local policy trained for the rod orientation that is closest
to the orientation observed in the initial state, in terms of
Euclidean distance. Successful trials are marked in green
and failures in red, and the overall success rate is 90.8%.
Although the nearest neighbor strategy is successful in this
case, it requires retaining the full set of local policies and
manual specification of the variables on which the nearest
neighbor queries should be performed (in this case, the rod
pose). Furthermore, the nearest neighbor strategy requires
us to know the rod pose, which in a physical experiment
would involve the use of external motion capture markers.
Can we instead learn a generalizable policy for grasping the
tube that does not use any external sensing, only the onboard
sensing available on the hand? In the next section, we will
discuss how deep neural network function approximators can
leverage the local policies to learn generalizable strategies.
12
Journal Title XX(X)
Demo playback
18
20
18
24
22
20
18
28
26
24
22
20
18
28
26
24
22
20
18
30
28
26
24
22
20
18
-1.65
0
-0.7
-1.3
-0.95
-1.65
-0.7
-0.35
-0.95
-0.35
Rod angle
condition1 (80%)
34
Failed
Success
32
Trajectory Cost. (15.0146)
Trajectory Cost. (18.2656)
30
condition2 (96%)
34
Failed
Success
32
Rod angle
Failed
Success
32
30
28
26
24
22
20
18
condition0 (67%)
34
30
28
26
24
22
20
18
30
28
26
24
22
20
18
Rod angle
0
0.95
0.75
0
0.35
0.95
0.75
0.35
0
-0.35
0.95
0.75
0.35
-0.95
0
14
-0.7
14
-0.35
14
0
14
0.35
16
14
-0.7
16
-0.35
16
-0.95
16
-1.3
16
Rod angle
Failed
Success
32
Trajectory Cost. (14.4766)
condition3 (73%)
34
Failed
Success
32
Rod angle
Trajectory Cost. (14.8559)
condition4 (86%)
34
-2
-2
Rod angle
-1.3
14
-1.65
14
-0.7
14
-1.3
14
-0.95
16
14
-1.65
16
-1.3
16
-1.65
16
-2
16
Rod angle
Trajectory Cost. (22.9098)
22
26
30
Rod angle
Rod angle
0
20
24
28
-0.7
22
26
30
-0.35
18
24
28
-0.95
20
26
30
0.95
22
28
Failed
Success
32
0.75
24
30
32
-1.3
26
32
condition5 (54%)
34
Failed
Success
0.35
28
condition6 (97%)
34
Failed
Success
Trajectory Cost. (21.4001)
30
34
Failed
Success
32
Trajectory Cost. (15.8836)
32
Trajectory Cost. (18.3819)
34
Failed
Success
condition7 (61%)
Trajectory Cost. (16.5734)
34
condition8 (52%)
Trajectory Cost. (18.8065)
condition9 (47%)
Rod angle
GPS neighbourhood.
Local Policies
(a) Testing robustness of the expert demonstration in its local
20
18
24
22
20
18
28
26
24
22
20
18
30
28
26
24
22
20
18
-1.65
0
-0.35
-0.7
-0.95
-1.3
-1.65
-0.35
-0.7
-0.95
-1.3
30
28
26
24
22
20
18
Rod angle
condition1 (87%)
34
Failed
Success
32
Trajectory Cost. (14.7869)
Trajectory Cost. (17.7907)
30
condition2 (93%)
34
Failed
Success
32
Rod angle
condition0 (63%)
34
Failed
Success
32
30
28
26
24
22
20
18
30
28
26
24
22
20
18
Rod angle
Rod angle
0.95
0.75
0
0.35
0.95
0.75
0.35
0
-0.35
0.95
0.75
0.35
0
-0.7
14
-0.35
14
-0.95
14
0
14
0.35
16
14
-0.35
16
-0.7
16
-0.95
16
-1.3
16
Rod angle
Rod angle
Failed
Success
32
Trajectory Cost. (14.4362)
condition3 (75%)
34
Failed
Success
32
Rod angle
Trajectory Cost. (14.6357)
condition4 (91%)
34
-2
-2
Rod angle
-1.65
14
-0.7
14
-1.3
14
-0.95
14
-1.65
16
14
-1.3
16
-1.65
16
-2
16
0
18
22
26
-0.35
20
24
28
16
Rod angle
Trajectory Cost. (20.6241)
22
26
30
-0.7
18
24
28
-0.95
20
26
30
0.95
18
22
28
0.75
20
24
30
Failed
Success
32
-1.3
22
26
Failed
Success
32
0
24
28
condition5 (86%)
34
0.35
26
30
condition6 (99%)
34
Failed
Success
32
Trajectory Cost. (18.9236)
28
Failed
Success
32
Trajectory Cost. (16.7368)
30
condition7 (81%)
34
Trajectory Cost. (18.072)
Failed
Success
32
Trajectory Cost. (18.3462)
condition8 (87%)
34
Trajectory Cost. (18.264)
condition9 (63%)
34
Rod angle
(b) Testing robustness of the local policies trained around the demonstrations in its local neighbourhood.
Figure 12. Different plots represent different conditions. Ticks on the X axis marks the different rod angles, corresponding to each
condition where expert demonstration was collected. The true angle of the plot is marked with solid black line. Each condition was
tested for 100 trials. Successful trials are marked in green and unsuccessful are marked in red. Success percentage are marked in
the title and the average cost are provided in the Y label.
Prepared using sagej.cls
Kumar, Gupta, Todorov and Levine
13
Average Cost
30
demo
GPS
20
10
Success %
0
100
50
0
1
2
3
4
5
6
7
8
9
10
condition
Figure 13. Performance comparison between the expert
demonstrations and the local policies trained around the expert
demonstrations in the local neighbourhood of the task (i.e. the
rod angle with the vertical)
8.4 Generalization with Deep Neural Networks
As discussed previously, although the nearest neighbor
strategy has an extremely high success rate for the
pick-up task being considered, it suffers from several
limitations. First, nearest neighbor methods suffer from
the curse of dimensionality, which limits their applicability
to high-dimensional spaces. Indeed, the nearest neighbor
technique described in the preceding section requires us to
manually specify that the nearest neighbor queries should
be performed with respect to the pose of the rod. This
information may not be readily available in general robotic
manipulation tasks, where the robot must choose the strategy
based on high-dimensional raw sensory information, such as
camera images or inputs from tactile sensors. Second, the
nearest neighbor method described above still corresponds
to a time-varying control strategy, which can be limiting in
cases where the robot might begin the task from arbitrary
initial states. Finally, it requires storing all of the individual
trajectory-centric controllers.
In this section, we evaluate whether a deep neural network
representation of the rod grasping manipulation skill can
be learned with comparable generalization capability and
limited sensing. Deep neural networks provide a powerful
and flexible class of function approximators that can be
readily combined with high-dimensional inputs and complex
sensing. Furthermore, the use of deep neural networks as the
final policy representation does not require us to manually
select a small set of state variables for nearest neighbor
queries, making them a more general choice that is likely
to extend to a wide variety of dexterous manipulation tasks.
To evaluate the use of deep neural networks for
learning generalizable dexterous manipulation strategies,
we used the collection of local policies trained from
initial demonstrations to produce training data for neural
network training. Each of the local policies discussed in
Section 8.2 was used to generate 50 sample trajectories
by executing each policy from its corresponding initial
state. These samples were then used to train a deep neural
network with 6 fully connected layers and 150 recitified
linear (ReLU) hidden units in each layer, as shown in
Prepared using sagej.cls
Figure 16. The network was trained either with the full state
information provided to the local policies (which includes
the pose and velocity of the rod), or with partial information
reflecting onboard sensing, with knowledge about the rod
pose excluded from the inputs to the neural network.
The results of the neural network policy with full state
observations are shown in Figure 17. This result indicates
that the neural network policy trained from the local policies
is generally not as successful as the nearest neighbor
strategy in the fully observed case. This is not particularly
surprising: the nearest neighbor strategy already uses a set
of very successful local policies that can succeed up to
one angle increment, and thus form overlapping regions of
effectiveness. However, the neural network must distill the
distinct strategies of different local policies into a single
coherent function, and therefore generally performs worse
in this case. Figure 18 and Figure 19 show the performance
of a large and small neural network trained without the rod
pose provided as input, instead using inputs from the hand’s
onboard tactile sensors in the fingertips. When the network
is trained without observations of the rod pose, it achieves
a success rate of 74%, nearly as high as the fully observed
condition, which is substantially higher than the best local
policies, as shown in Figure 14.† These results also show
that the larger network is essential for properly handling
this condition. Interestingly, when the neural network is
not provided with the tactile sensors either, as shown
in Figure 20, it achieves nearly the same performance,
indicating that the neural network is able to make use of
proprioceptive sensing to determine which strategy to use.
This is also not entirely surprising, since collisions and
contacts result in motion of the fingers that can be detected
from proprioception alone. These results indicate that the
neural network policy that is only allowed to use onboard
sensing learns a robust feedback behavior that can adjust
the grasping strategy based on proprioceptive feedback,
something that is not possible with the nearest neighbor
strategy, which must choose the local policy to deploy at the
beginning of the episode (before any proprioceptive sensing
can take place). This suggests one of the benefits of deep
neural networks for dexterous control.
8.5
Results on ADROIT Hardware Platform
In this section, we present our generalization results for the
pick up task on the ADROIT hardware platform. The details
of the task are the same as mentioned above. However, we
consider generalization both in the position and orientation
of the object. The orientation generalization is in a smaller
neighborhood than the simulated experiments above, and
we train a 6 layer fully connected neural network with 120
rectified linear (ReLU) units at each layer. The training
samples are collected by sampling the local controllers
learned around the 4 demonstrations provided by the expert
user. The expert demonstrations were collected at the initial
† It
is worth noting here that the hand is reset to the starting state of the
trajectory controller with the closest angle for each rod pose. While these
states are very similar, they may nonetheless provide additional cues to
the network. Our ongoing experiments, which we will include in the final
version, will address randomized initial poses to control for this potentially
confounding factor.
14
Journal Title XX(X)
30
20
0.95
0.75
0
0.35
-0.7
-0.35
-1.3
-0.95
-1.65
0.95
-2
Failed Trials
Successful Trial
50
40
30
20
Rod angle (500 samples)
Rod angle (500 samples)
0.95
0.75
0
0.35
-0.7
-0.35
-0.95
-1.3
0
-1.65
10
-2
0.95
0.75
0
0.35
-0.35
-0.7
-1.3
-0.95
-2
-1.65
0.95
0.75
0
0.35
-0.7
-0.35
-0.95
-1.3
-1.65
-2
0.95
0.35
-0.35
0.75
0
0
0
-0.7
0
-1.3
0
-0.95
0.75
20
10
-1.65
0
30
10
-2
0.35
40
10
Rod angle (500 samples)
Trajectory Cost
(Average=5.4164)
20
0.95
30
10
Rod angle (500 samples)
-0.7
0.95
0.75
0
0.35
40
0.75
20
Local Policy (condition0) (success:45.4%)
60
50
0
30
Rod angle (500 samples)
0.35
40
50
-0.7
20
20
0
Failed Trials
Successful Trial
-0.35
30
30
70
60
-1.3
40
Failed Trials
Successful Trial
60
50
40
Rod angle (500 samples)
Local Policy (condition1) (success:43.2%)
70
-0.95
Failed Trials
Successful Trial
Trajectory Cost
(Average=3.4108)
Trajectory Cost
(Average=2.9321)
50
-0.7
Rod angle (500 samples)
60
50
10
Local Policy (condition2) (success:43.8%)
70
Trajectory Cost
(Average=4.207)
Failed Trials
Successful Trial
60
-0.35
-1.3
Rod angle (500 samples)
Local Policy (condition3) (success:25.2%)
70
-0.95
-2
-1.65
0.95
0.75
0
0.35
-0.7
-0.35
-1.3
Rod angle (500 samples)
-0.95
-1.65
-2
0.35
-0.35
0.95
0
0.75
0
0
0
-0.7
0
-1.3
10
-0.95
10
-2
10
Local Policy (condition4) (success:35%)
Trajectory Cost
(Average=3.6139)
40
-0.35
20
-1.3
30
10
70
Trajectory Cost
(Average=2.4246)
40
-0.95
20
-2
30
-1.65
40
Failed Trials
Successful Trial
60
50
-1.65
20
50
Local Policy (condition5) (success:39.6%)
70
Failed Trials
Successful Trial
60
Trajectory Cost
(Average=4.3608)
30
50
Local Policy (condition6) (success:54.2%)
70
Failed Trials
Successful Trial
60
Trajectory Cost
(Average=3.9493)
Trajectory Cost
(Average=3.42)
40
Local Policy (condition7) (success:38.2%)
70
Failed Trials
Successful Trial
60
50
-1.65
Trajectory Cost
(Average=4.1713)
Local Policy (condition8) (success:42.6%)
70
Failed Trials
Successful Trial
60
-2
Local Policy (condition9) (success:46.2%)
70
Rod angle (500 samples)
Figure 14. Performance of the local policies across the task variations
Nearest Neighbor for Pick Up Task (success:90.8%)
Fully Observed Neural Network Policy using Large Network (success:75%)
14
Failed Trials
Successful Trial
30
Failed Trials
Successful Trial
25
Trajectory Cost
(Average=2.2729)
Trajectory Cost
(Average=1.9132)
12
20
15
10
10
8
6
4
5
Rod angle (500 samples)
Pressure
Tendon
Joint
Valve
(NI) (35 KHz)
(NI) (9 KHz)
µU (500 Hz)
(NI)(200 Hz)
0.95
0.75
0.35
0
-0.35
Figure 17. Neural network policy performance with fully
observed state, at random initial object poses. The network
consisted of 6 layers with 150 hidden units each, and was
trained on the local policies from conditions 1, 2, 3, 4, 5, 6, and
8.
70
Partially Observed Neural Network Policy without Rod Position/Velocity with Large Network (success:74%)
Failed Trials
Successful Trial
60
and
subsample
ADROIT State
Socket (200hz)
50
ADROIT Control
Socket (200hz)
Trajectory Cost
(Average=1.886)
Filte r
Driver
40
30
20
10
(100,1)
Figure 16. Over architecture of the system using the network
as controller.
condition as shown in Figure 21. Local controllers were
trained using 10 iterations of learning via imitation strategy
as mentioned in Section 7. The training set for the network
consists of 20 samples for each condition (i.e. 80 samples
in total). In figure Figure 22, we cross-validate each policy
for different condition, including random conditions. We
found that local controllers are partially successful in a
neighborhood wider than just their own. There is no local
controller that works well for all the conditions. Overall
the neural network policy performs at par with the local
controllers on the respective conditions. The neural network
policy, however, generalizes better than the individual local
policies as conveyed by its higher success in picking object
from random initial configurations.
Prepared using sagej.cls
0.95
0.75
0.35
0
-0.7
-0.35
(40,1)
-0.95
𝑉𝑎𝑙𝑣𝑒 𝑣𝑜𝑙𝑡𝑎𝑔𝑒
-1.3
Controls =
-2
State =
𝑗𝑜𝑖𝑛𝑡
𝑗𝑜𝑖𝑛𝑡 𝑣𝑒𝑙
𝑝𝑟𝑒𝑠𝑠𝑢𝑟𝑒
-1.65
Hardware
-0.7
Rod angle (500 samples)
Figure 15. Performance of the nearest neighbor policy with full
state information. Test trials are collected from random initial
conditions of the object, with the local policy corresponding to
the nearest rod orientation used in each trial.
Controller
-0.95
-1.3
-1.65
-2
0.95
0.75
0.35
0
-0.35
-0.7
-0.95
-1.3
-1.65
-2
2
Rod angle (500 samples)
Figure 18. Performance of a large neural network (6 layers,
150 units) with partial observations and touch sensors. The rod
position is not provided to the network, but instead the network
can use inputs from the tactile sensors on the fingertips. Note
that overall performance exceeds the best single local policy.
9
Discussion and Future Work
We demonstrated learning-based control of a complex, highdimensional, pneumatically-driven hand. Our results include
simple tasks such as reaching a target pose, as well as
dynamic manipulation behaviors that involve repositioning
a freely-moving cylindrical object. We also explored more
complex grasping tasks that require repositioning of an
object in the hand and handling delayed rewards, with
the use of expert demonstrations to bootstrap learning.
Aside from the high-level objective encoded in the cost
function and the demonstrations, the learning algorithms
do not use domain knowledge about the task or the
Kumar, Gupta, Todorov and Levine
15
Partially Observed Neural Network Policy without Rod Position/Velocity with Small Network (success:58.517%)
60
Failed Trials
Successful Trial
Trajectory Cost
(Average=2.4025)
50
40
30
20
0.95
0.75
0.35
0
-0.35
-0.7
-0.95
-1.3
-1.65
-2
10
Rod angle (499 samples)
Figure 19. Performance of a small neural network (4 layers, 80
units) with partial observations and touch sensors. The rod
position is not provided to the network, but instead the network
can use inputs from the tactile sensors on the fingertips. Note
that overall performance of the smaller network is substantially
degraded.
Partially Observed Neural Network Policy without Rod Position/Velocity/Touch Sensors (success:73%)
Failed Trials
Successful Trial
16
Trajectory Cost
(Average=4.2198)
14
12
10
8
6
4
0.95
0.75
0.35
0
-0.35
-0.7
-0.95
-1.3
-1.65
-2
2
Rod angle (1000 samples)
Figure 20. Performance of a large neural network (6 layers,
150 units) with partial observations and without touch sensors.
The rod position is not provided to the network. Note that the
large network performs well even without the touch sensors,
indicating that the network is able to use proprioceptive inputs
to determine the strategy.
Figure 21. Different conditions used for expert demonstrations
hardware. The experiments show that effective manipulation
strategies can be automatically discovered in this way. We
further evaluated the generalization capabilities of deep
neural network controllers and nearest neighbour switching
controllers. While both methods were able to generalize
to some extent, the performance of the nearest neighbour
method was higher. The two methods have relative strengths
Prepared using sagej.cls
Figure 22. Cross validation of different policies under different
conditions on ADROIT Hardware platform
and weaknesses and can be improved and perhaps combined
in future work, as described below.
The neural network controller was given access to
proprioceptive and tactile sensory input, but not to vision
input about the object state. Thus it was solving a harder
problem, with the caveat that the initial pose of the hand
was correlated with the initial pose of the object, and we
need to investigate in more detail to what extent the network
can indeed perform the task blindly. Blind manipulation is of
course not a requirement, given that many vision sensors now
exist and real-time state estimators (Schmidt, Newcombe and
Fox 2015) can provide object states. It will be interesting to
test the network performance with more complete sensory
input.
The nearest neighbour controller is currently timevarying: the switch happens at the beginning of the
movement based on the initial state of the object, and then
we use the selected local feedback controller for the rest
of the movement. Another approach would be to store the
dataset of states, controls and feedback gains along the
local trajectories, and use nearest neighbour queries at each
point in time. This requires more storage and processing
than evaluating a neural network, but should not be a
challenge for modern computing hardware. Furthermore,
it may be possible to store a reduced set of controllers
as opposed to storing the states along all trajectories. For
example, Simbicon (Yin, Loken and van de Panne 2007)
uses a small collection of time-invariant feedback controllers
together with a suitable switching policy, and performs
bipedal locomotion remarkably well. The Boston Dynamics
approach to control, while not described in detail, also
appears to be related.
It may be possible to combine the benefits of the two
generalization methods considered here. While deep learning
is currently popular, there is no reason to limit ourselves
to generic networks. Instead we could consider a mixtureof-experts architecture (Jordan and Jacobs 1994), where the
gating network corresponds to the switching mechanism
in our current nearest neighbor approach, while the expert
networks correspond to our local trajectory controllers.
We would still want to leverage the power of trajectory
optimization in training the experts, and perhaps use the
current switching mechanism to pre-train the gating network.
Another direction for future work is to expand the set of
tasks under consideration. While the present tasks are quite
complex, they have been selected for their intrinsic stability.
For example, consider our object-spinning task. If we were
16
to attempt an identical task but with the palm facing down,
our present approach would not work. This is because the
object would drop before we have had time to interact with
it and learn anything. In general, data-driven approaches to
robotics require either an intrinsically stable task (which may
be rare in practice), or a pre-existing controller that is able to
collect relevant data without causing immediate failure. In
this particular case we could perhaps use tele-operation to
obtain such a controller, but a more general and automated
solution is needed.
Finally, even though the focus of this paper is purely
data-driven learning, there is no reason to take such a onesided approach longer term. As with every other instance
of learning from limited and noisy data, the best results
are likely to be obtained when data is combined with
suitable priors. In the case of robotics, physics provide
a strong prior that can rule out the large majority of
candidate models and control policies, thereby improving
generalization from limited data on the physical system.
Existing physics simulators can simulate complex systems
such as the one studied here much faster than real-time,
and can be run in parallel. This functionality can be used
for model-predictive control, aided by a neural network
representing a controller and/or a value function (Zhong,
Johnson, Tassa, Erez and Todorov 2013). The model itself
could be a hybrid between a physics-based model with a
small number of parameters learned from data, and a neural
network with a larger number of parameters used to fit the
residuals that the physics-based model could not explain.
Acknowledgements
This work was supported by the NIH, NSF and DARPA. The
authors declare that there is no conflict of interest.
References
Abbeel P, Coates A, Quigley M and Ng A (2006) An application
of reinforcement learning to aerobatic helicopter flight. In:
Advances in Neural Information Processing Systems (NIPS).
Amend Jr JR, Brown E, Rodenberg N, Jaeger HM and Lipson
H (2012) A positive pressure universal gripper based on the
jamming of granular material. Robotics, IEEE Transactions on
28(2): 341–350.
Åström KJ and Wittenmark B (2013) Adaptive control. Courier
Corporation.
Bagnell JA and Schneider J (2003) Covariant policy search.
In: International Joint Conference on Artificial Intelligence
(IJCAI).
Bellman R and Kalaba R (1959) A mathematical theory of adaptive
control processes. Proceedings of the National Academy of
Sciences 8(8): 1288–1290.
Boyd S and Vandenberghe L (2004) Convex Optimization. New
York, NY, USA: Cambridge University Press.
Deisenroth M, Neumann G and Peters J (2013) A survey on policy
search for robotics. Foundations and Trends in Robotics 2(1-2):
1–142.
Deisenroth M, Rasmussen C and Fox D (2011) Learning to control
a low-cost manipulator using data-efficient reinforcement
learning. In: Robotics: Science and Systems (RSS).
Prepared using sagej.cls
Journal Title XX(X)
Gupta A, Eppner C, Levine S and Abbeel P (2016) Learning
dexterous manipulation for a soft robotic hand from human
demonstration. In: IROS 2016.
Jordan M and Jacobs R (1994) Hierarchical mixtures of experts and
the em algorithm. Neural Computation 6(2): 181–214.
Kober J, Bagnell JA and Peters J (2013) Reinforcement learning in
robotics: A survey. International Journal of Robotic Research
32(11): 1238–1274.
Kober J, Oztop E and Peters J (2010) Reinforcement learning
to adjust robot movements to new situations. In: Robotics:
Science and Systems (RSS).
Kumar V and Todorov E (2015) Mujoco haptix: A virtual reality
system for hand manipulation. In: Humanoids.
Kumar V, Todorov E and Levine S (2016) Optimal control with
learned local models: Application to dexterous manipulation.
In: International Conference on Robotics and Automation
(ICRA).
Kumar V, Xu Z and Todorov E (2013) Fast, strong and compliant
pneumatic actuation for dexterous tendon-driven hands. In:
International Conference on Robotics and Automation (ICRA).
Levine S and Abbeel P (2014) Learning neural network policies
with guided policy search under unknown dynamics. In:
Advances in Neural Information Processing Systems (NIPS).
Levine S, Finn C, Darrell T and Abbeel P (2015a) End-toend training of deep visuomotor policies. arXiv preprint
arXiv:1504.00702 .
Levine S and Koltun V (2013) Guided policy search.
In:
International Conference on Machine Learning (ICML).
Levine S, Wagener N and Abbeel P (2015b) Learning contact-rich
manipulation skills with guided policy search. In: International
Conference on Robotics and Automation (ICRA).
Li W and Todorov E (2004) Iterative linear quadratic regulator
design for nonlinear biological movement systems. In: ICINCO
(1). pp. 222–229.
Lioutikov R, Paraschos A, Neumann G and Peters J (2014) Samplebased information-theoretic stochastic optimal control. In:
International Conference on Robotics and Automation (ICRA).
Lynch K and Mason
MT (1999) Dynamic nonprehensile
manipulation: Controllability, planning and experiments.
International Journal of Robotics Research 18(1): 64–92.
Mitrovic D, Klanke S and Vijayakumar S (2010) Adaptive optimal
feedback control with learned internal dynamics models. In:
From Motor Learning to Interaction Learning in Robots,
volume 264. pp. 65–84.
Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare
MG, Graves A, Riedmiller M, Fidjeland AK and Ostrovski
G (2015) Human-level control through deep reinforcement
learning. Nature 518(7540): 529–533.
Pastor P, Hoffmann H, Asfour T and Schaal S (2009) Learning and
generalization of motor skills by learning from demonstration.
In: International Conference on Robotics and Automation
(ICRA).
Peters J, Mülling K and Altun Y (2010a) Relative entropy policy
search. In: AAAI. Atlanta.
Peters J, Mülling K and Altün Y (2010b) Relative entropy policy
search. In: AAAI Conference on Artificial Intelligence.
Peters J and Schaal S (2008) Reinforcement learning of motor skills
with policy gradients. Neural Networks 21(4).
Kumar, Gupta, Todorov and Levine
Schmidt T, Newcombe R and Fox D (2015) Dart: dense articulated
real-time tracking with consumer depth cameras.
In:
Autonomous Robots.
Sutton R and Barto A (1998) Reinforcement Learning: An
Introduction. MIT Press.
Tedrake R, Zhang T and Seung H (2004) Stochastic policy gradient
reinforcement learning on a simple 3d biped. In: International
Conference on Intelligent Robots and Systems (IROS).
Tesauro G (1994) Td-gammon, a self-teaching backgammon
program, achieves master-level play. Neural computation 6(2):
215–219.
Theodorou E, Buchli J and Schaal S (2010) Reinforcement
learning of motor skills in high dimensions. In: International
Conference on Robotics and Automation (ICRA).
Todorov E, Erez T and Tassa Y (2012) Mujoco: A physics engine
for model-based control. In: Intelligent Robots and Systems
(IROS), 2012 IEEE/RSJ International Conference on. IEEE,
pp. 5026–5033.
van Hoof H, Hermans T, Neumann G and Peters J (2015) Learning
robot in-hand manipulation with tactile features. In: Humanoid
Robots (Humanoids). IEEE.
Yin K, Loken K and van de Panne M (2007) Simbicon: Simple
biped locomotion control. ACM Trans. Graph. 26(3): Article
105.
Zhong M, Johnson M, Tassa Y, Erez T and Todorov E (2013)
Value function approximation and model-predictive control.
In: IEEE Symposium on Adaptive Dynamic Programming and
Reinforcement Learning.
Prepared using sagej.cls
17
| 3cs.SY
|
arXiv:1712.05288v3 [math.GR] 1 Feb 2018
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE
ALGEBRAS IN CHARACTERISTIC 5
ANASTASIA STAVROVA
Contents
1. Introduction
2. Graded Lie algebras
3. Main theorem
4. Applications to structurable algebras and Kantor pairs
References
1
2
5
6
10
Abstract. We observe that any finite-dimensional central simple 5-graded Lie algebra over
over a field k of characteristic 6= 2, 3 is necessarily classical, i.e. a twisted form of a Chevalley
Lie algebra. Consequently, the classification of central simple structurable algebras and Kantor
pairs over fields of characteristic 5 derives from the classification of simple algebraic groups.
1. Introduction
In 2008 A. Premet and H. Strade [PS08b] completed the classification of finite-dimensional
simple Lie algebras over an algebraically closed field of char 6= 2, 3. As a result, the Block–
Wilson–Premet–Strade classification theorem [Str09] establishes that every such Lie algebra is
one of the following:
• Chevalley (aka classical) simple Lie algebras, classified by Dynkin diagrams Al − Dl , E6 ,
E7 , E8 , F4 , G2 .
• Cartan type Lie algebras in characteristic p > 3.
• Melikian Lie algebras in characteristic p = 5.
We say that a finite-dimensional central simple Lie algebra L over a field k is of Chevalley
type, if it is the derived Lie algebra of the tangent Lie algebra of an adjoint simple algebraic
group. In [BDMSt] we have established that if L is a central simple Lie algebra of Chevalley
type with a non-trivial Z-grading over a field k of characteristic 6= 2, 3, then there is a simple
Kantor pair K over k such that L is the 5-graded Lie algebra to K. If, moreover, char k 6= 5,
then we showed that the converse is also true, namely, any central simple 5-graded Lie algebra
is of Chevalley type.
Previously, the non-degeneracy of simple Kantor pairs over a ring with invertible 2, 3, 5 was
established in [GGLN11]. Among the simple Lie algebras in the Block–Wilson–Premet–Strade
2010 Mathematics Subject Classification. 16W10, 17B60, 17B45, 17Cxx.
Key words and phrases. structurable algebra, Jordan algebra, Kantor pair, Kantor triple system, simple
algebraic group, 5-graded Lie algebra
The author is a winner of the contest “Young Russian Mathematics”. The results are supported by the Russian
Science Foundation grant 14-21-00035.
1
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
2
classification, only the Lie algebras of Chevalley type are non-degenerate [Sel67, p. 124], while
all simple Lie algebras of Cartan and Melikian type are degenerate (this observation goes back
to A. I. Kostrikin and A. Premet; see e.g. [Wil76] and [Str09, Corollary 12.4.7]). Thus, the
non-degeneracy of simple Kantor pairs implies that the associated 5-graded Lie algebras are of
Chevalley type.
In the present paper we extend the above results to the charateristic 5 case. Using the
Block–Wilson–Premet–Strade classification, we deduce that any central simple 5-graded Lie
algebra over a field of characteristic 6= 2, 3 is of Chevalley type. Consequently, for any central
simple structurable algebra or Kantor pair over such a field the associated 5-graded Lie algebra
is of Chevalley type, and hence their classification derives from the classification of isotropic
simple algebraic groups. This allows to extend to char = 5 the classification of central simple
structurable algebras over fields of characteristic 6= 2, 3, 5 due to Allison and Smirnov [All78,
Smi92].
All commutative rings we consider are assumed to be associative and unital. All algebras are
finite-dimensional over the respective scalars unless explicitly mentioned otherwise.
2. Graded Lie algebras
L
Definition 2.1. Let R be a commutative ring, and let L = i∈Z Li be a Z-graded Lie algebra
over R. We say that L is (2n + 1)-graded, if Li = 0 for all i ∈ Z such that |i| > n.
Construction 2.2. For any (2n+ 1)-graded Lie algebra L over R we define the (2n+ 1)-graded
Lie algebra
M
Le =
Lei
i∈Z
over R as follows.
For any i 6= 0, Lei = Li . For i = 0, we define Le0 to be the set of all
Q
φ = (φi ) ∈
EndR (Li ) satisfying the following conditions:
i∈Z\{0}
φi+j ([a, b]) = [φi (a), b] + [a, φj (b)]
for all − n ≤ i, j ≤ n, i, j 6= 0, i 6= −j; a ∈ Li , b ∈ Lj ;
φj ([[a, b], c]) = [[φi (a), b], c] + [[a, φ−i (b)], c] + [[a, b], φj (c)]
(1)
for all − n ≤ i, j ≤ n, i, j 6= 0, a ∈ Li , b ∈ L−i , c ∈ Lj .
In other words, φ ∈ Le0 behaves as a derivation of L that preserves grading, except that φ is
not defined on L0 . Clearly, Le0 is an R-module.
We define the Lie bracket [−, −]Le on Le in terms of the Lie bracket [−, −] on L as follows.
For any u ∈ Lei , v ∈ Lej , i, j ∈ Z, we let
[u, v]
if i 6= −j, i, j 6= 0;
ad([u, v])|Li i∈Z\{0} if i = −j, i 6= 0;
(2)
[u, v]Le = u(v)
if i = 0, j 6= 0;
−v(u)
if i 6= 0, j = 0;
uv − vu
if i = j = 0.
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
3
It is straightforward to check that Le is indeed a (2n + 1)-graded Lie algebra over R. Since it
is not likely to provoke confusion, in what follows we denote the Lie bracket on Le simply by
[−, −].
Note that the Lie algebra Le by construction contains an element ζ ∈ Le0 which acts as a
grading derivation:
[ζ, x] = ix for any i ∈ Z and x ∈ Lei .
It is also easy to see that there is a natural
graded Lie algebra homomorphism L → Le that
sends any element x ∈ L0 to ad(x)|Li i∈Z\{0} .
Lemma 2.3. [BDMSt, Lemma 4.1.4] Let L be a (2n + 1)-graded finite-dimensional Lie algebra
over a field k. Then for any commutative k-algebra R, there is a natural isomorphism of
(2n + 1)-graded R-Lie algebras
Le ⊗k R ∼
= (L ⊗k R)e.
To simplify the notation, under the assumptions of Lemma 2.3 we will consider the Lie
algebras L and Le as functors on the category of k-algebras R, so that, by definition,
e
L(R)
= Le ⊗k R ∼
= (L ⊗k R)e.
Similarly, Aut(L) and Der(L) will stand for the functors of Lie automorphisms and Lie derivations of L respectively:
Aut(L)(R) = AutR (L ⊗k R),
Der(L)(R) = DerR (L ⊗k R).
Note that Aut(L) and Der(L) are naturally represented by closed k-subschemes of the affine
k-scheme of linear endomorphisms of L. In particular, Lemmas ?? and 2.3 imply that there is
an isomorphism of functors
e
Le ∼
= Der(L).
(3)
Lemma 2.4. [BDMSt, Lemma 4.1.5] Let L be a central simple finite-dimensional (2n + 1)graded Lie algebra over a field k, and let K be a field extension of k.
(i) Let Ii ⊆ L(K)i , i ∈ Z \ {0}, be K-subspaces that satisfy the following conditions: for all
i, j ∈ Z \ {0} such that i 6= −j we have
(4)
[L(K)i , Ij ] ⊆ Ii+j ;
(5)
[[L(K)i , I−i ], L(K)i ] ⊆ Ii ;
(6)
[[L(K)i , I−i ], L(K)−i ] ⊆ I−i .
Then either Ii = 0 for all i ∈ Z \ {0}, or Ii = L(K)i for all i ∈ Z \ {0}.
e
(ii) If f ∈ Aut(L)(K)
satisfies f |L(K)i = idL(K)i for all i ∈ Z \ {0}, then f = idL(K)
.
e
L
Definition 2.5. Let L =
Li be a Z-graded Lie algebra over a field k. Consider the 1i∈Z
dimensional split k-subtorus S ∼
= Gm of Aut(L) defined as follows: for any k-algebra R, any
t ∈ Gm (R), and any i ∈ Z, v ∈ Li ⊗k R, we set t · v = ti v. We call S the grading torus of L.
Notation 2.6. Let L be a finite-dimensional 5-graded Lie algebra over a field k, char k 6= 2, 3.
For any commutative k-algebra R and any (x, s) ∈ (L ⊗k R)σ ⊕ (L ⊗k R)2σ = (Lσ ⊕ L2σ ) ⊗k R
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
4
we set
4
X
1
ad(x + s)i ∈ EndR (L ⊗k R).
eσ (x, s) =
i!
i=0
Definition 2.7. Let L be a finite-dimensional 5-graded Lie algebra over k. The grading derivation on L is the derivation ζ ∈ Derk (L) such that for any −2 ≤ i ≤ 2 and any x ∈ Li one
has
ζ(x) = i · x for any −2 ≤ i ≤ 2 and any x ∈ Li .
If L contains an element ζ such that adζ is the grading derivation, we call ζ a grading derivation
of L by abuse of language.
Definition 2.8. Let L be a finite-dimensional 5-graded Lie algebra over a commutative ring
4
P
1
R such that 2, 3 ∈ R× . We say that x ∈ L is algebraic, if the endomorphism
ad(x)i
i!
i=0
is an automorphism of L as an R-Lie algebra. We say that L is algebraic, if all elements
(x, s) ∈ Lσ ⊕ L2σ are algebraic.
Remark 2.9. By [BDMSt, Lemma 3.1.7] any 3-graded Lie algebra, and hence any Jordan
algebra over R with 2, 3 ∈ R× is algebraic; any 5-graded Lie algebra, and hence any structurable
algebra over R with 2, 3, 5 ∈ R× is algebraic. (Despite that lemma is stated for algebras over a
field, the proof is also valid over commutative rings.) By [BDMSt, Theorem 4.2.8] any central
simple structurable division algebra over a field of characteristic 6= 2, 3 is algebraic.
Lemma 2.10. Let L be a finite-dimensional 5-graded Lie algebra over a field k, char k 6= 2, 3.
(i) Let R be any commutative associative unital k-algebra. If (x, s) ∈ Lσ ⊕ L2σ is algebraic,
then (λx, µs) is an algebraic element of L ⊗k R for any λ, µ ∈ R.
(ii) Let L′ be another finite-dimensional 5-graded Lie algebra over k, and let f : L → L′
be a graded k-homomorphism of Lie algebras, such that f |Li : Li → L′i is a bijection for all
i ∈ {±1, ±2}. If (x, s) ∈ Lσ ⊕ L2σ is algebraic in L, then f (x, s) is algebraic in L′ .
Proof. The proof of (i) is the same as in [BDMSt, Lemma 3.1.6]. The proof of (ii) is the same
as in [BDMSt, Lemma 3.1.8].
The following results relate algebraicity to the classification of simple Lie algebras.
Theorem 2.11. [BDMSt, Theorem 4.1.8] Let L be an algebraic central simple 5-graded Lie
algebra over a field k of characteristic different from 2, 3, such that L =
6 L0 . Then the algebraic
k-group G = Aut(L)◦ is an adjoint absolutely simple group of k-rank ≥ 1, satisfying L =
e
[Lie(G), Lie(G)] and Lie(G) ∼
= L.
Lemma 2.12. [BDMSt, Lemma 4.2.4 (i)] Let k be a field, char k 6= 2, 3. Let G be an adjoint
2
L
Li be any
simple algebraic group over k. Let L = Lie(G) be its Lie algebra. Let L =
i=−2
5-grading on L such that L1 ⊕ L−1 6= 0. The 5-graded Lie algebra L is algebraic.
Theorem 2.13. Let L be a central simple 5-graded Lie algebra over a field k of characteristic
different from 2, 3, such that L =
6 L0 . Then L is of Chevalley type if and only if L is algebraic.
Proof. This follows from the two previous results and Lemma 2.10, since Lie algebras of adjoint
simple algebraic groups are by definition the central simple Lie algebras of Chevalley type.
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
3. Main theorem
Lemma 3.1. Let L 6= L0 be a simple 5-graded Lie algebra L
over a field k of characteristic
6= 2, 3, and let S be the corresponding grading torus. Let L = si=−r L[i] be another Z-grading
L
on L such that −1
i=−r L[i] is generated as a Lie algebra by L[−1] 6= 0. If the second grading is
preserved by S(k), then L[−1] 6⊆ L0 .
Proof. Since S preserves the second grading, we have
L[i] =
2
M
(L[i] ∩ Lj )
j=−2
for all −r ≤ i ≤ s. Assume that L[−1] ⊆ L0 . Then L[−i] ⊆ L0 for all 1 ≤ i ≤ r. For any x ∈ Lj ,
−2 ≤ j ≤ 2, we have t · x = tj x forLany t ∈ S(k). Since char(k) 6= 2, 3, there is t ∈ k such that
t±1 6= 1 and t±2 6= 1. L
Then Lj ⊆ i≥0 L[i] for all j 6= 0. Since L is simple, it is generated by
Lj , j 6= 0, hence L = i≥0 L[i] . However, this contradicts L[−1] 6= 0.
Lemma 3.2. Let L be one of the simple graded Cartan type Lie algebras X(m, n)(2) , X ∈
{W, S, H, K}, in the notation of [Str04], over an algebraically closed field k of characteristic 5.
Then L does not have a non-trivial 5-grading.
L
L
Proof. Let L = 2i=−2 Li a non-trivial 5-grading on L, and let L = si=−r L[i] be the standard
grading on L. Note that r = 1 for X = W, S, H, and r = 2, dim(L[−2] ) = 1 for X = K.
By [Str04, Theorem 7.4.1] we can assume that the grading torus S of the 5-grading preserves
the standard grading. Then by Lemma 3.1 the 5-grading on L[−1] is non-trivial. Since L[−1] is
contained in the restricted subalgebra X(m, 1)(2) of L, the induced 5-grading on the restricted
subalgebra is non-trivial, and hence we can assume that L = X(m, 1)(2) from the start. In
particular, ad(L[−2] )5 = 0.
Let x ∈ L[−1] be an element of non-zero 5-grading. Then ad(x)3 (L) is contained in a nilpotent
Lie subalgebra L±1 ⊕ L±2 of L. By [Wil75, Theorem 2] the space L[−1] is an irreducible
representation for the group of Lie algebra automorphisms of L preserving the standard grading.
Thus, L[−1] has a basis consisting of elements y such that ad(y)3 (L) is contained in a nilpotent
Lie subalgebra of L. Since L is restricted, we have ad(L[−2] )5 = 0. Then the Jacobi identity
readily implies that
ad(L[−1] )2·dim(L[−1] )+1+2·4·dim(L[−2] ) (L) ⊆ ad(y)3(L)
for an element y as above. On the other hand, we have L[0] ⊆ ad(L[−1] )s (Ls ). Since L[0] is not
nilpotent, this implies that
(7)
s < 2 · dim(L[−1] ) + 1 + 2 · 4 · dim(L[−2] ).
If X = W, S, H, then dim(L[−1] ) = m. Hence (7) becomes s < 2m + 1. If X = W , then
m ≥ 1 and s = 4m − 1. If X = S, then m ≥ 2 and s = 4m − 2. If X = H, then s = 4m − 3
and m ≥ 2. This contradicts (7).
If X = K, then dim(L[−1] ) = m − 1. Hence (7) becomes s < 2(m − 1) + 9 = 2m + 7. We
have s = 4m, if m + 3 ≡ 0 (mod 5), and s = 4m + 1 otherwise. Since m ≥ 3, both cases
contradict (7).
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
6
Theorem 3.3. Let L be a central simple 5-graded Lie algebra over an algebraically closed field k
of characteristic different from 2, 3, such that L =
6 L0 . Then L is a classical simple (Chevalley)
Lie algebra.
Proof. If char k 6= 5, then L is a Chevalley Lie algebra by Remark 2.9 combined with Theorem 2.13. Assume char k = 5. According to the Block–Wilson–Premet–Strade classification
theorem [Str04], it is enough to check that L is not of Cartan or Melikian type.
Assume first that L is a simple Lie algebra of Cartan type [Str04, Definition 4.2.4]. Let
L = L(−r) ⊇ . . . ⊇ L(s) be a standard filtration of L. By [Str04, Theorem 4.2.7 (3)] the
standard filtration is invariant under all automorphisms of L. In particular, it is invariant
under the grading torus S of the 5-grading. Let
s
M
Gr(L) =
Gr(L)[i]
i=−r
be the associated graded Lie algebra. Hence Gr(L) carries an induced 5-grading. The induced
5-grading is non-trivial, since there is 0 6= x ∈ Li , i 6= 0, and t ∈ F×
5 ⊆ S(k) such that
i
t · x = t x 6= x, and hence t acts non-trivially on the image of x in Gr(L).
The derived series of Gr(L) also inherits the 5-grading, hence Gr(L)(∞) is 5-graded. We
show that the induced 5-grading on Gr(L)(∞) is also non-trivial. Indeed, if it were trivial, then
Gr(L)(∞) ⊆ Gr(L)0 . Since Gr(L) is also a Lie algebra of Cartan type in the sense of [Str04,
Definition 4.2.4], Gr(L)(∞) is simple by [Str04, Theorem 4.2.7 (1)]. Then we have
Gr(L)(∞) ⊆ (Gr(L)0)(∞) ⊆ Gr(L)(∞) ,
which implies Gr(L)(∞) = (Gr(L)0 )(∞) . Then by [Str04, Lemma 4.2.5] we have Gr(L) =
Gr(L)0 , which contradicts the non-triviality of the 5-grading on Gr(L).
By [Str04, Theorem 4.2.7 (2)] we have Gr(L)(∞) = X(m, n)(2) , X = W, S, H, K. By
Lemma 3.2 these algebras do not have non-trivial 5-gradings.
Now let L = M(2, n1 , n2 ) be a simple Lie algebra of Melikian type. By [BKM15, Theorem
1.2] we can assume that the grading L
torus corresponding to the 5-grading under consideration
preserves the standard granding L = si=−3 L[i] of L. Then the grading drivation ζ corresponding to the 5-grading is a homogenous derivation of L. Let W (2, n1 , n2 ) be the standard simple
Witt subalgebra of L. Then by [Str04, Theorem 7.1.4] any homogeneous derivation of L acts
non-trivially on W (2, n1 , n2 ), and hence this subalgebra carries a non-trivial 5-grading induced
by ζ. However, this is not possible by Lemma 3.2.
Theorem 3.4. Let L be a central simple 5-graded Lie algebra over a field k of characteristic
different from 2, 3, such that L =
6 L0 . Then L is of Chevalley type.
Proof. Let k̄ be the algebraic closure of k. Then L ⊗k k̄ is a central simple Lie algebra over k̄
with a non-trivial 5-grading. Then by Theorem 3.3 L ⊗k k̄ is a Lie algebra of Chevalley type.
Then it is algebraic by Theorem 2.13. Hence L is algebraic, since L embeds into L ⊗k k̄. Again
by Theorem 2.13 we conclude that L is of Chevalley type.
4. Applications to structurable algebras and Kantor pairs
Definition 4.1. A structurable algebra over a field k of characteristic not 2 or 3 is a finitedimensional, unital k-algebra with involution (A,¯) such that
(8)
[Vx,y , Vz,w ] = V{x,y,z},w − Vz,{y,x,w}
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
7
for x, y, z, w ∈ A, where the left hand side denotes the Lie bracket of the two operators, and
where
Vx,y z := {x y z} := (xy)z + (zy)x − (zx)y.
For all x, y, z ∈ A, we write Ux,y z := Vx,z y and Ux y := Ux,x y. The trilinear map (x, y, z) 7→
{x y z} is called the triple product of the structurable algebra.
In [All78] and [All79], a structurable algebra is defined as an algebra with involution such
that
(9)
[Tz , Vx,y ] =VTz x,y − Vx,Tz y
for all x, y, z ∈ A with Tx := Vx,1 . The equivalence of (8) and (9) follows from [All79, Corollary
5.(v)].
Definition 4.2. Let (A,¯) be a structurable algebra; then A = H ⊕ S for
H = {h ∈ A | h = h} and S = {s ∈ A | s = −s}.
The elements of H are called hermitian elements, the elements of S are called skew-hermitian
elements or briefly skew elements. The dimension of S is called the skew-dimension of A.
As usual, the commutator and the associator are defined as
[x, y] = xy − yx,
[x, y, z] = (xy)z − x(yz),
for all x, y, z ∈ A. For each s ∈ S, we define the operator Ls : A → A by
Ls x := sx.
The following map is of crucial importance in the study of structurable algebras:
ψ : A × A → S : (x, y) 7→ xy − yx.
Definition 4.3. An ideal of A is a two-sided ideal stabilized by . A structurable algebra
(A, ) is simple if its only ideals are {0} and A, and it is called central if its center
Z(A, ) = Z(A) ∩ H
= {c ∈ A | [c, A] = [c, A, A] = [A, c, A] = [A, A, c] = 0} ∩ H
is equal to k1. The radical of A is the largest solvable ideal of A. A structurable algebra is
semisimple if its radical is zero.
If char(k) 6= 2, 3, 5, a semisimple structurable algebra is the direct sum of simple structurable
algebras [Smi92, Section 2].
Recall the generalization of the Tits–Kantor–Koecher construction that associates to any
structurable algebra A a 5-graded Lie algebra K(A).
Let End(A) be the ring of k-linear maps from A to A. For each A ∈ End(A), we define new
k-linear maps
Aǫ = A − LA(1)+A(1) ,
Aδ = A + RA(1) ,
where Lx and Rx denote left and right multiplication by an element x ∈ A, respectively. Define
the Lie subalgebra Strl(A, ) of End(A) as
(10)
Strl(A, ) = {A ∈ End(A) | [A, Vx,y ] = VAx,y + Vx,Aǫy }.
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
8
(This definition follows from [All78, Corollary 5].) It follows from the definition of structurable
algebras that Vx,y ∈ Strl(A, ), so we can define the Lie subalgebra
Inn(A, ) = Span{Vx,y | x, y ∈ A},
which is, in fact, an ideal of Strl(A, ).
Definition 4.4. Consider two copies A+ and A− of A with corresponding isomorphisms A →
A+ : x 7→ x+ and A → A− : x 7→ x− , and let S+ ⊂ A+ and S− ⊂ A− be the corresponding
subspaces of skew elements. Let
K(A) = S− ⊕ A− ⊕ Inn(A) ⊕ A+ ⊕ S+
as a vector space; as in [All79, §3], we make K(A) into a Lie algebra by extending the Lie
algebra structure of Inn(A) as follows:
• [Inn, K(A)]
[Va,b , Va′ ,b′ ] = V{a,b,a′ },b′ − Va′ ,{b,a,b′ } ∈ Inn(A),
[Va,b , x+ ] := (Va,b x)+ ∈ A+ ,
ǫ
[Va,b , y− ] := (Va,b
y)−
= (−Vb,a y)− ∈ A− ,
δ
[Va,b , s+ ] := (Va,b
s)+
ǫδ
[Va,b , t− ] := (Va,b
t)−
= −ψ(a, sb)+ ∈ S+ ,
= ψ(b, ta)− ∈ S− ,
• [S± , A± ]
[s+ , x+ ] := 0,
[t− , y− ] := 0,
[s+ , y− ] := (sy)+ ∈ A+ ,
[t− , x+ ] := (tx)− ∈ A− ,
• [A± , A± ]
[x+ , y− ] := Vx,y ∈ Inn(A),
[x+ , x′+ ] := ψ(x, x′ )+ ∈ S+ ,
′
[y− , y−
] := ψ(y, y ′)− ∈ S− ,
• [S± , S± ]
[s+ , s′+ ] := 0,
[t− , t′− ] := 0,
[s+ , t− ] := Ls Lt ∈ Inn(A),
for all x, x′ , y, y ′ ∈ A, all s, s′ , t, t′ ∈ S, and all Va,b , Va′ ,b′ ∈ Inn(A).
From the definition of the Lie bracket we clearly see that the Lie algebra K(A) has a 5-grading
given by K(A)j = 0 for all |j| > 2 and
K(A)−2 = S− ,
K(A)−1 = A− ,
K(A)0 = Inn(A),
K(A)1 = A+ ,
K(A)2 = S+ .
In the case where A is a Jordan algebra, we have S = 0, and thus the Lie algebra K(A) has a
3-grading; in this case K(A) is exactly the Tits–Kantor–Koecher construction of a Lie algebra
from a Jordan algebra.
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC 5
9
It is shown in [All79, §5] that the structurable algebra A is simple if and only if K(A) is a
simple Lie algebra, and that A is central if and only if K(A) is central.
Definition 4.5. A structurable algebra A is called algebraic, if K(A) is algebraic.
Definition 4.6. Let A be a structurable algebra over a field k of characteristic 6= 2, 3. An
element x ∈ A is called an absolute zero divisor if Ux y = 0 for any y ∈ A. The algebra A is
called non-degenerate if it has no non-trivial absolute zero divisors.
If an element x ∈ K(A)σ = Aσ is an absolute zero divisor of K(A), then it is represented by
an absolute zero divisor of A; this follows from the fact that by Definition 4.4,
[xσ , [xσ , y−σ ]] = −Vx,y x ∈ Aσ
for all x, y ∈ A.
The following theorem strengthens [BDMSt, Theorem 4.1.1].
Theorem 4.7. Let A be a central simple structurable algebra over a field k of characteristic
different from 2, 3. Then the algebraic k-group G = Aut(K(A))◦ is an adjoint absolutely simple
group of k-rank ≥ 1, and K(A) = [Lie(G), Lie(G)].
Proof. The same claim was established in [BDMSt, Theorem 4.1.1] under the additional assumption that A is algebraic. The Lie algebra K(A) is a central simple Lie algebra with a
non-trivial 5-grading, hence it is a Lie algebra of Chevalley type by Theorem 3.4. Hence it is
algebraic by Theorem 2.13.
Given a non-trivial 5-grading on a Lie algebra of Chevalley type, there may not be a
structurable algebra associated to it. However, we always obtain an associted Kantor pair
by [BDMSt, Lemma 4.3.3].
Definition 4.8 ([AF99]). A Kantor pair is a pair of vector spaces (K+ , K− ) over k equipped
with a trilinear product
{·, ·, ·}: Kσ × K−σ × Kσ → Kσ ,
σ ∈ {−1, 1},
satisfying the following two identities:
(KP1) [Vx,y , Vz,w ] = V{x,y,z},w − Vz,{y,x,w};
(KP2) Ka,b Vx,y + Vy,x Ka,b = KKa,bx,y ;
where Vx,y z := {xyz} and Ka,b z := {azb} − {bza}.
There is a tight connection L
between Kantor pairs and Lie triple systems. In [AF99], a ZTi is called sign-graded, if Ti = 0 for all i 6= ±1. By [AF99,
graded Lie triple system T =
i∈Z
Theorem 7] for any Kantor pair on T = K+ ⊕ K− there is a natural structure of a sign-graded
Lie triple system with two non-zero graded components T1 = K+ and T−1 = K− .
Let T = T1 ⊕ T−1 be a sign-graded Lie triple system over k. In [AF99, p. 532] the 5-graded
2
L
Lie algebra g(T ) =
g(T )i is defined, which is called the standard graded embedding of T .
i=−2
If T = K+ ⊕K− is the Lie triple system corresponding to a Kantor pair (K+ , K− ), the 5-graded
Lie algebra g(K+ ⊕ K− ) is also called the standard graded embedding of this Kantor pair.
According to the following result, every 5-grading on a simple Lie algebra of Chevalley type
over a field of characteristic 6= 2, 3 arises from a Kantor pair. In particular, every isotropic
adjoint simple algberaic group can be constructed from a simple Kantor pair.
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC10
5
Lemma 4.9. Let k be a field of characteristic different from 2, 3. Let G be an adjoint simple
2
L
algebraic group over k. Let L = Lie(G) be its Lie algebra, and let L =
Li be any 5-grading
i=−2
on L such that L1 ⊕ L−1 6= 0. Then (L1 , L−1) is a central simple Kantor pair in the sense of
Definition 4.8 with respect to the triple product operation Lσ × L−σ × Lσ → Lσ given by
{x, y, z} = −[[x, y], z].
Proof. It is clear that the central simple Lie algebra [L, L] envelopes the Kantor pair (L1 , L−1 )
in the sense of [AFS17]. Then by [AFS17, Theorem 3.5.5] (L1 , L−1) is central simple.
Now we can establish the converse as well, namely, that any simple Kantor pair over a field
of characteristic 6= 2, 3 arises from an isotropic simple algebraic group.
Theorem 4.10. Let (K+ , K− ) be a central simple Kantor pair over a field k of characteristic
6 2, 3, and let g(K+ ⊕ K− ) be its standard enveloping Lie algebra as defined in [AFS17]. Then
=
the algebraic k-group G = Aut(g(K+ ⊕ K− ))◦ is an adjoint absolutely simple group of k-rank
≥ 1, and g(K+ ⊕ K− ) = [Lie(G), Lie(G)].
Proof. By [AFS17, Theorem 3.5.5] L = g(K+ ⊕ K− ) is a central simple 5-graded Lie algebra.
Then by Theorem 3.4 L is of Chevalley type. Then by Theorem 2.13 it is algebraic. Then
by [BDMSt, Theorem 4.1.8] G = Aut(L)◦ is an adjoint absolutely simple group of k-rank ≥ 1,
and L = [Lie(G), Lie(G)].
Corollary 4.11. Let (K+ , K− ) be a simple Kantor pair over a commutative ring k such that
2, 3 ∈ k × . Then (K+ , K− ) is non-degenerate in the sense of [GGLN11].
Proof. Since (K+ , K− ) is simple, k does not have non-trivial ideals, and hence is a field. Extending k, we an assume that (K+ , K− ) is central simple. Then by Theorem 4.10 L = g(K+ ⊕ K− )
is a central simple Lie algebra of Chevalley type. Then L ⊗k k̄ is non-degenerate. Then L is
non-degenerate. Then by [GGLN11, Corollary 2.5] (K+ , K− ) is non-degenerate.
References
[All78]
B. N. Allison, A class of nonassociative algebras with involution containing the class of Jordan algebras, Math. Ann. 237 (1978), no. 2, 133–156. MR 507909 (81h:17003)
[All79]
, Models of isotropic simple Lie algebras, Comm. Algebra 7 (1979), no. 17, 1835–1875.
MR 547712 (81d:17005)
, Elementary groups and invertibility for Kantor pairs, Comm. Algebra 27 (1999), no. 2,
[AF99]
519–556. MR 1671930 (99k:17002)
[AFS17] Bruce Allison, John Faulkner, and Oleg Smirnov, Weyl images of Kantor pairs, Canad. J. Math. 69
(2017), no. 4, 721–766. MR 3679693
[BKM15] Yuri Bahturin, Mikhail Kochetov, and Jason McGraw, Gradings by groups on Cartan type Lie algebras, Lie algebras and related topics, Contemp. Math., vol. 652, Amer. Math. Soc., Providence, RI,
2015, pp. 15–46. MR 3453047
[BDMSt] L. Boelaert, T. De Medts, A. Stavrova, Moufang sets and structurable division algebras, 2016, to
appear in Memoirs of the AMS.
[GGLN11] E. García, M. Gómez Lozano, and E. Neher, Nondegeneracy for Lie triple systems and Kantor pairs,
Canad. Math. Bull. 54 (2011), no. 3, 442–455. MR 2857919 (2012m:17007)
[PS08b] A. Premet and H. Strade, Simple Lie algebras of small characteristic. VI. Completion of the classification, J. Algebra 320 (2008), no. 9, 3559–3604. MR 2455517 (2009j:17013)
[Sel67]
G. B. Seligman, Modular Lie algebras, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 40,
Springer-Verlag New York, Inc., New York, 1967. MR 0245627 (39 #6933)
ON THE CLASSIFICATION OF KANTOR PAIRS AND STRUCTURABLE ALGEBRAS IN CHARACTERISTIC11
5
[Smi92]
[Str04]
[Str09]
[Wil75]
[Wil76]
O. Smirnov, Simple and semisimple structurable algebras, Proceedings of the International Conference
on Algebra, Part 2 (Novosibirsk, 1989) (Providence, RI), Contemp. Math., vol. 131, Amer. Math.
Soc., 1992, pp. 685–694. MR 1175866 (93f:17008)
H. Strade, Simple Lie algebras over fields of positive characteristic. I, De Gruyter Expositions in
Mathematics, vol. 38, Walter de Gruyter & Co., Berlin, 2004, Structure theory. MR 2059133
H. Strade, Simple Lie algebras over fields of positive characteristic. II, de Gruyter Expositions in
Mathematics, vol. 42, Walter de Gruyter & Co., Berlin, 2009, Classifying the absolute toral rank two
case. MR 2573283 (2011c:17035)
Robert Lee Wilson, Automorphisms of graded Lie algebras of Cartan type, Comm. Algebra 3 (1975),
no. 7, 591–613. MR 0382374
, A structural characterization of the simple Lie algebras of generalized Cartan type over fields
of prime characteristic, J. Algebra 40 (1976), no. 2, 418–465. MR 0412239 (54 #366)
Chebyshev Laboratory, St. Petersburg State University, 14th Line V.O. 29B, 199178 Saint
Petersburg, Russia
E-mail address: [email protected]
| 4math.GR
|
Joint Attack Detection and Secure State Estimation
of Cyber-Physical Systems
Nicola Forti, Giorgio Battistelli, Luigi Chisci, and Bruno Sinopoli
arXiv:1612.08478v3 [cs.SY] 26 Apr 2017
Abstract
The paper addresses secure state estimation of cyber-physical systems subject to switching signal attacks and
injection of counterfeit measurements. In particular, the random set paradigm is adopted in order to model the
switching nature of the signal attack and the fake measurement injection via Bernoulli and/or Poisson random sets.
The problem of jointly detecting a signal attack and estimating the system state in presence of fake measurements
is then formulated and solved in the Bayesian framework leading to the analytical derivation of a hybrid Bernoulli
filter that updates in real-time the joint posterior density of the detection attack Bernoulli set and of the state vector.
A closed-form Gaussian-mixture implementation of the proposed hybrid Bernoulli filter is fully derived. Finally, the
effectiveness of the developed tools for joint attack detection and secure state estimation is tested on two case-studies
concerning a benchmark system with direct feedthrough and a standard IEEE power network application.
Index Terms
Cyber-physical systems; secure state estimation; signal attack; Bayesian state estimation; Bernoulli filter; extra
packet injection.
I. INTRODUCTION
Cyber-Physical Systems (CPSs) are complex engineered systems arising from the integration of computational
components and physical processes, tightly connected by a communication infrastructure. Typical examples of CPSs
include next-generation systems in electric power grids, transportation and mobility, building and environmental
monitoring/control, health-care, and industrial process control. While on one hand, the progress in CPS technology
will enable growing autonomy, efficiency, seamless interoperability and cooperation, on the other hand the increased
interaction between cyber and physical realms is inevitably introducing novel security vulnerabilities, which render
CPSs subject to non-standard malicious threats. Recent real-world attacks such as the Maroochy Shire sewage spill,
the Stuxnet worm sabotaging an industrial control system, and the lately reported massive power outage against
Ukrainian electric grid [1], have brought into particularly sharp focus the urgency of designing secure CPSs. In
presence of malicious threats against CPSs, standard approaches extensively used for control systems subject to
benign faults and failures are no longer suitable. Moreover, the design and implementation of defense mechanisms
usually employed for cyber security, can only guarantee limited layers of protection, since they do not take into
account vulnerabilities like the ones on physical components. This is why recent advances on the design of secure
systems have explored different paths. Preliminary studies addressed the problems of attack detection/identification,
and proposed attack monitors for deterministic control systems [2]. Secure strategies have been studied for replay
attacks [3],[4] where the adversary first records and then replays the observed data, as well as for denial-of-service
(DoS) attacks [5],[6] disrupting the flow of data. Furthermore, active detection methods have been designed in order
to reveal stealthy attacks via manipulation of e.g. control inputs [7] or dynamics [8]. In recent times, the problem
of secure state estimation, i.e. capable of reconstructing the state even when the CPS of interest is under attack, has
gained considerable attention [9]-[17]. Initial work considered a worst-case approach for the special family of SISO
systems [9]. Under the assumption of linear systems subject to an unknown but bounded number of false-data
injection attacks on sensor outputs, the problem for a noise-free system has been cast into an `0 −optimization
N. Forti, G. Battistelli, and L. Chisci are with the Dipartimento di Ingegneria dell’Informazione, Università degli Studi di Firenze, Via
Santa Marta 3, 50139 Firenze, Italy (e-mail: nicola.forti,giorgio.battistelli,[email protected]).
B. Sinopoli is with the Department of Electrical and Computer Engineering, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh,
PA 15213, USA (e-mail: [email protected]).
1
problem, which can be relaxed as a more efficient convex problem [10], and, in turn, adapted to systems with
bounded noise [11]. Further advances tried to tackle the combinatorial complexity of the problem by resorting to
satisfiability modulo theories [12] and investigated, in the same context, the case of Gaussian measurement noise
[13] and the concept of observability under attacks [14]. Deterministic models of the most popular attack policies
have been presented based on adversary’s resources and system knowledge [15], and resilient strategies have been
also proposed for noisy systems with direct feedthrough under both data injection and switching mode attacks [16].
Most recently, in [17] secure state estimation of CPSs has been addressed by modeling in a stochastic framework the
presumably complex decision-making of the attacker, in particular by assuming Markov (possibly noninformative)
decision processes instead of unknown or worst-case models.
In this work, three different classes of adversarial attacks on CPSs are considered: (i) signal attack, i.e. signal of
arbitrary magnitude and location injected (with known structure) to corrupt sensor/actuator data, (ii) packet substitution attack, describing an intruder that possibly intercepts and then replaces the system-generated measurement
with a fake (unstructured) one, and (iii) extra packet injection, a new type of attack against state estimation, already
introduced in information security (see, e.g., [18],[19]), in which multiple counterfeit observations (junk packets)
are possibly added to the system-generated measurement. Note that the key feature distinguishing signal attacks on
sensors from packet substitution, relies on the fact that the former are assumed to alter the measurement through
a given structure (i.e., known measurement function), whereas the latter mechanism captures integrity attacks that
spoof sensor data packets with no care of the model structure. By considering both structured and unstructured
injections, we do not restrict the type of attack the adversary can enforce on the sensor measurements.
The present paper aims to address the problem of simultaneously detecting a signal attack while estimating the
state of the monitored system, possibly in presence of fake measurements independently injected into the system’s
monitor by cyber-attackers. A random set attack modeling approach is undertaken by representing the signal attack
presence/absence by means of a Bernoulli random set (i.e. a set that, with some probability, can be either empty or
a singleton depending on the presence or not of the attack) and by taking into account possible fake measurements
by means of a random measurement set represented by a Bernoulli or Poisson random set for the packet substitution
or, respectively, extra packet injection attack. The joint attack detection-state estimation problem is then formulated
within the Bayesian framework as the recursive determination of the joint posterior density of the signal attack
Bernoulli set and of the state vector at each time given all the measurement sets available up to that time. Strictly
speaking, the posed Bayesian estimation problem is neither standard [20] nor Bernoulli filtering [21],[22],[23],[24]
but is rather a hybrid Bayesian filtering problem that aims to jointly estimate a Bernoulli random set for the signal
attack and a random vector for the system state. An analytical solution of the hybrid filtering problem has been found
in terms of integral equations that generalize the Bayes and Chapman-Kolmogorov equations of the Bernoulli filter.
In particular, the proposed hybrid Bernoulli Bayesian filter for joint attack detection-state estimation propagates in
time, via a two-step prediction-correction procedure, a joint posterior density completely characterized by a triplet
consisting of: (1) a signal attack probability; (2) a probability density function (PDF) in the state space for the
system under no signal attack; (3) a PDF in the joint attack input-state space for the system under signal attack.
Like standard Bernoulli filters, also the hybrid Bernoulli filter of this paper does not have, in general, a closed-form
solution so that it can be practically implemented as particle or Gaussian-mixture filter.
To the best of the authors’ knowledge, the proposed approach represents the first attempt to exploit stochastic
random set methods in the context of (secure) monitoring of cyber-physical systems. Moreover, unlike most previous
work cited above, in the present paper this problem is addressed from the estimator’s perspective, i.e. without
modeling any strategic interaction between estimator and attacker which, by contrast, characterizes game-theoretic
(worst-case) approaches. As compared to the existing literature, it is also worth to highlight that the adopted
approach brings together, under the same umbrella, the following positive features: 1) can encompass in a unique
and general framework different (and possibly combined) types of attacks (signal attacks, packet substitution, extra
packet injection, temporary DoS, etc.); 2) can deal with nonlinear systems; 3) takes into account the presence of
disturbances and noise; 4) propagates probability distributions of the system state, attack signal and attack existence,
and hence provides useful information for, respectively, real-time dynamic state estimation, attack reconstruction
and security decision-making. Preliminary results on this topic were presented in [25]. The rest of the paper is
organized as follows. Section II introduces the considered attack models and provides the necessary background
on joint input-and-state estimation as well as on random set estimation. Section III formulates and solves the
joint attack detection-state estimation problem of interest in the Bayesian framework. Section IV discusses the
2
Gaussian-mixture implementation of the joint attack detector-state estimator derived in Section III. Then, Section
V demonstrates the effectiveness of the proposed approach via two numerical examples. Finally, Section VI ends
the paper with concluding remarks and perspectives for future work.
II. PROBLEM SETUP AND PRELIMINARIES
A. System description and attack model
Let the discrete-time cyber-physical system of interest be modeled by
( 0
fk (xk ) + wk ,
under no attack
xk+1 =
1
fk (xk , ak ) + wk , under attack
(1)
where: k is the time index; xk ∈ Rn is the state vector to be estimated; ak ∈ Rm , called attack vector, is an
unknown input affecting the system only when it is under attack; fk0 (·) and fk1 (·, ·) are known state transition
functions that describe the system evolution in the no attack and, respectively, attack cases; wk is a random process
disturbance also affecting the system. For monitoring purposes, the state of the above system is observed through
the measurement model
( 0
hk (xk ) + vk ,
under no attack
yk =
(2)
h1k (xk , ak ) + vk , under attack
where: h0k (·) and h1k (·, ·) are known measurement functions that refer to the no attack and, respectively, attack
cases; vk is a random measurement noise. It is assumed that the measurement yk is actually delivered to the system
monitor with probability pd ∈ (0, 1], where the non-unit probability might be due to a number of reasons (e.g.
temporary denial of service, packet loss, sensor inability to detect or sense the system, etc.). The attack modeled
in (1)-(2) via the attack vector ak is usually referred to as signal attack. While for ease of presentation only the
case of a single attack model is taken into account, multiple attack models [16] could be accommodated in the
considered framework by letting (1)-(2) depend on a discrete variable, say νk , which specifies the particular attack
model and has to be estimated together with ak . Besides the system-originated measurement yk in (2), it is assumed
that the system monitor might receive fake measurements from some cyber-attacker. In this respect, the following
two cases will be considered.
1) Packet substitution - With some probability pf ∈ [0, 1), the attacker replaces the system-originated measurement
yk with a fake one ykf .
2) Extra packet injection - The attacker sends to the monitor one or multiple fake measurements indistinguishable
from the system-originated one.
System
Monitor
Physical
Plant
Man-in-the-middle
Fig. 1: Packet substitution attack.
For the subsequent developments, it is convenient to introduce the attack set at time k , Ak , which is either equal
to the empty set if the system is not under signal attack at time k or to the singleton {ak } otherwise, i.e.
∅,
if the system is not under signal attack
Ak =
{ak }, otherwise.
It is also convenient to define the measurement set
∅,
{yk },
Zk =
f
{yk },
at time k , Zk . For the packet substitution attack (Fig. 1):
with probability 1 − pd
with probability pd (1 − pf )
with probability pd pf
3
(3)
System
Monitor
Physical
Plant
Man-in-the-middle
Fig. 2: Extra packet injection attack.
where yk is given by (2) and ykf is a fake measurement provided by the attacker in place of yk . Conversely, for the
extra packet injection attack (Fig. 2) the definition (3) is replaced by
Z k = Yk ∪ F k
where
Yk =
∅,
with probability 1 − pd
{yk }, with probability pd
(4)
(5)
is the set of system-originated measurements and Fk the finite set of fake measurements.
The aim of this paper is to address the problem of joint attack detection and state estimation, which amounts to
4
jointly estimating, at each time k , the state xk and signal attack set Ak given the set of measurements Z k = ∪ki=1 Zi
up to time k .
B. Joint input and state estimation
In this section we review the formulation of the Joint Input and State Estimation (JISE) problem, already addressed
for linear [26],[27],[28] and nonlinear [29] systems in the Bayesian framework. To this end, let us consider a system
with direct feedthrough of the form
xk+1 = f (xk , uk ) + wk
(6)
yk = h(xk , uk ) + vk
where uk is the unknown input vector. The goal of stochastic Bayesian filtering is to recursively estimate the
time-varying posterior PDF of the unknown variables conditioned on all the information available up to that time.
Hence, when the objective is the simultaneous input and state estimation, at each time instant k , the estimates of
uk and xk can be obtained by solving the following problem.
JISE problem: For the system (6), given the measurement set y k = {y1 , y2 , . . . , yk }, sequentially compute the
joint conditional PDF p(uk , xk |y k ) from p(uk−1 , xk−1 |y k−1 ).
Assuming that the initial density p(u0 , x0 ) is given, the solution can be described as a two-step procedure of
prediction and correction. Let p(uk−1 , xk−1 |y k−1 ) denote the posterior PDF at k − 1. The prediction step computes
the conditional PDF p(xk |y k−1 ) via the Chapman-Kolmogorov equation:
ZZ
k−1
p(xk |y ) =
p(xk |uk−1 , xk−1 ) p(uk−1 , xk−1 |y k−1 ) duk−1 dxk−1
(7)
Then, at time instant k , the observed output yk is available and can be used to update p(xk |y k−1 ) and jointly
estimate the conditional PDF of uk , since yk is the first measurement containing information about the unknown
signal. The correction step is then performed by applying the Bayes rule:
p(uk , xk |y k ) =
p(yk |uk , xk ) p(xk |y k−1 ) p(uk )
p(yk |y k−1 )
(8)
Note that in (8) the unknown input is treated as a white stochastic process {uk }, independent of x0 , {wk } and {vk }
[29]. This means that uk and ul are independent random variables for k 6= l, and uk is independent of xk and
y k−1 . With the derived Bayesian solution to JISE in the presence of direct feedthrough, optimal (with respect to
4
any criterion) point estimates of the input and state can be obtained from this PDF, e.g. the Maximum A-posteriori
Probability (MAP) estimate. Notice that, in the case of no direct feedthrough of the unknown input uk to the output
yk in (6), the unknown input estimation will undergo a one-step delay, yk+1 being the first measurement containing
information on uk [26].
C. Random set estimation
An RFS (Random Finite Set) X over X is a random variable taking values in F(X), the collection of all finite
subsets of X. The mathematical background needed for Bayesian random set estimation can be found in [22]; here,
the basic concepts needed for the subsequent developments are briefly reviewed. From a probabilistic viewpoint,
an RFS X is completely characterized by its set density f (X ), also called FISST (FInite Set STatistics) probability
density. In fact, given f (X ), the cardinality probability mass function p(n) that X have n ≥ 0 elements and the
joint PDFs f (x1 , x2 , . . . , xn |n) over Xn given that X have n elements, are obtained as follows:
Z
1
p(n) =
f ({x1 , . . . , xn }) dx1 · · · dxn
n! Xn
1
f (x1 , x2 , . . . , xn |n) =
f ({x1 , . . . , xn }).
n! p(n)
In order to measure probability over subsets of X or compute expectations of random set variables, Mahler [22]
introduced the notion of set integral for a generic real-valued function g(X ) of an RFS X as
Z
Z
∞
X
1
g({x1 , . . . , xn }) dx1 · · · dxn
(9)
g(X ) δX = g(∅) +
n!
n=1
Two specific types of RFSs, i.e. Bernoulli and Poisson RFSs, will be considered in this work.
Bernoulli RFS
A Bernoulli RFS is a random set which can be either empty or, with some probability r ∈ [0, 1], a singleton {x}
whose element is distributed over X according to the PDF p(x). Accordingly, its set density is defined as follows:
(
1 − r,
if X = ∅
(10)
f (X ) =
r · p(x), if X = {x}
Poisson RFS
A Poisson RFS is a random finite set with Poisson-distributed cardinality, i.e.
e−ξ ξ n
, n = 0, 1, 2, . . .
(11)
n!
and elements independently distributed over X according to a given spatial density p(·). Accordingly, its set density
is defined as follows:
Y
f (X ) = e−ξ
ξ p(x).
(12)
p(n) =
x∈X
III. BAYESIAN RANDOM SET FILTER FOR JOINT ATTACK DETECTION AND STATE ESTIMATION
Let the signal attack input at time k be modeled as a Bernoulli random set Ak ∈ B(A), where B(A) = ∅ ∪ S(A)
is a set of all finite subsets of the attack probability space A ⊆ Rm , and S denotes the set of all singletons (i.e., sets
with cardinality 1) {a} such that a ∈ A. Further, let X ⊆ Rn denote the Euclidean space for the system state vector,
then we can define the Hybrid Bernoulli Random Set (HBRS) (A, x), as a new state variable which incorporates
the Bernoulli attack random set A and the random state vector x, taking values in the hybrid space B(A) × X. A
HBRS is fully specified by the (signal attack) probability r of A being a singleton, the PDF p0 (x) defined on the
state space X, and the joint PDF p1 (a, x) defined on the joint attack input-state space A × X, i.e.
(
(1 − r) p0 (x), if A = ∅
p(A, x) =
.
(13)
r · p1 (a, x),
if A = {a}
5
Moreover, since integration over B(A) × X takes the form
Z
Z
ZZ
p(A, x)δA dx = p(∅, x) dx +
p({a}, x) da dx
(14)
F (B)×X
where the set integration with respect to A is defined according to (9) while the integration with respect to x is
an ordinary one, it is easy to see that p(A, x) integrates to one by substituting (13) in (14), and noting that p0 (x)
and p1 (a, x) are conventional probability density functions on X and A × X, respectively. This, in turn, guarantees
that (13) is a FISST probability density for the HBRS (A, x), which will be referred to as hybrid Bernoulli density
throughout the rest of the paper. The notion of attack existence, embodied by parameter r in (13), is introduced so
as to detect the presence (existence) of a signal attack and hence initiate its estimation. Thanks to this concept, as
shown later on, the probability of attack existence can be directly computed from the filter.
A. Measurement models and correction
1) Packet substitution: Let us consider the packet substitution attack model introduced in Section II.A and denote
by λ(Zk |Ak , xk ) the likelihood function of the measurement set defined in (3), which has obviously two possible
forms, Ak being a Bernoulli random set. In particular, for Ak = ∅:
(
1 − pd ,
if Zk = ∅
λ(Zk |∅, xk ) =
(15)
pd (1 − pf ) `(yk |xk ) + pf κ(yk ) , if Zk = {yk }
where {yk } denotes the singleton whose element represents a delivered measurement, i.e. λ({yk }|Ak , xk ) is the
likelihood that a single measurement yk will be collected. Furthermore, `(yk |xk ) is the standard likelihood function
of the system-generated measurement yk when no signal attack is present, whereas κ(·) is a PDF modeling the
fake measurement ykf , assumed to be independent of the system state. Conversely, for Ak = {ak }:
(
1 − pd ,
if Zk = ∅
λ(Zk |{ak }, xk ) =
(16)
pd (1 − pf ) `(yk |ak , xk ) + pf κ(yk ) , if Zk = {yk }
where `(yk |ak , xk ) denotes the conventional likelihood of measurement yk , due to the system under attack ak
in state xk . Notice that, by using the definition of set integral (9), it is easy to check that both forms (15) and
(16) of the likelihood function λ(Zk |Ak , xk ) integrate to one. Using the aforementioned measurement model, it is
possible to derive the exact correction equations of the Bayesian random set filter for joint attack detection and
state estimation, in case of substitution attack.
Theorem 1: Suppose that the prior density at time k is hybrid Bernoulli of the form
(
(1 − rk|k−1 ) p0k|k−1 (xk ), if Ak = ∅
k−1
p(Ak , xk |Z
)=
.
rk|k−1 · p1k|k−1 (ak , xk ), if Ak = {ak }
(17)
Then, given the measurement random set Zk defined in (3), also the posterior density at time k turns out to be
hybrid Bernoulli of the form
(
(1 − rk|k ) p0k|k (xk ), if Ak = ∅
k
p(Ak , xk |Z ) =
(18)
rk|k · p1k|k (ak , xk ), if Ak = {ak }
completely specified by the triplet
rk|k , p0k|k (xk ), p1k|k (ak , xk ) = rk|k−1 , p0k|k−1 (xk ), p1k|k−1 (ak , xk )
if Zk = ∅ or, if Zk = {yk }, by:
(1 − pf ) Ψ1 + pf κ(yk )
r
(1 − pf )(Ψ0 − rk|k−1 Ψ) + pf κ(yk ) k|k−1
(1 − pf ) `(yk |xk ) + pf κ(yk ) 0
pk|k−1 (xk )
p0k|k (xk ) =
(1 − pf ) Ψ0 + pf κ(yk )
(1 − pf ) `(yk |ak , xk ) + pf κ(yk ) 1
p1k|k (ak , xk ) =
pk|k−1 (ak , xk )
(1 − pf ) Ψ1 + pf κ(yk )
rk|k =
(19)
(20)
(21)
6
where
4
Z
`(yk |xk ) p0k|k−1 (xk ) dxk
ZZ
4
=
`(yk |ak , xk ) p1k|k−1 (ak , xk ) dak dxk
Ψ0 =
(22)
Ψ1
(23)
4
Ψ = Ψ0 − Ψ1 .
(24)
Proof: The correction equation of the Bayes random set filter for joint attack detection and state estimation follows
from a generalization of (8), which yields
p(Ak , xk |Z k ) =
λ(Zk |Ak , xk ) p(Ak , xk |Z k−1 )
p(Zk |Z k−1 )
where λ(Zk |Ak , xk ) is given by (15) and (16), while
ZZ
k−1
p(Zk |Z
)=
λ(Zk |Ak , xk ) p(Ak , xk |Z k−1 ) δAk dxk
Z
ZZ
k−1
) dxk +
λ(Zk |{ak }, xk ) p({ak }, xk |Z k−1 ) dak dxk .
= λ(Zk |∅, xk ) p(∅, xk |Z
(25)
(26)
For the case Zk = ∅, the above reduces to
p(∅|Z k−1 ) = 1 − pd
(27)
RR 1
R 0
by substituting (15)-(16) and (17) in (26), and simply noting that pk|k−1 (xk )dxk = 1 and
pk|k−1 (ak , xk ) dak dxk =
1. The posterior probability of attack existence rk|k can be obtained from the posterior density (25) with Ak = ∅
via
Z
rk|k = 1 − p(∅, xk |Z k ) dxk
(28)
where - using (15), (17) and (27) in (25) - we have
p(∅, xk |Z k ) = (1 − rk|k−1 ) p0k|k−1 (xk ).
(29)
Moreover, p0k|k (xk ) = p(∅, xk |Z k )/(1−rk|k ), and the joint density for the system under attack can be easily derived
from the posterior density with Ak = {ak } by recalling that p1k|k (ak , xk ) = p({ak }, xk |Z k )/rk|k , where
p({ak }, xk |Z k ) = rk|k−1 · p1k|k−1 (ak , xk )
(30)
results from replacing R(16), (17) and (27) in
from the set integral definition (9), and densities
RR (25). Notice that
k
k
(29)-(30), it holds that p(∅, xk |Z ) dxk +
p({ak }, xk |Z ) dak dxk = 1. Hence, as stated, the Bayes correction
(18) provides a hybrid Bernoulli density.
Next, for the case Zk = {yk }, (26) leads to
p({yk }|Z k−1 ) = pd (1 − pf )(Ψ1 − rk|k−1 Ψ) + pf κ(yk )
(31)
so that from (25) one gets
(1 − pf ) `(yk |xk ) + pf κ(yk )
p(∅, xk |Z k ) =
(1 − pf )(Ψ1 − rk|k−1 Ψ) + pf κ(yk )
(1 − rk|k−1 ) p0k|k−1 (xk )
(32)
which, in turn, is used to obtain (19) through (28). Once rk|k is known, (20) immediately follows as previously
shown for the case Zk = ∅, while (21) comes from dividing the posterior
(1 − pf ) `(yk |xk ) + pf κ(yk )
p({ak }, xk |Z k ) =
r
p1
(ak , xk )
(33)
(1 − pf )(Ψ1 − rk|k−1 Ψ) + pf κ(yk ) k|k−1 k|k−1
by rk|k in (19).
7
2) Extra packet injection: Let us now consider the extra packet injection attack model introduced in Section
II.A, for which the measurement set defined in (4) is given by the union of two independent random sets. As it
is clear from (5), Yk is a Bernoulli random set (with cardinality |Yk | at most 1) which depends on whether the
system-originated measurement yk is delivered or not. Conversely, Fk is the random set of fake measurements that
will be modeled hereafter as a Poisson random set, such that the number of counterfeit measurements is Poissondistributed according to (11) and the FISST PDF of fake-only measurements γ(·) is given by (12) with spatial
distribution κ(·) in place of p(·). For the measurement set (4), the aim is to find the expression of the likelihood
function λ(Zk |Ak , xk ). To this end, let us first introduce the following FISST PDF for Ak = ∅:
(
1 − pd ,
if Yk = ∅
η(Yk |∅, xk ) =
(34)
pd `(yk |xk ), if Yk = {yk }
and for Ak = {ak }:
(
η(Yk |{ak }, xk ) =
1 − pd ,
if Yk = ∅
pd `(yk |ak , xk ), if Yk = {yk }
.
Then, using the convolution formula [22, p. 385], it follows that
X
λ(Zk |Ak , xk ) =
η(Yk |Ak , xk ) γ(Zk \ Yk ).
(35)
(36)
Yk ⊆Zk
Hence, the likelihood corresponding to Ak = ∅ is given by
X
η({yk }|∅, xk ) γ(Zk \ {yk })
λ(Zk |∅, xk ) = η(∅|∅, xk ) γ(Zk ) +
y ∈Z
k
k
X `(yk |xk )
= γ(Zk ) 1 − pd + pd
ξ κ(yk )
(37)
yk ∈Zk
where (34) and (12) have been used, while for Ak = {ak } we have
X
λ(Zk |{ak }, xk ) = η(∅|{ak }, xk ) γ(Zk ) +
η({yk }|{ak }, xk ) γ(Zk \ {yk })
(38)
yk ∈Zk
X `(yk |ak , xk )
.
= γ(Zk ) 1 − pd + pd
ξ κ(yk )
yk ∈Zk
Notice that the first term on the RHS of (38) accounts for the case of no system-originated measurement, i.e.
Fk = Zk , while the subsequent term in the summation considers the union of disjoint events that each observation
of Zk is system-originated and the rest are fake measurements, i.e. Fk = Zk \ {yk } for any yk ∈ Zk . Using the
above described measurement model, exact correction equations of the Bayesian random set filter for joint attack
detection and state estimation in the case of extra packet injection attack are obtained as follows.
Theorem 2: Assume that the prior density at time k is hybrid Bernoulli of the form (17). Then, given the
measurement random set Zk defined in (4), also the posterior density at time k turns out to be hybrid Bernoulli of
the form (18), with
1 − pd (1 − Γ1 )
r
1 − pd (1 − Γ0 + rk|k−1 Γ) k|k−1
X `(yk |xk )
1 − pd 1 −
ξ κ(yk )
yk ∈Zk
p0k|k (xk ) =
p0k|k−1 (xk )
1 − pd (1 − Γ0 )
X `(yk |ak , xk )
1 − pd 1 −
ξ κ(yk )
yk ∈Zk
p1k|k (ak , xk ) =
p1k|k−1 (ak , xk )
1 − pd (1 − Γ1 )
rk|k =
(39)
(40)
(41)
8
where
4
Γ0 =
R
X
`(yk |xk ) p0k|k−1 (xk ) dxk
ξ κ(yk )
yk ∈Zk
4
Γ1 =
RR
X
`(yk |ak , xk ) p1k|k−1 (ak , xk ) dak dxk
ξ κ(yk )
yk ∈Zk
(42)
(43)
4
and Γ = Γ0 − Γ1 .
Proof: Following the same rationale used for Theorem 1, from (26) it is first possible to derive
k−1
p(Zk |Z
) = γ(Zk ) 1 − pd + pd (1 − rk|k−1 )Γ0 + pd rk|k−1 Γ1
(44)
which is subsequently used together with (29) and (37) to obtain p(∅, xk |Z k ) from (25), and finally (39) via (28).
Then, the PDFs (40) and (41) can be derived from p(∅, xk |Z k ), p({ak }, xk |Z k ) and rk|k , by exploiting the likelihood
functions (37)-(38) of the measurement model under consideration.
From Theorems 1 and 2, it is evident that if pd = 1 and rk|k−1 = 1, then rk|k = 1 follows from both (19) and
(39). Moreover, if we further assume that no fake measurements are collected at time k , i.e. Zk = {yk }, then both
(21) and (41) simplify to the standard Bayes filter correction of the JISE problem (8). In an analogous way, if
rk|k−1 = 0, first we obtain rk|k = 0, then, from (20) and (40), the standard Bayes filter correction for an attack-free
system:
`(yk |xk ) p0k|k−1 (xk )
0
R
.
(45)
pk|k (xk ) =
`(yk |xk ) p0k|k−1 (xk ) dxk
B. Dynamic model and prediction
Let us next introduce the dynamic model of the HBRS (A, x) essential to derive the prediction equations. First,
it is assumed that, in the case of a system under normal operation at time k , an attack ak+1 will be launched to
the system by an adversary during the sampling interval with probability pb . On the other hand, if the system is
under attack (i.e., Ak is a singleton), it is supposed that the adversarial action will endure from time step k to time
step k + 1 with probability ps . It is further assumed that (A, x) is a Markov process with joint transitional density
π(Ak+1 , xk+1 |Ak , xk ) = π(xk+1 |Ak , xk ) π(Ak+1 |Ak )
(46)
which ensues from considering the attack as a stochastic process independent of the system state, as supposed for
the unknown input in Section II.B. Such an assumption is motivated by the fact that (i) ak may assume all possible
values, being completely unknown (we consider the most general model for signal attacks where any value can be
injected via the compromised actuators/sensors), and (ii) the knowledge of ak adds no information on al , if k 6= l.
In addition, note that
(
π(xk+1 |xk ),
if Ak = ∅
π(xk+1 |Ak , xk ) =
(47)
π(xk+1 |ak , xk ), if Ak = {ak }
are known Markov transition PDFs, while the dynamics of the Markov process Ak resulting from the aforestated
assumptions is Bernoulli, described by the following densities:
(
1 − pb ,
if Ak+1 = ∅
π(Ak+1 |∅) =
pb p(ak+1 ), if Ak+1 = {ak+1 }
(
1 − ps ,
if Ak+1 = ∅
π(Ak+1 |{ak }) =
ps p(ak+1 ), if Ak+1 = {ak+1 }
where p(ak+1 ) is the PDF of the attack input vector, assumed known in the Bayesian framework. Clearly, when the
attack vector is completely unknown, a non-informative PDF (e.g., uniform over the attack space) can be adopted for
9
p(ak+1 ), as usually done e.g. for the birth density in random set target tracking [22]. Under the above assumptions,
an exact recursion for the prior density can be obtained, as stated in the following theorem.
Theorem 3: Given the posterior hybrid Bernoulli
density p(Ak , xk |Z k ) at time k of the form (18), fully char
acterized by the triplet rk|k , p0k|k (xk ), p1k|k (ak , xk ) , also the predicted density turns out to be hybrid Bernoulli of
the form
(
(1 − rk+1|k ) p0k+1|k (xk+1 ), if Ak+1 = ∅
k
p(Ak+1 , xk+1 |Z ) =
(48)
rk+1|k · p1k+1|k (ak+1 , xk+1 ), if Ak+1 = {ak+1 }
with
rk+1|k = (1 − rk|k ) pb + rk|k ps
(1 − rk|k )(1 − pb ) pk+1|k (xk+1 |∅)
p0k+1|k (xk+1 ) =
1 − rk+1|k
rk|k (1 − ps ) pk+1|k (xk+1 |{ak })
+
1 − rk+1|k
(1 − rk|k ) pb pk+1|k (xk+1 |∅) p(ak+1 )
p1k+1|k (ak+1 , xk+1 ) =
rk+1|k
rk|k ps pk+1|k (xk+1 |{ak }) p(ak+1 )
+
rk+1|k
(49)
(50)
(51)
where
Z
π(xk+1 |xk ) p0k|k (xk ) dxk
ZZ
pk+1|k (xk+1 |{ak }) =
π(xk+1 |ak , xk ) p1k|k (ak , xk ) dak dxk .
pk+1|k (xk+1 |∅) =
(52)
(53)
Proof: The prediction equation of the Bayes random set filter is given by the following generalization of (7)
ZZ
k
p(Ak+1 , xk+1 |Z ) =
π(Ak+1 , xk+1 |Ak , xk ) p(Ak , xk |Z k ) δAk dxk
(54)
Z
ZZ
= (1 − rk|k ) π(Ak+1 , xk+1 |∅, xk ) p0k|k (xk ) dxk + rk|k
π(Ak+1 , xk+1 |{ak }, xk ) p1k|k (ak , xk ) δAk dxk
where the set integral definition (9) and (18) have been used. Then, we solve for Ak+1 = ∅. From (46), (47), and
(48), one has
Z
k
p(∅, xk+1 |Z ) = (1 − rk|k )(1 − pb ) π(xk+1 |xk ) p0k|k (xk ) dxk
(55)
ZZ
+ rk|k (1 − ps )
π(xk+1 |ak , xk ) p1k|k (ak , xk ) dak dxk .
Next, using (52) and (53), (55) becomes
p(∅, xk+1 |Z k ) = (1 − rk|k ) (1 − pb ) pk+1|k (xk+1 |∅) + rk|k (1 − ps ) pk+1|k (xk+1 |{ak }).
Analogously, for Ak+1 = {ak+1 } we obtain
k
p({ak+1 }, xk+1 |Z ) = (1 − rk|k ) pb pk+1|k (xk+1 |∅) + rk|k ps pk+1|k (xk+1 |{ak }) p(ak+1 ).
(56)
(57)
Thus, the output of the prediction step given by (55)-(57) is of the form (48) under the settings (49)-(53).
It is clear from (49) that the system is predicted to be under attack at time k+1 if either an existing malicious input
persists from time k , or a novel attack ak+1 starts affecting its dynamics. Similar to the standard Bernoulli filter,
the prediction step of the proposed filter involves two separate terms, here accounting for attack-birth and attacksurvival. Notice that, if pb = 0, ps = 1 and rk|k = 1, the prediction step (48) yields (7), which is the standard
Chapman–Kolmogorov equation for the system under attack, since from (49)-(51) it follows that rk+1|k = 1,
p0k+1|k (xk+1 ) = 0, and p1k+1|k (ak+1 , xk+1 ) = pk+1|k (xk+1 |{ak }) p(ak+1 ).
10
IV. G AUSSIAN - MIXTURE IMPLEMENTATION
Although no exact closed-form solution to the Bayes optimal recursion is admitted in general, for the special
class of linear Gaussian models it is possible to analytically propagate in time the posterior densities p0k|k (·) and
p1k|k (·, ·) in the form of Gaussian mixtures (weights, means and covariances), and the probability of signal attack
rk|k . Note that in the case of nonlinear models and/or non-Gaussian noises, the solution can be obtained via
nonlinear extensions of the GM approximation (e.g. Unscented/Extended GM) or sequential Monte Carlo methods
(i.e. particle filter).
Denoting by N (x; m, P ) a Gaussian PDF in the variable x, with mean m and covariance P , the closed-form
solution assumes linear Gaussian observation, transition, and (a priori) attack models, i.e.
`(yk |xk ) = N (y; Cxk , R)
(58)
`(yk |ak , xk ) = N (y; Cxk + Hak , R)
(59)
π(xk+1 |xk ) = N (x; Axk , Q)
(60)
π(xk+1 |ak , xk ) = N (x; Axk + Gak , Q)
Ja
X
p(a) =
ω̃ a,j N (a; ãj , P̃ a,j )
(61)
(62)
j=1
as well as state independent survival and measurement delivery probabilities ps and pd , i.e.
ps (x) = ps
(63)
pd (x) = pd .
(64)
Note that (62) uses given model parameters J a , ω̃ a,j , ãj , P̃ a,j , j = 1, . . . , J a , to define the a priori PDF of the
signal attack, here expressed as a Gaussian mixture and supposed time independent. In the GM implementation,
each probability density at time k is represented by the following set of parameters
1
0
1,j 1,j 1,j Jk|k
0,j 0,j 0,j Jk|k
(65)
, mk|k , Pk|k j=1
, mk|k , Pk|k j=1 , ωk|k
rk|k , p0k|k (xk ), p1k|k (ak , xk ) = rk|k , ωk|k
where ω and J indicate, respectively, weights and the number of mixture components, such that
J0
p0k|k (xk ) =
k|k
X
0,j
0,j
ωk|k
N (m0,j
k|k , Pk|k )
(66)
1,j
1,j
ωk|k
N (m1,j
k|k , Pk|k )
(67)
j=1
J1
p1k|k (ak , xk ) =
k|k
X
j=1
with
m0k|k
=
x̂0k|k ,
m1k|k
=
T
[x̂1k|k , âTk ]T ,
4
0 =
Pk|k
E[(xk −
4
x̂0k|k )(xk
−
x̂0k|k )T ],
1
Pk|k
4
1x
Pk|k Pkxa
4
1x =
=
, and Pk|k
Pkax Pka
E[(xk − x̂1k|k )(xk − x̂1k|k )T ], (Pkxa )T = Pkax = E[(ak − âk )(xk − x̂1k|k )T ], Pka = E[(ak − âk )(ak − âk )T ]. The weights
0
1
PJk|k
PJk|k
0,j
1,j
are such that j=1
ωk|k
= 1, and j=1
ωk|k
= 1.
The Gaussian Mixture implementation of the Hybrid Bernoulli Filter (GM-HBF) is described as follows.
1) GM-HBF correction for packet substitution:
Proposition 1: Suppose assumptions (58)-(64) hold, the measurement set Zk is defined
by (3), the predicted FISST
density at time k is fully specified by the triplet rk|k−1 , p0k|k−1 (xk ), p1k|k−1 (ak , xk ) , and p0k|k−1 (·), p1k|k−1 (·, ·) are
Gaussian mixtures of the form
0
Jk|k−1
p0k|k−1 (xk )
=
0,j
0,j
ωk|k−1
N (m0,j
k|k−1 , Pk|k−1 )
X
(68)
j=1
1
Jk|k−1
p1k|k−1 (ak , xk )
=
X
1,j
1,j
ωk|k−1
N (m1,j
k|k−1 , Pk|k−1 )
j=1
11
(69)
Then, the posterior FISST density rk|k , p0k|k (xk ), p1k|k (ak , xk ) is given by
rk|k
=
(1 − pf ) Ψ1 + pf κ(yk )
rk|k−1
(1 − pf )(Ψ0 − rk|k−1 Ψ) + pf κ(yk )
p0k|k (xk )
=
0,j
0,j
N (m0,j
ωk|k
k|k , Pk|k )
X
=
=
0,j
N (m0,j
k|k−1 , Pk|k−1 )
X
+
1,j
1,j
N (m1,j
ωk|k
k|k , Pk|k )
=
X
(71)
1,j
ωF̄1,j,k|k N (m1,j
k|k , Pk|k )
(72)
1
Jk|k−1
1,j
ωF,k|k
1,j
N (m1,j
k|k−1 , Pk|k−1 )
+
X
j=1
j=1
j=1
0,j
ωF̄0,j,k|k N (m0,j
k|k , Pk|k )
j=1
1
Jk|k−1
1
Jk|k
X
0,j
ωF,k|k
j=1
j=1
p1k|k (ak , xk )
0
Jk|k−1
0
Jk|k−1
0
Jk|k
X
(70)
where
i,j
ωF,k|k
=
ωF̄i,j,k|k =
i,j
pf κ(yk ) ωk|k−1
(1 − pf )Ψi + pf κ(yk )
,
i,j
(1 − pf ) ωk|k−1
(1 − pf )Ψi + pf κ(yk )
(73)
qki,j (yk )
(74)
for i = 0, 1, while
4
with C̃ = [C, H], Ψ0 =
0
PJk|k−1
j=1
0,j
T
qk0,j (yk ) = N (y; Cm0,j
k|k−1 , CPk|k−1 C + R)
(75)
1,j
T
qk1,j (yk ) = N (y; C̃m1,j
k|k−1 , C̃Pk|k−1 C̃ + R)
(76)
0,j
ωk|k−1
qk0,j (yk ), and Ψ1 =
1
PJk|k−1
j=1
1,j
ωk|k−1
qk1,j (yk ).
Proof: From Theorem 1, the corrected probability of signal attack existence can be directly written from (19) as
rk|k =
(1 − pf ) Ψ1 + pf κ(yk )
r
(1 − pf )(Ψ0 − rk|k−1 Ψ) + pf κ(yk ) k|k−1
(77)
where Ψ0 is obtained by substituting (58) and (68) into (22), so that
0
Jk|k−1
Z
N (y; Cxk , R)
Ψ0 =
X
0,j
0,j
N (m0,j
ωk|k−1
k|k−1 , Pk|k−1 ) dxk .
(78)
j=1
Then, by applying a standard result for Gaussian functions [30, Lemma 1], we can write
Z
0,j
0,j
N (y; Cxk , R) N (m0,j
k|k−1 , Pk|k−1 ) dxk = qk (yk )
(79)
where
0,j
T
qk0,j (yk ) = N (y; Cm0,j
k|k−1 , CPk|k−1 C + R)
(80)
and, hence, (78) takes the form
0
Jk|k−1
Ψ0 =
X
0,j
ωk|k−1
qk0,j (yk ).
(81)
j=1
Moreover, Ψ1 in (70) can be analogously
obtained by substituting (59) and (69) into (23), and by applying Lemma
RR
1,j
1 in [30] to the (double) integral
N (y; Cxk + Hak , R) N (m1,j
k|k−1 , Pk|k−1 ) dak dxk , so as to obtain
1
Jk|k−1
Ψ1 =
X
1,j
ωk|k−1
qk1,j (yk )
(82)
j=1
where
1,j
T
qk1,j (yk ) = N (y; C̃m1,j
k|k−1 , C̃Pk|k−1 C̃ + R)
j T T
1
T
and m1,j
k|k−1 = [(x̂k|k−1 ) , (âk ) ] .
12
(83)
Next, the posterior density p0k|k (·) can be derived from (20) in Theorem 1 as
p0k|k (xk ) =
(1 − pf ) `(yk |xk )
pf κ(yk )
p0k|k−1 (xk ) +
p0
(xk ).
(1 − pf ) Ψ0 + pf κ(yk )
(1 − pf ) Ψ0 + pf κ(yk ) k|k−1
(84)
By substituting (58) and (68) into (84), we obtain
0
Jk|k−1
p0k|k (xk )
=
X
0,j
pf κ(yk ) ωk|k−1
j=1
(1 − pf ) Ψ0 + pf κ(yk )
0,j
N (m0,j
k|k−1 , Pk|k−1 )
0
Jk|k−1
+
0,j
X (1 − pf ) ωk|k−1
N (y; Cxk , R)
(1 − pf ) Ψ0 + pf κ(yk )
j=1
0,j
N (m0,j
k|k−1 , Pk|k−1 ).
(85)
Then, by applying Lemma 2 in [30], we can write
0,j
0,j
0,j
0,j
N (y; Cxk , R) N (m0,j
k|k−1 , Pk|k−1 ) = qk (yk ) N (mk|k , Pk|k )
(86)
0,j
where qk0,j (yk ) has been defined in (80), while m0,j
k|k , Pk|k have been introduced in (66).
0,j
In the special case of linear Gaussian models, m0,j
k|k and Pk|k can be easily calculated following the standard
Bayes filter correction step, which in this case boils down to the standard Kalman filter for linear discrete-time
systems [26]:
0,j
0,j
0,j
m0,j
k|k = mk|k−1 + Lk (yk − Cmk|k−1 )
(87)
0,j
Pk|k
(88)
0,j
= (I − L0,j
k C)Pk|k−1
where
0,j
C T (Sk0,j )−1
= Pk|k−1
L0,j
k
Sk0,j
0,j
CT
CPk|k−1
=
(89)
+ R.
(90)
Thus, by substituting (86) into (85) with means and covariances given by (87)-(88), we can write
J0
p0k|k (xk ) =
k|k
X
0,j
0,j
N (m0,j
ωk|k
k|k , Pk|k )
(91)
j=1
0
which comprises 2 Jk|k−1
components, i.e.
0
Jk|k−1
p0k|k (xk ) =
X
0,j
0,j
ωF,k|k
N (m0,j
k|k−1 , Pk|k−1 )
j=1
J
+
0
k|k−1
X
0,j
ωF̄0,j,k|k N (m0,j
k|k , Pk|k )
(92)
j=1
with weights
0,j
ωF,k|k
ωF̄0,j,k|k
=
=
0,j
pf κ(yk ) ωk|k−1
(93)
(1 − pf ) Ψ0 + pf κ(yk )
0,j
(1 − pf ) ωk|k−1
qk0,j (yk )
(1 − pf ) Ψ0 + pf κ(yk )
.
(94)
0 = 2 J0
Note that, as it can be seen from (92), it turns out that Jk|k
k|k−1 , where the first legacy (not corrected)
components correspond to the hypothesis of the system-originated measurement being replaced by a fake one
ykf , while the remaining components are the ones corrected under the hypothesis of receiving yk with probability
(1 − pf ).
13
Following the same rationale, analogous results can be obtained for p1k|k (·, ·), with the exception that also signal
attack estimation has to be performed. By substituting (59) and (69) into (21) in Theorem 1, we obtain
1
Jk|k−1
p1k|k (ak , xk )
=
X
1,j
pf κ(yk ) ωk|k−1
j=1
(1 − pf ) Ψ1 + pf κ(yk )
1,j
N (m1,j
k|k−1 , Pk|k−1 )
1
Jk|k−1
1,j
X (1 − pf ) ωk|k−1
N (y; Cxk + Hak , R)
+
(1 − pf ) Ψ1 + pf κ(yk )
j=1
1,j
N (m1,j
k|k−1 , Pk|k−1 ).
(95)
Then, by applying Lemma 2 in [30], we can write
1,j
1,j
1,j
1,j
N (y; Cxk + Hak , R) N (m1,j
k|k−1 , Pk|k−1 ) = qk (yk ) N (mk|k , Pk|k )
(96)
1,j
where qk1,j (yk ) has been defined in (83), while m1,j
k|k , Pk|k have been introduced in (67). For linear Gaussian models,
1,j
m1,j
k|k and Pk|k can be calculated following the correction step of the filter for joint input and state estimation of
linear discrete-time systems [26], introduced in Section II.B. In particular, m1,j
k|k consists of:
1,j
1,j
1,j
j
x̂1,j
k|k = x̂k|k−1 + L̃k (yk − C x̂k|k−1 − Hâk )
1,j
+ L1,j
x̂k|k−1
k (yk −
j
Mk (yk − C x̂1,j
k|k−1 )
=
âjk
=
(97)
C x̂1,j
k|k−1 )
(98)
where
L1,j
k
L̃1,j
k
Sk1,j
Mkj
j
= L̃1,j
k (I − HMk )
=
=
=
(99)
1x,j
Pk|k−1
C T (Sk1,j )−1
1x,j
CT + R
CPk|k−1
T 1,j −1 −1 T 1,j −1
H (Sk ) H
H (Sk ) .
(100)
(101)
(102)
1,j
The elements composing Pk|k
can be computed as
1x,j
Pk|k
1x,j
= (I − L1,j
k C)Pk|k−1
(103)
Pka,j
= [H T (Sk1,j )−1 H]−1
(104)
Pkxa,j
=
(Pkax,j )T
=
a,j
−L̃1,j
k HPk .
(105)
Thus, by substituting (96) into (95) with means and covariances given by (97)-(98) and (103)-(105), we can write
J1
p1k|k (ak , xk ) =
k|k
X
1,j
1,j
ωk|k
N (m1,j
k|k , Pk|k )
(106)
j=1
which comprises
1
2 Jk|k−1
components, i.e.
1
Jk|k−1
p1k|k (ak , xk ) =
X
1
Jk|k−1
1,j
1,j
ωF,k|k
N (m1,j
k|k−1 , Pk|k−1 ) +
j=1
X
1,j
ωF̄1,j,k|k N (m1,j
k|k , Pk|k )
(107)
j=1
with weights
1,j
ωF,k|k
=
ωF̄1,j,k|k =
1,j
pf κ(yk ) ωk|k−1
(108)
(1 − pf ) Ψ1 + pf κ(yk )
1,j
(1 − pf ) ωk|k−1
qk1,j (yk )
(1 − pf ) Ψ1 + pf κ(yk )
14
.
(109)
2) GM-HBF correction for extra packet injection:
Proposition 2: Suppose assumptions (58)-(64) hold, the measurement set Zk is defined
by (4), the predicted FISST
density at time k is fully specified by the triplet rk|k−1 , p0k|k−1 (xk ), p1k|k−1 (ak , xk ) , and p0k|k−1 (·), p1k|k−1 (·, ·) are
Gaussian mixtures of the form (68) and (69), respectively. Then, the posterior FISST density rk|k , p0k|k (xk ), p1k|k (ak , xk )
is given by
rk|k
=
1 − pd + pd Γ1
rk|k−1
1 − pd + pd (1 − rk|k−1 )Γ0 + pd rk|k−1 Γ1
p1k|k (ak , xk )
=
=
X
0,j
0,j
N (m0,j
ωk|k
k|k , Pk|k )
X
=
j=1
j=1
1
Jk|k
1
Jk|k−1
X
0
Jk|k−1
0
Jk|k−1
0
Jk|k
p0k|k (xk )
(110)
1,j
1,j
ωk|k
N (m1,j
k|k , Pk|k )
X
=
j=1
0,j
ωD̄,k|k
0,j
N (m0,j
k|k−1 , Pk|k−1 )
+
X
X
yk ∈Zk
j=1
0,j
0,j
N (m0,j
ωD,k|k
k|k , Pk|k )(111)
1
Jk|k−1
1,j
ωD̄,k|k
1,j
N (m1,j
k|k−1 , Pk|k−1 )
j=1
+
X
X
yk ∈Zk
j=1
1,j
1,j
ωD,k|k
N (m1,j
k|k , Pk|k )(112)
where
i,j
ωD̄,k|k
=
i,j
=
ωD,k|k
i,j
(1 − pd ) ωk|k−1
1 − pd + pd Γi
i,j
pd ωk|k−1 qki,j (yk )
,
(113)
(114)
(1 − pd + pd Γi ) ξ κ(yk )
for i = 0, 1, while
J0
0,j
k|k−1
X X
ωk|k−1
Γ0 =
yk ∈Zk j=1
ξ κ(yk )
qk0,j (yk )
(115)
qk1,j (yk ).
(116)
J1
1,j
k|k−1
X X
ωk|k−1
Γ1 =
yk ∈Zk j=1
ξ κ(yk )
Proof: We first derive the corrected probability of signal attack existence, which can be directly written from (39)
as
1 − pd + pd Γ1
rk|k =
r
(117)
1 − pd + pd (1 − rk|k−1 )Γ0 + pd rk|k−1 Γ1 k|k−1
where Γ0 is obtained by substituting (58) and (68) into (42), so that
0
Jk|k−1
Z
N (y; Cxk , R)
0,j
0,j
ωk|k−1
N (m0,j
k|k−1 , Pk|k−1 ) dxk
j=1
X
Γ0 =
X
.
ξ κ(yk )
yk ∈Zk
(118)
Then, by applying (79), (118) takes the form (115). Moreover, Γ1 in (117) can be analogously obtained by
substituting (59) and (69) into (43), and by applying (96) which leads to (116).
Next, the posterior density p0k|k (·) can be derived from (40) in Theorem 2 as
p0k|k (xk ) =
X `(yk |xk )
1 − pd
pd
p0k|k−1 (xk ) +
p0
(xk ).
1 − pd + pd Γ0
1 − pd + pd Γ0
ξ κ(yk ) k|k−1
(119)
yk ∈Zk
By substituting (58) and (68) into (119), we obtain
0
Jk|k−1
p0k|k (xk ) =
X
j=1
1 − pd
0,j
ω 0,j N (m0,j
k|k−1 , Pk|k−1 )
1 − pd + pd Γ0 k|k−1
J0
+
k|k−1
X X
yk ∈Zk j=1
0,j
ωk|k−1
pd
N (y; Cxk , R)
0,j
N (m0,j
k|k−1 , Pk|k−1 ).
1 − pd + pd Γ0
ξ κ(yk )
15
(120)
Thus, by substituting (79) into (120), with means and covariances given by (87)-(88), we can write
J0
p0k|k (xk ) =
k|k
X
0,j
0,j
N (m0,j
ωk|k
k|k , Pk|k )
(121)
j=1
0
which comprises Jk|k−1
(1 + |Zk |) components, where |Zk | denotes the cardinality of the measurement set Z at
time k , i.e.
J0
0
Jk|k−1
p0k|k (xk )
=
X
0,j
ωD̄,k|k
0,j
N (m0,j
k|k−1 , Pk|k−1 )
k|k−1
X X
+
0,j
0,j
ωD,k|k
N (m0,j
k|k , Pk|k )
(122)
yk ∈Zk j=1
j=1
with weights
0,j
ωD̄,k|k
0,j
(1 − pd ) ωk|k−1
=
J0
0,h
k|k−1
X X
ωk|k−1
1 − pd + pd
ξ κ(yk )
qk0,h (yk )
yk ∈Zk h=1
0,j
qk0,j (yk )
pd ωk|k−1
0,j
ωD,k|k
=
.
J0
h
1 − pd + pd
0,h
k|k−1
X X
ωk|k−1
yk ∈Zk h=1
ξ κ(yk )
i
qk0,h (yk ) ξ κ(yk )
0 = J0
0
0
Note that, as it can be seen from (122), it turns out that Jk|k
k|k−1 + |Zk | Jk|k−1 = Jk|k−1 (1 + |Zk |), where
the first legacy components correspond to the fact that no measurement has been delivered and hence no update
is carried out, while the remaining components are the ones corrected when one or multiple measurements are
received.
Following the same rationale, analogous results can be obtained for p1k|k (·, ·). From (41) in Theorem 2:
p1k|k (ak , xk ) =
X `(yk |ak , xk )
1 − pd
pd
p1k|k−1 (ak , xk ) +
p1k|k−1 (ak , xk ).
1 − pd + pd Γ1
1 − pd + pd Γ1
ξ κ(yk )
(123)
yk ∈Zk
By substituting (59) and (69) into (123), we obtain
1
Jk|k−1
p1k|k (ak , xk ) =
X
j=1
1 − pd
1,j
1,j
ωk|k−1
N (m1,j
, Pk|k−1
)
k|k−1
1 − pd + pd Γ1
J1
+
k|k−1
X X
1,j
ωk|k−1
yk ∈Zk j=1
pd
N (y; Cxk + Hak , R)
1,j
N (m1,j
k|k−1 , Pk|k−1 ).
1 − pd + pd Γ1
ξ κ(yk )
(124)
Thus, by substituting (96) into (124), with means and covariances given by (97)-(98) and (103)-(105), we can write
J1
p1k|k (ak , xk ) =
k|k
X
1,j
1,j
ωk|k
N (m1,j
k|k , Pk|k )
(125)
j=1
which comprises
1
Jk|k−1
(1
+ |Zk |) components, i.e.
1
Jk|k−1
p1k|k (ak , xk )
=
X
J1
1,j
ωD̄,k|k
1,j
N (m1,j
k|k−1 , Pk|k−1 )
+
k|k−1
X X
yk ∈Zk j=1
j=1
16
1,j
1,j
ωD,k|k
N (m1,j
k|k , Pk|k )
(126)
with weights
1,j
(1 − pd ) ωk|k−1
1,j
ωD̄,k|k
=
J1
1,h
k|k−1
X X
ωk|k−1
1 − pd + pd
ξ κ(yk )
qk1,h (yk )
yk ∈Zk h=1
1,j
pd ωk|k−1
qk1,j (yk )
1,j
ωD,k|k
=
.
J1
h
1 − pd + pd
1,h
k|k−1
X X
ωk|k−1
yk ∈Zk h=1
ξ κ(yk )
i
1,h
qk (yk ) ξ κ(yk )
3) GM-HBF prediction:
Proposition 3: Suppose assumptions (58)-(64) hold, the posterior FISST density at time k is fully specified by
the triplet rk|k , p0k|k (xk ), p1k|k (ak , xk ) , and p0k|k (·), p1k|k (·, ·) are Gaussian mixtures of the form (66)-(67). Then
the predicted FISST density rk+1|k , p0k+1|k (xk+1 ), p1k+1|k (ak+1 , xk+1 ) is given by
rk+1|k = (1 − rk|k ) pb + rk|k ps
(127)
0
Jk+1|k
p0k+1|k (xk+1 ) =
X
0,j
0,j
ωk+1|k
N (m0,j
k+1|k , Pk+1|k )
(128)
1,j
1,j
N (m1,j
ωk+1|k
k+1|k , Pk+1|k )
(129)
j=1
1
Jk+1|k
p1k+1|k (ak+1 , xk+1 ) =
X
j=1
0 + J 1 components, i.e.
0
= Jk|k
where (128) comprises Jk+1|k
k|k
J0
p0k+1|k (xk+1 ) =
k|k
X
J1
0,j
0,j
ωB̄,k+1|k
N (m0,j
, PB̄,k+1|k
)+
B̄,k+1|k
j=1
k|k
X
0,j
0,j
0,j
)
, PS̄,k+1|k
N (mS̄,k+1|k
ωS̄,k+1|k
(130)
j=1
|
{z
}
no attack-birth
|
{z
no attack-survival
}
with
m0,j
= A m0,j
k|k
B̄,k+1|k
(131)
0,j
0,j T
A +Q
PB̄,k+1|k
= APk|k
0,j
ωB̄,k+1|k
=
(132)
(1 − rk|k )(1 − pb ) 0,j
ωk|k
1 − rk+1|k
(133)
m0,j
= Ã m1,j
k|k
S̄,k+1|k
(134)
0,j
1,j T
PS̄,k+1|k
= ÃPk|k
à + Q
0,j
ωS̄,k+1|k
=
(135)
rk|k (1 − ps ) 1,j
ω
1 − rk+1|k k|k
(136)
4
1
0 + J 1 ) components, i.e.
where à = [A, G]. Moreover, (129) comprises Jk+1|k
= J a (Jk|k
k|k
J0
p1k+1|k (ak+1 , xk+1 )
=
J1
a
k|k J
X
X
1,jh
ωB,k+1|k
1,jh
N (m1,jh
B,k+1|k , PB,k+1|k ) +
j=1 h=1
|
a
k|k J
X
X
1,jh
1,jh
ωS,k+1|k
N (m1,jh
S,k+1|k , PS,k+1|k ) (137)
j=1 h=1
{z
}
attack-birth
17
|
{z
attack-survival
}
where
"
m1,jh
B,k+1|k
=
"
A m0,j
k|k
ãh
#
(138)
=
0,j T
A +Q
0
APk|k
0
P̃ a,h
1,jh
ωB,k+1|k
=
(1 − rk|k ) pb 0,j a,h
ωk|k ω̃
rk+1|k
1,jh
PB,k+1|k
#
(139)
(140)
and
"
m1,jh
S,k+1|k
=
à m1,j
k|k
ãh
#
(141)
"
1,jh
PS,k+1|k
=
1,jh
ωS,k+1|k
=
1,j T
à + Q
0
ÃPk|k
0
P̃ a,h
rk|k ps 1,j a,h
ω ω̃ .
rk+1|k k|k
#
(142)
(143)
Proof: The predicted signal attack probability comes directly from (49). Let us now derive the predicted density
p0k|k (·). From (50) in Theorem 3:
Z
(1 − rk|k ) (1 − pb )
π(xk+1 |xk ), p0k|k (xk ) dxk
p0k+1|k (xk+1 ) =
1 − rk+1|k
ZZ
rk|k (1 − ps )
+
π(xk+1 |ak , xk ), p1k|k (ak , xk ) dak dxk .
(144)
1 − rk+1|k
Using (60), (66) in the first term and (61), (67) in the second term, we can rewrite
(1 − rk|k ) (1 − pb )
p0k+1|k (xk+1 ) =
1 − rk+1|k
rk|k (1 − ps )
+
1 − rk+1|k
J0
Z
N (x; Axk , Q)
k|k
X
0,j
0,j
ωk|k
N (m0,j
k|k , Pk|k ) dxk
j=1
J1
ZZ
N (x; Axk + Gak , Q)
k|k
X
1,j
1,j
ωk|k
N (m1,j
k|k , Pk|k ) dak dxk .
(145)
j=1
Hence, using Lemma 1 from [30] in both the above terms, we finally derive (130):
J0
p0k+1|k (xk+1 ) =
k|k
X
(1 − rk|k ) (1 − pb )
j=1
1 − rk+1|k
0,j
0,j T
ωk|k
N (x; Am0,j
k|k , APk|k A + Q)
J1
+
k|k
X
rk|k (1 − ps )
j=1
1 − rk+1|k
1,j
1,j T
ωk|k
N (x; Ãm1,j
k|k , ÃPk|k à + Q).
In a similar fashion, we can obtain p1k+1|k (·, ·). From (51) in Theorem 3:
Z
(1 − rk|k ) pb
1
pk+1|k (ak+1 , xk+1 ) =
π(xk+1 |xk ), p0k|k (xk ) dxk p(a)
rk+1|k
ZZ
rk|k ps
+
π(xk+1 |ak , xk ), p1k|k (ak , xk ) dak dxk p(a)
rk+1|k
18
(146)
(147)
which, using (60), (61), (62), (66) and (67), leads to
(1 − rk|k ) pb
p1k+1|k (ak+1 , xk+1 ) =
rk+1|k
rk|k ps
+
rk+1|k
J0
Z
N (x; Axk , Q)
k|k
X
a
0,j
ωk|k
0,j
N (m0,j
k|k , Pk|k ) dxk
j=1
1
k|k
J
ZZ
N (x; Axk + Gak , Q)
X
J
X
ω̃ a,h N (a; ãh , P̃ a,h )
h=1
a
1,j
ωk|k
1,j
N (m1,j
k|k , Pk|k ) dak dxk
j=1
J
X
ω̃ a,h N (a; ãh , P̃ a,h ).
(148)
h=1
Finally, by applying the same result on integrals of Gaussians used above, we obtain (137):
J0
p1k+1|k (ak+1 , xk+1 )
=
a
k|k J
X
X (1 − rk|k ) pb
j=1 h=1
1
k|k
J
+
rk+1|k
0,j T
0,j a,h
h
a,h
)
ω̃ N (x; Am0,j
ωk|k
k|k , APk|k A + Q) N (a; ã , P̃
a
J
XX
rk|k ps
j=1 h=1
rk+1|k
1,j T
1,j a,h
h
a,h
).
ω̃ N (x; Ãm1,j
ωk|k
k|k , ÃPk|k à + Q) N (a; ã , P̃
It is worth pointing out that, likewise other GM filters (e.g. [30]), also the proposed Gaussian Mixture Hybrid
Bernoulli Filter turns out to be affected by computational issues as the number of Gaussian components increases
with no bound over time. As already noticed in the above derivation, at time k the GM-HBF requires
(
0
2 Jk|k−1
,
packet substitution
0
Jk|k =
0
Jk|k−1
(1 + |Zk |), extra packet injection
(
1
2 Jk|k−1
,
packet substitution
1
Jk|k =
1
(1 + |Zk |), extra packet injection
Jk|k−1
components to exactly represent the posterior densities p0k|k (·) and p1k|k (·, ·), respectively. Here
0
0
1
Jk|k−1
= Jk−1|k−1
+ Jk−1|k−1
,
1
0
1
Jk|k−1
= J a (Jk−1|k−1
+ Jk−1|k−1
)
denote the number of components generated in the prediction step. In order to reduce the growing amount of
GM components, simple heuristic pruning and merging procedures presented in [30] can be directly performed at
each time step so as to remove low-weight components and combine statistically close components. For a better
understanding of the GM-HBF recursion presented in Propositions 2 and 3, detailed pseudocodes for both correction
(extra packet injection) and prediction steps are provided in Table I and, respectively, Table II. From Tables I and
II, it is clear that if the initial prior densities p01|0 (·), p11|0 (·, ·) are Gaussian mixtures, then all the following corrected
and predicted densities will be Gaussian mixtures as well.
19
TABLE I: Pseudocode for the GM-HBF correction step in the case of extra packet injection.
0
1
0,j
1,j
Jk|k−1
Jk|k−1
0,j
1,j
1: function C ORRECTION -GM-HBF-EPI( rk|k−1 , ωk|k−1
, m0,j
, ωk|k−1
, m1,j
, Zk , C, H, R, pd )
k|k−1 , Pk|k−1 j=1
k|k−1 , Pk|k−1 j=1
2: Construction of correction components:
0
3:
for j = 1, . . . , Jk|k−1
do
0,j
0,j
0,j
0,j −1
T
4:
ηk|k−1 = Cm0,j
Sk0,j = CPk|k−1
C T + R, L0,j
k = Pk|k−1 C (Sk )
k|k−1 ,
5:
end for
1
6:
for j = 1, . . . , Jk|k−1
do
−1 T 1,j −1
1,j
1x,j
1x,j
1,j −1
T
H (Sk )
7:
ηk|k−1 = C x̂1,j
Sk1,j = CPk|k−1
C T + R, L̃1,j
, Mkj = H T (Sk1,j )−1 H
k = Pk|k−1 C (Sk )
k|k−1 ,
xa,j
ax,j T
1,j
a,j
1,j
j
a,j
1,j −1
T
−1
8:
L1,j
=
L̃
(I
−
HM
),
P
=
[H
(S
)
H]
,
P
=
(P
)
=
−
L̃
HP
k
k
k
k
k
k
k
k
k
9:
end for
10: Correction step:
0
11:
for j = 1, . . . , Jk|k−1
do
0,j
0,j
0,j
0,j
0,j
0,j
12:
mk|k = mk|k−1 , Pk|k
= Pk|k−1
, ωk|k
= ωk|k−1
13:
end for
1
14:
for j = 1, . . . , Jk|k−1
do
1,j
1,j
1,j
1,j
1,j
1,j
15:
mk|k = mk|k−1 , Pk|k
= Pk|k−1
, ωk|k
= ωk|k−1
16:
end for
17:
`=0
18:
for each yk ∈ Zk do
19:
`=`+1
0
do
20:
for j = 1, . . . , Jk|k−1
21:
22:
23:
24:
25:
26:
0
0,`Jk|k−1
+j
0
0,`Jk|k−1
+j
Pk|k
0
0,`Jk|k−1
+j
ωk|k
Pk|k
32:
33:
34:
35:
36:
37:
38:
39:
40:
0,`J 0
+j
k|k−1
ωk|k
1
1,`Jk|k−1
+j
28:
31:
0,j
0,j
= ωk|k−1
N (y; ηk|k−1
, Sk0,j )
νk|k
=
ξκ(yk )
end for
1
for j = 1, . . . , Jk|k−1
do
mk|k
30:
0,j
= (I − L0,j
k C)Pk|k−1
0
0,`Jk|k−1
+j
27:
29:
0,j
0,j
= m0,j
k|k−1 + Lk (yk − ηk|k−1 )
mk|k
1
1x,`Jk|k−1
+j
1
a,`Jk|k−1
+j
Pk
"
#
1
1,`Jk|k−1
+j
1,j
x̂1,j
+ L1,j
(yk − ηk|k−1
)
x̂
k
k|k−1
k|k
=
= `J 1
1,j
+j
Mkj (yk − ηk|k−1
)
âk k|k−1
= [H T (Sk1,j )−1 H]−1
1
xa,`Jk|k−1
+j
Pk
1
1,`Jk|k−1
+j
ωk|k
1,`J 1
1x,j
= (I − L1,j
k C)Pk|k−1
a,j
= −L̃1,j
k HPk
1
`Jk|k−1
+j
1,j
1,j
= ωk|k−1
N (y; ηk|k−1
+ Hâk
, Sk1,j + HPka,j H T )
1,`J 1
+j
k|k−1
ωk|k
+j
νk|k k|k−1 =
ξκ(yk )
end for
end for
0,iJ 0
+h
k|k−1
0
1
νk|k−1
P
PJk|k−1
P
PJk|k−1
Γ0 = i∈Zk h=1
,
Γ
=
1
h=1
i∈Z
ξ κ(yk )
k
1−pd +pd Γ1
rk|k = 1−pd +pd (1−r
rk|k−1
k|k−1 )Γ0 +pd rk|k−1 Γ1
`=0
for each yk ∈ Zk do
`=`+1
0
for j = 1, . . . , Jk|k−1
do
0,j
41:
42:
43:
44:
0,j
ωk|k
=
(1−pd ) ωk|k
1−pd +pd Γ0
0
0,`Jk|k
+j
ωk|k
=
0,`J 0 +j
k|k
pd ωk|k
1−pd +pd Γ0 ξκ(yk )
end for
1
for j = 1, . . . , Jk|k−1
do
1,j
45:
46:
1,j
ωk|k
=
(1−pd ) ωk|k
1−pd +pd Γ1
1
1,`Jk|k
+j
ωk|k
=
1,`J 1 +j
k|k
pd ωk|k
1−pd +pd Γ1 ξκ(yk )
47:
end for
48:
end for
0
0
0
49:
Jk|k
= Jk|k−1
+ ` Jk|k−1
1
1
1
50:
Jk|k = Jk|k−1 + ` Jk|k−1
51: end function
20
1,iJ 1
+h
k|k−1
νk|k−1
ξ κ(yk )
TABLE II: Pseudocode for the GM-HBF prediction step.
0,j
0,j
1: function P REDICTION -GM-HBF( rk|k , ωk|k
, m0,j
k|k , Pk|k
2:
rk+1|k = (1 − rk|k ) pb + rk|k ps
3:
i=0
4:
`=0
0
5:
for j = 1, . . . , Jk|k
do
6:
i=i+1
0,j
7:
m0,i
k+1|k = A mk|k
0,i
0,j T
8:
Pk+1|k = APk|k A + Q
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:
(1−r
0
Jk|k
j=1
1,j
1,j
, ωk|k
, m1,j
k|k , Pk|k
j=1
, ω̃ a,j , ãj , P̃ a,j
Ja
, A, G, Q, pb , ps
j=1
)
. no attack-birth
)(1−p )
b
0,j
0,i
k|k
ωk|k
ωk+1|k
=
1−rk+1|k
a
for h = 1, . . . , J do
`=`+1 "
#
A m0,j
x̂1,`
1,`
k|k
k+1|k
mk+1|k =
=
h
`
ã
âk+1
1x,`
0,j T
Pk+1|k = APk|k A + Q
"
#
1x,`
Pk+1|k
0
1,`
Pk+1|k =
0
P̃ a,h
(1−r
. attack-birth
)p
1,`
0,j
ωk+1|k
= r k|k b ωk|k
ω̃ a,h
k+1|k
end for
end for
1
do
for j = 1, . . . , Jk|k
i=i+1
1,j
j
m0,i
k+1|k = A x̂k|k +"Gâk
#
1x,j
Pk|k
Pkxa,j AT
0,i
A
G
Pk+1|k =
+Q
GT
(Pkxa,j )T
Pka,j
r
1
Jk|k
. no attack-survival
(1−ps )
1,j
0,i
k|k
ωk|k
ωk+1|k
= 1−r
k+1|k
a
for h = 1, . . . , J do
`=`+1 "
#
j
A x̂1,j
x̂1,`
1,`
k|k + Gâk
k+1|k
mk+1|k =
=
h
`
ã
âk+1
"
#
1x,j
Pk|k
Pkxa,j AT
1x,`
Pk+1|k
+Q
= A G
GT
(Pkxa,j )T
Pka,j
"
#
1x,`
Pk+1|k
0
1,`
Pk+1|k
=
0
P̃ a,h
r
ps
1,j
1,`
ω̃ a,h
ωk+1|k
= rk|k ωk|k
k+1|k
end for
end for
0
Jk+1|k
=i
1
Jk+1|k = `
end function
. attack-survival
V. N UMERICAL EXAMPLES
The effectiveness of the developed tools, based on Bayesian random-set theory, for joint attack detection and
secure state estimation of cyber-physical systems has been tested on two numerical examples concerning a benchmark linear dynamical system and a standard IEEE power network case-study. Simulations have been carried out
in the presence of both signal and extra packet injection attacks as well as uncertainty on measurement delivery.
A. Benchmark linear system
Let us first consider the following benchmark linear system, already used in the JISE literature [16], [28]:
xk+1 = Axk + Gak + wk
yk = Cxk + Hak + vk
(149)
where A, C , R, and Q are the same as [16], while G = [e1 , e2 ] and H = [e3 , e1 ], where e1 , . . . , en denote the
canonical basis vectors.
21
For this numerical study, the probabilities of attack-birth and attack-survival are fixed, respectively, at pb = 0.2
and ps = 0.8. The system-generated measurement is supposed to be delivered at the monitor/control center with
probability pd = 0.98, while the initial signal attack probability is set to r1|0 = 0.1. The initial state has been set
equal to x0 = 0, whereas both densities p0 (·) and p1 (·, ·) have been initialized as single Gaussian components with
0 = 104 I . Moreover, the first estimate of the attack
first guess mean x̂01|0 = [10, 10, 0, 0, 0]T and covariance P1|0
a = 50I .
vector has been randomly initialized as ã1|0 = [15.1, 25.53]T , with associated initial covariance matrix P̃1|0
The extra fake measurements are modeled as a Poisson RFS uniformly distributed over the interval [−0.3, 140.3],
with average number ξ = 0.5. Finally, a pruning threshold γp = 10−3 and a merging threshold γm = 3 have been
chosen. As shown in Fig.3, at time k = 150 a signal attack vector a = [10, 20]T is injected into the system, persisting
for 200 time steps. The proposed GM-HBF promptly detects the unknown signal attack, by simply comparing the
attack probability rk|k obtained in (39) with the threshold 0.5. Fig. 4 provides a comparison between the true and
the estimated values of states x1 and x2 (clearly the only state components affected by the signal attack). Note
that the state estimate is obtained by means of a MAP estimator, i.e. by extracting the Gaussian mean with the
highest weight from the posterior density p0 (·) (40) or p1 (·, ·) (41), accordingly with the current value of the attack
probability.
true r k
rk|k
1
estimated rk|k
0.5
0
0
50
100
150
200
250
Time step
300
350
400
450
500
Fig. 3: True and estimated attack probability.
150
true x 1
x1
100
estimated x 1
50
0
-50
0
50
100
150
200
250
Time step
300
350
400
450
500
30
true x2
x2
20
estimated x2
10
0
-10
0
50
100
150
200
250
Time step
300
350
400
450
500
Fig. 4: True and estimated state components x1 and x2 .
Finally, Fig. 5 shows how the attack estimates extracted from p(a) of the two components of the attack vector,
coincide with the actual values inside the attack time interval [150, 350]. Note that outside that interval the estimates
of the attack vector are not meaningful because the attack probability rk|k is almost 0.
B. Energy delivery system
State estimation is of paramount importance to ensure the reliable operation of energy delivery systems since
it provides estimates of the power grid state by processing meter measurements and exploiting power system
models. Cyber attacks on energy delivery systems can alter available information at the control center and generate
fake meter and input data, potentially causing power outage and forcing the energy management system to make
erroneous decisions, e.g. on contingency analysis and economic dispatch. The proposed GM-HBF was tested on
the IEEE 14-bus system (Fig. 6) consisting of 5 synchronous generators and 11 load buses, with parameters taken
from [31]. The dynamics of the system can be described by the linearized swing equation [32] derived through
the Kron reduction [33] of the linear small-signal power network model. The DC state estimation model assumes
1 p.u. (per unit) voltage magnitudes in all buses and j1 p.u. branch impedance, with j denoting imaginary unit.
The system dynamics is represented by the evolution of n = 10 states comprising both the rotor angles δj and
22
20
true a 1
estimated a1
a1
15
10
5
0
0
50
100
150
200
250
Time step
300
350
400
450
500
30
a2
20
true a 2
10
estimated a2
0
0
50
100
150
200
250
Time step
300
350
400
450
500
Fig. 5: True and estimated attack components a1 and a2 .
Fake/undelivered measurements
Fig. 6: Single-line model of the IEEE 14-bus system. The true victim load buses 3 and 9 are circled in red.
20
15
10
5
0
0
10
20
30
40
50
60
70
80
90
100
Time step
Fig. 7: Number of extra fake measurements (ξ = 10) injected (blue circles) and undelivered (pd = 0.95) systemoriginated observations (red cross in −1) vs time. The proposed GM-HB filter turns out to be particularly robust
to extra packet injections.
the frequencies ωj of each generator j in the network. After discretization (with sampling interval T = 0.01s), the
model of the system takes the form (1)-(2), where the whole state is measured by a network Si of sensors. The
system is assumed to be corrupted by additive zero mean Gaussian white process and measurement noises with
2 = 0.01 and σ 2 = 0.01. At time k = 50 a signal attack vector a = [0.2, 0.1]T p.u. is injected into
variances σw
v
the system to abruptly increase the real power demand of the two victim load buses 3 and 9 with an additional
loading of 21.23% and, respectively, 33.9%. This type of attack, referred to as load altering attack in [34], can
provoke a loss of synchrony of the rotor angles and hence a deviation of the rotor speeds of all generators from
23
the nominal value. In addition, we fixed the following parameters: pb = 0.05, ps = 0.95, pd = 0.95, pruning and
merging thresholds γp = 10−2 and γm = 3 for the Gaussian-mixture implementation. The extra fake measurements
injected into the sensor channels are modeled as a Poisson RFS with average number ξ = 10 and probability
density uniformly distributed over the interval [−10, 5], suitably chosen to emulate system-originated observations.
Fake and missed packets are shown in Fig. 7 for a specific run. The joint attack detection and state estimation
performance of the GM-HBF algorithm has been analyzed by Monte Carlo simulations. Fig. 8 shows the true and
estimated probability of attack existence (a) and the Root Mean Square Error (RMSE), averaged over 1000 Monte
Carlo runs, relative to the rotor angle (b) and frequency (c) estimates. Fig. 8 (d) shows the RMSE of the estimated
components of the signal attack, extracted from p1k|k (a, x). As shown in the results (a)-(d), the proposed secure
state estimator succeeds in promptly detecting a signal attack altering the nominal energy delivery system behavior,
and hence in being simultaneously resilient to integrity attacks on power demand, and robust to extra fake packets
and undelivered measurements.
(a)
rk|k
1
0.5
true rk
estimated r k|k
0
0
10
20
30
40
10
20
30
40
10
20
30
40
δ RMSE
1
60
70
80
90
100
50
60
70
80
90
100
50
60
70
80
90
100
0.5
0
0
1
ω RMSE
50
(b)
Time
step
(c)step
Time
0.5
a RMSE
0
0
1
Time(d)
step
a1
a2
0.5
0
0
10
20
30
40
50
60
70
80
90
100
Time step
Fig. 8: Performance of the GM-HB filter in terms of joint attack detection (a) and estimation of rotor angles
δi , i = 1, . . . , 5 (b), frequencies ωi , i = 1, . . . , 5 (c), and signal attack (d).
Fig. 9 provides, for a single Monte Carlo trial, a comparison between the true and the estimated values of the two
rotor angles mainly affected by the victim load buses, and clearly shows how δ1 and δ3 lose synchrony once the
load altering attack enters into action. Nevertheless, the proposed secure filter keeps tracking the state evolution
with high accuracy even after time k = 50, once recognized that the system is under attack.
Finally, Fig. 10 shows the performance of the GM-HBF in estimating the generator frequencies ω1 and ω3 , before
and after the appearance of the signal attack on the victim loads.
VI. C ONCLUSIONS
This paper proposed a general framework to solve resilient state estimation for (linear/nonlinear) cyber-physical
systems considering both switching signal attacks and fake measurement injection. Random (Bernoulli and/or
Poisson) finite sets have been exploited in order to model the switching nature of the signal attack as well as
the possible presence of fake measurements, and a Bayesian random set estimation problem has been formulated
for jointly detecting a signal attack and estimating the system state. In this way, a hybrid Bernoulli filter for the
Bayes-optimal solution of the posed problem has been derived and implemented as a Gaussian-sum filter. Numerical
examples concerning both a benchmark system with direct feedthrough and a realistic energy delivery system have
24
2.5
2
Rotor angle [rad]
1.5
1
0.5
true δ1
estimated δ01
estimated δ11
true δ3
estimated δ03
estimated δ13
+- π/2
0
-0.5
-1
-1.5
-2
0
10
20
30
40
50
Time step
60
70
80
90
100
Fig. 9: Estimated vs true trajectory of rotor angles δj , j = 1, 3. Note that, if |δj | is sufficiently large (values close
to π/2), the linear small signal approximation significantly deviates from the nonlinear dynamics of the system,
and hence the assumed dynamic model becomes inaccurate.
1.2
1
Frequency [rad s-1]
0.8
0.6
0.4
true ω1
estimated ω01
estimated ω11
true ω3
estimated ω03
estimated ω13
0.2
0
-0.2
-0.4
-0.6
-0.8
0
10
20
30
40
50
Time step
60
70
80
90
100
Fig. 10: Estimated vs true trajectory of frequencies ω1 and ω3 .
been presented so as to demonstrate the potentials and the real-world applicability of the proposed approach. Future
work will concern worst-case performance degradation analysis for the developed filter and its application to resilient
state estimation in distributed settings.
R EFERENCES
[1] “The Industrial Control Systems Cyber Emergency Response Team (ICS-CERT).” https://ics-cert.us-cert.gov/.
[2] F. Pasqualetti, F. Dörfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE Transactions on Automatic
Control, vol. 58, no. 11, pp. 2715–2729, 2013.
[3] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in Proc. 47th Allerton Conference on Communication, Control, and
Computing, pp. 911–918, 2009.
[4] F. Miao, M. Pajic, and G. J. Pappas, “Stochastic game approach for replay attack detection,” in Proc. 52nd IEEE Conference on
Decision and Control, pp. 1854–1859, 2013.
[5] C. D. Persis and P. Tesi, “Input-to-state stabilizing control under denial-of-service,” IEEE Transactions on Automatic Control, vol. 60,
no. 11, pp. 2930–2944, 2015.
[6] H. Zhang, P. Cheng, L. Shi, and J. Chen, “Optimal denial-of-service attack scheduling with energy constraint,” IEEE Transactions on
Automatic Control, vol. 60, no. 11, pp. 3023–3028, 2015.
[7] Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication of control systems: Designing watermarked control inputs to detect
counterfeit sensor outputs,” IEEE Control Systems Magazine, vol. 35, no. 1, pp. 93–109, 2015.
[8] S. Weerakkody and B. Sinopoli, “Detecting integrity attacks on control systems using a moving target approach,” in Proc. 54th IEEE
Conference on Decision and Control, pp. 5820–5826, Osaka, Japan, 2015.
[9] Y. Mo and B. Sinopoli, “Secure estimation in the presence of integrity attacks,” IEEE Transactions on Automatic Control, vol. 60,
no. 4, pp. 1145–1151, 2015.
25
[10] H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and control for cyber-physical systems under adversarial attacks,” IEEE
Transactions on Automatic Control, vol. 59, no. 6, pp. 1454–1467, 2014.
[11] M. Pajic, I. Lee, and G. J. Pappas, “Attack-resilient state estimation for noisy dynamical systems,” IEEE Transactions on Control of
Network Systems, vol. 4, no. 1, pp. 82–92, 2017.
[12] Y. Shoukry, A. Puggelli, P. Nuzzo, A. L. Sangiovanni-Vincentelli, S. A. Seshia, and P. Tabuada, “Sound and complete state estimation
for linear dynamical systems under sensor attacks using satisfiability modulo theory solving,” in Proc. American Control Conference,
pp. 3818–3823, Chicago, Illinois, USA, 2015.
[13] S. Mishra, Y. Shoukry, N. Karamchandani, S. N. Diggavi, and P. Tabuada, “Secure state estimation against sensor attacks in the presence
of noise,” IEEE Transactions on Control of Network Systems, vol. 4, no. 1, pp. 49–59, 2017.
[14] M. S. Chong, M. Wakaiki, and J. P. Hespanha, “Observability of linear systems under adversarial attacks,” in Proc. American Control
Conference, pp. 2439–2444, 2015.
[15] A. Teixeira, I. Shames, H. Sandberg, and K. H. Johansson, “A secure control framework for resource-limited adversaries,” Automatica,
vol. 51, no. 1, pp. 135–148, 2015.
[16] S. Yong, M. Zhu, and E. Frazzoli, “Resilient state estimation against switching attacks on stochastic cyber-physical systems,” in Proc.
54th IEEE Conference on Decision and Control, pp. 5162–5169, Osaka, Japan, 2015.
[17] D. Shi, R. J. Elliott, and T. Chen, “On finite-state stochastic modeling and secure estimation of cyber-physical systems,” IEEE
Transactions on Automatic Control, vol. 62, no. 1, pp. 65–80, 2017.
[18] Q. Gu, P. Liu, S. Zhu, and C.-H. Chu, “Defending against packet injection attacks in unreliable ad hoc networks,” in Proc. IEEE Global
Telecommunications Conference, vol. 3, pp. 1837–1841, St. Louis, Missouri, USA, 2005.
[19] X. Zhang, H. Chan, A. Jain, and A. Perrig, “Bounding packet dropping and injection attacks in sensor networks,” Tech. Rep. 07-019,
CMU-CyLab, Pittsburgh, PA, USA, 2007. Available online at: https://www.cylab.cmu.edu/files/pdfs/tech reports/cmucylab07019.pdf.
[20] Y. Ho and R. Lee, “A Bayesian approach to problems in stochastic estimation and control,” IEEE Transactions on Automatic Control,
vol. 9, no. 4, pp. 333–339, 1964.
[21] B. Ristic, B.-T. Vo, B.-N. Vo, and A. Farina, “A tutorial on Bernoulli filters: theory, implementation and applications,” IEEE Transactions
on Signal Processing, vol. 61, no. 13, pp. 3406–3430, 2013.
[22] R. P. S. Mahler, Statistical multisource multitarget information fusion. Norwood, MA, USA: Artech House, Inc., 2007.
[23] B.-T. Vo, D. Clark, B.-N. Vo, and B. Ristic, “Bernoulli forward-backward smoothing for joint target detection and tracking,” IEEE
Transactions on Signal Processing, vol. 59, no. 9, pp. 4473–4477, 2011.
[24] B.-T. Vo, C. M. See, N. Ma, and W. T. Ng, “Multi-sensor joint detection and tracking with the Bernoulli filter,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 48, no. 2, pp. 1385–1402, 2012.
[25] N. Forti, G. Battistelli, L. Chisci, and B. Sinopoli, “A Bayesian approach to joint attack detection and resilient state estimation,” in
Proc. 55th IEEE Conference on Decision and Control, pp. 1192–1198, Las Vegas, Nevada, USA, 2016.
[26] S. Gillijns and B. D. Moor, “Unbiased minimum-variance input and state estimation for linear discrete-time systems with direct
feedthrough,” Automatica, vol. 43, no. 5, pp. 934–937, 2007.
[27] S. Yong, M. Zhu, and E. Frazzoli, “A unified filter for simultaneous input and state estimation of linear discrete-time stochastic systems,”
Automatica, vol. 63, no. 1, pp. 321–329, 2016.
[28] Y. Cheng, H. Ye, Y. Wang, and D. Zhou, “Unbiased minimum-variance state estimation for linear systems with unknown input,”
Automatica, vol. 45, no. 2, pp. 485–491, 2009.
[29] H. Fang, R. A. De Callafon, and J. Cortés, “Simultaneous input and state estimation for nonlinear systems with applications to flow
field estimation,” Automatica, vol. 49, no. 9, pp. 2805–2812, 2013.
[30] B.-N. Vo and W. K. Ma, “The Gaussian mixture probability hypothesis density filter,” IEEE Transactions on Signal Processing, vol. 54,
no. 11, pp. 4091–4104, 2006.
[31] R. Zimmerman, C. Murillo-Sanchez, and R. Thomas, “MATPOWER: Steady-state operations, planning, and analysis tools for power
systems research and education,” IEEE Transactions on Power Systems, vol. 26, no. 1, pp. 12–19, 2011.
[32] P. Kundur, N. Balu, and M. Lauby, Power System Stability and Control. McGraw-Hill, 1994.
[33] F. Pasqualetti, A. Bicchi, and F. Bullo, “A graph-theoretical characterization of power network vulnerabilities,” in Proc. American
Control Conference, pp. 3918–3923, 2011.
[34] S. Amini, H. Mohsenian-Rad, and F. Pasqualetti, “Dynamic load altering attacks in smart grid,” in Proc. Innovative Smart Grid
Technologies Conference, pp. 1–5, 2015.
26
| 3cs.SY
|
On Extended Long Short-term Memory and Dependent
Bidirectional Recurrent Neural Network
Yuanhang Sua,∗, Yuzhong Huanga , C.-C. Jay Kuoa
arXiv:1803.01686v1 [cs.LG] 27 Feb 2018
a University
of Southern California, Ming Hsieh Department of Electrical Engineering, 3740
McClintock Avenue, Los Angeles, CA, United States
Abstract
In this work, we investigate the memory capability of recurrent neural networks
(RNNs), where this capability is defined as a function that maps an element
in a sequence to the current output. We first analyze the system function
of a recurrent neural network (RNN) cell, and provide analytical results for
three RNNs. They are the simple recurrent neural network (SRN), the long
short-term memory (LSTM), and the gated recurrent unit (GRU). Based on
the analysis, we propose a new design to extend the memory length of a cell,
and call it the extended long short-term memory (ELSTM). Next, we present
a dependent bidirectional recurrent neural network (DBRNN) for the sequencein-sequence-out (SISO) problem, which is more robust to previous erroneous
predictions. Extensive experiments are carried out on different language tasks
to demonstrate the superiority of our proposed ELSTM and DBRNN solutions.
Keywords: recurrent neural networks, long short-term memory, gated
recurrent unit, bidirectional recurrent neural networks, convolutional sequence
to sequence, natural language processing
1. Introduction
The recurrent neural network (RNN) has proved to be an effective solution
for natural language processing (NLP) through the advancement in the last
∗ Corresponding Author
Email address: [email protected]
Preprint submitted to Journal of LATEX Templates
March 6, 2018
three decades [1, 2]. At the cell level, the long short-term memory (LSTM) [3]
and the gated recurrent unit (GRU) [4] are often adopted by an RNN as its lowlevel building element. Built upon these cells, various RNN models have been
proposed to solve the SISO problem. To name a few, there are the bidirectional
RNN (BRNN) [5], the encoder-decoder model [4, 6, 7, 8] and the deep RNN [9].
LSTM and GRU cells were designed to enhance the memory length of RNNs
and address the gradient vanishing/exploding issue [3, 10, 11], yet thorough
analysis on their memory decay property is lacking. The first objective of this
research is to analyze the memory length of three RNN cells - simple RNN
(SRN) [1, 2], LSTM and GRU. It will be conducted in Sec. 2. Our analysis
is different from the investigation of gradient vanishing/exploding problem in
the following sense. The gradient vanishing/exploding problem occurs in the
training process while memory analysis is conducted on a trained RNN model.
Based on the analysis, we further propose a new design in Sec. 3 to extend
the memory length of a cell, and call it the extended long short-term memory
(ELSTM).
As to the macro RNN model, one popular choice is the BRNN [5]. Since
elements in BRNN output sequences should be independent of each other, the
BRNN cannot be used to solve dependent output sequence problem alone. Nevertheless, most language tasks do involve dependent output sequences. Another
choice is the encoder-decoder system, where the attention mechanism was introduced to improve its performance in [7, 8]. We show that the encoder-decoder
system is not an efficient learner by itself. A better solution is to exploit the
encoder-decoder and the BRNN jointly so as to overcome their individual limitations. Following this line of thought, we propose a new multi-task model,
called the dependent bidirectional recurrent neural network (DBRNN), in Sec.
4.
To demonstrate the performance of the DBRNN model with the ELSTM cell,
we conduct a series of experiments on the part of speech (POS) tagging and the
dependency parsing (DP) problems in Sec. 5. Finally, concluding remarks are
given and future research direction is pointed out in Sec. 6.
2
2. Memory Analysis of SRN, LSTM and GRU
For a large number of NLP tasks, we are concerned with finding semantic patterns from input sequences. It was shown by Elman [1] that an RNN
builds an internal representation of semantic patterns. The memory of a cell
characterizes its ability to map input sequences of certain length into such a
representation. Here, we define the memory as a function that maps elements
of the input sequence to the current output. Thus, the memory of an RNN is
not only about whether an element can be mapped into the current output but
also how this mapping takes place. It was reported by Gers et al. [12] that
an SRN only memorizes sequences of length between 3-5 units while an LSTM
could memorize sequences of length longer than 1000 units. In this section, we
conduct memory analysis on SRN, LSTM and GRU cells.
2.1. Memory of SRN
For ease of analysis, we begin with Elman’s SRN model [1] with a linear
hidden-state activation function and a non-linear output activation function
since such a cell model is mathematically tractable while its performance is
equivalent to Jordan’s model [2] and Tensorflow variations.
The SRN model can be described by the following two equations:
ct
= Wc ct−1 + Win Xt ,
(1)
ht
= f (ct ),
(2)
where subscript t is the time unit index, Wc ∈ RN ×N is the weight matrix for
hidden-state vector ct−1 ∈ RN , Win ∈ RN ×M is the weight matrix of input
vector Xt ∈ RM , ht ∈ RN in the output vector, and f (·) is an element-wise
non-linear activation function. Usually, f (·) is a hyperbolic-tangent or a sigmoid
function. Throughout this paper, we omit the bias terms by including them in
the corresponding weight matrices. The multiplication between two equal-sized
vectors in this paper is element-wise multiplication.
3
By induction, ct can be written as
ct = Wct c0 +
t
X
Wct−k Win Xk ,
(3)
k=1
where c0 is the initial internal state of the SRN. Typically, we set c0 = 0. Then,
Eq. (3) becomes
ct =
t
X
Wct−k Win Xk .
(4)
k=1
Let λmax be the largest singular value of Wc . Then, we have
||Wct−k Xk || ≤ ||Wc |||t−k| ||Xk || = λ|t−k|
max ||Xk ||.
(5)
Here, we are only interested in the case of memory decay when λmax < 1. Since
|t−k|
the contribution of Xk , k < t, to output ht decays at least in form of λmax ,
we conclude that SRN’s memory decays exponentially with its memory length
|t − k|.
2.2. Memory of LSTM
Figure 1: The diagram of a LSTM cell.
By following the work of Hochreiter et al. [3], we plot the diagram of the
LSTM cell in Fig. 1. In this figure, φ, σ and ⊗ denote the hyperbolic tangent
function, the sigmoid function and the multiplication operation, respectively.
All of them operate in an element-wise fashion. The LSTM cell has an input
4
gate, an output gate, a forget gate and a constant error carousal (CEC) module.
Mathematically, the LSTM cell can be written as
ct
=
σ(Wf It )ct−1 + σ(Wi It )φ(Win It ),
(6)
ht
=
σ(Wo It )φ(ct ),
(7)
where ct ∈ RN , column vector It ∈ R(M +N ) is a concatenation of the current
input, Xt ∈ RM , and the previous output, ht−1 ∈ RN (i.e., ItT = [XtT , hTt−1 ]).
Furthermore, Wf , Wi , Wo and Win are weight matrices for the forget gate, the
input gate, the output gate and the input, respectively.
Under the assumption c0 = 0, the hidden-state vector of the LSTM can be
derived by induction as
ct =
t
t Y
X
σ(Wf Ij ) σ(Wi Ik )φ(Win Ik ).
(8)
j=k+1
k=1
|
{z
}
forget gate
By setting f (·) in Eq. (2) to the hyperbolic-tangent function, we can compare
outputs of the SRN and the LSTM below:
X
t
t−k
W
W
X
hSRN
=
φ
in k ,
c
t
(9)
k=1
M
hLST
t
= σ(Wo It )φ
X
t Y
t
k=1
σ(Wf Ij ) σ(Wi Ik )φ(Win Ik ) . (10)
j=k+1
|
We see from the above that Wct−k and
{z
forget gate
Qt
j=k+1
}
σ(Wf Ij ) play the same memory
role for the SRN and the LSTM, respectively.
If Wf in Eq. (10) is selected such that
min |σ(Wf Ij )| ≥ λmax ,
then
t
Y
∀λmax ∈ [0, 1),
|t−k|
σ(Wf Ij ) ≥ λmax
.
(11)
j=k+1
As given in Eqs. (5) and (11), the impact of input Ik on the output of the
LSTM lasts longer than that of the SRN. This is the case if an appropriate
weight matrix, Wf , of the forget gate is selected.
5
2.3. Memory of GRU
The GRU was originally proposed for neural machine translation [4]. It
provides an effective alternative for the LSTM. Its operations can be expressed
by the following four equations:
zt
= σ(Wz Xt + Uz ht−1 ),
(12)
rt
= σ(Wr Xt + Ur ht−1 ),
(13)
h̃t
= φ(W Xt + U (rt ⊗ ht−1 )),
(14)
ht
= zt ht−1 + (1 − zt )h̃t ,
(15)
where Xt , ht , zt and rt denote the input, the hidden-state, the update gate
and the reset gate vectors, respectively, and Wz , Wr , W , are trainable weight
matrices. Its hidden-state is also its output, which is given in Eq. (15). By
setting Uz , Ur and U to zero matrices, we can obtain the following simplified
GRU system:
zt
= σ(Wz Xt ),
(16)
h̃t
= φ(W Xt ),
(17)
ht
= zt ht−1 + (1 − zt )h̃t .
(18)
For the simplified GRU with the initial rest condition, we can derive the following by induction:
ht =
t Y
t
X
k=1
σ(Wz Xj ) (1 − σ(Wz Xk ))φ(W Xk ).
(19)
j=k+1
|
{z
update gate
}
By comparing Eqs. (8) and (19), we see that the update gate of the simplified
GRU and the forget gate of the LSTM play the same role. One can control the
memory decay behavior of the GRU by choosing the weight matrix, Wz , of the
update gate carefully.
3. Extended Long Short-Term Memory (ELSTM)
As discussed above, the LSTM and the GRU have longer memory lengths by
introducing the forget and the update gates, respectively. However, from Eqs.
6
(10) and (11), we see that the impact of the proceeding element to the current
output at time step t still decays quickly. However, this does not have to be the
case. To demonstrate this point, we will present a new model and call it the
extended long short-term memory (ELSTM) in this section. In particular, we
propose the following two ELSTM cells:
• ELSTM-I: the ELSTM with trainable input scaling vectors si ∈ RN , i =
1, · · · , t − 1, where si and sj (with i 6= j) are independent.
• ELSTM-II: the ELSTM-I with no forget gate.
They are depicted in Figs. 2 (a) and (b), respectively.
(a)
(b)
Figure 2: The diagrams of (a) the ELSTM-I cell and (b) the ELSTM-II cell.
7
The ELSTM-I cell can be described by
ct
=
σ(Wf It )ct−1 + st σ(Wi It )φ(Win It ),
(20)
ht
=
σ(Wo It )φ(ct + b).
(21)
where b ∈ RN is a trainable bias vector. The ELSTM-II cell can be written as
ct
=
ct−1 + st σ(Wi It )φ(Win It ),
(22)
ht
=
σ(Wo It )φ(ct + b).
(23)
As shown above, we introduce scaling factor, si , i = 1, · · · , t−1, to the ELSTM-I
and the ELSTM-II to increase or decrease the impact of input Ii in the sequence.
To prove that the ELSTM-I has longer memory than the LSTM, we first
derive a closed form expression of ht as
ht = σ(Wo It )φ
X
t
sk
Y
t
k=1
σ(Wf Ij ) σ(Wi Ik )φ(Win Ik ) + b .
(24)
j=k+1
Then, we select sk such that
sk
t
Y
σ(Wf Ij ) ≥
t
Y
σ(Wf Ij ) .
(25)
j=k+1
j=k+1
By comparing Eq. (25) with Eq. (11), we conclude that the ELSTM-I has
longer memory than the LSTM. It is important to emphasize that we only need
to show the sufficient but not necessary conditions such as those in Eq. (25)
to explain the memory capability of an RNN system. In other words, there is
no need to argue whether such a system is always capable in retaining longer
memory as compared to some other systems. As a matter of fact, retaining
longer memory is not always desired, which is the reason why a forgetting gate
[12] is introduced for LSTM. On the other hand, if the memory of a RNN model
is limited by design, it would have negative impact on its performance.
To examine the memory capability problem from this viewpoint, we find that
scaling factors sk play a role similar to the attention score in various attention
models such as Vinyals et al. [7]. The impact of proceeding elements to the
current output can be adjusted (either increased or decreased) by sk . The
8
memory capability of the ELSTM-II can be proven in a similar fashion. We
should point out that, even the ELSTM-II does not have a forget gate, it can
attend or forget a particular position of a sequence as the ELSTM-I through the
scaling factor. On the other hand, fewer parameters are used in the ELSTM-II
than the ELSTM-I. The numbers of parameters used by various RNN cells are
compared in Table 1, where Xt ∈ RM , ht ∈ RN and t = 1, · · · , T .
Table 1: Comparison of Parameter Numbers.
Cell
Number of Parameters
LSTM
4N (M + N + 1)
GRU
3N (M + N + 1)
ELSTM-I
4N (M + N + 1) + N (T + 1)
ELSTM-II
3N (M + N + 1) + N (T + 1)
As shown in Table 1, the number of parameters of the ELSTM cell depends
on the maximum length, T , of the input sequences, which makes the model size
uncontrollable. To address this problem, we choose a fixed Ts (with Ts < T ) as
the upper bound on the number of scaling factors, and set sk = s(k−1) mod Ts +1 ,
if k > Ts and k starts from 1, where mod denotes the modulo operator. In other
words, the sequence of scaling factors is a periodic one with period Ts , so the
elements in a sequence that are distanced by the length of Ts will share the
same scaling factor.
The ELSTM-I cell with periodic scaling factors can be described by
ct
= σ(Wf It )ct−1 + sts σ(Wi It )φ(Win It ),
(26)
ht
= σ(Wo It )φ(ct + b),
(27)
where ts = (t − 1) mod Ts + 1. Similarly, the ELSTM-II cell with periodic
scaling factors can be written as
ct
= ct−1 + sts σ(Wi It )φ(Win It ),
(28)
ht
=
(29)
σ(Wo It )φ(ct + b).
9
We observe that the choice of Ts affects the network performance. Generally
speaking, a small Ts value is suitable for simple language tasks that demand
shorter memory while a larger Ts value is desired for complex ones that demand
longer memory. For the particular sequence-to-sequence (seq2seq [6, 7]) RNN
models, a larger Ts value is always preferred. We will elaborate the parameter
settings in Sec. 5.
4. Dependent BRNN (DBRNN) Model
0
We are interested in the problem of predicting an output sequence, {Yt }Tt=1
with Yi ∈ RN , based on an input sequence, {Xt }Tt=1 with Xi ∈ RM , where T
and T 0 are lengths of the input and the output sequences, respectively. To solve
this problem, we investigate the macro RNN model and propose a multi-task
model, called the dependent BRNN (DBRNN), in this section. Our design is
inspired by pros and cons of two RNN models; namely, the bidirectional RNN
(BRNN) [5] and the encoder-decoder design [4]. We will review the BRNN and
the encoder-decoder in Sec. 4.1 and, then, propose the DBRNN in Sec. 4.2 in
this section.
4.1. BRNN and Encoder-Decoder
The BRNN is used to model the conditional probability density function in
form of
P (Yt |{Xi }Ti=1 ).
Its output is a combination of the output of a forward RNN and the output of
a backward RNN. Due to the bidirectional design, the BRNN can utilize the
information of the entire input sequence to predict each individual output element. However, the BRNN does not exploit the predicted output in predicting
Yt , and elements in the predicted sequence
Yˆt = argmax P (Yt |{Xi }Tt=1 )
Yt
are generated from the input sequences only. Since the prediction of each individual output is conducted independently, the inherent correlation among
10
elements in the output sequence is not utilized during the prediction stage. The
correlation can be maximally captured by explicitly feeding the previous predicted outputs back to the system to avoid the mis-alignment [8] problem. This
idea is exploited by the encoder-decoder system in form of
T
Yˆt = argmax P (Yt |{Ŷi }t−1
i=1 , {Xi }i=1 ).
Yt
On the other hand, the encoder-decoder system is vulnerable to previous erroneous predictions in the forward path. Recently, the BRNN was introduced
to the encoder by Bahdanau et al. [8], yet their design does not address the
erroneous prediction problem.
4.2. DBRNN Model and Training
Being motivated by observations in Sec. 4.1, we propose a multi-task BRNN
model, called the dependent BRNN (DBRNN), to achieve the following objectives:
pt
= W f pft + W b pbt
(30)
argmax pft ,
(31)
argmax pbt ,
(32)
argmax pt
(33)
Ŷtf
=
Ŷtb
=
Ŷt
=
Yt
Yt
Yt
where
pft
=
P (Yt |{Xi }Ti=1 , {Ŷif }ti=1 ),
0
(34)
pbt
=
P (Yt |{Xi }Ti=1 , {Ŷib }Ti=t ),
(35)
pt
=
P (Yt |{Xi }Ti=1 ),
(36)
and W f and W b are trainable weights. As shown in Eqs. (31), (32) and (33),
the DBRNN has three learning objectives: 1) the target sequence for the forward RNN prediction, 2) the reversed target sequence for the backward RNN
prediction, and 3) the target sequence for the bidirectional prediction.
11
The DBRNN model is shown in Fig. 3. It consists of a lower and an upper
BRNN branches. At each time step, the input to the forward and the backward
parts of the upper BRNN is the concatenated forward and backward outputs
from the lower BRNN branch. The final bidirectional prediction is the pooling
of both the forward and the backward predictions. We will show later that this
design will make the DBRNN robust to previous erroneous predictions.
Figure 3: The DBRNN model.
Let F (·) be the cell function. The input is fed into the forward and backward
RNN of the lower BRNN branch as
hft = Flf xt , cfl(t−1) ,
hbt = Flb xt , cbl(t+1) ,
hft
ht = ,
hbt
(37)
where c and l denote the cell hidden state and the lower BRNN, respectively.
The final output, ht , of the lower BRNN is the concatenation of the output, hft ,
of the forward RNN and the output, hbt , of the backward RNN. Similarly, the
upper BRNN generates the final output pt as
pft = Fuf ht , cfu(t−1) , pbt = Fub ht , cbu(t+1) ,
pt = W f pft + W b pbt ,
(38)
ˆ
where u denotes the upper BRNN. To generate forward prediction Ytf and
backward prediction Yˆtb , the forward and backward paths of the upper BRNN
branch are separately trained by the original and the reversed target sequences,
respectively. The results of forward and backward predictions of the upper RNN
branch are then combined to generate the final result.
12
ˆ
There are three errors: 1) forward prediction error ef for Ytf , 2) backward
prediction error eb for Yˆtb , and 3) bidirectional prediction error e for Yˆt . To
train the proposed DBRNN, ef is backpropagated through time to the upper
forward RNN and the lower BRNN, eb is backpropagated through time to the
upper backward RNN and the lower BRNN, and e is backpropagated through
time to the entire model.
To show that DBRNN is more robust to previous erroneous predictions than
one-directional models, we compare their cross entropy defined as
l=−
K
X
pt,k log(p̂t,k ),
(39)
k=1
where K is the total number of classes (e.g. the size of vocabulary for the
language task), p̂t is the predicted distribution, and pt is the ground truth
distribution with k 0 as the ground truth label. It is in form of one-hot vector.
That is,
pt = (δ1,k , · · · , δk,k0 , · · · , δK,k )T ,
k = 1, · · · , K,
where δk,k0 is the Kronecker delta function. Based on Eq. (30), l can be further
expressed as
l
=
−
K
X
pt,k log(Wkf p̂ft,k + Wkb p̂bt,k ),
(40)
k=1
=
−log(Wkf0 p̂ft,k0 + Wkb0 p̂bt,k0 ).
(41)
We can select Wkf0 and Wkb0 such that Wkf0 p̂ft,k0 + Wkb0 p̂bt,k0 is greater than p̂ft,k0
and p̂bt,k0 . Then, we obtain
l
< −
K
X
log(p̂ftk ),
(42)
log(p̂btk ).
(43)
k=1
l
< −
K
X
k=1
The above two equations indicate that the DBRNN can have a cross entropy
lower than those of one-directional predictions by pooling opinions from both
the forward and the backward predictions.
13
It is worthwhile to compare the proposed DBRNN and the bi-attention
model in Cheng et al. [13]. Both of them have bidirectional predictions for
the output, yet there are three main differences. First, the DBRNN provides a
generic solution to the SISO problem without being restricted to dependency
ˆ
parsing. The target sequences in training (namely, Ytf , Yˆtb and Yˆt ) are the same
for the DBRNN while the solution in [13] has different target sequences. Second, the attention mechanism is used in [13] but not in the DBRNN. Third,
The encoder-decoder design is adopted by [13] but not by the DBRNN.
5. Experiments
5.1. Experimental Setup
In the experiments, we compare the performance of five RNN macro-models:
1. basic one-directional RNN (basic RNN);
2. bidirectional RNN (BRNN);
3. sequence-to-sequence (seq2seq) RNN [6] (a variant of the encoder-decoder);
4. seq2seq with attention [7];
5. dependent bidirectional RNN (DBRNN), which is proposed in this work.
For each RNN model, we compare four cell designs: LSTM, GRU, ELSTM-I
and ELSTM-II.
We conduct experiments on two problems: part of speech (POS) tagging and
dependency parsing (DP). We report the testing accuracy for the POS tagging
problem and the unlabeled attachment score (UAS) and the labeled attachment
score (LAS) for the DP problem. The POS tagging task is an easy one which
requires shorter memory while the DP task demands much longer memory. For
the latter, there exist more complex relations between the input and the output.
For the DP problem, we compare our solution with the GRU-based bi-attention
model (bi-Att). Furthermore, we compare the DBRNN using the ELSTM cell
with two other non-RNN-based neural network methods. One is transitionbased DP with neural network (TDP) proposed by Chen et al. [14]. The other
14
is convolutional seq2seq (ConvSeq2seq) proposed by Gehring et al. [15]. For the
proposed DBRNN, we show the best results from the three outputs (namely, pft ,
pbt and pt ). We adopt Ts = 1 in the basic RNN, BRNN, and DBRNN models
and Ts = 100 in the other two seq2seq models for the POS tagging problem.
We use Ts = 100 in all models for the DP problem.
The training dataset used for both problems are from the Universal Dependency 2.0 English branch (UD-English). It contains 12,543 sentences and 14,985
unique tokens. The test dataset in both experiments is from the test English
branch (gold, en.conllu) of CoNLL 2017 shared task development and test data.
The input to the POS tagging and the DP problems are the stemmed and lemmatized sequences (column 3 in CoNLL-U format). The target sequence for the
POS tagging is the universal POS tag (column 4). The target sequence for the
DP is the interleaved dependency relation to the headword (relation, column
8) and its headword position (column 7). As a result, the length of the target
sequence for the DP is twice of the length of the input sequence.
Table 2: Network parameters and training details.
Number of RNN layers
1
Embedding layer vector size
512
Number of RNN cells
512
Batch size
20
Training steps
11 epochs
Learning rate
0.5
Optimizer
AdaGrad[16]
The input is first fed into a trainable embedding layer [17] before it is sent to
the actual network. Table 2 shows the detailed network and training specifications. We do not finetune network hyper-parameters or apply any engineering
trick (e.g. feeding additional inputs other than the raw embedded input sequences) for the best possible performance since our main goal is to compare
the performance of the LSTM, GRU, ELSTM-I and ELSTM-II cells under var-
15
ious macro-models.
Table 3: POS tagging test accuracy (%)
LSTM
GRU
ELSTM-I
ELSTM-II
BASIC RNN
87.30
87.51
87.44
86.92
BRNN
89.55
89.39
89.29
86.89
Seq2seq
24.43
35.27
50.42
57.84
Seq2seq with Att
31.34
34.60
81.72
57.35
DBRNN
89.86
89.06
89.28
87.97
Table 4: DP test results (UAS/LAS %)
LSTM
GRU
ELSTM-I
ELSTM-II
BASIC RNN
43.24/25.28
45.24/29.92
58.49/36.10
58.87/36.50
BRNN
37.88/25.26
16.86/8.95
55.97/35.13
52.15/31.72
Seq2seq
29.38/6.05
36.47/13.44
48.58/24.05
53.84/34.08
Seq2seq with Att
31.82/16.16
43.63/33.98
64.30/52.60
55.13/35.60
DBRNN
52.19/40.29
52.23/37.25
61.35/43.32
61.94/42.89
Bi-Att [13]
1
59.97/44.94
5.2. Comparison of RNN Models
The results of the POS tagging and the DP problems are shown in Tables
3 and 4, respectively. We see that the DBRNN outperforms the BRNN and
the seq2seq in both the POS tagging and the DP problems regardless of the
cell types. This shows the advantage of expert opinion pooling from the input
as well as the predicted output. The DBRNN achieves a training loss that is
similar or better than the seq2seq model with attention as shown in Figs. 4 and
5. However, the DBRNN can overfit to the training data more easily due to a
1 The
result is generated by using exactly the same settings in Table. 2. We do not feed in
the network with information other than input sequence itself.
16
larger model size. To overcome it, one can use a proper regularization scheme
in the training process.
1.5
POS Tagging: Different Architectures with LSTM Cell (Zoom in)
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
1.45
1.4
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
1.45
1.4
1.35
1.35
1.3
Perplexity
Perplexity
POS Tagging: Different Architectures with GRU Cell (Zoom in)
1.5
1.25
1.2
1.3
1.25
1.2
1.15
1.15
1.1
1.1
1.05
1.05
1
1
0
2
4
6
8
10
0
12
Training Steps
10
2
4
4
(a)
8
10
12
104
(b)
POS Tagging: Different Architectures with ELSTM-I Cell (Zoom in)
POS Tagging: Different Architectures with ELSTM-II Cell (Zoom in)
1.5
1.5
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
1.45
1.4
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
1.45
1.4
1.35
1.35
1.3
Perplexity
Perplexity
6
Training Steps
1.25
1.2
1.15
1.3
1.25
1.2
1.15
1.1
1.1
1.05
1.05
1
1
0
2
4
6
8
10
Training Steps
12
0
104
(c)
2
4
6
8
Training Steps
10
12
104
(d)
Figure 4: The training perplexity of different models with the LSTM (top left), the GRU (top
right), the ELSTM-I (bottom left) and the ELSTM-II (bottom right) for the POS tagging
task.
The proposed ELSTM-I and ELSTM-II cells outperform the LSTM and
GRU cells in most RNN models. This is especially true for complex language
tasks, where the two ELSTM cells outperform traditional cell designs by a
significant margin. This demonstrates the effectiveness of the sequence of scaling
factors adopted by the ELSTM cells. It allows the network to retain longer
memory with better attention.
17
DP: Different Architectures with LSTM Cell
50
45
40
35
35
30
30
25
25
20
20
15
15
10
10
5
5
0
2
4
6
8
10
12
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
45
Perplexity
Perplexity
40
DP: Different Architectures with GRU Cell
50
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
14
0
2
4
104
Training Steps
(a)
10
12
14
104
8
DP: Different Architectures with ELSTM-II Cell (Zoom in)
10
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
9
Basic RNN
BRNN
Seq2Seq
Seq2Seq Attention
DBRNN
9
8
7
Perplexity
7
Perplexity
8
(b)
DP: Different Architectures with ELSTM-I Cell (Zoom in)
10
6
Training Steps
6
5
6
5
4
4
3
3
2
2
1
1
0
2
4
6
8
Training Steps
10
12
14
104
(c)
0
2
4
6
8
10
12
Training Steps
14
104
(d)
Figure 5: The training perplexity of different models with the LSTM (top left), the GRU (top
right), the ELSTM-I (bottom left) and the ELSTM-II (bottom right) cells for the DP task.
18
The ELSTM-I cell even outperforms the bi-Att model, which was designed
specifically for the DP task. For the POS tagging problem, the advantage
of the ELSTM cells is not as obvious. This is probably due to the shorter
memory requirement in this simple task. In this context, ELSTM cells are
over-parameterized, and they converge slower and tend to overfit the training
data.
The ELSTM-I and ELSTM-II cells with large Ts value perform particularly
well for the seq2seq (with and without attention) model. The hidden state,
ct , of ELSTM cells is more expressive in representing patterns over a longer
distance. Since the seq2seq design relies on the expressive power of a hidden
state, ELSTMs have a clear advantage.
To substantiate our claim in Sec. 2, we conduct additional experiments
to show the robustness of the ELSTM cells and the DBRNN. Specifically, we
compare the performance of the same five models with LSTM, ELSTM-I and
ELSTM-II with It = Xt for the same language tasks. We do not include the
GRU cell since it inherently demands ItT = [XtT , hTt−1 ]. The convergence behaviors of It = Xt and ItT = [XtT , hTt−1 ] with the LSTM, ELSTM-I and ELSTM-II
cells for the DP problem are shown in Fig. 6. We see that the ELSTM-I and the
ELSTM-II do not behave much differently between It = Xt and ItT = [XtT , hTt−1 ]
while the LSTM does. This shows the effectiveness of the ELSTM-I and the
ELSTM-II design regardless of the input. More performance comparison will
be provided in the Appendix.
5.3. Comparison between ELSTM and Non-RNN-based Methods
As stated earlier, the ELSTM design is more capable of extending the memory and capturing complex SISO relationships than other RNN cells. In this subsection, we compare the DP performance of two models built upon the ELSTM-I
cell (namely, the DBRNN and the seq2seq with attention) and two non-RNNbased neural network based methods (i.e., the TDP [14] and the convseq2seq
[15]). The TDP is a hand-crafted method based on a parsing tree, and its neural
network is a multi-layer perceptron with one hidden layer. Its neural network is
19
DP: Basic RNN Architecture with LSTM Cell
50
DP: Basic RNN Architecture with ELSTM-II Cell
50
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
40
It = X t
40
It = X t
40
It = X t
30
25
20
35
Perplexity
35
Perplexity
35
Perplexity
DP: Basic RNN Architecture with ELSTM-I Cell
50
45
30
25
20
30
25
20
15
15
15
10
10
10
5
5
0
2
4
6
8
Training Steps
10
12
5
0
#10
2
4
4
(a)
6
8
Training Steps
(b)
10
12
0
#10
4
2
4
6
8
Training Steps
10
12
#104
(c)
Figure 6: Training perplexity of the basic RNN with It = Xt and ItT = [XtT , hT
t−1 ] for the
DP problem.
used to predict the transition from a tail word to its headword. The convseq2seq
is an end-to-end convolutional neural network (CNN) with an attention mechanism. We used the default settings for the TDP and the convseq2seq as reported
in [14] and [15], respectively. For the TDP, we do not use the ground truth POS
tags but the predicted dependency relation labels as the input to the parsing
tree for the next prediction.
We see from Table 5 that the ELSTM-based models learn much faster than
the CNN-based convseq2seq model with fewer parameters. The convseq2seq
uses dropout while the ELSTM-based models do not. It is also observed that
convseq2seq does not converge if Adagrad is used as its optimizer. The ELSTMI-based seq2seq with attention even outperforms the TDP, which was specifically designed for the DP task. Without a good pretrained word embedding
scheme, the UAS and LAS of TDP drop drastically to merely 8.93% and 0.30%
respecively.
6. Conclusion and Future Work
Although the memory of the LSTM and GRU celles fades slower than that
of the SRN, it is still not long enough for complicated language tasks such
as dependency parsing. To address this issue, we proposed the ELSTM-I and
the ELSTM-II to enhance the memory capability of an RNN cell. Besides, we
20
Table 5: DP test accuracy (%) and system settings
Seq2seq-E-I
DBRNN-E-II
Convseq2seq
TDP
UAS
64.30
61.94
52.55
62.29
LAS
52.60
42.89
44.19
52.18
Training steps
11 epochs
11 epochs
11 epochs
11 epochs
# parameters
12,684,468
16,460,468
22,547,124
950,555
Pretrained embedding
No
No
No
Yes
End-to-end
Yes
Yes
Yes
No
Regularization
No
No
No
Yes
Dropout
No
No
Yes
Yes
AdaGrad
AdaGrad
NAG [18]
AdaGrad
Learning rate
0.5
0.5
0.25
0.01
Embedding size
512
512
512
50
Encoder layers
1
N/A
4
N/A
Decoder layers
1
N/A
4
N/A
Kernel size
N/A
N/A
3
N/A
Hidden layer size
N/A
N/A
N/A
200
Optimizer
21
presented a new DBRNN model that has the merits of both the BRNN and
the encoder-decoder. It was shown by experimental results that the ELSTM-I
and the ELSTM-II outperforms other RNN cell designs by a significant margin
for complex language tasks. The DBRNN model is superior to the BRNN and
the seq2seq models for simple and complex language tasks. Furthermore, the
ELSTM-based RNN models outperform the CNN-based convseq2seq model and
the handcrafted TDP. There are interesting issues to be explored furthermore.
For example, is the ELSTM cell also helpful in more sophisticated RNN models
such as the deep RNN? Is it possible to make the DBRNN deeper and better?
They are left for future study.
7. Declarations of interest
Declarations of interest: none
8. Acknowledgements
This research did not receive any specific grant from funding agencies in the
public, commercial, or not-for-profit sectors.
References
References
[1] J. Elman, Finding structure in time, Cognitive Science 14 (1990) 179–211.
[2] M. Jordan, Serial order: A parallel distributed processing approach, Advances in Psychology 121 (1997) 471–495.
[3] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (1997) 1735–1780.
[4] K. Cho, B. v. Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares,
H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoderdecoder for statistical machine translation, In Proceedings of The
Empirical Methods in Natural Language Processing (EMNLP 2014).
22
[5] M. Schuster, K. K. Paliwal, Bidirectional recurrent neural networks, Signal
Processing 45 (1997) 2673–2681.
[6] I. Sutskever, O. Vinyals, Q. V. Le, Sequence to sequence learning with neural networks, Advances in Neural Information Processing Systems (2014)
3104–3112.
[7] O. Vinyals, L. Kaiser, T. Koo, S. Petrov, I. Sutskever, G. Hinton, Grammar
as a foreign language, Advances in Neural Information Processing Systems
(2015) 2773–2781.
[8] D. Bahdanau, K. Cho, Y. Bengio, Neural machine translation by jointly
learning to align and translate, In Proceedings of the International Conference on Learning Representations (ICLR 2015).
[9] R. Pascanu, C. Gulcehre, K. Cho, Y. Bengio, How to construct deep recurrent neural networks, arXiv:1312.6026.
[10] P. Razvan, T. Mikolov, Y. Bengio, On the difficulty of training recurrent
neural networks, In Proceedings of The International Conference on Machine Learning (ICML 2013) (2013) 1310–1318.
[11] Y. Bengio, P. Simard, P. Frasoni, Learning long-term dependencies with
gradient descent is difficult, Neural Networks 5 (1994) 157–166.
[12] F. A. Gers, J. Schmidhuber, F. Cummins, Learning to forget: Continual
prediction with lstm, Neural Computation (2000) 2451–2471.
[13] H. Cheng, H. Fang, X. He, J. Gao, L. Deng, Bi-directional attention with
agreement for dependency parsing, In Proceedings of The Empirical Methods in Natural Language Processing (EMNLP 2016).
[14] D. Chen, M. Christopher, A fast and accurate dependency parser using
neural networks, in: In Proceedings of The Empirical Methods in Natural
Language Processing (EMNLP 2014), 2014, pp. 740–750.
23
[15] J. Gehring, G. Auli M, D. Yarats, Y. Denis, D. Yann N., Convolutional
sequence to sequence learning, in: arXiv preprint, no. 1705.03122, 2017.
[16] Duchi, Adaptive subgradient methods for online learning and stochastic
optimization, The Journal of Machine Learning Research (2011) 2121–2159.
[17] Y. Bengio, R. Ducharme, P. Vincent, C. Jauvin, A neural probabilistic
language model, Journal of Machine Learning Research (2003) 1137–1155.
[18] Y. Nesterov, A method of solving a convex programming problem with
convergence rate o (1/k2), in: Soviet Mathematics Doklady, Vol. 27, 1983,
pp. 372–376.
24
Appendix A: More Experimental Results
In the appendix, we provide more experimental results to shed light on the
convergence performance in the training of various models with different cells for
the POS tagging and the DP tasks. First, we compare the training perplexity
between It = Xt and ItT = [XtT , hTt−1 ] for various models with the LSTM,
the ELSTM-I and the ELSTM-II cells in Figs. .7-.12. Then, we examine the
training perplexity with ItT = [XtT , hTt−1 ] for various models with different cells
in Figs. .13-.15.
DP: BRNN Architecture with LSTM Cell
50
DP: BRNN Architecture with ELSTM-II Cell
50
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
40
It = X t
40
It = X t
40
It = X t
30
25
20
35
Perplexity
35
Perplexity
35
Perplexity
DP: BRNN Architecture with ELSTM-I Cell
50
45
30
25
20
30
25
20
15
15
15
10
10
10
5
5
0
2
4
6
8
Training Steps
10
12
5
0
#10
2
4
4
(a)
6
8
Training Steps
10
12
0
#10
2
4
4
(b)
6
8
Training Steps
10
12
#104
(c)
Figure .7: The training perplexity of the BRNN model with It = Xt and ItT = [XtT , hT
t−1 ] for
the DP task.
DP: DBRNN Architecture with LSTM Cell
50
DP: DBRNN Architecture with ELSTM-II Cell
50
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
40
It = X t
40
It = X t
40
It = X t
30
25
20
35
Perplexity
35
Perplexity
35
Perplexity
DP: DBRNN Architecture with ELSTM-I Cell
50
45
30
25
20
30
25
20
15
15
15
10
10
10
5
5
0
2
4
6
8
Training Steps
(a)
10
12
5
0
#104
2
4
6
8
Training Steps
(b)
10
12
0
#104
2
4
6
8
Training Steps
10
12
#104
(c)
Figure .8: The training perplexity of the DBRNN model with It = Xt and ItT = [XtT , hT
t−1 ]
for the DP task.
25
DP: Seq2Seq Architecture with LSTM Cell
50
DP: Seq2Seq Architecture with ELSTM-II Cell
50
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
40
It = X t
40
It = X t
40
It = X t
30
25
20
35
Perplexity
35
Perplexity
35
Perplexity
DP: Seq2Seq Architecture with ELSTM-I Cell
50
45
30
25
20
30
25
20
15
15
15
10
10
10
5
5
0
2
4
6
8
Training Steps
10
12
5
0
#10
2
4
4
(a)
6
8
Training Steps
10
12
0
#10
2
4
4
(b)
6
8
Training Steps
10
12
#104
(c)
Figure .9: The training perplexity of the seq2seq model with It = Xt and ItT = [XtT , hT
t−1 ]
for the DP task.
DP: Seq2Seq with Attention with LSTM Cell
50
DP: Seq2Seq with Attention with ELSTM-II Cell
50
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
45
ITt = [XTt , hTt-1]
40
It = X t
40
It = X t
40
It = X t
30
25
20
35
Perplexity
35
Perplexity
35
Perplexity
DP: Seq2Seq with Attention with ELSTM-I Cell
50
45
30
25
20
30
25
20
15
15
15
10
10
10
5
5
0
2
4
6
8
Training Steps
10
12
5
0
2
4
#104
(a)
6
8
Training Steps
10
12
0
2
4
#104
(b)
6
8
Training Steps
10
12
#104
(c)
Figure .10: The training perplexity of the seq2seq with attention model with It = Xt and
ItT = [XtT , hT
t−1 ] for the DP task.
POS Tagging: Seq2Seq Architecture with LSTM Cell
18
ITt = [XTt , hTt-1]
16
I =X
t
POS Tagging: Seq2Seq Architecture with ELSTM-I Cell
20
ITt = [XTt , hTt-1]
18
It = X t
16
14
t
Perplexity
Perplexity
14
12
10
8
POS Tagging: Seq2Seq Architecture with ELSTM-II Cell
20
ITt = [XTt , hTt-1]
18
It = X t
16
14
Perplexity
20
12
10
8
12
10
8
6
6
6
4
4
4
2
2
0
2
4
6
8
Training Steps
10
12
(a)
2
0
#10
4
2
4
6
8
Training Steps
(b)
10
12
0
#10
4
2
4
6
8
Training Steps
10
12
#104
(c)
Figure .11: The training perplexity of the seq2seq model with It = Xt and ItT = [XtT , hT
t−1 ]
for the POS tagging task.
26
10
8
POS Tagging: Seq2Seq with Attention with ELSTM-II Cell
20
ITt = [XTt , hTt-1]
18
I =X
16
t
t
14
Perplexity
12
POS Tagging: Seq2Seq with Attention with ELSTM-I Cell
20
ITt = [XTt , hTt-1]
18
I =X
16
t
t
14
Perplexity
Perplexity
POS Tagging: Seq2Seq with Attention with LSTM Cell
20
ITt = [XTt , hTt-1]
18
It = X t
16
14
12
10
8
12
10
8
6
6
6
4
4
4
2
2
0
2
4
6
8
Training Steps
10
12
2
0
#104
(a)
2
4
6
8
Training Steps
(b)
10
12
0
#104
2
4
6
8
Training Steps
10
12
#104
(c)
Figure .12: The training perplexity of the seq2seq with Att model with It = Xt and ItT =
[XtT , hT
t−1 ] for the POS tagging task.
27
POS Tagging: Basic RNN Architecture with Different Cells
5
4
4
3.5
3.5
3
LSTM
GRU
ELSTM-I
ELSTM-II
4.5
Perplexity
Perplexity
4.5
POS Tagging: BRNN Architecture with Different Cells
5
LSTM
GRU
ELSTM-I
ELSTM-II
2.5
3
2.5
2
2
1.5
1.5
1
1
0
2
4
6
8
10
12
0
2
4
104
Training Steps
(a)
8
10
12
104
(b)
POS Tagging: Seq2Seq Attention Architecture with Different Cells
LSTM
GRU
ELSTM-I
ELSTM-II
9
8
POS Tagging: Basic RNN Architecture with Different Cells
5
10
LSTM
GRU
ELSTM-I
ELSTM-II
4.5
4
7
3.5
Perplexity
Perplexity
6
Training Steps
6
5
3
2.5
4
2
3
1.5
2
1
1
0
2
4
6
8
Training Steps
10
0
12
104
(c)
2
4
6
8
10
Training Steps
12
104
(d)
Figure .13: The training perplexity for the basic RNN (top left), the BRNN (top right), the
seq2seq with Att (bottom left) and the DBRNN (bottom right) for the POS tagging.
28
DP: Basic RNN Architecture with Different Cells
20
18
16
LSTM
GRU
ELSTM-I
ELSTM-II
18
16
14
Perplexity
14
Perplexity
DP: BRNN Architecture with Different Cells
20
LSTM
GRU
ELSTM-I
ELSTM-II
12
10
12
10
8
8
6
6
4
4
2
2
0
2
4
6
8
10
12
14
0
2
4
104
Training Steps
(a)
10
12
14
104
16
DP: DBRNN Architecture with Different Cells
20
LSTM
GRU
ELSTM-I
ELSTM-II
18
LSTM
GRU
ELSTM-I
ELSTM-II
18
16
14
Perplexity
14
Perplexity
8
(b)
DP: Seq2Seq Attention Architecture with Different Cells
20
6
Training Steps
12
10
12
10
8
8
6
6
4
4
2
2
0
2
4
6
8
Training Steps
10
12
14
104
(c)
0
2
4
6
8
10
12
Training Steps
14
104
(d)
Figure .14: The training perplexity for the basic RNN (top left), the BRNN (top right), the
seq2seq with Att (bottom left) and the DBRNN models (bottom right) for the DP task.
29
POS Tagging: Seq2Seq Architecture with Different Cells
10
9
8
LSTM
GRU
ELSTM-I
ELSTM-II
18
16
14
Perplexity
7
Perplexity
DP: Seq2Seq Architecture with Different Cells
20
LSTM
GRU
ELSTM-I
ELSTM-II
6
5
12
10
8
4
6
3
4
2
2
1
0
2
4
6
8
10
12
0
2
4
104
Training Steps
(a)
8
10
12
14
104
(b)
POS Tagging: Seq2Seq Attention Architecture with Different Cells
LSTM
GRU
ELSTM-I
ELSTM-II
9
8
DP: Seq2Seq Attention Architecture with Different Cells
20
10
LSTM
GRU
ELSTM-I
ELSTM-II
18
16
14
Perplexity
7
Perplexity
6
Training Steps
6
5
12
10
8
4
6
3
4
2
2
1
0
2
4
6
8
Training Steps
10
0
12
2
4
104
(c)
6
8
Training Steps
10
12
14
104
(d)
Figure .15: The training perplexity for the seq2seq model (top), and the seq2seq with Att
model (bottom), for the POS task (left) and the DP (right) task.
30
| 9cs.NE
|
EURETILE 2010-2012 summary: first
three years of activity of the European
Reference Tiled Experiment.
Pier Stanislao Paolucci, Iuliana Bacivarov, Gert Goossens, Rainer
Leupers, Frédéric Rousseau, Christoph Schumacher, Lothar Thiele,
Piero Vicini
www.euretile.eu
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
ACKNOWLEDGMENT
The authors summarized in this document the work performed by the EURETILE team during the
first three years of the project.
During the period 2010-2012 the full team was composed by:
Istituto Nazionale di Fisica Nucleare - sezione di Roma (INFN)
Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Michela Giovagnoli,
Francesca Lo Cicero, Alessandro Lonardo, Pier Stanislao Paolucci, Francesco Simula,
Davide Rossetti, Laura Tosoratto, Piero Vicini
RWTH Aachen University – ISS & SSS (RWTH)
Jovana Jovic, Rainer Leupers, Luis Murillo, Christoph Schumacher, Jan Henrik
Weinstock
ETH Zurich - The Swiss Federal Institute of Technology Zurich (ETHZ)
Iuliana Bacivarov, Devendra Rai, Lars Schor, Lothar Thiele, Hoeseok Yang
Université Joseph Fourier – Grenoble Institute of Technology - TIMA Laboratories
(UJF/INP TIMA)
Ashraf El Antably, Ikbel Belaid, Clément Deschamps, Nicolas Fournel, Mohamad
Jaber, Julian Michaud, Etienne Ripert, Frédéric Rousseau
Target Compiler Technologies (TARGET)
Gert Goossens, Werner Geurts
This publication is a derivative of the EURETILE 2010-2012 Publishable Executive Summary. All
project deliverable documents have been peer reviewed by a committee of four experts of the field.
The EURETILE project is funded by the European Commission, through the Grant Agreement no.
247846, Call: FP7-ICT-2009-4 Objective FET-ICT-2009.8.1 Concurrent Tera-device Computing.
www.euretile.eu
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Table of Contents
1.
Abstract .......................................................................................................................................... 5
1.1.
Participating Institutions ...................................................................................................... 6
2.
Structure of the Document ............................................................................................................. 6
2.1.
Bibliographical Notes .......................................................................................................... 6
3.
Publishable summary – 2010-2012 revision .................................................................................. 9
3.1.
Project Objectives and Expected Final Results ................................................................... 9
3.1.1. Software Tool-Chain ........................................................................................................ 9
3.1.2. Experimental Hardware and Simulation Platforms for Embedded Systems and HPC .. 12
3.1.3. Definition of a Hierarchical Brain-Inspired Many-Tile Many-Process Architecture .... 13
3.1.4. Fault-Tolerance and Scalability ..................................................................................... 14
3.1.5. Versatile Distributed Network Processor ....................................................................... 14
3.1.6. ASIP Design ................................................................................................................... 14
3.1.7. Application Benchmarks ................................................................................................ 15
3.1.8. Main expected final results ............................................................................................ 16
3.1.9. Work-packages............................................................................................................... 16
3.2.
Work Performed and Achieved Results in 2010 (First Year)............................................ 17
3.2.1. HW Architecture Design ................................................................................................ 17
3.2.2. DNP Design ................................................................................................................... 17
3.2.3. LQCD and PSNN ASIP Design ..................................................................................... 18
3.2.4. System-Level Programming Framework ....................................................................... 18
3.2.5. Scalable Simulation Environment .................................................................................. 18
3.2.6. Hardware-Dependent Software ...................................................................................... 19
3.3.
Work Performed and Achieved Results in 2011 (Second Year) ....................................... 20
3.3.1. Definition of Fault Aware Architecture ......................................................................... 20
3.3.2. Challenging Application Benchmarks ........................................................................... 20
3.3.3. Innovation on HW Intellectual Properties and design tools .......................................... 21
3.3.4. System-Level Programming Framework (DAL) ........................................................... 23
3.3.5. Scalable Embedded Computing Simulation Environment (VEP) ................................. 24
3.3.6. AED integration with DNP driver software ................................................................... 25
3.3.7. Hardware-Dependent Software (HdS) ........................................................................... 25
3 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.4.
Work Performed and Achieved Results in 2012 (Third Year) .......................................... 27
3.4.1. System-Level Programming Framework (DAL) ........................................................... 27
3.4.2. Hardware-Dependent Software (HdS) ........................................................................... 27
3.4.3. Scalable Embedded Computing Simulation Environment (VEP) ................................. 28
3.4.4. HW support to Fault Awareness .................................................................................... 29
3.4.5. DNP and APENet+ Innovations .................................................................................... 29
3.4.6. ASIP support to RDMA ................................................................................................. 30
3.4.7. ASIP Design Tool Improvement.................................................................................... 31
3.4.8. HW Experimental Platform Prototype (QUonG)........................................................... 31
3.4.9. DPSNN-STDP: Distributed Polychronous and Plastic Spiking Neural Net .................. 32
3.5.
Plans for 2013 - 2014 activities ......................................................................................... 33
3.6.
The SHAPES Project Background (2006-2009) ............................................................... 34
3.6.1. 2006 SHAPES Project Abstract ..................................................................................... 34
3.6.2. Investigation of the tiled HW paradigm ......................................................................... 34
3.6.3. Experimentation of real-time, communication aware system SW ................................. 35
3.6.4. DIOPSIS940HF MPSOC and SHAPES RISC + mAgic VLIW DSP + DNP Tile. ....... 37
4.
Dissemination............................................................................................................................... 39
4.1.
Papers - Posters - Technical Press – Newsletters .............................................................. 39
4.2.
Books/Book Chapters ........................................................................................................ 42
4.3.
Presentations ...................................................................................................................... 42
4.4.
CASTNESS’11 .................................................................................................................. 44
CASTNESS’11 Agenda ............................................................................................................... 44
5.
APPENDIX: Euretile Project Glossary........................................................................................ 47
References ............................................................................................................................................ 50
4 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
1. Abstract
This is the summary of first three years of activity of the EURETILE FP7 project 247846.
EURETILE investigates and implements brain-inspired and fault-tolerant foundational innovations to
the system architecture of massively parallel tiled computer architectures and the corresponding
programming paradigm. The execution targets are a many-tile HW platform, and a many-tile
simulator. A set of SW process - HW tile mapping candidates is generated by the holistic SW toolchain using a combination of analytic and bio-inspired methods. The Hardware dependent Software
is then generated, providing OS services with maximum efficiency/minimal overhead. The many-tile
simulator collects profiling data, closing the loop of the SW tool chain. Fine-grain parallelism inside
processes is exploited by optimized intra-tile compilation techniques, but the project focus is above
the level of the elementary tile. The elementary HW tile is a multi-processor, which includes a fault
tolerant Distributed Network Processor (for inter-tile communication) and ASIP accelerators.
Furthermore, EURETILE investigates and implements the innovations for equipping the elementary
HW tile with high-bandwidth, low-latency brain-like inter-tile communication emulating 3 levels of
connection hierarchy, namely neural columns, cortical areas and cortex, and develops a dedicated
cortical simulation benchmark: DPSNN-STDP (Distributed Polychronous Spiking Neural Net with
synaptic Spiking Time Dependent Plasticity). EURETILE leverages on the multi-tile HW paradigm
and SW tool-chain developed by the FET-ACA SHAPES Integrated Project (2006-2009).
The APE Parallel Computing Lab of INFN Roma is in charge of the EURETILE HW Design
(QUonG system/APENet+ board/DNP (Distributed Network Processor) and Scientific Application
Benchmarks. The Computer Engineering and Networks Laboratory (TIK) of ETH Zurich (Swiss
Federal Institute of Technology) designs the high-level explicit parallel programming and automatic
mapping tool (DOL/DAL) and a set of “Embedded Systems” benchmarks. The Software for Systems
on Silicon (SSS) of the ISS institute of RWTH Aachen, investigates and provides the parallel
simulation technology and scalable simulation-based profiling/debugging support. The TIMA
Laboratory of the University Joseph Fourier in Grenoble explores and deploys the HdS (Hardware
dependent Software) including the distributed OS architecture. TARGET Compiler Technologies, the
Belgian leading provider of retargetable software tools and compilers for the design, programming,
and verification of application-specific processors (ASIPs), is in charge of the HW/SW Co-design
tools for custom components of the EURETILE architecture.
Grant Agreement no. 247846 Call: FP7-ICT-2009-4 Objective FET-ICT-2009.8.1 Concurrent Teradevice Computing
Scientific Coordinator: Pier Stanislao Paolucci, Istituto Nazionale di Fisica Nucleare, Roma, Italy
Administrative Coordinator: Michela Giovagnoli, Istituto Nazionale di Fisica Nucleare, Roma, Italy
5 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
1.1.Participating Institutions
INFN (Istituto Nazionale di Fisica Nucleare) Sezione di Roma (coordinator): the APE Parallel
Computing Lab;
The Computer Engineering and Networks Laboratory (TIK) of ETH Zurich (Swiss Federal
Institute of Technology);
The Software for Systems on Silicon (SSS) of the ISS institute of RWTH Aachen;
UJF-TIMA: the TIMA Laboratory of the University Joseph Fourier in Grenoble;
TARGET Compiler Technologies
2. Structure of the Document
This document summarizes the activities of the European Reference Tiled Experiment, the
EURETILE FET (Future Emerging Technologies) FP7 project 247836, during its first three years of
activity (2010-2012).
The document should provide a global view of the project framework, and it should simplify the
reading of individual publications produced during the project time frame.
Another purpose is to create an historical record of the project development flow. Therefore,
individual sections are dedicated to each year of activity, and they describe the project status as
reported at the end of 2010, 2011 and 2012.
As the EURETILE project is an evolution of the SHAPES FP7 project, we inserted also a short
summary of the SHAPES results that has been relevant for the EURETILE project.
2.1.Bibliographical Notes
The “References” section at the end of the document is composed of four groups of citations.
The first group contains a restricted set of “prior art” publications produced by other groups before
2010, the start date of the EURETILE project. Without any claim for completeness, we list a few
publications that surely contributed to define the starting point of the project.
The second group contains a selection of publications produced by members of our team before the
start of the project, mainly during the time frame of the SHAPES project (2006-2009).
The third group is a selection of publications produced by other groups during the 2010-2012 period
about topics that are similar to those investigated by the EURETILE project.
The last set is a selection of the publications produced by the members of the EURETILE team as a
result of the project activities. The complete list of presentations and publication produced by the
project is included in the “Dissemination” section.
6 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
The software design for specific platforms follows the methodologies described in the reference
paper from Sangiovanni-Vincentelli, Martin (2001). The software is generated from a flow that
respects the Y-chart approach proposed by Kienhuis and all (2002), taking into account the
architecture specification. It includes few layers, and the lowest layer is called the Hardware
dependant Software (HdS) as it is related to hardware access from the software. HdS development
process is described in Schirner , Domer and Gerstlauer (2009).
The Kahn process networks model of computation (Khan 1974) is used to express the computational
kernels that constitute the lower levels of hierarchy of EURETILE applications, leveraging on the
Distributed Operating Layer (DOL) (Thiele, Bacivarov, Haid, Huang, 2007). developed in the
framework of the SHAPES project (Paolucci, Jerraya, Leupers, Thiele, Vicini., 2006). The KPN
model explicitly separates computation and communication in the parallel application specification:
an application kernel is expressed as a set of concurrently executing processes that only communicate
via point-to-point FIFO channels. In EURETILE, the individual KPN processes are specified in
C/C++, while the topology of the KPN is described using XML-based formats (Haid et al. 2009), that
specify the instantiation of processes, their connection by FIFO channels, and their mapping onto the
target platforms. This way, the processes that compose the KPN can be compiled for any processor
for which a C/C++ compiler is available. Moreover, the specification of the mapping in an XML file
rather than in the source code of processes makes the description of the application kernel more
platform-independent. The expression of dynamic application scenarios and the specification of
upper level of application hierarchies are among the targets addressed by the development of the
EURETILE Distributed Application Layer (DAL) (see section 3.1.1).
In the multi-tile context, one challenge is to manage as much efficiently as possible the
communication between tiles. We focus the work in EURETILE on the software driver development
for HdS generation, while some other works focus on the mapping quality (Lin, Gerstlauer, Evan,
2012) in order to reduce the impact on performances. Our software driver generation depends highly
on the hardware architecture, which is a costly approach when dealing with many-tile system with
several kinds of tiles. One way to facilitate the re-use of existing drivers is probably the use of
virtualization environment (Heiser, 2011). Such a virtualization environment may help as well for
multi-application support.
The simulation of the EURETILE many-tile system is a key aspect of the project. The work
“Exploiting parallelism and structure to accelerate the simulation of chip multi-processors”
experiments with a custom parallel simulation framework (Penry et al., 2006). It discusses the
impact of varying system sizes and accompanying cache effects on the simulation host machine, and
suggests taking these cache effects into account for the scheduling of logical processes to simulation
host cores. Ezudheen et al (2009) “Parallelizing SystemC kernel for fast hardware simulation on SMP
machines” is one of the early works describing parallel SystemC simulation that relies on parallel
execution of process executions within a delta-cycle. Various methods to schedule SystemC logical
processes to physical simulation host OS threads are discussed. Lu et al. (2008) “Learning from
mistakes: A comprehensive study on real world concurrency bug characteristics” presented an
exhaustive analysis and classification of concurrency bugs (e.g., atomicity violations, order
violations) observed in mature real-world parallel applications, as well as a set of conclusions and
facts which set a precedent and a guide for research on modern debug methodologies and tools for
parallel systems.
7 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
One of the goals of EURETILE is to develop the DPSNN-STDP benchmark (Distributed
Polychronous Spiking Neural Net with synaptic Spiking Time Dependent Plasticity). The recent
paper “Cognitive Computing” (Modha et al., 2011) optimally describes the boundaries of the arena
where we intend to play. We resume here a few key references that contributed to define the startingpoint of our work. Song (2000) “Competitive Hebbian learning through spike-timing-dependent
synaptic plasticity” described the importance of precise time ordering between pre and post-synaptic
spiking as a way to capture causal/anti-causal relations between pairs of neurons and decide if to
potentiate or depress an individual synapse. In our opinion, this is one of the key features to be
captured if brain-inspired model of computations have to be applied to real-world problems. The
importance of polychronization, i.e. the relevance of individual axonal delays on computation, is
described by Izhikevich (2006). This is another feature that we deemed essential in our cortical
simulation benchmark, to create an interconnection topology that could inspire advanced models of
hardware interconnection and brain-inspired computational models. We followed Izhikevich (2003,
2004, 2006) also for what concern the model of the spiking neuron. We surely acknowledge the
importance of the SpiNNaker Massively-Parallel Neural Net Simulator (Steve Furber et al. 2008) as
source of thought about the characteristics of a brain-inspired hardware interconnection systems.
As the EURETILE project is a continuation of the SHAPES project (2006-2009), we listed a few key
references about the SHAPES hardware and software tool-chain architecture (Thiele et al, 2006,
2007; Paolucci et al. 2006, Rousseau, Jerraya et al. 2008). The simulation techniques developed in
the SHAPES time-framework are summarized by a few papers (Kraemer, Leupers et al., 2007;
Schumacher, Leupers et al. 2010). A key role in the hardware architecture of the SHAPES tile has
been played by Diopsis MPSOC (Paolucci et al. 2003, 2006) architecture and by the mAgic VLIW
DSP (Paolucci et al. 2001, 2002, 2004, 2008).
8 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3. Publishable summary – 2010-2012 revision
3.1.Project Objectives and Expected Final Results
The EURETILE project (http://www.euretile.eu) investigates brain-inspired, foundational innovations
to the software and hardware architecture of future fault-tolerant and dynamic many-tile systems, to
be applied to those Embedded Systems and High Performance Computers requiring extreme
numerical and DSP computation capabilities.
The project will deliver:
Experimental Many-Tile Platforms for the study of many-tile systems in the scenarios of
Embedded Systems and HPC (QUonG Hardware Platform and VEP-EX simulation platform);
A many-tile programming/optimization environment, to be applied to dynamic, fault-tolerant,
many-process numerical/DSP applications, with foundational innovations;
A set of application benchmarks, representative of both HPC and Embedded System domains,
coded using the new programming environment.
3.1.1. Software Tool-Chain
From a software perspective, EURETILE aims at providing a scalable and efficient extensive
programming framework for many-tile platforms, by exploiting the underlying parallelism and
investigating some key brain-inspired architectural enhancements specific to EURETILE. The
proposed programming environment is based on a new model of computation that explicitly exposes
the coarse-grain and fine-grain parallelism present in applications. Then, automatic tools are aiming
at obtaining predictable and efficient system implementations by optimally matching the concurrency
and parallelism present in applications, with the underlying many-tile hardware. The figure sketches
the envisioned software EURETILE tool-chain. The software tool-chain follows the well-known Ychart approach and in EURETILE each design phase will be enriched with novel concepts for
programming efficiently the three layers of hierarchy present in EURETILE platform. The proposed
software tool-chain aims at optimizing issues related to system throughput while still guaranteeing
real-time constraints for applications that operate under such restrictions. Additionally, faulttolerance aspects are considered, from system-level programming. The way of representing
parallelism, concurrency, and fault tolerance at system level, the implementation of the distributed
real-time operating system, and the efficient fast simulation environment have all a strong impact on
the overall optimality of the system implementation.
The proposed programming framework developed by ETHZ is based on a new model of computation
that matches well the many-tile EURETILE architecture. We have identified three levels of hierarchy
in the “network of processes” describing the structure of the application, in analogy to three levels
identified in the cortical-inspired organization of the hardware, i.e., 1-cortical columns, 2-cortical
areas, and 3-neo cortex. The two upper levels of hierarchy in the network of processes are used to
represent the coarse-grain concurrency. The first level is associated to finer grain parallelism.
Practically, the concept of a distributed operation layer (DOL), applied with success to the multi-tile
SHAPES platform (see SHAPES FP6 project http://apegate.roma1.infn.it/SHAPES and
http://www.tik.ee.ethz.ch/~shapes) will be extended to describe the tremendous concurrency of
different sets of parallel applications running simultaneously (dependently or independently) on the
new brain-inspired many-tile architecture. In SHAPES, DOL was applied to describe static streamoriented applications and for mapping only a single application to a multi-core architecture. In other
9 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
words, the third level of hierarchy as described above was missing. Moreover, in EURETILE we aim
at executing concurrently several dynamic applications, taking advantage of the many-tile distributed
platform.
applications
scenarios FSM
scenario
3
scenario
1
internal
event
scenario
2
external
event
scenario 3
scenario 2
scenario 1
many-core
platform
dynamic
applications
event
dynamic selection
of mappings
(cf. FSM scenarios)
Dynamic
remapping:
- semi-dynamic
(intra-cluster)
- static (processorlevel duplication)
- fully dynamic
fault
fault
event
mapping 3
mapping 2
mapping 1
fault-tolerance
(dynamic)
remapping
design space
exploration
software synthesis
virtual platform simulation/
hardware emulation
Figure 3-1. Schematic View of the Holistic and Scalable EURETILE Software Tool-Chain.
10 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
To answer all these challenges, the DOL programming environment needs to be enriched with
innovative concepts. First, a new programming model that we will call Distributed Application Layer
(DAL) is being developed at ETHZ, corresponding to the third level of hierarchy of EURETILE
platform.
This adds one level of hierarchy above the current process network layer in DOL, allowing the
application programmer to specify concurrent execution of several dynamically instantiated
applications on the same hardware.
Second, the DOL programming model, which still relates to the first two levels of hierarchy in the
architecture, needs to match well the underlying brain-inspired hardware platform. In particular,
planned and unplanned internal and external events will be considered as initiators of dynamic
remapping, exploiting the hierarchical nature of the underlying execution platform and the dynamism
and concurrency inherent to the applications. (Semi-) automated off-line and efficient on-line
mapping methods are to be investigated, including run-time monitoring and optimizations that
consider timing predictability of individual applications as well as efficiency in terms of system
throughput and accelerated executions.
Then, in addition, fault tolerance needs to be part of all aspects of the EURETILE system, i.e.,
hardware platform, programming model, compilation, hardware-dependent software, operating
system, and many-tile simulation. Fault-tolerance aspects are concretely taken into consideration
starting from system-level, in the EURETILE programming environment and dynamic mapping
strategies on many-tiles. In this sense, the DAL environment allows designers expressing faulttolerance aware mechanisms by dedicated programmable routines.
11 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.1.2. Experimental Hardware and Simulation Platforms for Embedded Systems and HPC
The project develops two platforms, which will be used to experiment innovations to future many-tile
software and hardware architectures:
VEP-EX, Embedded Systems many-tile simulation platform. The simulation framework will
be developed by RWTH-AACHEN and will include multiple RISCs networked through a
custom interconnect mesh composed of INFN DNPs (Distributed Network Processors).
The QUonG Hardware Platform for scientific high performance computing. The hardware
platform is networked through a custom interconnect mesh composed of DNPs hosted on
FPGA boards (APENet+, providing innovative fault-awareness and fault-injection features).
The platform includes off-the-shelf boards mounting INTEL multi-core CPUs. INFN will also
explore the addition of GPGPUs.
Figure 3-2. Simplified view of the EURETILE platforms (VEP and QUonG), which will be used for experiments
on hardware and software innovations on future many-tile Embedded Systems (VEP) and HPC (QUonG).
The project will also explore the addition of software programmable hardware accelerators in the
form of ASIPs, using the framework developed by TARGET Compiler Technologies.
12 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.1.3. Definition of a Hierarchical Brain-Inspired Many-Tile Many-Process Architecture
The plan is to match the coarse-grain and fine-grain concurrency of applications with that of the
underlying hardware platform, to obtain predictable and efficient systems.
Figure 3-3 Hierarchical Brain-Inspired Architecture: Conceptual View, applied to the Embedded System case. At
the first (bottom) level of hierarchy, the elementary tile includes the DNP (Distributed Network Processor) for
inter-tile communications, general-purpose processor(s) and local memories. Depending on the target application
domains, the tile can include DSP(s) and ASIP(s). Upper levels of the hierarchy are constructed according to the
distributed memory paradigm.
In particular, we will research on reflecting the three levels of hierarchy of the brain-inspired
architecture (1-cortical columns, 2-cortical areas, 3-neo cortex), in the dynamic network of processes
that describes the structure of the application.
First level of the hierarchy: "Cortical columns" vs. EURETILE elementary tile. The
computational kernel carried on by each process can access in read/write mode all the
variables stored in the tile memories, in analogy with the “columnar computation” enabled by
the all-to-all network among the neurons inside the column. The fine-grain concurrency
(instruction-level parallelism and shared-memory threads) will be exploited at the level of
abstraction corresponding to complex numerical computations, elementary signal or image
processing operators, filters, small scale FFT or matrix computations, inversions, state
machines or any other single-thread of computation, including the simulation of a set of
neighbouring neurons.
Second level of the hierarchy: "Cortical areas" vs. EURETILE arrays of hardware tiles
and processes. In EURETILE computations, it will be possible to declare arbitrary small and
large-scale communication patterns between processes (such as n-dimensional arrays of
13 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
message-passing connected computational processes), which will then be mapped on arrays
of computational hardware tiles. This kind of n-dimensional decomposition of data-structures
into smaller domains is a classical strategy of DSP and numerical algorithms.
Third level of the hierarchy: “Neo-cortex” vs. EURETILE hardware system and
complex software application description. On EURETILE, it will be possible to program a
broad range of complex numerical, DSP and control applications coding them as a network of
high abstraction computational models, as described in the second level of hierarchy. Several
of these mostly independent, but possibly communicating applications, will be concurrently
executed and share the hardware platform. This layer is dynamic in the sense that applications
may dynamically enter or exit the system.
The figure graphically represents the different levels of brain-inspired hierarchy and the associated
concepts in concurrency management, applied to the Embedded System case.
3.1.4. Fault-Tolerance and Scalability
Elementary processors are designed to guarantee a failure rate low enough to sustain the execution of
platforms based on few components. The scaling to many processor systems requires the
management of more frequent failures. Current parallel systems manage faults with a primitive
approach based on period check-pointing (i.e. periodically saving the full state of the application on
external storage) managed by the application programmer. If a hardware fault happens, the hardware
is repaired or excluded, and the previous checkpoint is reloaded, as soon as the hardware system or a
partition is available. In EURETILE, the aim is to automate the handling of the faults. From hardware
point-of-view, the DNP (Distributed Network Processor) will be able to discover critical events,
faulty lines and/or faulty tiles, and to inject synthetic critical and fault events, stimulating a systemic
fault-management reaction. The DNP will monitor the system behaviour, and will signal the failure
to the software environment, for automation of higher-level strategies. We will also investigate the
routing of messages around faulty components.
A key topic is the investigation of the scaling behaviour of the brain inspired system to many-tile
configurations. To such purpose it will be useful to raise the level of abstraction and use a simulator
capable to show the behaviour of high abstraction models of DNPs and processors under different
kind of synthetic traffic workloads and failures.
3.1.5. Versatile Distributed Network Processor
The DNP will be compatible with the platforms above described through appropriate processor
interfaces. A more flexible software-programmable DNP can be designed in the form of an ASIP
(using TARGET’s ASIP design tool-suite). The hardware prototype delivered by EURETILE will be
integrated using a custom interconnection integrated on FPGAs. In principle, DAL (ETHZ) will be
compatible with all proposed platforms, due to the platform-independent characteristics of this
system-level environment. However, in order to apply DAL to effectively programming EURETILE
platforms, fundamental platform-specific support need to be implemented in the hardware-dependent
software (UJF-TIMA), and ensued by simulation capabilities of the fast simulation environment.
3.1.6. ASIP Design
The HW/SW co-design tools operate at the level of elementary tiles in the brain-inspired many-tile
architecture. Its work is based on its retargetable tools-suite for the design and programming of
14 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Application-Specific Processors (ASIPs). Building on this technology, Target contributes to the
project in two major ways. First, new software-programmable accelerators will become part of the
elementary tiles. These accelerators take the form of ASIPs optimised for the typical numerical
kernels of the applications envisaged in the project. Architectural exploration is based on profiling of
application code in the retargetable SDK. In addition, the tools generate efficient RTL hardware
models of the ASIP, enabling a quick implementation on FPGAs. Secondly, Software Development
Kits (SDKs) will be developed for any processors in the elementary tiles that are intended for
numerical data processing, which may include existing DSPs or new ASIP accelerators. Key
elements of these SDKs are an efficient C compiler and an on-chip debugger. Third, ASIP will
contribute to the acceleration of the DNP/APENet+ interconnection system.
3.1.7. Application Benchmarks
A few characteristics of the project determined the selection of the set of benchmarks:
The software environment and the hardware architecture conceived by EURETILE should be
qualitatively and quantitatively driven by the solution of problems selected from both the
embedded systems and scientific computing scenarios;
EURETILE is about future many-tile architectures applied to dynamic scenarios that include
numerically intensive digital signal processing and scientific kernels, individually represented
by Kahn many-processes networks;
We should explore dynamism and fault-tolerance of many-processes applications;
We also expected to exploit hints coming from the brain architecture, a system capable of
extreme parallelism and low power operation.
We decided to fix the following set of benchmarks, selected from 3 application domains:
1. Dataflow oriented / digital signal processing benchmarks. The new DAL environment will be
used to describe the expected dynamism (changing scenarios, e.g. for mobile appliances),
using an explicitly tiled description of the available parallelism. This set of benchmarks is
developed by ETHZ, a partner with a consolidated experience in the solution of optimization
techniques for embedded systems.
2. The DPSNN-STDP (Distributed Polychronous Spiking Neural Net with synaptic Spiking
Time Dependent Plasticity). cortical simulation benchmark, used with 3 purposes:
a. as a source of requirements and architectural inspiration towards extreme parallelism;
b. as a parallel/distributed coding challenge;
c. as a scientific grand challenge.
The study of the brain simulation benchmark will be developed under the direct
responsibility of the coordinator (Pier Stanislao Paolucci – INFN Roma).
3. LQCD (Lattice Quantum Chromo Dynamics), a classic HPC grand challenge, which since
1983 had been driving the development of several generations of massive parallel/distributed
computers. This benchmark will be maintained by the APE group (INFN Roma).
The Brain Benchmark will be coded during the temporal framework of the EURETILE project both
using a standard C/C++ plus MPI environment, as well as using the C/C++ plus XML DAL
environment. This will allow to compare the features of the new environment in comparison to a
classical environment on a benchmark which will be coded “from scratch” using an explicit
description of parallelism.
15 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.1.8. Main expected final results
To achieve this goal, our main expected final results are:
The definition of a hierarchical, scalable many-tile HW architecture, taking inspiration from
the cortical structure, and searching for brain-inspired and fault-tolerant foundational
innovations;
A hardware prototype (QUonG) integrating at least 64 tiles and 128 cores, with support for
fault awareness and tolerance;
A scalable and flexible simulator, enabling the exploration of behaviour under faults and an
easier debugging of many-process applications;
A set of highly representative benchmarks: LQCD (Lattice Quantum Chromo Dynamics),
PSNN (Polychronous Spiking Neural Networks), multi-dimensional FFT (Fast Fourier
Transforms) and DSP/Data Flow Oriented, demonstrating the potential of EURETILE
architecture on a broad class of numerical, DSP, and control applications on both scenarios
(Embedded Systems and HPC);
A many-tile programming environment, where applications can be expressed as dynamic
network of processes and can be dynamically and mapped and controlled in an efficient
manner on the 3-levels brain-inspired architecture (cortical columns, cortical areas, neocortex);
The implementation of many-tile awareness of critical events and faults, and fault-tolerance at
system-level through the support of all software and architecture layers.
3.1.9. Work-packages
The project is structured in ten work-packages:
WP1
WP2
WP3
WP4
WP5
WP6
WP7
WP8
WP9
WP10
Brain-Inspired Many-Process System Software Requirements
Many-Tile Hardware Architecture Specification
Foundational Many-Process Programming Environment
Distributed Hardware Dependent Software Generation
Many-Tile Simulation/Profiling
Innovations on Hardware Intellectual Properties
Challenging Tiled Applications
Software Tool-Chain Integration
Training, Exploitation and Dissemination
Management
16 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.2.Work Performed and Achieved Results in 2010 (First Year)
This section is a record of activities and results obtained during the first year of the project, and the
technical content has been maintained as in the original yearly reports to maintain an historical
track of the project flow. Next sections (3.3. and 3.4) describe the project evolution in 2011 and
2012.
3.2.1. HW Architecture Design
During the first year the HW Architecture design has been performed by INFN in cooperation with
the consortium partners.
WP2 led by INFN and including contributions from other partners, started and concluded during this
first period and results are reported in the deliverable D2.1.
The overall objective of WP2 was to produce a preliminary system specification of the EURETILE
many-tile hardware platforms. The collaboration decided to go for 2 parallel and synergic
development lines: the first in the area of embedded systems while the other in the area of HPC
systems.
From the software point of view, the unifying elements are the common software tool-chain (from
the high level programming mechanism to the HDS) and the simulation framework while the main
hardware related developments is in the area of scalable, fault tolerant, brain-inspired network
capable to be interfaced to custom ASIP, optimized for specific scientific applications (specifically
for LQCD and PSSN in the EURETILE case studies), and/or commodities computing accelerators.
3.2.2. DNP Design
This area of activity is covered by INFN. In particular, in the area of DNP enhancement main goals
for 2010 were:
Design of a DNP-based 3D Torus network for HPC systems through:
o VHDL Coding of DNP interface to HPC PCI Express based host system
o DNP-uP integration to optimize RDMA support for INTEL-based host system
Porting of DNP 3D Torus network controller on FPGA platform
Design of a 6 channel FPGA-based electronic card for HPC system
During 2010, DNP activities were executed in collaboration with the APEnet+ project. The INFN
initiative called APEnet aims to build high performances custom networks for HPC commodities PC
Clusters using ideas, know-how and IPs developed in the framework of the APE project.
APEnet+ is the last generation of NIC boards
(FPGA based) implementing a 3D Torus
network for PC Cluster.
During 2010, we progressed toward the
EURETILE HPC platform, which foresees the
development of an FPGA-based network card
(a 3D Torus network router using the DNP
architecture) to interconnect commodities PC
clusters. During this first year, we synthesized
the current DNP on the target FPGA ALTERA
STRATIX IV EP4SGX 290 device obtaining
preliminary but encouraging results (clock
Figure 3-4. DNP test board (APEnet+) developed during frequency and footprint). Then, we redesigned
2010.
the DNP serial-link interface to be compliant
17 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
with the target Ser/Des, for a total speed of 34Gbps per toroidal link. The DNP-PCI Express interface
was developed and a preliminary test of integration of microprocessor for better support RDMA
operations was performed. A complete design of the APEnet+ prototype board was the result.
3.2.3. LQCD and PSNN ASIP Design
During the first year of activity, the ASIP SDK technology has been extended by TARGET with new
methods for checking the suitability of a processor model for C compiler support, advanced on-chip
debugging capabilities and a new graphical integrated development environment. Then, an initial
feasibility study was made on the design of two ASIPs for the scientific numerical processing tasks
envisaged in the project (LQCD and PSNN).
3.2.4. System-Level Programming Framework
In 2010, DAL and its main concepts have been specified and the corresponding DAL API has been
established. The DAL specification document, including the basic DAL description, API, and
interfaces with other work-packages has been sent to all partners, and the key issues are summarized
in deliverable D1.1 “Report on Brain-Inspired Many-Process System SW Requirements” (delivered
at M12). During 2010, we developed a first implementation in a distributed functional SystemC
simulation that can check the correct functionality of multiple concurrent applications, dynamic
mapping and re-mapping strategies, as well as some fault-tolerance features. For the following
EURETILE project periods, (i.e. starting from 2011) we planned an advanced implementation of the
distributed functional simulation to be used by all partners, as basis for a bug-free application
development, support for the final HdS implementation, and source of first results in terms of
performance analysis and system optimization.
3.2.5. Scalable Simulation Environment
This area of activity is led by RWTH, in cooperation with the other partners. The main target of the
scalable simulation environment is to support the development and validation of system
hardware/software and applications. This activity comprises four different areas: Abstract simulation,
parallel simulation, debugging support and fault injection. Since the beginning of the project, work
on the following topics has been performed:
Abstract simulation: The Hybrid Processor Simulation (HySim) has initially been deployed on
ARM and magic processors within the SHAPES project. In the meantime, the HySim engine has
been restructured for easier porting and improved to support more complex applications, such as
those that will be part of EURETILE. In addition, an extension to this concept for the purpose of fast
Network-on-Chip simulation has been investigated on an exemplary tiled platform. A satisfactory
initial model has been tested and a full integration in the simulation flow will follow.
Parallel simulation: In 2010, research on parallel simulation that was started during the FP7
SHAPES project has been continued. Furthermore, a mixed-level parallelized simulation of a tiled
computing platform has been created. It consists of basic tiles with processing elements and local
memory, which are connected by packet routers. The simulation was exhibited at DAC 2010, and the
technical details were presented at CODES/ISSS 2010.
Debugging support: Novel debugging techniques to lower the complexity of MPSoC software
development on the EURETILE system were specified and presented to the project partners. Such
techniques are based on the advanced visibility, controllability and determinism of virtual
prototyping, and aim at finding software malfunctions caused by concurrent inter-tile and intra-tile
18 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
activity. The implementation of debugging support to cope with current simulation technology used
in EURETILE will be covered during next project stages.
Fault injection: Discussions are on going with hardware and system software partners, regarding
which kinds of faults can be handled by the EURETILE system. These faults will then be modelled
and supported for injection using the system simulator implemented.
3.2.6. Hardware-Dependent Software
UJF-TIMA is responsible of providing HdS for the SW tool-chain able to generate binary codes for
processors on the HW platform. We define hardware-dependent software (HdS) to be “software
directly dependent on the underlying hardware.” UJF-TIMA is in charge of providing HdS, which
includes OS services and low-level software.
The main challenges for HdS generation come from the brain-inspired paradigm and the hierarchical
architecture. This may imply some dynamic changes during execution, such as new task running on
processor for new application or a failure detected somewhere in the architecture, as the brain is
naturally fault-tolerant with redundancy. For that reason, we are dealing with fault tolerance aware
software services provided by HdS and real time properties as well as new control mechanisms. New
control mechanisms concern the overall architecture and control organization: distributed, centralized
or a hybrid version. In 2010, this has been still an on-going work.
An emerging solution for fault tolerance seems tasks migration. But this is really challenging in
NUMA architecture, non-SMP, and in a hierarchical architecture. Task to be migrated is supposed to
have reached a specific point of execution, and a copy of all what is needed to restart this task on
another processor (stack, registers, memory, task code, cache?) will be sent. The starting point of
EURETILE is an in-house OS (called DNA-OS developed for the previous FP6 SHAPES project)
running on each processor of the architecture. During 2010, DNA-OS has been extended to include
all what was missing for SMP system and task migration, that was the first step to move forward the
EURETILE context. DNA-OS is running on ARM processors, as well as other architectures (SPARC
V8, MIPS R3000, mAgicV,)
Control mechanisms and task migration are based on an efficient communication infrastructure and
SW drivers. This has been consolidated with an automatic driver selection and specialization
regarding communication protocol selected and requirements.
19 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.3.Work Performed and Achieved Results in 2011 (Second Year)
This section is a record of activities and results obtained during the second year of the project, and
the technical content has been maintained as in the original yearly reports to maintain an historical
track of the project flow. Next sections (3.4) describe the project evolution in 2012.
3.3.1. Definition of Fault Aware Architecture
During 2011, INFN proposed to the consortium a novel HW design paradigm, named “LO|FA|MO”,
which creates a systemic awareness of faults and critical events, thanks to a distributed approach that
uses additional hardware components on each DNP, and needs dedicated software components
running on each tile to create a local awareness of faults and critical events. This local awareness is
then propagated along the system hierarchy. Onto such local and systemic fault awareness, a software
approach to fault reactivity can be grounded, which has been proposed by ETHZ at the end of 2011.
From a system-level perspective, two fault reactivity strategies have been investigated in WP3 and
first ideas have been implemented in the Distributed Application Layer (DAL), namely fault recovery
and fault tolerance.
Figure 3-5 - The Network Processor of each leaf in the many-tile HW system is equipped with its own LO|FA|MO
components. The Local Awareness of faults and critical events is propagated towards the upper hierarchy levels,
creating Systemic Awareness. Reactions to faults and critical events could be autonomously initiated by sub-system
controllers or could require a systemic response.
3.3.2. Challenging Application Benchmarks
We have three benchmarking areas: DSP/Data Flow Oriented, PSNN-Brain Simulation and Lattice
Quantum Chromo Dynamics.
20 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
During 2011, several application benchmarks have been implemented as distributed application layer
(DAL) specifications.
About DSP/Data Flow Oriented, at ETH Zurich, several DAL-compliant applications have been
specified, of which the most significant are a picture-in-picture application consisting of two parallel
MJPEG decoders, and the kernels of a fast-Fourier transformation, and matrix multiplication. These
applications are included in the DAL distribution and can be downloaded from
http://www.tik.ee.ethz.ch/~euretile.
Topologically, the applications consist of a task-level parallel specification in the form of individual
process networks and an application-level parallel specification, expressing dynamic application
scenarios, in the form of a finite state machine. The DAL framework is then capable of deciding on
the mapping or remapping of applications/processes as well as to investigate some key features
related to application’s fault management. In particular, case studies of the picture-in-picture
application have been designed to demonstrate the parallel execution of multiple applications, as well
as basic DAL actions such as starting, stopping, pausing, and resuming the different videos, together
with the implementation of the control infrastructure. Moreover, distributed mapping and remapping
of individual component processes are investigated. Basic fault management can be illustrated at
mapping level, where applications residing on damaged cores are smoothly redirected towards
reserved healthy cores.
These investigations open the way for further research and design options, such as process level
redundancy for critical applications, and the on-line exploitation of pre-calculated performance tradeoffs.
PSNN. About the study of Brain Simulation, used as benchmark and as a drive source of
architectural concepts, during 2011 we produced a first version of D-PSNN, the natively Distributed
– Polychronous Spiking Neural Network Benchmark. The benchmark has been compiled and run to
initialization a “toy-scale” network in two frameworks: the DAL environment (C++ plus XML) and a
“standard” C++ plus MPI context. The coding has been executed after:
1- The production of a list of “computationally Inspiring” review of biological facts from
experimental and theoretical neuroscience;
2- A complete “reverse engineering” of a sequential reference PSNN code, to grasp all
details and understand the blocking point against parallelization.
LQCD. The INFN team has a multi-decennial experience on Lattice Quantum Chromo Dynamics,
and a multi-process representation of the problem available. Starting from 2012, LQCD will be used
to validate the new features associated to dynamism and fault tolerance developed by ETHZ in 2011.
During this year, we performed a quantitative reassessment of bandwidth and latency requirements,
taking in account the characteristics of the EURETILE HW platform.
3.3.3. Innovation on HW Intellectual Properties and design tools
In 2011, in addition to the definition and first implementation of the LO|FA|MO fault awareness
system, INFN main achievements were the delivery of a refined DNP component and the
introduction of APENet+ and QUonG hardware prototypes. On the ASIP development side TARGET
designed a specific ASIP architectures, optimised for application domains targeted by the project.
The DNP released at the end of 2011 is a powerful, robust and faster design thanks to the
introduction of specialized hardware to speed-up PCIe side transactions as well as the 3DTorus
transfers.
The APENet+ final board, integrating the last release of DNP IP, was produced in pre-series (4
boards) and validated with extensive test sessions.
21 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Figure 3-6. APENet+ Board produced in 2011.
INFN also put in production a reduced prototype of QUonG system, the hybrid parallel cluster
selected as the HPC Euretile platform.
3.3.3.1. DNP Innovation.
During 2011 INFN executed several refinements of the DNP IP in order to get a more robust and
faster design.
In particular we focused on the optimization and improvement of the “network interface” block
adding dedicated hardware logic to improve RDMA tasks execution and speeding-up the recently
introduced hardware block implementing GPU peer-to-peer protocol (a very brand new concept for
the current commercial network cards).
Some work was also devoted to the enhancement of physical links interface and to the optimization
of the FPGA embedded microprocessor (NIOSII) firmware.
In addition we also started to define the architecture and planning the design of the hardware blocks
required to implement the LO-FAMO fault tolerant mechanism in the EURETILE platform
demonstrator.
APEnet+ innovation. INFN introduced the first release of the APEnet+ board, the DNP-based, PCIe
board implementation of the 3D Torus network, featuring 6 fully bidirectional links with 34 Gbps of
raw bandwidth per direction. With the integration of the P2P Enabled DNP into its high-end FPGA,
APEnet+ is the first non-NVidia device with specialized hardware blocks to support the NVidia GPU
direct peer-to-peer inter-GPU protocol. During the 2011 we produced and tested a set of 4 APEnet+
prototypes on which we performed a preliminary bandwidth and latency analysis obtaining very good
results in terms of access latency reduction.
QUonG innovation. QUonG, the successor of the former APE supercomputer and the prototype of
the HPC EURETILE platform, is a hybrid INTEL server x86-based cluster accelerated using NVidia
GPUs and interconnected by APEnet+ 3D Torus network.
22 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
In 2H11 we initiated the procurement phase of QUonG reduced prototype (25 TFlops) to be upgraded
to a full “QUonG tower” system, capable of a peak performance of ~65 TFlops, during 2012 and
2013. The QUonG tower will be made available to the Euretile collaboration to perform tests and
software development.
3.3.3.2. ASIP Design Tool Improvement.
The Euretile consortium’s approach towards the design of ASIP is to use a retargetable tool-suite
based on a processor description language. The tool-suite supports architectural exploration, the
generation of a software development kit, and the generation of an RTL hardware implementation of
the ASIP. TARGET’s IP Designer tool-suite is used for this purpose.
In project year 2011, TARGET has developed a number of new extensions and optimizations for its
retargetable tool-suite. These activities are part of WP6. These developments focussed on two main
topics:
A new methodology has been defined to model memory and communication interfaces of
ASIPs.
New techniques have been developed and implemented to provide enhanced feedback to
users of the retargetable tool-suite during the architectural optimisation process.
3.3.3.3. LQCD ASIP Design.
In project year 2011, we started the development of specific ASIP architectures, optimised for
application domains targeted by the project. These activities are part of WP6, and were carried out
by TARGET in cooperation with INFN. More specifically, an efficient ASIP for LQCD has been
designed using the IP Designer tool-suite. The ASIP has been named “VCFLX” (referring to Vector
Complex Floating-point – Flexible”). Benchmarking shows that a single VCFLX ASIP can
implement LQCD 33x faster than a 32-bit floating-point CPU and 4x faster than the mAgicV DSP.
About 50 instances of the VCFLX ASIP can fit on a single Virtex-7 FPGA, thus speeding up LQCD
even further, provided that the FPGA’s RapidIO fast communication channels can be used to achieve
a sufficiently high communication bandwidth.
3.3.4. System-Level Programming Framework (DAL)
During 2011, in the frame of WP3 led by ETHZ, the second complete prototype of the distributed
application layer (DAL) has been implemented and the consolidation of DAL underlying concepts
has been done in line with all project work-packages and tools.
Distributed application layer (DAL) is a system-level programming environment for many-core
architectures that considers dynamic application scenarios, i.e. applications that may appear and
disappear dynamically at runtime. At the same time, the approach allows designers to handle at
system-level faults that occur at runtime. In particular, besides mapping optimization of applications
to the EURETILE platform, DAL provides the environment for remapping or restarting applications
in case of changes in the application scenarios or in case of permanent system faults. However,
providing timing guarantees for applications remapped/restarted due to faults that occur arbitrarily at
runtime is hard. Therefore, in DAL additional methods are investigated such that specification,
generation, and analysis of redundant task graphs for so-called critical applications with real-time
constraints.
As a result of 2011 efforts, the second DAL prototype has been made available including the
complete specification, a functional simulation that is automatically generated from the initial
specification, several case studies demonstrating the capabilities of DAL, and a complete
23 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
documentation, see http://www.tik.ee.ethz.ch/~euretile. As well, to make the tool-chain complete,
basic tools are included for performance analysis, search in the design space, and mapping
optimization under specific platform constraints.
The key challenges encountered by application programmers when designing applications using DAL
are in a fist instance exposing the process-level parallelism in the application, and second, exposing
application-level concurrency and dynamism in a formal programming model. Additional challenges
are in exposing fault-awareness of the system early in the design and deciding among proposed fault
management strategies. To answer all these challenges, the DAL functional simulator has been
developed as a tool for application programmers, to test and profile their applications early in the
design, on top of an ordinary computer system.
Technically, the DAL functional simulator version made available at the end of 2011 has been
running on top of Linux and uses POSIX threads to execute DAL processes that simply communicate
via the shared memory of the Linux system. Additionally, a specific process network designed for the
hierarchical control of application-level dynamism is first computed out of the specification,
automatically generated during the synthesis step, together with the functional simulator, and then
loaded when the system is booting-up. The design space exploration tool provides the mapping
information. It consists of a set of mappings, each one valid for a set of application states. This set is
compiled into the master controller that provides the basic process management instructions for the
slave controller, such as starting, pausing or restarting a task as well as installing communication
channels and the corresponding memory.
In the proposed functional simulator virtually all main DAL concepts can easily be tested. Therefore,
such a functional simulator is enabling the fast development and test before porting the DAL design
to the target EURETILE platform. Moreover, such a fast development prototype is opening the way
for further research in the areas of performance analysis, mapping/remapping optimization, and faultawareness offering all functional capabilities of the EURETILE system, and hiding all low-level
details and parallel development of the hardware-dependant software.
3.3.5. Scalable Embedded Computing Simulation Environment (VEP)
During 2011, work has been performed by RWTH related to the simulation environment on the
following topics:
VEP-EX simulator: The base version of the VEP-EX simulator has been realized and distributed to
the partners. The VEP-EX simulator offers two different selectable computing elements: Either the
IRISC processor instruction set simulator is used, or alternatively an abstract execution device (AED)
is employed.
The AED is a SystemC module that can load host-compiled shared libraries containing code to
execute inside the simulator environment. This mechanism is used to ramp up DNP drivers and test
software before the actual tool-chain for the RISC processor becomes available. Introducing the AED
is additional effort not originally part of the project plan. It has become necessary as a bridging
solution until the RISC software tool-chain becomes available.
Debugging support: A debug technique for concurrent software running on heterogeneous multicore systems was developed in order to set the grounds for a preliminary debugger framework for the
EURETILE system. This new technique, named Event-based Bug Pattern Descriptions (EBPD), adds
automation for debugging by using assertional techniques similar to those in formal verification and
can be easily retargeted to cover different programming models and systems. Some building blocks
of the underlying debugger architecture, like the BPD grammar, the BPD compiler, the Event
Monitor (EV) and the Sequential Consistency Analyser (SCA), were developed and evaluated in
24 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
simpler multi-core systems. Common bug patterns in concurrent software were also described using
the BPD grammar in order to start populating a database of well-known patterns.
Abstract simulation: The HySim technology, legacy of SHAPES, has been enhanced in two ways.
Firstly, the framework has been redesigned to support a class of conceptually different processor
architectures, namely customizable cores, and a more profound way of processor state
synchronization between the native and the target processors has been introduced. Secondly, a time
synchronization mechanism has been introduced, to enable the support for multi-core or any
platforms relying on timing, like those including watchdogs or timeouts. Apart from that, some
preliminary experiments were conducted on a multi-core TLM2.0 platform, to evaluate the hybrid
NoC simulation.
Parallel simulation: In collaboration with Synopsys, Inc., parallel simulation has been further
researched, with focus on applicability to systems currently used by industry.
Fault injection: Faults possible to simulate have been investigated and identified, and there are no
known conflicts with other work packages. As a first step, packet probe blocks have been introduced
into the VEP-EX simulation model. Such probes allow dropping or corrupting packets on links
between two given DNPs in a custom fashion.
3.3.6. AED integration with DNP driver software
During 2011, RWTH and INFN ported DNP device drivers previously written by INFN to the AED.
First, the basic DNP RDMA API was ported, which enabled AED software to conduct DNP
PUT/GET/SEND operations. After the low-level RDMA API was put into place, the next-higher
level Presto communication API has also been ported to the AED, which now enables AED software
to conduct synchronous blocking send and receive operations using the DNP 3D torus network.
Communication partners are addressed by their respective communication rank in the Presto API.
Further on, being able to run DNP driver code natively on the host using the AED simplifies
debugging of such driver code. This is especially true in the event of many cores communicating with
each other simultaneously.
AED test applications are currently used as software to further develop the fault-aware DNP model.
3.3.7. Hardware-Dependent Software (HdS)
We can split the TIMA contribution and work in 2011 in two main points: port of DNA-OS on top of
the target processors and platforms, and specification of research development to reach the overall
objective of EURETILE.
Port of DNA-OS: We have ported DNA-OS on top of two new processors: iRiSC (RISC) and x86
(CISC). The port on top of the iRiSC has been started in the last quarter of 2011, after getting
development tools from Synopsys for the iRiSC processor, including compiler, linker, simulator, As
no architecture has already been provided during 2011, only the kernel of DNA-OS has been ported.
C library and peripherals (drivers) are still an on-going work and remain to be done in 2012.
For x86, the kernel has been ported on the x86 architecture with standard peripherals like timers.
Standard library is also de facto available after re-writing specific functions (printf,) to fit with
peripherals. Moreover, we manage to provide a booting version on USB stick or on x86 simulator
(QEMU based) with a simple application using semaphores, multi-threading, timers
Regarding TIMA main contribution on EURETILE, providing task migration on top of NUMA (and
non-SMP) architectures, we are looking at virtualization to provide multi-application support as well
as other features previously planned for this project. We are following two different tracks:
improving DNA-OS to support virtualization, or using an existing virtualization solution (OKL4
25 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
provided by NICTA-lab, Sydney, Australia) and adapting it to our requirements. Unfortunately, the
OKL4 version we found freely available in 2011 has been an old version that needs an old software
tool chain version and it runs only on ARM. As a conclusion, we are heading for a modification of
DNA-OS, to provide virtualization support, as well as support for the DAL API described in the
deliverable D_WP1_R1. The integration of DAL API for task migration support was still under
development at the end of 2011.
26 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.4.Work Performed and Achieved Results in 2012 (Third Year)
3.4.1. System-Level Programming Framework (DAL)
During 2012, in the frame of WP3 led by ETHZ, the distributed application layer (DAL) has been
finalized and its key concepts published. DAL is a scenario-based design flow that supports design,
optimization, and simultaneous execution of multiple applications targeting heterogeneous manycore systems. In particular, the deliverable document produced by WP3 provides detailed insights in
the high-level specification model for dynamic systems targeting many-core systems. Afterwards, a
multi-objective mapping optimization algorithm is proposed for DAL. As the optimization problems
seen in dynamic systems are relatively complex, the proposed technique uses problem decomposition
to provide scalability. As many-core systems are prone to high chip temperatures, providing
guarantees on maximum temperature is as important as functional correctness and timeliness when
designing many-core systems. Thus, two thermal analysis methods for many-core systems are
proposed in the WP3 2012 deliverable document. The first method estimates the temperature based
on a set of application-specific calibration runs and associated temperature measurements using
available built-in sensors. The second method is a formal worst-case real-time analysis method to
provide safe bounds on the execution time and the maximum chip temperature. Finally, a hierarchical
control mechanism is proposed and it is shown how the hierarchical control mechanism can be used
to implement an automatic fault recovery mechanism.
All the discussed concepts have been implemented in a prototype of the distributed application layer.
In order to model the high-level specification, DALipse, a software development environment for
DAL, has been developed in the frame of WP3. As a result of the 2012 efforts, DALipse is available
as a plug-in for Eclipse. In addition, SADEXPO, a state- and architecture-based decomposition
framework for mapping optimization is integrated in DALipse. In order to execute applications
specified using the DAL paradigm on a distributed Linux cluster, a run-time environment has been
implemented using the message-passing interface (MPI) as a communication layer. Finally, a
software tool-chain has been developed that uses as input the high-level specification of a DAL
benchmark and automatically generates code for the selected target architecture, i.e., a distributed
Linux cluster.
3.4.2. Hardware-Dependent Software (HdS)
The HdS activity during 2012 could be split in 3 different parts: basic architecture features (for x86based QUONG architecture and iRiSC-based VEP architecture), the DAL2binary tool chain
generation code, and the state of the art of task migration for embedded systems.
We can summarize in few words the main differences between the 2 architectures:
- x86-based QUONG platform: This architecture is based on Intel multi-core x86 architecture.
The DNP is accessible through the PCIe interface, as well as peripherals. This supposes the
support of PCIe for our own operating system DNA-OS.
- iRiSC based VEP architecture: This architecture is based on the iRiSC processor, with one
processor per tile. Each tile includes the DNP, which allows connecting several tiles. There is
no PCIe anymore, so the driver for the DNP is directly called by DNA-OS.
For the x86-based QUONG platform, we have done several developments in 2012.
- DNA-OS is now able to boot on multi-core Intel x86 processor.
27 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
-
The PCIe support is now available in that context.
We have developed a command interface (as an added DNA-OS service) to facilitate
debugging or in-line command in a terminal. This includes a keyboard management.
We have implemented of the DAL API: start, stop, pause, resume allowing to start, stop,
pause and resume task on one processor.
The support of Ethernet connection has been validated for x86 architecture in DNA-OS. It
includes the lightweight IP (open source TCP/IP stack for embedded).
Interrupt management support by DNA-OS for x86 standard architecture.
1 other specific development for the demos: graphical windows on bare metal x86 hardware
architecture used by DNA-OS for frame buffer management of video screen.
For the iRiSC-based VEP architecture, we have done one important development in 2012 as well.
- DNA has been ported on iRiSC architecture and validated.
- The DNP driver has been developed and validated on the iRiSC-based architecture. It is worth
mentioning that this driver has been first validated on our own ARM-based simulation
platform that includes DNP for multi-tile simulation environment.
During 2012, the first version of the DAL2binary tool chain has been provided. This tool chain takes
as input the DAL model, and depending on the target, it generates C code that should be compiled
with DNA-OS and requested services to get the final binary code for each processor.
Finally, the state of the art regarding the task migration for embedded system has been done as well
to fit with the requirements of the EURETILE project depending on the target architecture.
3.4.3. Scalable Embedded Computing Simulation Environment (VEP)
During 2012, research activities in simulation and debugging technologies were carried out that
improved the VEP environment in terms of efficiency, scalability and usability for the project. The
following concrete activities were performed in the simulation ecosystem:
VEP-EX: On one hand, the VEP was extended to facilitate the validation of EURETILE’s main
brain-inspired concepts. The simulator has been provisioned with a service network that completes
the necessary artefacts to develop fault monitoring and awareness features across different
EURETILE layers (i.e. DNP and HdS). On the other hand, new features to improve the VEP’s
usability during software development tasks were added. User-configurable loggers and a
visualization interface have been developed in order to provide better tracing, analysis and
visualization of the software under execution and its output. Finally, some internal components of the
VEP-EX were optimized in order to achieve higher simulation speed. Included optimizations help to
avoid overhead of unnecessary events inside the SystemC kernel and data transactions in intercomponent communication routines. With these new optimizations alone the simulator executes ca.
2.2X faster than the previous version.
Fault Injection: A fault injection framework has been created and coupled to the VEP-EX. The
framework allows users to inject faults to the system with specific temporal and spatial
characteristics, while avoiding the manipulation of complex simulator source code. By using a
formalized XML specification of faults, it is possible to force the manifestation of the same faults
28 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
repeatedly across different simulation executions. This guarantees deterministic faulty scenarios and
facilitates the development of fault aware components in different EURETILE layers.
Parallel Simulation: The parSC kernel, which was conceived in SHAPES and matured during the
previous years of EURETILE, has been used together with the VEP-EX in order to accelerate its
execution speed. The use of parSC in EURETILE posed new challenges to guarantee the simulator
determinism and its compatibility with legacy sequential models. New technologies were developed
(i) to facilitate the integration of legacy models in a practical way and (ii) analyse nondeterministic
traits in SystemC code that might lead to functional anomalies (i.e., the SCandal tool). These new
technologies were presented in FDL’12 and HLDVT’12, and a joint paper by RWTH, INFN and
Synopsys was accepted for publication at VIPES’13. After guaranteeing the simulator determinism,
the VEP-EX with parSC was found to achieve a speed-up of ca. 7.77X over the previous version
(including the other speed optimizations added to the VEP).
Debugging and Profiling: The event-based multi-tile debugger architecture developed in previous
years was improved to attain higher retargetability and scalability. Event monitoring in the
framework was re-designed as a new component-based system, which facilitates the definition of
custom monitors for different OSs and target processors. A well-defined event-based intermediate
representation (EIR) was also developed that facilitates the abstraction of complex low-level details
of target hardware and OSs into events at a level of abstraction useful for systematic debug. To attain
higher scalability, the debugger’s structure was organized into a tree-like network of event monitors
that allows distributing and parallelizing the debugger itself when dealing with massively parallel
systems. Some of the components of the debugger infrastructure have been ported to target the VEPEX processing components. The new multi-core debugger architecture was presented in S4D’12.
Additionally, a new scriptable, interactive command line interface (CLI) was added to the VEP-EX
to allow defining custom automatic debug and profiling tasks. The CLI allows extending basic debug
and inspection APIs of the VEP through the TCL programming language.
3.4.4. HW support to Fault Awareness
Fault-tolerance issues has been addressed in EURETILE since 2011 with the LO|FA|MO HW/SW
approach, whose principles presented in deliverable D6.1 (to be published) have been applied in the
DNP SystemC model and lately demonstrated over the VEP simulator at the 2012 review meeting
(activity reported by D5.2).
The “fault awareness” activities in 2012 focused on the implementation of LO|FA|MO mechanism
for the EURETILE HPC experimental platform, and resulted in a consolidated specification of the
LO|FA|MO approach, the integration of the VHDL code of DNP Fault Manager in APEnet+
hardware, the introduction of a controller of the diagnostic messages flow over the 3D Torus
Network (the Link Fault Manager, LiFaMa), and the first implementation of the Host Fault Manager
as a Linux daemon. In the current implementation the Nios II micro-controller performs the reading
process of the FPGA sensors over the Avalon Bus and we foresee to implement a technique to
transfer the information to the DNP status registers. The new functional blocks were also verified
through an extensive test session.
3.4.5. DNP and APENet+ Innovations
During 2012, the integration of the complete DNP IP in APEnet+ hardware system allowed real tests
of APEnet+ architecture in the final environment.
29 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
First, INFN released a library of test programs based on synthetic and real application kernels and
performed extensive test sessions to measure bandwidth and latency in different conditions.
Benchmark results were used to optimize specific hardware blocks impacting on system
performances. In particular we executed some re-design activities on 3D links sensible obtaining
channel bandwidth enhancement and a more robust transfer protocol.
Further activities were also devoted to optimization of the APEnet+ host interface where the PCIe
protocol shows high latency that has to be controlled and minimized: the Host Interface TX block has
been completely re-designed and the software interface accelerated moving some critical tasks to
FPGA custom hardware blocks. In cooperation with TARGET, preliminary evaluation of effects of
introduction on RX RDMA path of a TLB (Translation Look aside Buffer) for virtual to physical
address translation has been performed.
The P2P GPU interface was improved and heavily tested achieving good results in terms of
bandwidth enhancement and small packets latency reduction.
In order to explore future systems development activities in 2012 we executed the porting of our
DNP IP design on state-of-the-art FPGA devices that, built using a 28nm silicon process, show
impressive amount of internal user-available resources and huge number of high speed transceivers.
The preliminary evaluation and synthesis activities showed very encouraging results: the DNP logic
occupation is limited to 15% of a medium size devices (allowing architecture improvement through
introduction of new computing/specialized hardware) while its I/O throughput can double thanks to
the use of the 28nm devices improved transceivers.
Software activities focused on system software bug fixing, performance optimization and
development of new modules supporting hardware innovations.
Highlights here are the release of a VEP interface for Presto programming environment and the
development of APEnet+ API OpenMPI layer to support DAL on APEnet+ platform. Furthermore, in
collaboration with UJF, a DNA-OS prototype driver specific to the APEnet+ hardware has been
analysed and a simple program produced by the DNA-OS tool chain, running on x86 and setting
selected APEnet+ board registers through the PCIe was coded.
Several exploitation activities are also in place from 2012. In particular we demonstrated the
advantages of DNP/APEnet+ architecture introduction in high-level trigger systems of the current
and future High Energy Physics colliders. In this framework the APEnet+ features, one for all the
GPU Direct P2P mechanism assuring low access latency to numerical accelerators, and its
implementation FPGA-based allowing IP re-configurability and specialization, can reduce data
transport overheads and can increase the time budget for trigger computing.
3.4.6. ASIP support to RDMA
EURETILE’s many-core platforms critically depend on efficient packet-based communication
between tiles in the 3D network. Efficiency refers to both high bandwidth and low latency. Our
observation for the HPC platform is that the current DNP implementation, which includes firmware
running on a NIOS soft-core from Altera, offers high-bandwidth communication, but with a too high
latency. To reduce the latency, a significant acceleration of networking functions (e.g. buffer search,
virtual-to-physical address translation) is needed.
During 2012, TARGET in cooperation with INFN started the development of a softwareprogrammable DNP architecture using ASIP technology. These activities are part of WP6. A first
version of a DNP ASIP has been modelled and optimised, using the IP Designer tool-suite. The basic
30 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
architecture is a 32-bit microprocessor. Adding dedicated hardware and instructions to accelerate the
buffer-search functionality for RDMA tasks has further optimised the architecture. As a result, the
cycle count of typical buffer-search firmware programs has been reduced by more than 85%
compared to the existing implementation on NIOS. Currently the clock frequency of the ASIP is still
lower than that of NIOS. Nonetheless, buffer search already executes significantly faster on the
ASIP compared to NIOS, and moreover further cycle-time improvements are within reach.
3.4.7. ASIP Design Tool Improvement.
EURETILE’s approach towards the design of ASIPs is to use a retargetable tool-suite based on a
processor description language. The tool-suite supports architectural exploration, the generation of a
software development kit, and the generation of an RTL hardware implementation of the ASIP.
TARGET’s IP Designer tool-suite is used for this purpose.
During 2012, TARGET developed new extensions and optimisations for the retargetable tool-suite,
with specific relevance to the other Euretile activities. These tool development activities are part of
WP6. In particular, we focussed on the following topics:
We continued the work on the development of advanced capabilities for modelling and
synthesis of I/O interfaces of ASIPs. While the basic concepts behind this new approach
were already reported last year, this year we have delivered a first implementation of the new
concepts, and tested it by modelling and synthesising first example I/O interfaces.
The tools have been extended with
capabilities to model multithreaded architectures.
This
includes
the
generation
of
hardware support for storing the
context of a thread upon a context
switch, as well as C programming
aspects.
Finally we developed and implemented
additional capabilities to provide feedback
to users of the retargetable tool-suite
during the ASIP architectural optimisation
process. Specifically, the tools can now
analyse and report about connectivity
issues within an ASIP architecture.
3.4.8. HW Experimental Platform
Prototype (QUonG)
During 2012, INFN completed the
procurement and assembly phase of a 16
nodes QUonG system, a 16K core, 32
TFlops peak single precision computing
platform with a 4x2x2 topology. All
required servers and GPU accelerators
were procured from H1-2012 while a
batch of 15 APEnet+ boards were
delivered in September 2012 and put in Figure 3-7. 16 nodes QUonG prototype in 2012
31 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
place in H2-2012.
Furthermore, a 12 APEnet+ boards addictive production was launched in H2-2012 and the boards
will be delivered in February 2013.
In parallel a reduced QUonG platform prototype (8 nodes, 2x2x2 topology) was heavily tested to
validate systems hardware and software and to provide a prototype system to develop applications.
In addition to the EURETILE scientific applications and benchmarks, the QUonG prototype (with
and without APEnet+ network) was validated through the successful execution of several grand
challenges application kernels among them the Heisenberg Spin Glass code (HSG), the Graph500
benchmark, Laser-Plasma Interaction simulation and preliminary HEP trigger related algorithms.
3.4.9. DPSNN-STDP: Distributed Polychronous and Plastic Spiking Neural Net
Figure 3-8. Collective spiking rastergram and individual neuron activity produced by DPSNN-STDP.
32 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
During 2012, we designed and implemented the C++ code of a complete prototype of the DPSNNSTDP simulator (Distributed Polychronous Spiking Neural Net with synaptic Spiking Time
Dependent Plasticity).
Then we run the full simulation cycle (initialization and dynamic phases of a network including 10^5
synapses) in the C++ plus MPI environment on a QUONG prototype, using an industry standard
interconnection system (InfiniBand switched fabric), reproducing an identical behavior with different
number of processes.
We ported and run the initialization phase of the DPSNN-STDP code under the DAL plus C++
environment, and executed it on a prototype environment where two QUONG nodes were
interconnected by two APENet+ card. We used the same C++ classes as building blocks to construct
the DAL and MPI network of processes. This will permit, during 2013, to start a comparison activity
between the APENet+ interconnect system and the InfiniBand system, and among the performances
offered by the DNA-OS, MPI and Presto message passing software layers, when applied to the
DPSNN-STDP benchmark. Then we will identify the middleware and hardware bottlenecks, and
derive architectural improvements for future generations of interconnect systems.
3.5.Plans for 2013 - 2014 activities
The prototypes of the Experimental hardware platform (QUonG), of the simulator (VEP), and of the
integrated software tool-chain based on the DAL programming environment and on the EURETILE
HdS should be delivered before the end of 2013. By the end of the 2013, each work-package will
have demonstrated working prototypes about fault management, description of application dynamism
and support of scalability. During 2013, the consortium will elaborate a more complete picture about
those subjects and, taking as example the fault management area, reduce the distance between what
each partner is doing to detect and manage faults and what is actually done at integrated level. By the
end of 2013, we plan to have completed the porting of the full DPSNN-STDP cortical simulator on
the EURETILE platform and started a campaign of measurements to identify the most promising
areas of architectural improvements. A possible extension of the project end until September 2014 is
under evaluation. At project level, we foresee a few main topics that should be covered during 2014:
Qualitative assessments and quantitative measurements. The availability of the integrated
EURETILE system would enable a more extensive campaign of qualitative evaluations and
quantitative measurements of our integrated software and hardware platforms and
development environment. This way, the consortium could elaborate about the main
EURETILE merits and further developments that we perceive necessary on the bases of an
adequate temporal period dedicated to the evaluation of the platform developed.
Better integration of management’s methods of fault, application dynamism and
scalability. During 2014, the existing demonstrators would be improved toward greater
system integration. We will also develop benchmarks dedicated to show application
dynamism and use the benchmarks to highlight the fault-tolerant behaviours.
Brain-Inspired enhancements. During 2014, we will start experiments about hardware and
software improvements derived from the experience gained thanks to the run of the DPSNNSTDP benchmark.
33 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.6. The SHAPES Project Background (2006-2009)
As the EURETILE project is an evolution of the SHAPES FP7 project, we insert here a summary of
the features of that project that has been relevant for the EURETILE project.
3.6.1. 2006 SHAPES Project Abstract
There is no processing power ceiling for low consumption, low cost, dense Numerical Embedded
Scalable
Systems
dedicated to future
human-centric
applications,
which
will manage multichannel audio, video
and
multi-sensorial
input/outputs.
Nanoscale systems on
chip will integrate
billion-gate designs.
The challenge is to
find
a
scalable
HW/SW design style
for future CMOS
technologies.
The
main
HW
problem is wiring,
which
threatens
Moore’s law. Tiled
architectures suggest a
possible HW path:
“small”
processing Figure 3-9. SHAPES Architecture template
tiles connected by
“short wires”. A second HW problem is the management of the design complexity of billion gate
designs. A tiled design style extensively reuses processing tiles, each tile composed of stable
Intellectual Properties requiring only a few million gates: a manageable complexity. The SW
challenge is to provide a simple and efficient programming environment for a (massive) tiled parallel
architecture. The proposed approach must be experimented through applications.
3.6.2. Investigation of the tiled HW paradigm
SHAPES investigated a ground-breaking HW/SW architecture paradigm. A typical heterogeneous
SHAPES tile is composed of a VLIW floating-point DSP (Digital Signal Processor), a RISC
controller, a DNP (Distributed Network Processor), distributed on chip memory, a POT (a set of
Peripherals On Tile) plus an interface for DXM (Distributed External Memory). Each Tile includes a
few million gates, for optimal balance among parallelism, local memory, and IP reuse on future
technologies.
The SHAPES routing fabric connects on-chip and off-chip tiles, weaving a distributed packet
switching network. 3D next-neighbours engineering methodologies have been studied for off-chip
34 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
networking and maximum system density, see the Figure 3 on previous page. SHAPES opened the
path towards new density records, with multi-Teraflops single-board Tiled computers and Petaflops
systems.
3.6.3. Experimentation of real-time, communication aware system SW
SHAPES adopted layered system software, which does not destroy the information about algorithmic
parallelism, data and workload distribution and real-time requirements provided by the programmer.
The system software should be fully
aware of the tiled hardware paradigm.
For efficiency and predictability, the
system software manages intra-tile
and inter-tile latencies, bandwidths,
computing
resources,
using
application- and architecture-level
profiling.
(a) Application description using the
Process Network formalism and a
simplified representation of the
SHAPES tile. (b) Mapping of the
application process network on the
SHAPES tile
The application is described, using a
model based approach, in terms of a
network of actors, with explicit realtime constraints. Figure 4a shows an
illustration of the application
description). The application is
mapped to the SHAPES architecture
by an iterative automated (or semi- Figure 3-10. (a) Application description using the Process Network
automated)
multi-objective formalism and a sumplified representation of the SHAPES tile. (b)
optimization. The mapping procedure Mapping og the application process network onthe SHAPES tile.
uses performance estimations obtained at different levels of abstraction, i.e. system-level analytic
predictions based on results of the low-level simulation environment. Figure 4b shows a mapping
representation.
The layered structure of the software separates the application code from the Hardware-dependent
Software (HdS). Therefore it is possible to debug the application and system software layers
independently and obtain higher simulation speeds. The application can be debugged using a virtual
architecture, which is agnostic of the HdS details and the real hardware properties. The HdS
generator refines the system software, first through a transaction-accurate step, and then down to the
virtual prototype level (i.e. using the virtual Shapes platform – VSP), where the generated system
software contains all the details needed in order to be executed on the actual hardware
platform.Figure 5 shows the different components of the SHAPES SW environment as well as their
interactions. In EURETILE, this concept of the SW environment, the interfaces between the various
tools and the software design flow will be adapted towards the new challenges imposed by the threelayer hierarchy and the dynamic adaptation capabilities.
35 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
application
analysis, mapping, code generation
Model Compiler
Functional Simulator
mapping
information
Analytic Estimator
trace
information
component interaction,
properties and
constraints
Exploration
HdS Generator
component
source code
Virtual Shapes
Platform
HdS source code
memory mapping
source code
binary
Compiler
Link
Dispatch
OS services
binary libraries
composed
binary
optimised compilation, simulation
Figure 3-11. SHAPES software environment: overview.
The software accesses on-chip and off-chip networks through a homogeneous interface. The same
hardware and software interface can be adopted for integration with signal acquisition and
reconfigurable logic tiles. Generation after generation, the number of tiles on a single-chip will grow,
but the application will be portable.
The SHAPES HW and SW platform has been benchmarked through a set of applications
characterized by a large inherent parallelism and, with one exception, by real-time constraints: wave
field synthesis for array of sound sources reproduced by large arrays of loud-speakers, treatment of
audio signals acquired by arrays of microphones, Ultrasound Scanners, and simulation of Theoretical
Physics (Lattice Quantum-Chromo Dynamics).
SHAPES proposed a programmable, high performance, low power, dense system solution designed
for interfacing with reconfigurable logic and signal acquisition and generation systems. The
architecture is designed to scale from:
low end single module hosting 1-8 tiles for mass market applications
classic digital signal processing systems like radar and medical equipment (2 K tiles)
high-end systems requiring massive numerical computation (32 K tiles)
36 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
3.6.4. DIOPSIS940HF MPSOC and SHAPES RISC + mAgic VLIW DSP + DNP Tile.
ATMEL ROMA, partner of
the
SHAPES
project,
delivered
(2008)
the
DIOPSIS940HF 130 nm
CMOS silicon (RISC +
mAgicV
VLIW
DSP
MPSoC) and development
board, which constituted
the HW reference model
for the SW simulator of the
computational tile and led
place & route trials on
advanced technology (65
and/or 45 nm) to assess the
scalability toward HW
multi-tiling. During 2009,
the consortium finalized the
tape-out trial of an eight
tiles chip using a 45 nm
CMOS target technology, Figure 3-14. The SHAPES elementary tile: RISC + VLIW DSP + Distributed
this way getting final Network Processor.
figures for Area (56 mm2), Computational Power (25 GigaFlops), and Power Consumption (2.9 W).
Each elementary hardware tile is a multi-processor, which includes a distributed network processor
(for inter-tile communications), a floating-point
Figure 3-13. Board mounting the DIOPSIS 940 HF
RISC + mAgic VLIW DSP multi-processor.
Figure 3-12. Eight SHAPES tiles placed&routed in 45
nm.
VLIW processor (for numerical intensive
computations) and a RISC processor (for control, user interface and sequential computations).
The SHAPES Hardware tile was a joint design effort of the APE team of INFN (the Italian Istituto
Nazionale di Fisica Nucleare) and ATMEL Roma (here, we take to opportunity to acknowledge the
work of Federico Aglietti, Antonio Cerruto, Antonio Costa, Maurizio Cosimi, Andrea Michelotti,
37 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Elena Pastorelli, Andrea Ricciardi and of all the designers of the DIOPSIS design center). University
of Roma (Prof. Alessandro Trifiletti and Giuseppe Scotti, among the others) played a key role in the
place & route on the target 45 nm CMOS technology of the 8 tiles SHAPES chip shown in the
picture. ST Microelectronics (Marcello Coppola and his team), University of Cagliari (prof. Luigi
Raffo and his team) and Pisa (Prof. Luca Fanucci and his team), evolved the Spidergon Network-onChip to be used for inter-tile communication, attached to the Distributed Network Processor designed
by INFN.
38 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
4. Dissemination
During this period, the EURETILE project produced:
A set of publications and presentations describing the EURETILE results, listed by the first
sections of this chapter;
The CASTNESS’11 workshop (Roma, 17-18 January 2011). Twenty speakers, representing
the four Teradevice Computing projects (EURETILE, TERAFLUX, TRAMS and SooS)
presented their activities to about 60 researchers. The last section of this document details the
lists of presentations and participants.
4.1. Papers - Posters - Technical Press – Newsletters
ETHZ
P. Kumar, and L. Thiele. Behavioural Composition: Constructively Built Server Algorithms. Proc. 5th Workshop on
Compositional Theory and Technology for Real-Time Embedded Systems, San Juan, Puerto Rico, p. 9-12, Dec.
2012.
P. Kumar and L. Thiele. Quantifying the Effect of Rare Timing Events with Settling-Time and Overshoot. Proc.
IEEE Real-Time Systems Symposium (RTSS), San Juan, Puerto Rico, p. 149-160, Dec. 2012.
S.-H. Kang, H. Yang, L. Schor, I. Bacivarov, S. Ha, and L. Thiele. Multi-Objective Mapping Optimization via
Problem Decomposition for Many-Core Systems. Proc. IEEE Symposium on Embedded Systems for Real-Time
Multimedia (ESTIMedia), Tampere, Finland, p. 28-37, Oct. 2012.
L. Schor, I. Bacivarov, D. Rai, H. Yang, S.-H. Kang, and L. Thiele. Scenario-Based Design Flow for Mapping
Streaming Applications onto On-Chip Many-Core Systems. Proc. Int'l Conf. on Compilers Architecture and
Synthesis for Embedded Systems (CASES), Tampere, Finland, p. 71-80, Oct. 2012.
D. Rai, H. Yang, I. Bacivarov, and L. Thiele. Power Agnostic Technique for Efficient Temperature Estimation of
Multicore Embedded Systems. Proc. Int'l Conf. on Compilers Architecture and Synthesis for Embedded Systems
(CASES), Tampere, Finland, p. 61-70, Oct. 2012.
L. Schor, H. Yang, I. Bacivarov, and L. Thiele. Thermal-Aware Task Assignment for Real-Time Applications on
Multi-Core Systems. Proc. Int'l Symposium on Formal Methods for Components and Objects (FMCO) 2011, Turin,
Italy, Volume 7542 of LNCS, p. 294-313, Oct. 2012.
L. Schor, H. Yang, I. Bacivarov, and L. Thiele. Worst-Case Temperature Analysis for Different Resource Models.
IET Circuits, Devices & Systems, Volume 6, Issue 5, p. 297-307, Sep. 2012.
K. Huang, W. Haid, I. Bacivarov, M. Keller, L. Thiele. Embedding Formal Performance Analysis into the Design
Cycle of MPSoCs for Real-time Streaming Applications. ACM Transactions in Embedded Computing Systems
(TECS Journal), ACM, Volume 11, Issue 1, p. 8:1-8:23, 2012.
L. Schor, I. Bacivarov, H. Yang, and L. Thiele. Fast Worst-Case Peak Temperature Evaluation for Real-Time
Applications on Multi-Core Systems. Proc. IEEE Latin American Test Workshop (LATW), Quito, Ecuador, p. 1-6,
Apr. 2012.
L. Schor, I. Bacivarov, H. Yang, and L. Thiele. Worst-Case Temperature Guarantees for Real-Time Applications on
Multi-Core Systems. Proc. IEEE Real-Time and Embedded Technology and Applications Symposium (RTAS), Proc.
IEEE Real-Time and Embedded Technology and Applications Symposium (RTAS), Beijing, China, p. 87-96, Apr.
2012.
P. Kumar and L. Thiele. Timing Analysis on a Processor with Temperature-Controlled Speed Scaling. Proc. IEEE
Real-Time and Embedded Technology and Applications Symposium (RTAS), Beijing, China, p. 77-86, Apr. 2012.
I. Bacivarov, I. Belaid, A. Biagioni, A. El Antably, N. Fournel, O. Frezza, J. Jovic, R. Leupers, F. Lo Cicero, A.
Lonardo, L. Murillo, P.S. Paolucci, D. Rai, D. Rossetti, F. Rousseau, L. Schor, C. Schumacher, F. Simula, L. Thiele,
L. Tosoratto, P. Vicini, H. Yang – “DAL: Programming Efficient and Fault-Tolerant Applications for Many-Core
Systems” - Poster at HIPEAC12 - Jan 23-25, 2012 Paris, France
39 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
P. Kumar, J.-J. Chen, and L. Thiele. Demand Bound Server: Generalized Resource Reservation for Hard Real-Time
Systems. Proc. Int'l Conference on Embedded Software (EMSOFT), pages 233-242, Oct. 2011.
L. Schor, H. Yang, I. Bacivarov, and L. Thiele. Worst-Case Temperature Analysis for Different Resource
Availabilities: A Case Study. Proc. Workshop on Power and Timing Modeling, Optimization and Simulation
(PATMOS), Lecture Notes on Computer Science (LNCS), Springer, Vol. 6951, pages 288-297, Sep. 2011.
P. Kumar, J.-J. Chen, L. Thiele, A. Schranzhofer, and G. C. Buttazzo. Real-Time Analysis of Servers for General Job
Arrivals. Proc. Intl. Conf. on Embedded and Real-Time Computing Systems and Applications (RTCSA), pages 251258, Aug. 2011.
L. Thiele, L. Schor, H. Yang, and I. Bacivarov. Thermal-Aware System Analysis and Software Synthesis for
Embedded Multi-Processors. Proc. Design Automation Conference (DAC), pages 268-273, Jun. 2011.
D. Rai, H. Yang, I. Bacivarov, JJ. Chen, L. Thiele. Worst-Case Temperature Analysis for Real-Time Systems. In
Proceedings of Design, Automation and Test in Europe (DATE), Grenoble, France, March 2011.
S. Perathoner, K. Lampka, L. Thiele. Composing Heterogeneous Components for System-wide Performance
Analysis. In Proceedings of Design, Automation and Test in Europe (DATE), Grenoble, France, March 2011 (invited
paper).
I. Bacivarov, H. Yang, L. Schor, D. Rai, S. Jha, L. Thiele, Poster: Distributed Application Layer - Towards Efficient
and Reliable Programming of Many-Tile Architectures. Design, Automation and Test in Europe (DATE) Friday
Workshop, Grenoble, France, March 2011.
K. Huang, L. Santinelli, JJ. Chen, L. Thiele, and G. C. Buttazzo. Applying Real-Time Interface and Calculus for
Dynamic Power Management in Hard Real-Time Systems. Real-Time Systems Journal, Springer Netherlands, Vol.
47, No. 2, pages 163-193, Mar. 2011.
A. Schranzhofer, JJ. Chen, L. Thiele. Dynamic Power-Aware Mapping of Applications onto Heterogeneous MPSoC
Platforms. IEEE Transactions on Industrial Informatics, IEEE, Vol. 6, No. 4, pages 692 -707, November, 2011.
RWTH
C. Schumacher, J. H. Weinstock, R. Leupers, G. Ascheid, L. Tossorato, A. Lonardo, D. Petras, T. Groetker.
“legaSCi: Legacy SystemC Model Integration into Parallel SystemC Simulators”. 1st Workshop on Virtual
Prototyping of Parallel and Embedded Systems (ViPES), 2013, Boston, USA (accepted for publication).
C. Schumacher, J. H. Weinstock, R. Leupers and G. Ascheid: Cause and effect of nondeterministic behavior in
sequential and parallel SystemC simulators. IEEE International High Level Design Validation and Test Workshop
(HLDVT'12). Nov 2012, Huntington Beach (California-USA).
C. Schumacher, J. H. Weinstock, R. Leupers and G. Ascheid: Scandal: SystemC Analysis for NonDeterminism
AnomaLies. Forum on Specification and Design Languages (FDL '12), Sep 2012, Vienna (Austria)
L. G. Murillo, J. Harnath, R. Leupers and G. Ascheid: Scalable and Retargetable Debugger Architecture for
Heterogeneous MPSoCs. System, Software, SoC and Silicon Debug Conference (S4D '12), Sep 2012, Vienna
(Austria)
L. G. Murillo, J. Eusse, J. Jovic, S. Yakoushkin, R. Leupers and G. Ascheid: Synchronization for Hybrid MPSoC
Full-System Simulation. Design Automation Conference (DAC '12), Jun 2012, San Francisco (USA)
R. Leupers: More Real Value for Virtual Platforms. Design, Automation and Test in Europe (DATE '12), Mar 2012,
Dresden (Germany)
J. Jovic, S. Yakoushkin, L. G. Murillo, J. Eusse, R. Leupers and G. Ascheid: Hybrid Simulation for Extensible
Processor Cores. Design, Automation and Test in Europe (DATE '12), Mar 2012, Dresden (Germany)
S. Kraemer, R. Leupers, D. Petras, T. Philipp, A. Hoffmann: Checkpointing SystemC-Based Virtual Platforms.
International Journal of Embedded and Real-Time Communication Systems (IJERTCS), vol. 2, no. 4, 2011
L. G. Murillo, W. Zhou, J. Eusse, R. Leupers, G. Ascheid: Debugging Concurrent MPSoC Software with Bug Pattern
Descriptions. System, Software, SoC and Silicon Debug Conference (S4D '11), Oct 2011, Munich (Germany)
R. Leupers, G. Martin, N. Topham, L. Eeckhout, F. Schirrmeister, X. Chen: Virtual Manycore Platforms: Moving
Towards 100+ Processor Cores. Design Automation & Test in Europe (DATE), Mar 2011, Grenoble (France)
J. Castrillon, A. Shah, L. G. Murillo, R. Leupers, G. Ascheid: Backend for Virtual Platforms with Hardware
Scheduler in the MAPS Framework. 2nd IEEE Latin America Symp. on Circuits and Systems, Feb 2011, Bogota
(Colombia)
S. Kraemer, Design and analysis of efficient MPSoC simulation techniques, dissertation, 2011, Aachen, Germany
(http://darwin.bth.rwth-aachen.de/opus3/frontdoor.php?source_opus=3769&la=de)
40 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
C. Schumacher, R. Leupers, D. Petras and A. Hoffmann. parSC: Synchronous Parallel SystemC Simulation on MultiCore Host Architectures. In proceedings of CODES/ISSS '10, October, 2010, Scottsdale, Arizona, USA
(http://dx.doi.org/10.1145/1878961.1879005)
TIMA
A. Chagoya-Garzon, F. Rousseau, F. Pétrot, Multi-Device Driver Synthesis Flow for Heterogeneous Hierarchical
Systems, Euromicro Conference on Digital System Design, Sept 2012, pp. 389 – 396, Izmir, Turkey.
Ashraf Elantably, Frédéric Rousseau, Task migration in multi-tiled MPSoC: Challenges, state-of-the-art and
preliminary solutions, Journée National du Réseau Doctoral en Microélectronique, Marseille, France, June 2012,
Poster and 4 pages paper (in English)
Chagoya-Garzon, N. Poste, F. Rousseau, Semi-Automation of Configuration Files Generation for Heterogeneous
Multi-Tile Systems, Computer Software and Application Conference (COMPSAC 2011), Munich, Germany, 18-21
July 2011.
H. Chen, G. Godet-Bar, F. Rousseau, F. Petrot, Me3D : A Model-driven Methodology expediting Embedded Device
Driver Development, International Symposium on Rapid System Prototyping (IEEE RSP 2011), pp. 171-177, May
2011, Karlsruhe, Germany.
INFN
Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Francesca Lo Cicero, Alessandro Lonardo, Pier Stanislao
Paolucci, Davide Rossetti, Francesco Simula, Laura Tosoratto, Piero Vicini - QUonG: A GPU-based HPC System
Dedicated to LQCD Computing - Application Accelerators in High-Performance Computing, Symposium on, pp.
113-122, 2011 Symposium on Application Accelerators in High-Performance Computing, 2011 [[1]]
Pier Stanislao Paolucci - FP7 EURETILE Project: EUropean REference TILed architecture Experiment - HipeacInfo,
Quarterly Newsletter, Number 24, page 11, October 2010
(http://www.Hipeac.net/newsletter) File:PaolucciEuretileHipeacInfo24October2010.pdf
Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Francesca Lo Cicero, Alessandro Lonardo, Pier Paolucci,
Roberto Petronzio, Davide Rossetti, Andrea Salamon, Gaetano Salina, Francesco Simula, Nazario Tantalo, Laura
Tosoratto, Piero Vicini - APEnet+: a 3D toroidal network enabling Petaflops scale Lattice QCD simulations on
commodity clusters - High Energy Physics - Lattice 2010 (hep-lat); Distributed, Parallel, and Cluster Computing
(cs.DC) arXiv:1012.0253v1 [hep-lat] (proceedings)
R. Ammendola, A. Biagioni, O. Frezza, F. Lo Cicero, A. Lonardo, P.S. Paolucci, D. Rossetti, A. Salamon, G. Salina,
F. Simula, L. Tosoratto, P. Vicini - apeNET+: High Bandwidth 3D Torus Direct Network for PetaFLOPS Scale
Commodity Clusters, International Conference on Computing in High Energy and Nuclear Physics (CHEP), October
2010, Taipei, Taiwan - Proceedings on J. Phys.: Conf. Ser. 331 052029 doi:10.1088/1742-6596/331/5/052029
R. Ammendola, A. Biagioni, O. Frezza, F. Lo Cicero, A. Lonardo, P.S. Paolucci, D. Rossetti, A. Salamon, G. Salina,
F. Simula, L. Tosoratto, P. Vicini - Mastering multi-GPU computing on a torus network - GPU Technology
Conference 2010 (GTC) - http://www.nvidia.com/content/GTC/posters/2010/I09-Mastering-Multi-GPU-Computingon-a-Torus-Networki.pdf (poster)
R. Ammendola, A. Biagioni, G. Chiodi, O. Frezza, A.Lonardo, F. Lo Cicero, R. Lunadei, D. Rossetti, A. Salamon,
G. Salina, F. Simula, L. Tosoratto, P. Vicini - High speed data transfer with FPGAs and QSFP+ modules - Topical
Workshop on Electronics for Particle Physics, Aachen, Germany / September 20-24, 2010 - Proceedings on JINST 5
C12019 doi:10.1088/1748-0221/5/12/C12019
R. Ammendola et al. - High speed data transfer with FPGAs and QSFP+ modules - Nuclear Science Symposium
Conference Record (NSS/MIC) 2010 IEEE, Publication Year: 2010, Page(s): 1323 1325, November 2010, Knoxville,
Tennessee. DOI: 10.1109/NSSMIC.2010.5873983
41 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
4.2. Books/Book Chapters
ETHZ
Book chapter: I. Bacivarov, W. Haid, K. Huang, L. Thiele. Methods and Tools for Mapping Process Networks onto
Multi-Processor Systems-On-Chip. Handbook of Signal Processing Systems, Springer, pages 1007-1040, October,
2010.
RWTH
Book: R. Leupers and O. Temam (Eds.), Processor and System-On-Chip Simulation, Springer, September
2010, ISBN 978-1441961747
Book: T. Kempf, G. Ascheid, R. Leupers: Multiprocessor Systems on Chip: Design Space Exploration, Springer, Feb
2011, ISBN 978-1441981523
TIMA
Katalin Popovici, Frederic Rousseau, Ahmed A. Jerraya, Marilyn Wolf: Embedded Software Design and
Programming of Multiprocessor System-on-Chip, Simulink and SystemC Case Studies, Springer, April 2010, ISBN
978-1-4419-5566-1
Book chapter: Xavier Guerin, Frederic Petrot, Operating System Support for Applications targeting Heterogeneous
Multi-Core System)on-Chip in the book Multi-Core Embedded Systems, CRC Press, Chapter 9, 24 pages, April
2010.
4.3. Presentations
ETHZ
Iuliana Bacivarov, Thermal-Aware Design of Real-Time Multi-Core Embedded Systems, Invited talk at Mapping
Applications to MPSoCs, Jun. 2011.
Iuliana Bacivarov, Temperature Predictability in Multi-Core Real-Time Systems, Invited talk at DAC Workshop on
Multiprocessor System-on-Chip for Cyber Physical Systems: Programmability, Run-Time Support, and Hardware
Platforms for High Performance Embedded Applications, Jun. 2011.
Iuliana Bacivarov, Distributed Application Layer – Towards Seamless Programming of Many-Tile Architectures,
CASTNESS 2011, 17 and 18 January 2011, Rome,
Italy,http://euretile.roma1.infn.it/mediawiki/img_auth.php/8/88/EURETILE-2-IulianaBacivarov.pdf.
Iuliana Bacivarov, Distributed Operation Layer: An Efficient and Predictable KPN-Based Design Flow, invited talk
at Workshop on Compiler-Assisted System-On-Chip Assembly 2010, in conjunction with Embedded Systems Week,
Scottsdale, AZ, US, October 2010, http://www12.cs.fau.de/ws/casa10.
Iuliana Bacivarov, Efficient Execution of Kahn Process Networks on CELL BE, invited talk at Summer School on
Models for Embedded Signal Processing Systems at Lorentz Center, Leiden, Netherlands, 30 Aug - 3 Sep
2010, http://www.lorentzcenter.nl/lc/web/2010/427/presentations/Iuliana-cell.pdf.
Iuliana Bacivarov, Distributed Operation Layer: A Practical Perspective, tutorial at Summer School on Models for
Embedded Signal Processing Systems at Lorentz Center, Leiden, Netherlands, 30 Aug - 3 Sep
2010, http://www.lorentzcenter.nl/lc/web/2010/427/presentations/Iuliana-demo.pdf.
Iuliana Bacivarov, Distributed Operation Layer: Efficient Design Space Exploration of Scalable MPSoC, invited talk
at Combinatorial Optimization for Embedded System Design workshop 2010 in conjunction with CPAIOR2010, 7th
International Conference on Integration of Artificial Intelligence and Operations Research techniques in Constraint
Programming, Bologna, Italy, June 2010, http://www.artist-embedded.org/artist/Overview,2022.html.
Iuliana Bacivarov, invited talk at Efficient Execution of Kahn Process Networks on MPSoC, Mapping Applications
to MPSoCs 2010, June 29-30, 2010, St. Goar, Germany, http://www.artist-embedded.org/artist/Program,1822.html.
RWTH
Jovana Jovic, Simulation Challenges in the EURETILE Project, CASTNESS 2011, January 17-18, 2011, Rome, Italy
Special session at DATE 2011: Virtual Manycore Platforms: Moving Towards 100+ Processor Cores, organized by
R. Leupers and G. Martin
42 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
R. Leupers, H. Meyr: Embedded Processor Design, Block lecture, ALARI, Lugano, Feb 2011
ICT Technology Transfer Workshop targeting Horizon 2020, Brussels, Apr 2011, organized by R. Leupers
S. Yakoushkin: Advanced Simulation Techniques, Joint RWTH/TU Tampere Seminar, June 2011
R. Leupers: SoC Design Research in the UMIC Excellence Cluster, Seminar, TU Berlin, Sep 2011
Juan Eusse: Hybrid Simulation Technology for Extensible Cores and Full System Simulation of Complex MPSoCs,
Nov 2011 (Presentation at HiPEAC Computing Systems Week)
Rainer Leupers, HiPEAC Cluster Meeting (Design and Simulation Cluster), October 2010, Barcelona, Spain
Rainer Leupers, MPSoC Design for Wireless Multimedia, Tutorial, MIXDES, June 2010, Wroclaw, Poland
Rainer Leupers, Cool MPSoC Design, ASCI Winter School on Embedded Systems, March 2010, Soesterburg,
Netherlands
Rainer Leupers, Design Technologies for Wireless Systems-On-Chip, Huawei ESL Symposium, September 2010,
Shenzhen, People's Republic of China
Rainer Leupers, Embedded Processor Design and Implementation, course in MSc in Embedded Systems track at
ALaRI Institute, March 1-4, 2010, University of Lugano, Switzerland
Stefan Kraemer, Advanced Simulation Techniques for Virtual Platforms, May 26, 2010, Imperial College London,
London, United Kingdom
Christoph Schumacher, Stefan Kraemer and Rainer Leupers, demonstration at DAC 2010 exhibition: parSC: parallel
SystemC simulation, deterministic, accurate, fast, June 14-16, 2010, Anaheim, USA
Christoph Schumacher, Virtual Platform Technologies for Multi-core Platforms, UMIC Day, 19 October, 2010,
RWTH Aachen
INFN
P. Vicini, QUonG: A GPU-based HPC System Dedicated to LQCD Computing - Symposium on Application
Accelerators in High-Performance Computing, Knoxville, TN - USA, July
2011http://doi.ieeecomputersociety.org/10.1109/SAAHPC.2011.15
D. Rossetti, Remote Direct Memory Access between NVIDIA GPUs with the APEnet 3D Torus Interconnect - SC11
- International Conference for High Performance Computing, Networking, Storage and Analysis - Seattle,
WA http://nvidia.fullviewmedia.com/fb/nv-sc11/tabscontent/archive/304-wed-rossetti.html
P.S. Paolucci, EURETILE: Brain-Inspired many-tile SW/HW Experiment, CASTNESS 2011, 17 and 18 January
2011, Rome, Italy,http://euretile.roma1.infn.it/mediawiki/img_auth.php/3/3a/EURETILE-1PierStanislaoPaolucci.pdf
P. Vicini, EURETILE: The HPC and Embedded Experimental HW Platform, CASTNESS 2011, 17 and 18 January
2011, Rome, Italy,http://euretile.roma1.infn.it/mediawiki/img_auth.php/6/69/EURETILE-6-PieroVicini.ppt
R. Ammendola, apeNET+: a 3D toroidal network enabling petaFLOPS scale Lattice QCD simulations on commodity
clusters, Lattice 2010, THE XXVIII INTERNATIONAL SYMPOSIUM ON LATTICE FIELD THEORY,
Villasimius, Italy, June
2010, http://agenda.infn.it/contributionDisplay.py?contribId=335&sessionId=70&confId=2128
D. Rossetti, apeNET+ Project Status, Lattice 2010, THE XXIX INTERNATIONAL SYMPOSIUM ON LATTICE
FIELD THEORY, Squaw Village, Lake Tahoe, CA, USA, July 2011 - to be published
on https://latt11.llnl.gov/html/proceedings.php
Target
W. Geurts, G. Goossens, "Ideas for the Design of an ASIP for LQCD", CASTNESS 2011, Rome (Italy), January 1718, 2011,http://euretile.roma1.infn.it/mediawiki/img_auth.php/3/3c/EURETILE-5-WernerGeurts.ppt
G. Goossens, “Why Compilation Tools are the Catalyst for Multicore SoC Design”, Electronic Design and Solutions
Fair, Yokohama (Japan), January 27-28, 2011.
G. Goossens, “Why Compilation Tools are a Catalyst for Multicore SoC Design”, Third Friday Workshop on
Designing for Embedded Parallel Computing Platforms: Architectures, Design Tools, and Applications, Design
Automation and Test in Europe (DATE-2011), Grenoble (France), March 18, 2011.
G. Goossens, P. Verbist, “Enabling the Design and Programming of Application-Specific Processors”, SophiaAntipolis Micro-Electronics Conference (SAME-2011), Sophia-Antipolis (France), October 12-13, 2011.
G. Goossens, “Building Multicore SoCs with Application-Specific Processors”, Electronic Design and Solutions
Fair, Yokohama (Japan), November 16-18, 2011.
43 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
P. Verbist, “Building software-programmable accelerators for ARM-based subsystems”, ARM Technical
Symposium, Taipei and Hsinchu (Taiwan), November 17-18, 2011.
G. Goossens, “Building Multicore SoCs with Application-Specific Processors”, Intl. Conf. on IP-Based SoC Design
(IP-SoC-2011), Grenoble (France), December 7-8, 2011.
G. Goossens, “Design Tools for Building Software-Programmable Accelerators in Multicore SoCs”, Workshop on
Tools for Embedded System Design, Sint-Michielsgestel (Netherlands), December 13, 2011.
G. Goossens, E. Brockmeyer, W. Geurts, “Application-Specific Instruction-set Processors (ASIPs) and related design
tools for tiled systems”, CASTNESS 2012, Paris (France), January 26, 2012.
G. Goossens, "How ASIP Technology can Make your RTL Blocks More Flexible", Electronic Design and Solutions
Fair, Yokohama (Japan), January 28-29, 2010.
S. Cox, G. Goossens, "Hardware Accelerator Performance in a Programmable Context: Methodology and Case
Study", Embedded Systems Conference, San Jose (CA, USA), April 26-29, 2010.
G. Goossens, "Design of Programmable Accelerators for Multicore SoCs", First Artemis Technology Conference,
Budapest (Hungary), June 29-30, 2010.
TIMA
Frédéric Rousseau, Requirements in Communication Synthesis for EURETILE: The use of Communication Path
Formalization, CASTNESS 2012, January 26th, 2012, Paris, France
Frédéric Rousseau, presentation of the EURETILE project in front of the Board of directors of the University Joseph
Fourier, March 2011, Grenoble, France
Frédéric Rousseau, Communication Synthesis in Low Level Software for Hierarchical Heterogeneous Systems,
CASTNESS 2011, January 17-18, 2011, Rome, Italy
4.4. CASTNESS’11
A key action conducted during this period in the framework of WP9 has
been the organization of the CASTNESS’11 Workshop (Computer
Architectures, Software tools and nano-Technologies for Numerical and
Embedded Scalable Systems), hosted in Roma on 17-18 January 2011 at
the Frentani Conference Center. http://www.congressifrentani.it
Twenty speakers, representing the four Teradevice Computing projects
(EURETILE, TERAFLUX, TRAMS and SooS) presented their activities
to about 60 researchers.
An archive, hosting all the presentations, is available for public access,
hosted on the http://www.euretile.eu web site.
CASTNESS’11 Agenda
Monday 17 January 2011: Day 1
Event
Speaker Name
Affiliation
REGISTRATION
OPENING and WELCOME
SOOS: Scaling Issues in Current
OS Architectures and
Approaches to Overcome them
Duration
Start
End
00:40
13:00
13:40
Speranza
Falciano
INFN Roma
00:20
13:40
14:00
Daniel Rubio
Bonilla
USTUTT-HLRS,
Stuttgart
00:20
14:00
14:20
44 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
SOOS: Composition vs
Concurrency
Mihai Letia
EPFL, Lausanne
SHORT BREAK
SOOS: Issues in Large Scale
Scheduling of Distributed
Applications
SOOS: Resource Discovery and
Modelling in Heterogeneous
Computing Environments
00:20
14:20
14:40
00:20
14:40
15:00
Tommaso
Cucinotta
Scuola Superiore
Sant'Anna, Pisa
00:20
15:00
15:20
Christiaan Baaij
University of
Twente
00:20
15:20
15:40
00:20
15:40
16:00
SHORT BREAK
TRAMS: Introduction to the
TRAMS Project Objectives
Antonio Rubio
UPC Barcelona
00:20
16:00
16:20
TRAMS Variability at Device
Level
Andrew Brown
University of
Glasgow
00:20
16:20
16:40
00:30
16:40
17:10
COFFEE BREAK
TRAMS: New Tools and
Methods in Robust SRAM
Design
Paul Zuber
Imec
00:20
17:10
17:30
TRAMS: Temperature, Voltage
and Process Variations in
SRAMs
Shrikanth
Ganapathy
UPC Barcelona
00:13
17:30
17:43
TRAMS: Robustness of SRAM
Memories
Ioana Vatajelu
UPC Barcelona
00:13
17:43
17:56
TRAMS: Carbon Nanotub
Technology an Alternative in
Future RAM Memories
Carmen Garcia
UPC Barcelona
00:14
17:56
18:10
SOCIAL DINNER
20:30
Tuesday 18 January 2011: Day 2
Event
Speaker Name
Affiliation
Duration
Start
End
Pier S. Paolucci
INFN Roma
00:20
09:00
09:20
EURETILE: Distributed
Application Layer - Towards
Seamless Programming of
Many-Tiles Architectures
Iuliana
Bacivarov
ETH Zurich
00:20
09:20
09:40
EURETILE: Simulation
Challenges in the EURETILE
Project
Jovana Jovic
RWTH Aachen
00:20
09:40
10:00
00:30
10:00
10:30
EURETILE: Brain-Inspired
many-tile SW/HW Experiment
COFFEE BREAK
45 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
EURETILE: Communication
Synthesis in Low Level
Software for Hierarchical
Heterogeneous Systems
EURETILE: Ideas for the design
of an ASIP for LQCD
Frederic
Rousseau
TIMA/UJF
Grenoble
00:20
10:30
10:50
Werner Geurts
TARGET
Leuven
00:20
10:50
11:10
00:20
11:10
11:30
SHORT BREAK
EURETILE: The HPC and
Embedded Experimental HW
Platform
Piero Vicini
INFN Roma
00:20
11:30
11:50
TERAFLUX: Intro
Roberto Giorgi
Università di
Siena
00:20
11:50
12:10
TERAFLUX: OS
Doron Shamia
Microsoft
00:20
12:10
12:30
01:40
12:30
14:10
LUNCH
TERAFLUX: On Transactional
Memory in TERAFLUX
Berham Khan
University of
Manchester
00:20
14:10
14:30
TERAFLUX: On Computational
Models
Salman Khan
University of
Manchester
00:20
14:30
14:50
00:20
14:50
15:10
SHORT BREAK
TERAFLUX: On State-of-theart and Plans on the Loop Nest
Optimization
Konrad
Trifunovic
INRIA Saclay,
France
00:20
15:10
15:30
FET Plans
Jean-Marie
Auger
EU FET Officer
00:20
15:30
15:50
CLOSURE
15:50
46 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
5. APPENDIX: Euretile Project Glossary
A
ABI - Application Binary Interface
ALUT - Adaptive LookUp Table
AED - Abstract Execution Device
API - Application Programming Interface
APEnet+ - An FPGA-based card for low latency, high bandwidth direct network interconnection based on the
DNP
ASIP - Application Specific Instruction Set Processor
BER - Bit Error Rate
BML - Byte Management Layer, framework of the OpenMPI library
BPDLang - Bug Pattern Description Language
BTL - Byte Transfer Layer, framework of the OpenMPI library
CA - Cycle Accurate
CDR - Clock Data Recovery
CLI - Command Line Interface
CQ - Completion Queue
CRC - Cyclic Redundancy Check
DAL - Distributed Application Layer
DFM - DNP Fault Manager
DMI - Direct Memory Interface
DNAOS - DNA is Not just Another Operating System
DNP - Distributed Network Processor
DOL - Distributed Operating Layer
DPSNN-STDP - Distributed Polychronous Spiking Neural Networks with synaptic Spiking Time Dependent
Plasticity(a PSNN-STDP code natively redesigned and rewritten to exploit parallel/distributed computing
systems)
DWARF - Debugging with Attributed Record Formats, a standard format for debug information in a binary
object
DWR - DNP Watchdog Register
ECC - Error correcting code
EIR - Event-based Intermediate Representation for multi-core debugging
EURETILE - EUropean REference TILed architecture Experiment - (all)
FIT - Failure In Time
FM - Fault manager FPGA - Field-Programmable Gate Array
GUI - Graphic User Interface
HAL - Hardware Abstraction Layer
HdS - Hardware dependent Software
HFM - Host Fault Manager
HLEM - High-level Event Monitor for multi-core debugging
HPC - High Performance Computing
HyNoC - Hybrid NoC simulation technology
HySim - Hybrid Simulation technology
HWR - Host Watchdog Register
B
C
D
E
F
G
H
47 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
I
IA - Instruction Accurate
ICE - Institute for Communication technologies and Embedded systems
IDE - Integrated Development Environment (tools to develop and debug embedded software, integrated in a
GUI)
IMC - Interface Method Call
INFN - Istituto Nazionale di Fisica Nucleare (National Institute for Nuclear Physics)
I/O - Input/Output
IOCTL - Input/Output Control, is a system call for device-specific input/output operations and other operations
which cannot be expressed by regular system calls.
IP - Intellectual Property
IP Designer - TARGET's tool-suite for the design and programming of ASIPs
IRISC - ICE Reduced Instruction Set Computer (RWTH's proprietary processor with that name)
ISR - Interrupt Service Routine
ISS - Instruction Set Simulator
JIT - Just-in-time compilation
KPN – Khan Process Network
LDM - LiFaMa Diagnostic Message
LiFaMa - Link Fault Manager
LISA - Language for Instruction Set Architecture Design
LO|FA|MO - Local Fault Monitor
LP - Logical Process
LQCD - Lattice Quantum-ChromoDynamics
LSB - Least Significant Bit
LTD - Long Term synaptic Depression
LTL - Linear Temporal Logic
LTP - Long Term synaptic Potentiation
LUT - Look-Up Table
LwIP - Lightweight IP
MM - Memory Management
MMU - Memory Management Unit
MPI - Message Passing Interface
MSB - Most Significant Byte
MTL - Matching Transport Layer, framework of the OpenMPI library
NIC - Network Interface Controller
NIOS II - 32-bit microprocessor available as soft-core in Altera FPGAs (often shorthanded as "NIOS")
nML - Not a Modelling Language (a processor architectural description language)
OMPI - Open MPI, an implementation of the MPI standard.
OPAL - Open Portability Access Layer, part of the OpenMPI library
OpenGL - The Open Graphics Library multi-platform API for rendering computer graphics
ORTE - Open Run-time Environment, part of the OpenMPI library
OS - Operating System
OSCI - Open SystemC Initiative
OSI - Open Systems Interconnection
parSC - parallel SystemC (a parallelizing SystemC execution engine)
J
K
L
M
N
O
P
48 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Q
PCI/PCIe - Peripheral Component Interconnect - TIMA/UJF
PCS - Physical Coding Sublayer -INFN
PDC - Pin-Down Cache - INFN
PEO - Process Execution Order
PIC - Programmable Interrupt Controller
PMA - Physical Medium Attachment
PML - Point-to-point Message Layer, framework of the OpenMPI library
PRBS - Pseudorandom Binary Sequence
Presto - MPI-like library for APEnet+/DNP
PSB - Peripheral Subsystem Bus in the VEP-EX
PSNN - Polychronous Spiking Neural Network (A neural network which takes in consideration the delay
introduced by the axonal arborization, in principle different for each synapse, and reproduces the spiking
behaviour of a neural network)
Python - A general-purpose interpreted programming language
P2P - peer-to-peer
QUonG - LQCD on GPU platform
RB - Ring Buffer
RDMA - Remote Direct Memory Access
RDMA GET - RDMA READ operation that implies an handshake between the sender and the receiver
RDMA PUT - RDMA WRITE operation that implies an handshake between the sender and the receiver
RTE - Runtime Environment - RWTH
RTL - Register Transfer Level (also used to refer to register-transfer languages, such as VHDL or Verilog)
RWTH - Rheinisch-Westfaelische Technische Hochschule Aachen
SC - SystemC
SCandal - SystemC Analysis for NonDeterminism Anomalies
SCML - SystemC Modelling Language (a collection of convenience modelling objects for SystemC)
SIMD - Single Instruction Multiple Data (also known as vector processing, a processor architectural concept to
implement data-level parallelism)
SNDNP - Service Network DNP
SNET - Service Network
STDP - synaptic Spiking Time Dependent Plasticity (a mechanism of synaptic evolution that depends on the
relative timing between the spike incoming to a neuron, and the spike that the neuron emits. The synapses can be
potentiated or depressed)
SIP - Software Interface Protocol
TCL - Tool Command Language, a general-purpose interpreted programming language
TLM - Transaction Level Modelling
TSBE - Target-specific Back-end for multi-core debugging
V2P - Virtual to Physical
VBE - VESA BIOS Extensions
VEP - Virtual EURETILE Platform
VEP-Ex - Virtual EURETILE Platform Experimental
VHDL - VHSIC Hardware Description Language
VLIW - Very Long Instruction Word (a processor architectural concept to implement instruction-level
parallelism)
WD - WatchDog
XML - Extensible Markup Language
XSD - XML Schema Definition
R
S
T
V
W
X
49 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
References
A. Sangiovanni-Vincentelli, G. Martin, Platform-based design and software design methodology for
embedded systems. IEEE Design & Test 18, 6, pp. 23–33, 2001.
G. Kahn, “The semantics of a simple language for parallel programming,” in Proc. IFIP
Congr.,Stockholm, Sweden, Aug. 1974, pp. 471–475.
E. A. Lee and A. Sangiovanni-Vincentelli, “A framework for comparing models of computation,”
IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst., vol. 17, no. 12, pp. 1217–1229, Dec.
1998.
S. Edwards, L. Lavagno, E. A. Lee, and A. Sangiovanni-Vincentelli, “Design of embedded systems:
Formal models, validation, and synthesis,” Proc. IEEE, vol. 85, no. 3, pp. 366–390, Mar. 1997.
B. Kienhuis, Ed F. Deprettere, P. Van der Wolf, and K. Vissers, A methodology to design
programmable embedded systems - the Y-Chart approach. In Proceedings of the International
Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, pages
18–37. Springer-Verlag, 2002.
G. Schirner, R. Dömer, A. Gerstlauer, High level development, modeling and automatic generation
of hardware-dependent software, Hardware-dependent Software Principles and Practice, Springer
2009, Chapter 8, pp. 203-231.
Lothar Thiele, Iuliana Bacivarov, Wolfgang Haid, and Kai Huang. 2007. Mapping Applications to
Tiled Multiprocessor Embedded Systems. In Proceedings of the Seventh International Conference
on Application of Concurrency to System Design (ACSD '07). IEEE Computer Society, 29-40.
DOI = 10.1109/ACSD.2007.53 http://dx.doi.org/10.1109/ACSD.2007.53
Paolucci, P.S.; Jerraya, A.A.; Leupers, R.; Thiele, L.; Vicini, P., "SHAPES:: a tiled scalable software
hardware architecture platform for embedded systems, "Hardware/Software Codesign and System
Synthesis, 2006. CODES+ISSS '06. Proceedings of the 4th International Conference , pp.167,172,
22-25 Oct. 2006 doi: 10.1145/1176254.1176297
Jing Lin, Andreas Gerstlauer, Brian L. Evans, Communication-aware Heterogeneous Multiprocessor
Mapping for Real-time Streaming Systems. Signal Processing Systems 69(3): 279-291 (2012)
Gernot Heiser, Virtualizing embedded systems: why bother? DAC 2011, pp. 901-905
P. Ezudheen, P. Chandran, J. Chandra, B. P. Simon, and D. Ravi,
"Parallelizing SystemC kernel for fast hardware simulation on SMP
machines," workshop on principles of advanced and distributed simulation, PADS 2009
S. Lu, S. Park, E. Seo, and Y. Zhou, "Learning from mistakes: A
comprehensive study on real world concurrency bug characteristics," in proceedings of the 13th
international conference on Architectural support for programming languages and operating
50 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
systems, ASPLOS, 2008.
D. A. Penry, D. Fay, D. Hodgdon, R. Wells, G. Schelle, D. I. August, and D. Connors. "Exploiting
parallelism and structure to accelerate the simulation of chip multi-processors," in proceedings of
the 12th international symposium on high-performance computing architecture, HPCA, 2006
Song, S. et al. Competitive Hebbian learning through spike-timing-dependent synaptic plasticity.
Nature Neuroscience 3, 919-926, (2000).
Izhikevich, E. “Polychronization: Computation with Spikes.” Neural Computation, 18, 245-282
(2006) .
Izhikevich, E. Dynamical Systems in Neuroscience: the Geometry of Excitability and Bursting.
Cambridge, MA: The MIT Press (2006) .
Izhikevich, E. M. Simple Model of Spiking Neurons. IEEE Transactions on Neural Networks, Vol.
14, No. 6, November 2003 .
Izhikevich, E. M. Which Model to use for cortical spiking neurons? IEEE Transaction on Neural
Networks, 15, no. 5 1063-1070 (2004) .
Xin Jun, Steve B. Furber, John V. Woods. Efficient Modelling of Spiking Neural Networks on a
Scalable Chip Multi-processor. Int. Joint Conf. on Neural Networks 2008 (IJCNN 2008), 28122819 (2008) .
L. A. Plana, J. B. An On-Chip and Inter-Chip Communications Network for the SpiNNaker
Massively-Parallel Neural Net Simulator. IEEE 2008.
Paolucci, P.S. "The Diopsis Multiprocessor Tile of SHAPES," MPSOC'06, 6th Int.Forum on
Application-Specific MPSoC, 2006.
Katalin Popovici, Xavier Guerin, Frédéric Rousseau Pier Stanislao Paolucci, Ahmed Amine
Jerraya: Platform-based software design flow for heterogeneous MPSoC. ACM Trans. Embedded
Comput. Syst. 7(4) (2008) DOI=10.1145/1376804.1376807
Xavier Guerin, Frédéric Pétrot: A System Framework for the Design of Embedded Software
Targeting Heterogeneous Multi-core SoCs. ASAP 2009: pp. 153-160
W. Haid, K. Huang, I. Bacivarov, and L. Thiele, “Multiprocessor SoC Software Design Flows”, in
IEEE Signal Processing Magazine, 26(6):64—71, Nov. 2009.
W. Haid, M. Keller, K. Huang, I. Bacivarov, and L. Thiele, “Generation and Calibration of
Compositional Performance Analysis Models for Multi-Processor Systems”, in Proc. Int'l Conf. on
Embedded Computer Systems: Architectures, Modeling and Simulation (SAMOS), pages 92—99,
Samos, Greece, July 2009. - awarded the Stamatis Vassiliadis Best Paper Award.
51 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
T. Sporer, M. Beckinger, A. Franck, I. Bacivarov, W. Haid, K. Huang, L. Thiele, P. S. Paolucci, P.
Bazzana, P. Vicini, J. Ceng, S. Kraemer, and R. Leupers, “SHAPES — A Scalable Parallel
HW/SW Architecture Applied to Wave Field Synthesis”, in Proc. Audio Engineering Society
(AES) Int'l Conference, pages 175-187, Hillerød, Denmark, Sept. 2007.
C. Schumacher, R. Leupers, D. Petras and A. Hoffmann. “parSC: Synchronous Parallel SystemC
Simulation on Multi-Core Host Architectures”, in proceedings of international conference on
hardware/software codesign and system synthesis (CODES+ISSS) 2010, October, 2010,
Scottsdale, Arizona, USA
S. Kraemer, L. Gao, J. H. Weinstock, R. Leupers, G. Ascheid and Meyr H., “HySim: A Fast
Simulation Framework for Embedded Software Development,” in proceedings of the 5th
Conference on Hardware/Software Codesign (CODES+ISSS '07) and System Synthesis (Salzburg,
Austria), 2007.
Chagoya-Garzon, A.; Guerin, X.; Rousseau, F.; Petrot, F.; Rossetti, D.; Lonardo, A.; Vicini, P.;
Paolucci, P.S., "Synthesis of Communication Mechanisms for Multi-tile Systems Based on
Heterogeneous Multi-processor System-On-Chips," Rapid System Prototyping, 2009. RSP '09.
IEEE/IFIP International Symposium on , vol., no., pp.48,54, 23-26 June 2009
doi:10.1109/RSP.2009.12
P. S. Paolucci et al. "Complex Domain Floating Point VLIW DSP WithData/Program Bus
Multiplexer and Microprocessor Interface", US Patent 7437540, (2008)
Popovici, K.; Guerin, X.; Rousseau, F.; Paolucci, P.S.; Jerraya, A., "Efficient Software Development
Platforms for Multimedia Applications at Different Abstraction Levels," Rapid System
Prototyping, 2007. RSP 2007. 18th IEEE/IFIP International Workshop on, pp.113,122, 28-30
May 2007 doi: 10.1109/RSP.2007.21
Paolucci, P.S. “Apparatus and Method for Dynamic Program Decompression”, US Patent 6766439
(2004), filed 2002.
P. S. Paolucci, P. Kajfasz, P. Bonnot, B. Candaele, D. Maufroid, E. Pastorelli, A. Ricciardi, Y.
Fusella, and E. Guarino, “mAgic-FPU and MADE: A customizable VLIW core and the modular
VLIW processor architecture description environment,” Computer Physics Communications, vol.
139, no. 1, pp. 132–143, September 2001.
P.S. Paolucci et al. “Janus: A gigaflop VLIW+RISC Soc Tile”, Hot Chips 15 IEEE Stanford
Conference (2003). http://www.hotchips.org (note: Janus was the development name of Diopsis).
G. Mariani, P. Avasare, G. Vanmeerbeeck, C. Ykman-Couvreur, G. Palermo, C. Silvano, V.
Zaccaria, “An Industrial Design Space Exploration Framework for Supporting Run-Time
Resource Management on Multi-Core Systems”, in Proc. DATE , pages 196–201, 2010
S. Stuijk, M. Geilen, Twan Basten, “A Predictable Multiprocessor Design Flow for Streaming
Applications with Dynamic Behaviour”, in Proc. DSD, pages 548–555, 2010
52 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
P. van Stralen and A. Pimentel, “Scenario-Based Design Space Exploration of MPSoCs”, in Proc.
ICCD, pages 305-312, 2010.
Z. Wang and J. Henkel: “HyCoS: hybrid compiled simulation of embedded
software with target dependent code,” in proceedings of the 8th Conference
on Hardware/software codesign and system synthesis (CODES+ISSS '12), Oct
2012, Tampere (Finland)
W. Chen, X. Han, and R. Domer, "Out-of-order parallel simulation for
ESL design," Design, Automation and Test in Europe Conference and Exhibition (DATE), 2012
Modha, S. D., & al. (2011). Cognitive Computing. Communications of the ACM , 54 (08 (pag. 6271)).
I. Bacivarov, W. Haid, K. Huang, and L. Thiele, “Methods and Tools for Mapping Process Networks
onto Multi-Processor Systems-On-Chip”, in “Handbook of Signal Processing Systems”, Springer,
pages 1007-1040, Oct. 2010
L. Schor, H. Yang, I. Bacivarov, and L. Thiele, “Worst-Case Temperature Analysis for Different
Resource Models”, in IET Circuits, Devices & Systems Journal, Volume 6, Issue 5, p. 297-307,
Sep. 2012
K. Huang, W. Haid, I. Bacivarov, M. Keller, and L. Thiele, “Embedding Formal Performance
Analysis into the Design Cycle of MPSoCs for Real-time Multimedia Applications”, in ACM
Transactions on Embedded Computing Systems (TECS), Volume 11, Issue 1, p. 8:1-8:23, 2012
S.-H. Kang, H. Yang, L. Schor, I. Bacivarov, S. Ha, and L. Thiele, “Multi-Objective Mapping
Optimization via Problem Decomposition for Many-Core Systems”, in Proc. IEEE Symposium on
Embedded Systems for Real-Time Multimedia (ESTIMedia), Tampere, Finland, p. 28-37, Oct.
2012
L. Schor, I. Bacivarov, D. Rai, H. Yang, S.-H. Kang, and L. Thiele, “Scenario-Based Design Flow
for Mapping Streaming Applications onto On-Chip Many-Core Systems”, in Proc. Int'l Conf. on
Compilers Architecture and Synthesis for Embedded Systems (CASES), Tampere, Finland, p. 7180, Oct. 2012
D. Rai, H. Yang, I. Bacivarov, and L. Thiele, “Power Agnostic Technique for Efficient Temperature
Estimation of Multicore Embedded Systems”, in Proc. Int'l Conf. on Compilers Architecture and
Synthesis for Embedded Systems (CASES), Tampere, Finland, p. 61-70, Oct. 2012 – best paper
award
Katalin Popovici, Frédéric. Rousseau, Ahmed Jerraya, Marilyn Wolf, editors et authors, Embedded
Software Design and Programming of Multiprocessor System-on-Chip: Simulink and SystemC
Case Studies, Springer, p 290, ISBN-13: 978-1441955661, april 2010.
Alexandre Chagoya-Garzon, Nicolas Poste, Frédéric Rousseau: Semi-automation of Configuration
53 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Files Generation for Heterogeneous Multi-tile Systems. COMPSAC 2011, pp. 157-166
Alexandre Chagoya-Garzon, Frédéric Rousseau, Frédéric Pétrot: Multi-device Driver Synthesis Flow
for Heterogeneous Hierarchical Systems, DSD 20012, pp. 389-396.
C. Schumacher, J. H. Weinstock, R. Leupers, G. Ascheid, L. Tossorato, A. Lonardo, D. Petras, T.
Groetker: "legaSCi: Legacy SystemC Model Integration into Parallel SystemC Simulators". 1st
Workshop on Virtual Prototyping of Parallel and Embedded Systems (ViPES), 2013, Boston,
USA
(accepted for publication)
L. G. Murillo, J. Eusse, J. Jovic, S. Yakoushkin, R. Leupers and G. Ascheid: “Synchronization for
Hybrid MPSoC Full-System Simulation,” Design Automation Conference (DAC '12), Jun 2012,
San Francisco (USA)
L. G. Murillo, J. Harnath, R. Leupers and G. Ascheid: “Scalable and Retargetable Debugger
Architecture for Heterogeneous MPSoCs,” System, Software, SoC and Silicon Debug Conference
(S4D '12), Sep 2012, Vienna (Austria)
Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Francesca Lo Cicero, Alessando Lonardo,
Pier Stanislao Paolucci, Davide Rossetti, Francesco Simula, Laura Tosoratto, Piero Vicini.
APEnet+: a 3D Torus network optimized for GPU-based HPC Systems 2012 J. Phys.: Conf.
Ser. 396 042059 doi:10.1088/1742-6596/396/4/042059
Andrea Biagioni, Francesca Lo Cicero, Alessandro Lonardo, Pier Stanislao Paolucci, Mersia Perra,
Davide Rossetti, Carlo Sidore, Francesco Simula, Laura Tosoratto, Piero Vicini. The Distributed
Network Processor: a novel off-chip and on-chip interconnection network architecture
http://arxiv.org/abs/1203.1536 (2012)
Roberto Ammendola, Andrea Biagioni, O Prezza, F Lo Cicero, Alessandro Lonardo, Pier Stanislao
Paolucci, Davide Rossetti, Andrea Salamon, Gaetano Salina, Francesco Simula, Laura Tosoratto,
Piero Vicini. APEnet+: high bandwidth 3D torus direct network for petaflops scale commodity
clusters 2011 J. Phys.: Conf. Ser. 331 052029 doi:10.1088/1742-6596/331/5/052029,
http://iopscience.iop.org/1742-6596/331/5/052029
Ammendola, R.; Biagioni, A.; Frezza, O.; Lo Cicero, F.; Lonardo, A.; Paolucci, P.S.; Rossetti, D.;
Simula, F.; Tosoratto, L.; Vicini, P., "QUonG: A GPU-based HPC System Dedicated to LQCD
Computing," Application Accelerators in High-Performance Computing (SAAHPC), 2011
Symposium on , vol., no., pp.113,122, 19-21 July 2011 doi: 10.1109/SAAHPC.2011.15, IEEE
2011
Davide Rossetti, Roberto Ammendola, Piero Vicini, Andrea Biagioni, Ottorino Frezza, Francesco Lo
Cicero, Alessandro Lonardo, Pier Stanislao Paolucci, Francesco Simula, Laura Tosoratto.
APEnet+ project status. Proceedings of the XXIX International Symposium on Lattice Field
Theory (Lattice 2011). July 10-16, 2011. Squaw Valley, Lake Tahoe, California, id.45 Online
at http://pos.sissa.it/cgi-bin/reader/conf.cgi?confid=139
54 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
Ammendola, R.; Biagioni, A.; Chiodi, G.; Frezza, O.; Lo Cicero,F.; Lonardo, A.; Lunadei, R.;
Paolucci, P.S.; Rossetti, D.; Salamon, A.; Salina, G.; Simula, F.; Tosoratto, L.; Vicini. P., “Highspeed data transfer with FPGAs and QSFP+ modules”, 2010 JINST 5 C12019 doi:10.1088/17480221/5/12/C12019
55 of 56
Project: EURETILE – European Reference Tiled Architecture Experiment
Grant Agreement no.: 247846
Call: FP7-ICT-2009-4 Objective: FET - ICT-2009.8.1 Concurrent Tera-device Computing
EURETILE 2010-2012 summary: first three years of activity of the European Reference
Tiled Experiment.
Pier Stanislao Paolucci, Iuliana Bacivarov, Gert Goossens, Rainer Leupers, Frédéric Rousseau,
Christoph Schumacher, Lothar Thiele, Piero Vicini
The EURETILE project is funded by the European Commission through the Grant Agreement no.
247846 Call: FP7-ICT-2009-4 Objective FET-ICT-2009.8.1 Concurrent Tera-device Computing.
www.euretile.eu
| 6cs.PL
|
Co-evolving Real-Time Strategy Game Micro
Navin K Adhikari∗ , Sushil J. Louis† Siming Liu‡ , and Walker Spurgeon§
arXiv:1803.10314v1 [cs.NE] 27 Mar 2018
Department of Computer Science and Engineering
University of Nevada, Reno
Email: ∗ [email protected], † [email protected] ‡ [email protected], and § [email protected]
Abstract—We investigate competitive co-evolution of unit micromanagement in real-time strategy games. Although good longterm macro-strategy and good short-term unit micromanagement
both impact real-time strategy games performance, this paper focuses on generating quality micro. Better micro, for example, can
help players win skirmishes and battles even when outnumbered.
Prior work has shown that we can evolve micro to beat a given
opponent. We remove the need for a good opponent to evolve
against by using competitive co-evolution to evolve high-quality
micro for both sides from scratch. We first co-evolve micro to
control a group of ranged units versus a group of melee units.
We then move to co-evolve micro for a group of ranged and
melee units versus a group of ranged and melee units. Results
show that competitive co-evolution produces good quality micro
and when combined with the well-known techniques of fitness
sharing, shared sampling, and a hall of fame takes less time to
produce better quality micro than simple co-evolution. We believe
these results indicate the viability of co-evolutionary approaches
for generating good unit micro-management.
Index Terms—Co-evolutionary genetic algorithm, Influence
map, Potential field, Real-time strategy game,Micro
I. I NTRODUCTION
Real-Time Strategy (RTS) games have become a new research frontier in the field of Artificial Intelligence (AI) as they
represent a challenging environment for an autonomous agent.
An RTS game player needs to collect resources, use those
resources to construct a base, train units, research technologies
and control different types of units to defeat the opponent
while at the same time defending their own base from opponent attacks in a complex dynamic environment. All of
these different actions that can be executed in any given state
make a huge decision space for a player. RTS players usually
divide these decision spaces into two different levels of tasks:
macromanagement and micromanagement. Macromanagement
encompasses a wide variety of tasks such as collecting more
resources, constant unit production, technology upgrades, and
scouting. In contrast, micromanagement is the ability of a
player to control a group of units to beat an opponent. Better
micro, for example, can help players win skirmishes and
battles even when outnumbered or minimize damage received.
Although good long-term macro-strategy and good shortterm unit micromanagement both impact real-time strategy
games performance, this paper focuses on generating quality
micro. More specifically, we focus on two aspects of the
micromanagement: tactics and reactive control [1]. Tactics deal
with the overall positioning and movement of a group of units
while reactive control deals with controlling a specific unit to
achieve commonly used micro techniques: concentrating fire
on a target, retreating seriously damaged units from the front
line of the battle, and kiting (hit and run).
We build on prior work [2] and represent micro-behaviors
of units on both sides with a set of parameters. We use a
commonly used technique called Influence Maps (IMs) to
represent enemy distribution over the game map. An IM is
a grid placed over the map with a value assigned to each grid
cell using an IM function that depends on the number of enemy
units in the vicinity. Good IMs can tell us where the enemy
is strongest and where they are weakest (That is, the best
target position for friendly units to go). To navigate a group
of units to a target location on the map given by an IM, we
use Potential Fields (PFs). PFs are used widely in multi-agent
systems for coordinated movement of multiple agents [3] [4].
We use two IMs parameters and four PF parameters for tactics
and six parameters for reactive control. We then use a Coevolutionary Genetic Algorithm (CGA) to search and find
good combinations of these twelve parameters that lead to
a good micro behavior on both sides.
Prior work has shown a Genetic Algorithm (GA) can
evolve good micro to beat a given opponent but that micro
performance depends on having a good opponent to play
against. Furthermore, it is non-trivial to hard-code a good
opponent to play against. We remove the need for a good
opponent to evolve against by using competitive co-evolution
to evolve high-quality micro for both sides from scratch.
In competitive co-evolution, two populations evolve against
each other. That is, individuals in one population play against
individuals in the other population for evaluating each other’s
fitness. Note that the fitness of an individual in one population
depends on the fitness of the opponents drawn from the other
population. As both population evolve, individuals from each
population must compete against more and more challenging
opponents leading to an arms race. This simple model of
co-evolution suffers from several well known problems [5].
Although even this simple model of co-evolution works well
enough to produce better than random micro, we use three
techniques: competitive fitness sharing, shared sampling and
hall of fame from Rosin and Belew [6] to produce better
quality micro in less time than using simple co-evolution.
We first co-evolve micro to control a group of ranged
units versus a group of melee units. We then move to coevolve micro for a group of ranged and melee units versus
an opponent group of ranged and melee units. Results show
that we can co-evolve good micro for both opponents in both
scenarios. In addition, we tested generalizability of the co-
evolved micro by evaluating performance of co-evolved micro
in different initial configurations and different initial positions
of units. Results show that micro co-evolved in one scenario
work well in other scenarios as well.
The remainder of this paper is organized as follows. Section
II describes related work in RTS AI research, generating
game players using co-evolutionary techniques and common
techniques used in RTS micro. The next section describes
our RTS research environment. Section III explains our
CGA implementation. Section IV presents preliminary results.
Finally, the last section provides conclusions and discusses
possible directions for future work.
II. R ELATED W ORK
Traditionally, much work has been done in Computational
Intelligence (CI) and AI in games revolving around board
games, using a variety of techniques [7] [8]. More recently,
research has shifted away from board game towards more
complex computer games and real-time strategy games like
starcraft pose several challenges for computational intelligence
research [9]. In addition, challenges in RTS games are strikingly similar to real-world challenges making RTS games a
good research platform.
Much work has been done on RTS games addressing different aspects of developing an AI player for such games [1].
In this paper, we are interested in one aspect of an RTS game:
“Micro.” Micro stands for micromanagement, the reactive and
tactical control of a group of units to maximize their effectiveness (usually) in combat. Game tree search techniques and
machine learning approach have been explored for micro tactics and reactive control. Churchill and Buro explored a game
tree search algorithm for tactical battles in RTS games [10].
Synnaeve and Bessiere applied bayesian modeling to inverse
fusion of the sensory inputs of the units for integration of
tactical goals directly in micro-management [11]. Wender and
Watson evaluated different reinforcement learning for micromagement [12].
A number of micro techniques use influence maps and
potential fields for tactical and reactive control of units in an
RTS game. An Influence Map (IM) tiles a map into square tiles
with each tile or grid cell getting a value provided by an IM
function. Grid cell values determine enemy unit locations and
concentrations and can be used to provide a variety of useful
information for unit maneuvering. We describe influence maps
and potential fields later in this paper. Miles evolved the
parameters of an influence map using a genetic algorithm in
order to evolve an RTS game player [13]. Sweetser and Wiles
used IMs to help a decision-making agent in their EmerGEnt
game [14]. Bergsma and Spronck generated adaptive AI for
a turn-based strategy game using IMs [15]. Jan and Cho
used information provided by layered IMs to evolve nonplayer
characters’ strategies in the strategy game Conqueror [16].
Preuss investigated flocking-based and IM-based path-finding
algorithms to optimize group movement in the RTS game
Glest [17] [18]. Uriarte and Ontanon used an IM-based
approach to evolve kiting (similar to hit-and-run) behavior in
the Starcraft bot Nova [19]. Danielsiek investigated influence
maps to support flanking the opponent in a RTS game [20].
This paper uses influence maps to determine a target location
to move towards and attack. Potential fields guide our unit
movement.
Potential fields (PFs) of the form cde where d is distance and
c and e are tunable parameters have been used in robotics and
games for generating smooth group movement [21]. Relevent
to RTS games, Jonas and Kostler used PFs to control units
optimally in StarCraft II for simulating optimal fights [22].
Sabdberg and Togelius Hagel investigated multi agent potential
field based AI approach for small scale combat in RTS
games [23]. Rathe and Svendsen did unit micromanagement
in Starcraft using potential fields [24]. They all used genetic algorithm to tune multiple potential fields’ parameters.
Hagelback and Johansson applied potential fields in their RTS
games research [25]. They proposed a multiagent PF-based
bot architecture for the RTS games ORTS and applied PFs for
tactical and reactive unit movement. Closer to our work, Liu
and Louis used parameterized algorithms that determined unit
positioning, movement, target selection, kiting, and fleeing.
Then a genetic algorithm tuned these parameters by evolving
against a hand-coded opponent or an existing Starcraft BWAPI
bot [2]. We build on this prior work and use the same
representation (parameterized algorithms) but co-evolve, rather
then evolve, micro without the need for a good opponent to
play against.
Coevolution in games goes back to Shannon’s work on
checkers in the 50s with the most recent notable example being
Alpha-go Zero [26] [27]. In RTS games, Ballinger and Louis
showed that coevolution led to more robust build orders. Buildorder optimization enables players to generate the right mix
and numbers of units meeting a strategic need [28]. Avery and
Louis coevolved team-tactics using a set of IMs, navigating a
group of friendly units to move and attack enemy units on
the basis of the opponent’s position [29]. More relevant to our
coevolutionary approach, Rosin and Belew improved the performance of a coevolution using three techniques: competitive
fitness sharing, shared sampling and hall of fame [6]. We use
these techniques and show their effectiveness in coevolving
good micro in less time than simple coevolution.
The next section describes our game engine and provides
details on how we simulate skirmishes in this game engine
for fitness evaluation. We then describe our representation and
evolutionary algorithm tuned parameters and our methodology
for measuring coevolutionary progress.
III. M ETHODOLOGY
Apart from the Starcraft BWAPI and Starcraft II API, there
now exist a number of other RTS game-like engines that
can be used for RTS game research [30] [31]. The opensource FastEcslent game engine which runs game graphics
in a separate thread is especially suitable for evolutionary
computing research in games since we can run the underlying
game simulation without graphics and thus more easily do
Fig. 1: Screenshot of a skirmish in FastEcslent
multiple parallel evaluations. Figure 1 shows a screenshot from
FastEcslsent running with graphics.
We use unit health, weapons, and speed values from Starcraft to create the equivalent of Vultures and Zealots in
FastEcslent [32] [30]. A Vulture is StarCraft unit that is fast
but fragile and can attack form a longer distance which helps
such units “kite,” during a skirmish while a Zealot is slower
but stronger and has a shorter attack distance. To evaluate
the fitness of a chromosome, we decode the chromosome and
use the twelve resulting parameters to control our units in the
game simulation. The simulation ends when all the units on
one side are destroyed or time runs out. After each simulation,
FastEcslent returns a score for each side in the skirmish based
on how much damage was done and how much damage was
received and this score is used to compute a fitness.
The goal of our work is to evolve good micro for opponents in an RTS game without the need of an opponent
to evolve against. We therefore use a coevolutionary algorithm to achieve this goal. In coevolution, two populations
of individuals play each other to compute fitnesses that drive
evolution [6]. Extending prior work, we represent micro by
a set of parameterized algorithms and the genetic or coevolutionary algorithm tunes these parameters to find good
micro. These algorithms specify group positioning, movement,
target selection, kiting, and fleeing. Table 2 details the twelve
parameters in our representation which is identical to the
representation used by Liu [2]. Tuning these parameters results
in micro for one type of friendly unit against one type of
enemy unit. We explain these parameters below.
Good positioning during a skirmish can reduce damage
received and increase damage dealt to opponent units. We
use Influence Maps (IMs) to try and find vulnerable positions
to attack. An influence map is a grid of cells placed over
the map, where each cell has a value determined by an IM
function. In our work, the IM function specifies a weight
parameter (We ) for each cell occupied by an enemy entity.
The entitys influence decreases as a function of distance and
ceases after Re distance (in number of cells) units. We sum
the influence of all enemy entities in range of cell to compute
the cells IM value. The cell with the lowest value and that is
closest to the enemy determines our attack location. The GA
Fig. 2: Parameters tuned by coevolution
or coevolutionary algorithm determines We and Re .
How a group of units moves to the target location also
determines skirmish performance. We use attractive and repulsive potential fields to control group movement []. The typical
representation of an attractive and a repulsive potential field
is given by equation 1
PF = ca dea + cr der
(1)
where Ca and Ea are parameters of the attractive force and Cr
and Er are parameters of repulsive force. However, balancing
these forces to achieve smooth, effective unit movement is
difficult and we therefore use the CGA to find the best values
for these parameters. Once we reach the target location, target
selection, kiting, and fleeing become important. Good target
selection can significantly affect performance since selecting a
weaker target, to destroy more quickly, can thus more quickly
reduce damage being received. The CGA evolves the two
parameters, HPe f and Rn t, defined in Table 2 to guide a
target selection algorithm [2].
Kiting by longer ranged units is an effective tactic used by
good human players in skirmishes with short ranged melee
units. Three parameters that determine kiting behavior are 1)
how far away from a target the unit needs to start kiting (Dk ),
2) the waiting time before moving after each firing (st ), and
3) how far a unit should retreat before attacking (Dk b). We
use a parameterized kiting algorithm which uses these three
parameters to kite [2]. Finally, removing weakened units from
the front line to save them for later is determined by a hit-point
threshold HPf b, also coevolved by the CGA. Good values
for all these parameters can lead to micro that beats state of
the art BWAPI competition bot micro [9] [2] when evolved
against such micro. This paper seeks to use CGAs to reach
high levels of performance without the need for good micro
to evolve against.
A. Coevolution and Fitness Evaluation
In coevolution, individual fitnesses result from direct competition between individuals in two populations. We want
to maximize damage done and minimize damage received.
More precisely, when an individual i, from one population
competes against individuals from the other population, i gets
a score given by equation 2, based on damage done by Nf
friendly units to Ne enemy units and damage received in each
competition.
Score =
+
V1
V2
N
Pf
(HPf /HP Fmax )
n=1
Ne
P
(2)
(HP Emax − HPe )
n=1
HP Fmax is the starting hitpoints corresponding to maximum
health for each friendly unit. Similary HP Emax specifies the
starting hitpoints of each enemy unit. HPf represents the
remaining hitpoints for friendly units at the end of a fitness
simulation while HPe represents the same parameter for
enemy units. V1 and V2 are scores for saving friendly hitpoints
(health) or reducing enemy hitpoints. We obtain these values
from the Starcraft BWAPI. We explain how this score leads
to an individual’s fitness after describing the coevolutionary
algorithm.
Since we are coevolving both sides, we refer to the two
sides coevolving in their distinct populations as red and blue.
Figure 3 shows how individuals in the blue and red populations
are evaluated and how a single evaluation determines the
fitness of two individuals - one from the blue and one from
the red population.
B. One unit type versus one unit type
Our first experimental scenario coevolved 5 red Vultures
against 25 blue Zealots. For each individual in the blue
population, we send the 12 parameters specified by that
individual to control micro for the 25 blue side zealots against
every individual in the red population. Each red individual’s
chromosome controls the 5 vultures. For a population size,
p, and assuming both red and blue have the same population
size, we need a total of p2 evaluations to obtain a fitness for
every individual in both populations.
Equation 2 specified the score received during one evaluation; an individual’s fitness is the average of all the scores
obtained by playing one red individual (for example) against
all p members of the blue population. V1 and V2 differ for
the red and blue populations since each is trying to micro
a different type of unit. V1 = 400, V2 = 160, HP Fmax =
80, HP Emax = 160, Ne = 25 (zealots), and Nf = 5 vultures
for the red population. From the blue population’s point of
view, these parameter values are different. Blue friend Zealots
compete against red enemy Vultures and V1 = 160, V2 =
80, HP Fmax = 160, HP Emax = 80, Ne = 5 (vultures),
Fig. 3: The coevolutionary algorithm plays individuals from
the blue population against individuals in the red population
to obtain damage done and received and thus determine
relative fitness.
and Nf = 25 zealots for the blue population. Except for the
number of enemies, Ne , and number of friends, Nf , the values
of all other parameters are obtained from the Starcraft1.
C. Two unit types versus two unit types
Good results from coevolving micro for groups composed
from one type of unit versus groups also composed from one,
albeit differnt, type of unit led us to consider a second set
of experiments where we investigated coevolving micro for
groups composed from two types of units against an opponent
group also composed from two types of units. Specifically, we
coevolved micro for a group of 5 vultures and 25, say on the
red side against an identical group of 5 vultures and 25 zealots
on the blue side. Our chromosomes doubled in size from 12 to
24 parameters and the first 12 parameters controlled vultures
while the second set of 12 parameters controlled zealots.
We also generalized Equation 2 to handle multiple types of
friend and enemy units. Essentially this means that there are
two values for V1 , one for vultures (400) and one for zealots
(160). Similarly there are two values for V2 when considering
damage to enemy vultures (80) and zealots (160). Maximum
values for hitpoints also depend on the unit type.
For simple competitive coevolution, the fitness of an individual is the average of scores obtained from playing against
all individuals in the opponent population. An individual plays
against another by being placed in our game and running the
game until either all the units in one side are destroyed or
time runs out. Once we have such a measure of fitness, the
two populations can potentially coevolve leading to an armsrace of increasing fitness.
Although this model of coevolution works well enough to
produce better than random micro, we use three techniques:
competitive fitness sharing, shared sampling, and hall of fame
as described by Rosin and Belew [6] to produce better quality
micro in less time than using simple coevolution [33]. We
provide brief descriptions of these three methods below.
The idea of fitness sharing is to prevent diverse niches from
prematurely going extinct. Sharing an individuals score from
defeating a specific individual i drawn from the opponent
population among all the individuals that defeated i, leads to
higher fitness for individuals that defeat opponent individuals
that no one else can. This decreases the probability of important innovations going extinct.
The usual way to evaluate an individual is to play against
all the individuals in the opponent population. To reduce
computational effort, shared sampling evaluates an individual
by playing against a sample of individuals drawn from the
opponent population. In order to increase the diversity in
this opponent sample, first select an opponent individual A
that defeated the most individuals in your population. Then
an individual defeating those individuals that defeated A are
selected, and so on, until the sample size becomes full.
A finite population means that a high fitness individual
from one generation may not stay high fitness in a different
context provided by an evolving opponent population. To
ensure against permanent loss of such strong individuals and
to prevent cycling caused by intransitive superiority, we keep
such current strong individuals in a hall of fame so that we can
use a strong diverse sample of past (hall of fame) and current
individuals to play against in order to gain a better measure
of an individuals fitness. This helps evolve individuals that are
more robust.
IV. R ESULTS AND D ISCUSSION
For all evaluations, we ran for a maximum of 2500 frames
which, despite running without graphics, took an average of 5
seconds per evaluation. We therefore parallelized evaluations
to get reasonable run times and achieved approximately linear
speedup. Coevolution run results are averaged over ten runs
with different random seed.
First, we coevolved micro for a group of ranged units versus
a group of melee units with simple coevolution - that is
coevolution without any shared sampling, shared fitness, or
hall of fame. Second, we compared the results produced using
simple coevolution with the results produced when using all
three techniques.
Specifically, in the first set of experiments we coevolve 5
red vultures versus 25 blue zealots. Both populations used a
population size of 50 and we ran for 60 generations. Crossover
and mutation probabilities were 0.95 and 0.03 respectively.
Since the fitness of an individual, i, depends on the quality
of individuals from the opponent population that i competes
against in coevolution, plotting fitness over time does not
measure progress in the same that such plots do for standard
genetic algorithms. Instead, we use a different approach and
start by generating a baseline individual. Every coevolutionary
generation, we take the best individual from the blue (or
red) population and play this best individual against the
Fig. 4: Performance of coevolving vultures against the
baseline zealots
fixed baseline. As coevolution progresses, we expect the best
individual in subsequent generations to improve performance
over the fixed baseline.
Figure 4 plots the best coevolving vulture (red population)
against such a baseline zealot. This baseline zealot beats
1996/2000 randomly generated individuals for a 90% win
rate. The solid line shows simple coevolution while the dashed
line shows coevolution augmented with fitness sharing, shared
sampling, and hall of fame. We can see improvement over time
for both and we can see that the three techniques do improve
micro quality faster. To reduce the computational effort, the
size of shared sample and hall of fame should be as low as
possible. But, decreasing the size too much may reduce needed
diversity in the set of individuals selected for playing against.
We thus need a delicate balance between maintaining diversity
in the shared sample and hall of fame to play against, and low
computational effort. In these and subsequent experiments the
shared sample size and hall of fame size are both set to five (5)
- a value found through experimentation. With these settings
we get 50 population size ×10, shared sampling size plus hall
of fame size, ×2 for the two populations for a total of 1000
evaluations per coevolutionary generation. This equates to a
savings of 2500−1000
= 60% in terms of computational effort
2500
measured in number of evaluations.
Figure 5 shows a similar patterns when comparing the
coevolving zealots against a baseline vulture micro. Note
that in both sets of results using the three methods result in smoother performance curves. Videos of gameplay
show complex patterns of movement. Zealots learn to herd
vultures into a corner while vultures learn kiting and to
stay out of range of zealots. These videos are available at
https://www.cse.unr.edu/ navin/coevolution.
Building on these results, we next investigated more complex micro for groups composed from zealots and vultures
versus an opponent also built from the same two types of units.
Fig. 5: Performance of coevolving zealots against the
baseline vultures
Fig. 6: Performance of coevolving zealots and vultures
versus the baseline for the Red population
With two types of units we can also look to see whether, and
what kind of, cooperative behavior emerges between the two
types of units.
A. Two types of units versus two types of units
We investigate coevolving micro for a group of 5 vultures
and 25 zealots versus an identical opponent group (5 zealots,
25 vultures). We did not test simple coevolution, preferring
to use the coevolution with the three methods since the
augmented coevolutionary algorithm performs better and due
to a lack of time.
Again, we used a population size of 50 running for 60
generations with the same crossover and mutation probabilities
of 0.95 and 0.03 as before. Note that the chromosome size
needs to double so that micro for the two unit types coevolve to
take advantage of each unit type’s unique properties. Progress
is again measured against a baseline group of 5 vultures and
25 zealots that beats 99% of randomly generated opponents.
Figure 6 shows coevolutionary progress against this baseline
for the red population. Again we see fairly smooth (for coevolution) progress in finding increasingly good micro. Figure 7
shows, unsurprisingly, that the coevolving blue population
has similar performance improvement. These results seem
to indicate the potential for a coevolutionary approach to
coevolve good micro from scratch.
Next we consider the robustness of this coevolved micro
by testing the coevolved micro in scenarios hitherto unseen.
First, we looked at the micro coevolved for the one unit type
versus one unit type experiments and selected the best coevolved individual from both populations. We then played
these two best individuals in three different starting formations
(or scenarios) and ten different starting locations. We did the
same for the best individuals in the two unit types versus two
unit types experiments.
Figure 8 shows screenshots of these three scenarios. The
first distributes units within a circle (labeled 1), the second
Fig. 7: Performance of coevolving zealots and vultures
versus the baseline for the Blue population
uses a line formation (2), and the third distributes units
randomly (3). Blue and Red indicate side in the screenshot. We
describe our experiments and results with respect to these three
formations next. Red bars represent red population (vulture)
performance versus baseline zealots and blue bars represent
blue (zealot) performance versus baseline zealots.
B. Scenario 1: Circular
This was our training scenario in that coevolution took
place with units placed within this circle and always started in
the same initial positions during a fitness evaluation. To test
robustness we randomly changed the starting positions and
generated 10 randomly generated sets of starting positions
for the units. We tested the coevolved micro against our
baseline player on these 10 different scenarios and computed
the average score. Figure 9 shows that coevolved micro seems
Fig. 10: Performance of co-evolved player against baseline
in different scenarios; Red and Blue sides
Fig. 8: Snapshot of circular formation (1), line formation (2),
and random formation (3)
over 10 different initial position and shown in Figure 9. With
this formation, we can see that vultures do not fare well. We
address this in our future work.
Finally, Figure 10 shows the same information for the two
unit types versus two unit types experiments. We can see the
same trend. These figures indicate the potential for our CGA
approach to find good robust micro from scratch.
V. C ONCLUSIONS AND F UTURE W ORK
Fig. 9: Performance of co-evolved player against baseline in
different scenarios; Co-evolved Vulture (Red), Co-evolved
Zealot (Blue)
robust to starting position with performance similar to those
in Figure 4 and Figure 5.
C. Scenario 2: Line formation
In this scenario, units from both sides area placed in a line
opposite each other on the game map. With this formation,
we want to see how the coevolved micro does when changing
both the formation and the initial positions on this formation.
Again, we randomly generated 10 different sets of unit starting
positions on the line and averaged the score obtained by the
best coevolved micro against our baseline. The second set of
bars in Figure 9 shows that co-evolved micro does just as well
in this new formation over multiple sets of starting locations.
D. Scenario 3: Random starting locations
In this scenario, rather than putting units into any particular
formation, we place them randomly in the game. The score of
the best individual against baseline player is again averaged
Our research focuses on exploring coevolutionary approaches to finding good micro in RTS games. This eliminates
the need for a good opponent to evolve against. We compactly
represented micro with 12 parameters that control simple
algorithms for target selection, kiting, and unit movement and
used a coevolutionary algorithm to tunes these parameters
values. We measured the performance of two independent
coevolving populations by playing the best individual from
each generation and each population against a baseline player
seperately. Results show that we can coevolve a group of
ranged units versus a group of melee units using simple
coevolution. We also compare these results using three different techniques for improving competitive coevolution as
described by Rosin and Belew [6]. Results also indicates that
we can coevolve a better micro in less time than using simple
coevolution.
We then coevolved micro for units composed from two types
of units versus similar opponents. For a mix of ranged and
melee units results show that we can coevolve good microbehavior using coevolution augmented with shared sampling,
fall of fame, and shared fitness.
Both sets of solutions seem to be robust. We checked the
robustness of our co-evolved micro in three different unseen
scenarios and ten different sets of starting positions. Results
shows that our approach can find micro that performs well in
unseen scenarios. We believe, using a combination of random
and structured scenarios during coevolution will lead to more
robustness.
The main constraints with a coevolution in an RTS game
is computational effort for evaluations. As a single fight
simulation takes significant time and each individual needs
multiple evaluations, required computational effort tends to
outstrip available resources. Although shared sampling and
hall of fame result in good reduction, it still takes days to get
significant results. Given more computational resources, we
may be able to use much larger population sizes and much
longer run times to get significantly higher quality.
We believe these results indicate the viability of coevolutionary approaches for generating good unit micromanagement
and we plan to build on this in our future work. We would
like to investigate other representations and coevolve within
Starcraft II using the recently released Starcraft II API.
R EFERENCES
[1] S. Ontaon, G. Synnaeve, A. Uriarte, F. Richoux, D. Churchill, and
M. Preuss, “A survey of real-time strategy game ai research and competition in starcraft,” in IEEE Trans. Comput. Intell. AI Games,vol.5,no.4,
Dec 2013, pp. 1–19.
[2] S. Liu, S. J. Louis, and C. A. Ballinger, “Evolving effective microbehavior in real-time strategy games,” in IEEE Trans. Comput. Intell. AI
Games,vol.8,no.4, Dec 2016.
[3] R. Olfati-Saber, J. A. Fas, and R. M. Murry, “Consensus and cooperation
in networked multi-agent system,” in Proceedings of the IEEE vol.95,
no.1, 2007, pp. 215–233.
[4] M. Egerstedt and X. Hu, “Formation constrained multi-agent control,”
in IEEE tramsactions on, vol.17, no.6, 2001, pp. 947–951.
[5] J. Schonfeld, S. J. Louis, and J. Doe, “A survery of coevolutionary
algorithms research,” evolution, vol. 32, no. 67, p. 63.
[6] C. D. Rosin and R. K. Belew, “New methods for competitive coevolution,” in Evolutionary computation, 1997 - MIT Press, 1997, pp. 1–29.
[7] J. Furnkranz, “Machine learning in games: A survey,” in Machine that
Learn to Play Games, Huntington, NY, USA: Nova Science, 2001, pp.
11–59.
[8] J. Rubin and I. Watson, “Computer poker: A review,” in Artif. Intell.,
vol.175,no.5, 2011, pp. 958–987.
[9] M. Buro, “Real-time strategy games: A new ai research challenge,” in
IJCAI, 2003 - skatgame.net, 2003.
[10] D. Churchill, A. Saffidine, and M. Buro, “Fast heuristic search for rts
game combat scenarios,” in AIIDE, 2012.
[11] G. Synnaeve and P. Bessiere, “A bayesian model for rts units control
applied to starcraft,” in Proceedings of IEEE CIG, 2011.
[12] S. Wender and I. Watson, “Applying reinforcement learning to small
scale combat in the real-time strategy game starcraft:broodwar,” in IEEE
CIG, 2012.
[13] C. Miles, J. Quiroz, R. Leigh, and S. Louis, “Co-evolving influence map
tree based strategy game players,” in Proc. IEEE symp. Comput. Intell.
Game, April 2007, pp. 88–95.
[14] P. 6th Int. Conf. Intell. Data Eng. Autom. Learning, “Combining
influence maps and cellular automata for reactive game agents,” in Proc.
IEEE symp. Comput. Intell. Game, 2005, pp. 209–215.
[15] M.Bergsma and P. Spronck, “Adaptive spatial reasoning for turn-based
strategy games,” in Proc. Artif. Intell. Interactive Digital Entertain.
Conf., 2008.
[16] S.-H. Jang and S.-B. Cho, “Evolving neural npcs with layered influence
map in the real-time simulation game conqueror,” in Proc. IEEE Symp.
Comput. Intell. Games, Dec 2008, pp. 385–385.
[17] M. P. et al, “Towards intelligent team composition and maneuvering in
real-time strategy games,” in IEEE Trans. Comput. Intell. AI Games,
vol.2,no.2, Jan 2010, pp. 82–98.
[18] H. D. et al, “Intelligent moving of groups in real-time strategy games,”
in IEEE Sypm. Comput. Intell. Games, 2008, pp. 71–78.
[19] A. Uriarte and S. Ontann, “Kiting in rts games using influence maps,”
in Proc. 8th Artif. Intell. Interactive Digital Entertain. Conf., 2012.
[20] H. Danielsiek, R. Stuer, A. Thom, N. B. B. Naujoks, and M. Preuss,
“Intelligent moving of groups in real-time strategy games,” in IEEE
Symposium On Computational Intelligence and Games, 2008, pp. 71–
78.
[21] O. khatib, “Real-time obstacle avoidance for manipulators and mobile
rotobs,” in Int. J. Robot. Res. vol.5,no.1, Jun 1986, pp. 90–98.
[22] J. Schmitt and H. Kstler, “A multi-objective genetic algorithm for
simulating optimal fights in starcraft ii,” in IEEE CIG, 2016.
[23] T. W. Sandberg and J. Togelius, “Evolutionary multi-agent potential field
based ai approach for ssc scenarios in rts games,” in PHD thesis, 2011.
[24] E. A. Rathe and rgen Be Svendsen, “Micromanagement in starcraft using
potential fields tuned with a multi- objective genetic algorithm,” in Maste
thesis, 2012.
[25] J. Hagelback and S. Johansson, “Using multi-agent potential fields
in real-time strategy games,” in Proc. Artif. Intell. Interactive Digital
Entertainment Conf., 2008.
[26] C.E.Shannon, “Game playing machines,” in Journal of the Franklin
Institute Vol. 260 no. 6, 1995.
[27] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van
Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam,
M. Lanctot et al., “Mastering the game of go with deep neural networks
and tree search,” nature, vol. 529, no. 7587, pp. 484–489, 2016.
[28] C. Ballingers and S. Louis, “Comparing coevolution, genetic algorithms,
and hill-climbers for finding real-time strategy game plans,” in GECCO
’13 Companion Proceedings of the 15th annual conference companion
on Genetic and evolutionary computation, July 2013, pp. 47–48.
[29] P. Avery and S. Louis, “Coevolving influence maps for spatial team
tactics in a rts game,” in Proc. 12th -Annu. Conf. Genetic Evol. Comput.,
2010, pp. 783–790.
[30] A. Heinermann et al., “Bwapian api for interacting with starcraft:
Broodwar,” available from (accessed 2014-02-03), 2012.
[31] Starcraft ii api. [Online]. Available: http://us.battle.net/sc2/en/blog/
20944009/the-starcraft-ii-api-has-arrived-8-9-2017
[32] An application programming interface for interacting with starcraft:
Broodwar. [Online]. Available: https://bwapi.github.io
[33] W. D. Hillis, “Co-evolving parasites improve simulated evolution as an
optimization procedure,” Physica D: Nonlinear Phenomena, vol. 42, no.
1-3, pp. 228–234, 1990.
| 9cs.NE
|
1
Packet Reception Probabilities in Vehicular
Communications Close to Intersections
arXiv:1509.00399v2 [cs.SY] 11 Oct 2016
Erik Steinmetz Student Member, IEEE, Matthias Wildemeersch Member, IEEE,
Tony Q.S. Quek, Senior Member, IEEE, and Henk Wymeersch, Member, IEEE
Abstract
Vehicular networks allow vehicles to share information and are expected to be an integral part
in future intelligent transportation system (ITS). In order to guide and validate the design process,
analytical expressions of key performance metrics such as packet reception probabilities and throughput
are necessary, in particular for accident-prone scenarios such as intersections. In this paper, we analyze
the impact of interference in an intersection scenario with two perpendicular roads using tools from
stochastic geometry. We present a general procedure to analytically determine the packet reception
probability and throughput of a selected link, taking into account the geographical clustering of vehicles
close to the intersection. We consider both Aloha and CSMA MAC protocols, and show how the
procedure can be used to model different propagation environments of practical relevance. We show
how different path loss functions and fading distributions can be incorporated in the analysis to model
propagation conditions typical to both rural and urban intersections. Our results indicate that the
procedure is general and flexible to deal with a variety of scenarios. Thus, it can serve as a useful design
tool for communication system engineers, complementing simulations and experiments, to obtain quick
insights into the network performance.
E. Steinmetz and H. Wymeersch are with the Department of Signals and Systems, Chalmers University of Technology,
Gothenburg, Sweden, e-mails: {estein,henkw}@chalmers.se. E. Steinmetz is also with SP Technical Research Institute of
Sweden, Borås, Sweden. M. Wildemeersch is with the International Institute for Applied Systems Analysis (IIASA), Laxenburg,
Austria, e-mail: [email protected]. T.Q.S. Quek is with Singapore University of Technology and Design, Singapore, email: [email protected]. This research was supported, in part, by the European Research Council under Grant No. 258418
(COOPNET), the EU project HIGHTS (High precision positioning for cooperative ITS applications) MG-3.5a-2014-636537,
and VINNOVA under the program “Nationell Metrologi vid SP Sveriges Tekniska Forskningsinstitut”. Part of this work was
presented in [1].
2
I. I NTRODUCTION
Vehicular networks have gained considerable attention in the past years and are regarded as
one of the key components in future intelligent transportation systems (ITS) [2]–[7]. By the use
of wireless communication they allow vehicles to continuously share information with each other
and their surrounding (e.g., roadside infrastructure), in order to perceive potentially dangerous
situations in an extended space and time horizon [3]. This enables a new set of applications
that are expected to enhance both traffic safety and efficiency. These applications include lane
change assistance, cooperative collision avoidance, emergency vehicle warning, traffic condition
warning, tolling, hazardous location warning, speed management.
The IEEE 802.11p standard has been defined to meet the communication demand of these
applications, and 5G cellular networks standards are being developed to support device-todevice (D2D) communication [8]–[13]. However, different ITS applications clearly have different
requirements on the communication links, with the most stringent demands imposed by safetyrelated applications, with extremely low latencies (below 50 ms in pre-crash situations), high
delivery ratios (for full situational awareness), and relatively long communication ranges (to
increase the time to react in critical situations) [14]–[16]. These requirements, in combination
with a possible high density of vehicles, makes the design of vehicular communication systems
challenging. This is further exacerbated by high mobility and passing vehicles, which leads to
rapidly changing signal propagation conditions (including both severe multipath and shadowing)
and constant topology changes.
In order to guide and validate the communication system design, extensive simulations and
measurements are often used [16], [17], which are both time consuming and scenario-specific.
In order to obtain insight in scalability and performance, analytical expressions of key performance metrics are necessary, in particular for high velocity scenarios (in particular highways)
and accident-prone scenarios (e.g., intersections). Stochastic geometry is a tool to obtain such
expressions, and has been widely used in the design and analysis of wireless networks [18]. In
2-D planar networks, the analysis is well developed and a multitude of approaches to consider
both geographical and medium access control (MAC) induced clustering [19], [20] as well as
different types of fading [21]–[23] exist. However, in vehicular networks, where the location
of the nodes are restricted by the roads, previous work that includes the spatial statistics of
3
vehicles typically considers one-dimensional roads [24]–[28]. For these vehicular scenarios,
geographical clustering has been addressed in [24], while effects due to the 802.11p carrier
sense multiple access (CSMA) MAC protocol was studied in [25], [26], [29]. Hence, these
works enable communication system design for one-dimensional highway scenarios, but do not
capture well the salient effects of intersections. Intersections were considered explicitly in [1],
[30], which found that it is important to properly model the interference from different roads and
account for the distance of receivers to the intersection, i.e., to take into account the clustering
of cars around the intersection and the non-stationarity of the spatial distribution.
In this paper, we present a general procedure for the evaluation of packet reception probability
and throughput in intersection scenarios, and provide a model repository that can be used to adapt
to a variety of different environments of importance in the vehicular context. This includes both
rural and urban scenarios, different propagation conditions, and different MAC protocols. Latency
and mobility are not treated in this paper.
The remainder of the paper is organized as follows. Section II introduces the system model.
In Section III, we discuss typical characteristics of the vehicular channel and show how the
model can be tailored to different environments. In Section IV, we present a general procedure
to calculate the packet reception probability near an intersection, as well as the throughput.
Section V shows how the proposed procedure can be used to calculate these performance metrics
for a number of cases of practical relevance, and how different assumptions on loss function,
fading, and MAC protocols affect the analytical tractability of the packet reception probability.
Finally, Section VI summarizes and concludes the paper.
II. S YSTEM M ODEL
We consider an intersection scenario with two perpendicular roads, as shown in Figure 1. For
simplicity, we assume that the width of the two roads indicated by H and V can be neglected, and
that the roads each carry a stream of vehicles, modeled as one-dimensional homogeneous Poisson
point processes (PPPs). The intensity of vehicles on both roads is denoted by λH and λV , and the
point processes describing the location of the vehicles on the two roads are represented by ΦH ∼
PPP(λH ) and ΦV ∼ PPP(λV ). The positions of individual vehicles (also referred to as nodes)
on the two roads H and V are denoted by xi = [xi , 0]T and xi = [0, yi ]T , respectively, assuming
the roads are aligned with the horizontal and vertical axes. As both vehicle-to-vehicle (V2V)
4
and infrastructure-to-vehicle (I2V) communication is of interest, we consider a transmitter (Tx)
with arbitrary location xtx = [xtx , ytx ]T .1 The Tx broadcasts with a fixed transmission power
P . Without loss of generality, we consider a receiver (Rx) on the H-road at location xrx =
[xrx , 0]T , i.e., at a distance d = |xrx | away from the the intersection.2 The signal propagation
comprises power fading S and path loss l(xtx , xrx ). At the Rx, the signal is further affected by
white Gaussian noise with noise power N and interference from other concurrently transmitting
vehicles on the H- and V-road. The amount of interference experienced by the Rx depends on
the choice of MAC protocol. For a given MAC scheme, the position of interfering vehicles at a
given time can be represented by the thinned point processes ΦMAC
and ΦMAC
.3 We can express
H
V
the signal-to-interference-plus-noise ratio (SINR) as
P S0 l(xtx , xrx )
P
,
x∈ΦMAC P Sx l(x, xrx ) +
x∈ΦMAC P Sx l(x, xrx ) + N
SINR = P
H
(1)
V
where S0 denotes the fading on the useful link and Sx denotes the fading on an interfering link
for an interferer at location x. A packet is considered to be successfully received if the SINR
exceeds a threshold β.
Our aim is to analytically characterize (i) the probability that the Rx successfully receives
a packet sent by the Tx; (ii) the throughput of the link between Tx and Rx. This problem is
challenging due to the specific propagation conditions and interference levels experienced in
these intersection scenarios. In the next section, we will describe these in more detail.
Remark 1. While the scenario considered here is simple, it can easily be extended to cases
where the width of the roads can not be ignored (without introducing significant modeling
errors) by splitting the road into several lanes, each modeled as a new road/lane, as discussed
in Section V-A. Furthermore, general multi road/lane extensions can be used to explicitly model
interference from other roads in the surrounding.
1
Note that in the case the Tx is on one of the roads, it can belong to either ΦH or ΦV (but does not necessarily have to) as
the results still hold due to Slivnyak’s Theorem [18, Theorem A.5]
2
Note that due to the symmetry of the scenario this also captures the case when Rx is on the V-road
3
For a general MAC scheme, the thinning process is inhomogeneous.
5
V
V
V
Tx
Tx
xtx
d
H
H
Rx
Rx
(a) intersection scenario
H
xrx
(b) abstraction
Figure 1. Illustration of considered scenario: (a) A two-way intersection scenario in which each road carries a stream a vehicles,
(b) the abstraction used in modeling. The Tx (indicated by the blue car) can be at any location, while the target Rx (green car)
is located on road H. Other vehicles on the roads H and V, of which some transmit concurrently and cause interference, are
shown as grey cars.
III. M ODELS
IN
V EHICULAR C OMMUNICATION
Vehicular communication systems must be able to function in a large variety of conditions,
including in urban canyons and in rural settings. In this section, we will discuss characteristics
for vehicular channels that are important from an SINR point of view, and detail different models
regarding path loss, fading, and MAC protocol.
A. Power decay and blockage
Extensive measurement campaigns [16], [31]–[34] have been performed to characterize the
vehicular channel in a variety of propagation environments such as rural, highway, suburban,
and urban scenarios. As it is important to understand how the power decays with distance, much
efforts have been put into finding large-scale path loss models, which characterize the slope of
distance-dependent power loss in decibels (dB). We will distinguish between line-of-sight (LOS)
and non-line-of-sight (NLOS) propagation, depending on whether or not the direct LOS signal
between a Rx and a Tx is blocked. For LOS propagation, conventional path loss models, where
power decays approximately with the squared Euclidean distance between Rx and Tx are well-
6
accepted [16]. For NLOS propagation, e.g., in urban canyons, measurements indicate increased
loss over LOS propagation, with complex dependencies on the absolute position of Tx and Rx,
widths of the roads, and different loss exponents for own and orthogonal road [33], [34]. The
complexity of these models render them intractable when it comes to mathematical analysis,
so we rely on the simpler and more tractable Manhattan model, which was first proposed for
modeling of similar scenarios in the well-known WINNER II project [35]. In particular, to allow
for mathematical analysis, we suggest the path loss of the following form.
•
For NLOS propagation, where the direct line-of-sight (LOS) between the Rx and the Tx is
blocked by buildings and the signals have to propagate along the urban canyons formed by
the orthogonal streets, we use the Manhattan distance:
lM (xtx , xrx ) = A kxrx − xtx k−α
1 ,
(2)
where kxrx − xtx k1 is the ℓ1 norm, α > 0 is the path loss exponent, and A is a constant that
depends on several factors such as antenna characteristics, carrier frequency, and propagation
environment.
•
For LOS propagation, where the direct LOS between the Rx and the Tx is unobstructed,
we use the Euclidean distance
lE (xtx , xrx ) = A kxrx − xtx k−α
2 ,
(3)
where kxrx − xtx k2 is the ℓ2 norm. Note that the values of α and A might be different in
(3) and (2).
B. Random power variations due to fading
Fading refers to random fluctuations in the received power around the average received
power, given by the path loss. The fading experienced on a link depends on the scenario
and the environment, and is typically modeled as a random variable [36]. For example, near
a rural intersection, vehicles are likely to communicate via LOS links, and exponential fading
is considered an appropriate model [8], [34]. On the contrary, if the intersection is located in
an urban environment with tall buildings, the fading for NLOS links is modeled using a lognormal model [33], [34], with typical values on power variations with respect to the path loss for
NLOS intersections are in the range of 3–6 dB. Based on these empirical results, we will model
7
the fading as log-normal (and approximated by an Erlang random variable for mathematical
tractability – see Section IV-A2) for NLOS links and as exponential for LOS links.
C. MAC protocols
The MAC protocol governs when a user can access the channel, and aims to control the
interference in the network. The two most common MAC protocols for ad-hoc networks are
Aloha and CSMA. In Aloha, which is the simpler of the two, nodes that have a packet to send,
access the channel during a time slot with a probability p ∈ [0, 1]. In contrast, in CSMA, before
sending a packet, a node verifies that the channel is free by listening to the channel. Only if
the channel is free, the node transmits the packet. If the channel is busy, the node is forced
to wait a random back-off time before it can try again [17]. The 802.11p standard, which has
been designed for the first generation vehicular networks, will rely on a CSMA/CA (collision
avoidance) MAC. We will consider CSMA as well as Aloha, as Aloha is easier to analyze and
has been argued to exhibit similar performance as CSMA, especially for dense networks [26].
IV. S TOCHASTIC G EOMETRY A NALYSIS
From Section III, it is apparent that vehicular communication systems will operate under a
variety of propagation conditions. In this section, we describe a general and unified methodology
to compute the communication performance for all these conditions, as well as different MAC
protocols. In particular, we will determine (i) the packet reception probability P(β, xrx, xtx ),
i.e., the probability that a receiver located at xrx can successfully decode a transmission from a
transmitter located at xtx , in the presence of interferers on the H- and V-road; (ii) the throughput
T (β, xrx, xtx ), i.e., the expected rate for the link between the Rx and Tx at locations xrx and xtx ,
accounting for both the packet reception probability and the probability of gaining access to the
channel. Both P(β, xrx, xtx ) and T (β, xrx , xtx ) depend on the loss function, fading distribution,
and the MAC protocol. Note that the loss function and fading distribution relate to the power
decay and blockage as well as the random signal variations in the specific scenario, while the
MAC protocol relates to number of interferers and their locations. Several applications of this
methodology will be discussed in Section V.
8
A. Packet reception probability
To derive the packet reception probability for the intersection scenario, we start by accounting
for the fading distribution of the useful link. We express
P(β, xrx , xtx ) = Pr(SINR ≥ β)
= Pr S0 ≥ IH + IV + Ñ β/l(xtx , xrx )
in which Ñ = N/P and IH =
P
x∈ΦMAC
H
Sx l(x, xrx ) while IV =
P
x∈ΦMAC
V
(4)
Sx l(x, xrx ). Condi-
tioning on the path loss, we can now write the packet reception probability as
P(β, xrx, xtx )
o
n
= EIH ,IV F̄S0 IH + IV + Ñ β/l(xtx , xrx )
¨
=
F̄S0 t1 + t2 + Ñ β̃ fIH ,IV (t1 , t2 )dt1 dt2 ,
(5)
where β̃ = β/l(xtx , xrx) and F̄So (s0 ) is the complementary cumulative distribution function (CCDF)
of the random variable S0 , evaluated in s0 .
The expression (5) can be interpreted in two ways: (i) as the expectation of F̄S0 ((IH + IV +
Ñ )β/l(xtx , xrx )) with respect to the interference distribution; and (ii) as the transformation of the
interference distribution with a kernel function determined by the CCDF of the fading distribution
of the useful link. In either interpretation, the distributions of the interference and the fading
play an important role. Note that for all relevant fading distributions of the useful link, (5) will
result in the Laplace transform (LT) of the interference distribution or a function of LTs of the
interference distribution. It is therefore convenient to express these distributions through their
(LT) or, equivalently, their moment generating function (MGF).
1) LT of the interference: From (5), we see that the packet reception probability P(β, xrx, xtx )
is a function of the interference distribution, which itself depends on the location of the Rx and
the interferers, as well as their fading distributions and path loss. For a general MAC protocol
the interference from the H- and V-road are not independent. However, for the MAC protocols
studied in this paper the interference distribution factorizes as fIH ,IV (t1 , t2 ) = fIH (t1 )fIV (t2 ). In
fact, the interference is independently thinned on the H- and V-road in the case of Aloha, while
for the CSMA scheme we can approximate the joint interference distribution as the product of
the marginals, where the dependence is captured by a location dependent thinning of the original
9
PPPs [37]. This means that the interfering point processes ΦMAC
and ΦMAC
either are, or are
H
V
approximated as PPPs (for more details see Section IV-A3), and that we can focus on a single
road R ∈ {H, V }, with interference distribution fIR . The Laplace transform of fIR is defined as
LIR (s) = E[exp(−sIR )],
(6)
in which
IR =
X
Sx l(x, xrx ).
(7)
x∈ΦMAC
R
Substitution of (7) into (6) then yields
Y
(a)
LIR (s) = EΦ
ESx {exp (−s Sx l(x, xrx ))}
(8)
x∈ΦMAC
R
= EΦ
Y
x∈ΦMAC
R
ˆ
= exp −
(b)
+∞
LSx (s l(x, xrx))
λMAC
R
−∞
(x (z) , xtx ) (1 − LSx (s l(x(z), xrx ))) dz ,
(9)
(10)
where (a) holds due to the independence of the fading parameters, EΦ [·] is the expectation
operator with respect to the location of the interferers, and LSx (·) is the LT of the fading
distribution of the interfering link; (b) is due to the probability generating functional (PGFL) for
a PPP [18, Definition A.5], in which λMAC
(x(z), xtx ) represents the intensity of the PPP ΦMAC
,
R
R
which depends on the specific MAC protocol and in some cases on the transmitter’s location.
Note that in (10), the intensity is defined over z ∈ R, which represents the position along the
road R ∈ {H, V }, where
[z 0]T , R = H
x (z) =
.
[0 z]T , R = V
(11)
To determine LIR (s), we must be able to compute the integral (10), which involves knowledge
of λMAC
(x (z) , xtx ) and LSx (s).
R
Remark 2. The Laplace transform of the interference can also be computed using the principle
of stochastic equivalence [23], where the LT in case of an arbitrary fading distribution can be
found based on the LT in case of Rayleigh fading, given an appropriate scaling of the system
parameters.
10
2) LT of fading: For many relevant fading distributions, the LT is known, including for
exponential, Gamma, Erlang, and χ2 random variables. While the log-normal distribution is
harder to deal with, it can be approximated by the Erlang distribution [38], which combines
tractability with expressiveness. When Sx ∼ E (k, θ), i.e., an Erlang distribution with shape
parameter k ∈ N and rate parameter 1/θ > 0, then
LSx (s) = (1 + sθ)k .
(12)
As special cases, (i) k = 1 corresponds to an exponential distribution with mean θ; (ii) θ = 1/k
corresponds to Nakagami-m power fading.
Remark 3. When the fading of the useful link is exponentially distributed, (5) allows us to inter
pret P(β, xrx, xtx ) as the LT of the interference, so that P(β, xrx, xtx )= exp −Ñ β̃/θ LIH (β̃/θ)LIV (β̃/θ).
3) Intensity of the interfering PPPs: The intensity λMAC
(x (z) , xtx ) of the interference deR
pends on the type of MAC that is utilized. We distinguish between two cases: Aloha with transmit
probability p ∈ [0, 1], and CSMA with contention region with radius δ ≥ 0.
•
Aloha: For an Aloha MAC, the vehicles on each road will transmit with a probability
p. This leads to an independent thinning of the PPPs, so that λMAC
(x (z) , xtx ) = pλR ,
R
irrespective of z or xtx .
•
CSMA: For a CSMA MAC, a vehicle will transmit if it has the lowest random timer within
its sensing range (contention region). This means that (i) the intensity is in this case also
a function of xtx as other nodes in its contention region are forced to be silent when it is
active; (ii) the interference from the H- and V-road is not independent. The timer process and
the corresponding dependent thinning result in a Matérn hard-core process type II, which
can be approximated by a PPP with independently thinned node density. The approximation
of the hard-core process by a PPP is shown to be accurate in [37] and has been applied in
the context of heterogeneous cellular networks, for instance in [39].4 When the transmitter
at xtx is active the resulting intensity of the PPPs used to approximate the point process of
4
The extension to CSMA schemes with discrete back-off timers has been proposed in [25], which retains concurrent transmitters
due to the non-zero probability of nodes with the same timer value.
11
interferers can be expressed as
λMAC
(x (z) , xtx ) =
R
pA (x (z)) λR
0
kx (z) − xtx k > δ
.
(13)
kx (z) − xtx k ≤ δ
In (13), pA (x (z)) is the access probability of a node. The access probability (which is used
to thin the original process) is the probability that the given node has the smallest random
timer in the corresponding contention region (in this case modeled as a 2-dimensional ball
B2 (x (z) , δ) with radius δ centered at location x (z)), and can for one of the roads be
expressed as
pA (x (z)) =
=
where
Λ(B2 (x (z) , δ)) =
ˆ
1
exp(−tΛ(B2 (x (z) , δ)))dt
1 − exp(−Λ(B2 (x (z) , δ)))
,
Λ(B2 (x (z) , δ))
2δλR
(14)
0
q
2δλR + 2 δ 2 − kx (z)k2 λR′
kx (z)k > δ
(15)
(16)
kx (z)k ≤ δ
represents the average number of nodes in the contention region. Note that the average
number of nodes, and thus the access probability depends on the position z along the
road and the intensities λR and λR′ , which here represent the intensities of the unthinned
processes on the relevant road R and the other road, respectively.
B. Throughput
From a system perspective, the packet reception probability is not sufficient to characterize
the performance, since a MAC that allows few concurrent transmissions leads to high packet
reception probabilities but low throughputs. Thus, to be able to compare the impact of different
MAC protocols, we characterize the throughput for the intersection scenario, i.e., the number
of bits transmitted per unit time and bandwidth on a specific link. For the general case with a
receiver and transmitter located at xrx and xtx , respectively, we express the throughput as
T (β, xrx, xtx ) = pA (xtx )P(β, xrx, xtx ) log2 (1 + β)
(17)
where pA (xtx ) is the access probability of a transmitter located at xtx , i.e., the probability that
the transmitter obtains access to the channel to transmit a packet. For the Aloha MAC, the access
12
probability is simply pA (xtx ) = p, while for the CSMA case the access probability is given in
(15) and depends on the void probability in the 2-dimensional ball used to model the contention
region around xtx .
C. General Procedure
Given the analysis in the previous subsections, the general procedure for determining the
packet reception probability P(β, xrx, xtx ) and the throughput T (β, xrx , xtx ) is thus as follows:
•
Step 1: Determine the fading LT LSx (s) for the interfering links, as described in Section IV-A2.
•
Step 2: Determine the intensity of the interference PPP λMAC
(x (z) , xtx ) for R ∈ {H, V},
R
as described in Section IV-A3.
•
Step 3: From step 1 and step 2, determine the LT of the interference LIR (s) for R ∈ {H, V}
using (10).
•
Step 4: Determine the fading LT LS0 (s) for the useful link, as described in Section IV-A2.
•
Step 5: From step 4 and step 3, determine P(β, xrx, xtx ) using (5), either by drawing samples
from the interference, or by considering the CCDF of the fading on the useful link as a
kernel in a transformation (i.e., evaluating a function of LTs of the interference distribution).
Finally, use the obtained packet reception probability P(β, xrx, xtx ) in conjunction with the
access probability pA (xtx ) used in step 2 to determine the throughput T (β, xrx , xtx ).
Whether or not each step is tractable depends on the assumptions we make regarding the loss
function, the fading distribution, and the MAC protocol, which will be further discussed in
Section V.
V. C ASE S TUDIES
In this Section we present three case studies to show how the different models presented in the
paper can be used to model both rural and urban intersection scenarios, and how shadowing, LOS
blockage, and different MAC protocols affect the performance of the communication system. In
Case I, we present the most basic case which corresponds to the rural setting, while in Case II,
we show how an urban intersection can be modeled. Finally, in Case III, we will study the
impact of the different MAC protocols. In each case study, we will discuss the tractability of the
13
resulting expressions, validate modeling assumptions through simulations, and provide numerical
performance examples.
A. Case I - Rural intersection with Aloha
In the rural intersection scenario [1], [30], vehicles are assumed to communicate via LOS
links. Hence, path loss is described by the Euclidean distance loss function lE (·), defined in (3)
with path loss exponent α = 2, while power fading is modeled with an exponential distribution
(i.e., S ∼ E [1, 1]), for both useful and interfering links. Furthermore, we consider an Aloha
MAC with transmit probability p.
1) Packet reception probability : Using the procedure from Section IV-C, the packet reception
probability for the rural intersection scenario is given in Proposition 4 (see also [1], [30]).
Proposition 4. Given a slotted Aloha MAC with transmit probability p, exponential fading (i.e,
S ∼ E(1, 1)) for each link, Euclidean loss function lE (·) with path loss exponent α = 2, and a
scenario as outlined in Section II, the packet reception probability can be expressed as
!
Nβ kxrx − xtx k22
P(β, xrx, xtx ) = exp −
PA
2
p
pλV πβ kxrx − xtx k2
× exp −pλH π β kxrx − xtx k2 exp − q
β kxrx − xtx k22 + d2
(18)
Proof: See Appendix A.
We note that the packet reception probability comprises three factors: the first factor corresponds to the packet reception probability in the absence of interferers; the second factor captures
the reduction of the packet reception probability due to interferers on the H-road; the third factor
captures the additional reduction of packet reception probability due to interferers on the V-road.
Remark 5. Proposition 4 can be extended in a number of ways:
•
As was noted in [1] additional roads/lanes with arbitrary orientations can be accounted for,
each road contributing with an additional factor to the packet reception probability. This
approach can for example be used to take into account interference from surrounding roads.
Furthermore, it can be used to handle cases where the width of the roads can no longer be
ignored, by splitting the road into several lanes.
14
•
Extensions to scenarios with non-homogeneous PPPs are also possible, in order to model,
e.g., clustering of vehicles due to traffic congestions. In general this requires numeric
integration to evaluate the LTs of the interference distribution, but for special cases such as
piecewise linear intensity functions, closed-form expressions can be found.
2) Numerical results: Throughout the remainder of the paper, we consider an intersection
where the intensity of vehicles on the two roads are λH = λV = 0.01 (i.e., with an average
inter-vehicle distance of 100 m). Furthermore, we assume a noise power N of −99 dBm,
an SINR threshold of β = 8 dB [17], and that A = 3 · 10−5 , approximately matching the
conditions in [32]. We set the transmit power to P = 100 mW, corresponding to 20 dBm. For the
purpose of visualization, we show the outage probability POut (β, xrx, xtx ) = 1 − P(β, xrx, xtx )
instead of the packet reception probability. Figure 2 shows the analytical outage probability
for the rural intersection scenario as a function of distance between transmitter and receiver
kxrx − xtx k2 for different distances to the intersection d ∈ {0 m, 100 m, 500 m} and different
transmit probabilities p ∈ {0, 0.005, 0.1}. We observe that the outage probability increases with
the distance between the receiver and the transmitter, and that interference has a negative impact
on the performance as the outage probability is higher for increased transmit probabilities. In
the absence of interferers (p = 0) the system achieves an outage probability of 10 % when
the receiver and transmitter are spaced approximately 600 m apart. When p is increased to
0.005 the communication range is drastically reduced to about 130 m, due to the interference.
Furthermore, the figure reflects the location dependence of the outage probability with respect
to the intersection, and we can see that the outage probability increases when the receiver is
closer to the intersection. For the purpose of validation, we have added Monte Carlo simulation
with 10,000 realizations of the PPPs and fading parameters, perfectly matching the analytical
expressions.
B. Case II - Urban Intersection with Aloha
This case study, model an urban intersection scenario, with the Tx on the V-road and the Rx
on the H-road. Signals arriving to the Rx from the V-road are assumed to be in NLOS, modeled
through Manhattan path loss and Erlang fading (which serves as an approximation of log-normal
fading). Signals arriving to the Rx from the own H-road are in LOS, modeled through Euclidean
path loss and exponential fading.
15
100
Analytical - Exponential
Simulated - Exponential
rS
bC
bC
outage probability, POut (β, xrx , xtx )
lDbCrSlDbCrS
lD
lD
rS
rS
lD
bC
lDbCrS rS
lD
lD
bC
lDbCrSlDbCrS
bC
rS
rS
lDbC
rS
lDbC
lD
lD
p = 0.1
10−1
lDbCrS
lDbCrS lDbCrS
lDbCrS
lDbCrS
lDbC
p = 0.005
rS
p=0
bC
rS
lDbC
bC
lD
rS
bC
rS
rSlD
lDbCrS
lD
lD
bC
rSbC
lD
lDbC
rS
bC
lD
lD
rS
lD
rS
rS
bC
bC bC
lDrS
bC
rS
10−2
100
101
102
103
distance between receiver and transmitter, kxrx − xtx k2 [m]
Figure 2.
Comparison of analytical and simulated outage probability POut (β, xrx , xtx ) versus distance between transmitter
and receiver kxrx − xtx k2 for different distances to the intersection d as well as different transmit probabilities p. The distances
d are 0 m (diamonds), 100 m (circles) and 500 m (squares).
1) Packet reception probability: The packet reception probability for the urban intersection
scenario is given in Proposition 6.
Proposition 6. Given a slotted Aloha MAC with transmit probability p, Erlang fading (i.e.,
S ∼ E(k0 , θ0 )) and Manhattan loss function lM (·) for the useful link, Erlang fading (i.e., S ∼
E(kV , θV )) and Manhattan loss function for the interfering links from the V-road, exponential
fading (i.e, S ∼ E(1, 1)) and Euclidean loss function lE (·) for the interfering links from the Hroad, and a scenario as outlined in Section II, the packet reception probability can be expressed
as
− ζN
P
P(β, xrx, xtx ) = e
kX
i
0 −1 X
i=0 j=0
i ζ i (j) (i,j)
C D ,
j i!
(19)
where
C
(j)
=
j
X
N
j
n=0
n
P
j−n
n
X
√
n −κ ζ −n
(−1) e
ζ
l=0
√ l 2−m+l−2n
l
X
(−1)m −κ ζ
2
n
,
m!
(−m
+
l)!
m=0
(20)
16
and
D (i,j) =
(21)
kX
V −1
−q
Aζ
1
1
di−j
kV
exp
−2pλ
+q
Γ
(−1)
V
d(ζ)i−j
αΓ
[k
]
θ
α
q
V
V
q=0
− α1 +q
!!
α
Aζ
1
1
1
d
×−
Γ − + kV − q +d1+αq Γ[kV ] 2 F1 kV , +q, 1+ + q, −
.
θV
α
α
α
AζθV
i−j
In which κ = 2pλH A1/α π/α csc (π/α) and ζ = β kxrx − xtx kα1 / (Aθ0 ) .
Proof: See Appendix B.
We observe that the analytical expressions become more involved when changing the loss
function as well as the fading distribution for the links to the V-road, but in contrast to the
rural intersection scenario it is possible to obtain closed form expressions for a general α (this
is because Manhattan path loss for the interferers from the V-road is easier to handle than
Euclidean path loss). Furthermore, it should be noted that if the Tx is assumed to be on the
H-road, the expressions become more compact. Moreover, similarly as for the model presented
in [34], Proposition 6 only gives realistic results when the Rx and the Tx are at least a few
meters away from the intersection. This is because when the Rx is at the intersection, all links
become LOS, while when the Tx is at the intersection, the useful link becomes LOS. In either
case, the corresponding links should be modeled with exponential fading, rather than Erlang
fading.
2) Numerical results: In this section we intend to validate the accuracy of the Erlang approximation. We consider the same parameters for the LOS propagation as in Section V-A. We set the
Aloha transmit probability to p ∈ {0.002, 0.02}. For all NLOS links, we use the same value of
A as in the LOS links, set α = 2 , and consider the fading to be log-normal with 3.2 dB standard
deviation, as in one of the intersections studied in [33]. Maximum likelihood fitting of the Erlang
distribution to the log-normal distribution yielded k0 = kV = k = 2 and θ0 = θV = θ ≈ 0.66. In
Figure 3, a comparison between the outage probability obtained by evaluating Proposition 6 under
the Erlang approximation, and Monte Carlo simulations for the same scenario but with 3.2 dB
log-normal shadowing is shown. First, we observe that the analytical results based on the Erlang
approximation agrees well with the simulations, i.e., in terms of outage probability the Erlang
fading provides a good approximation to the actual log-normal fading. Furthermore, we see that
17
100
lDbCrS
lD
outage probability, POut (β, xrx , xtx )
bC
ytx = 150m
lD
lD
lD
bC
rS
bC
ytx = 100m
bC
rS
lD
p = 0.02
bC
rS
ytx = 50m
rS
rS
10−1
lD
bC
lD
bC
ytx = 100m
bC
rS
lD
ytx = 150m
lD
bC
p = 0.002
rS
ytx = 50m
Analytical-Erlang
Simulated-LogN
rS
rS
10−2
100
101
102
103
distance between receiver and intersection, d [m]
Figure 3. Comparison between analytical and simulated outage probability POut (β, xrx , xtx ) versus distance between receiver
and intersection d for different Tx positions and transmit probabilities. Lines show analytical results based on Proposition 6,
while markers show simulations for the same scenario but with 3.2 dB log-normal fading. The different transmit probabilities p,
which the outage probability is plotted for are 0.002 (solid) and 0.02 (dashed) lines.
as expected, lower transmit probability, as well as placing the Tx closer to the intersection results
in lower outage probabilities. Even though the results shows very good agreement between the
analytical results and the simulations, it should be mentioned that the approximation become less
accurate when the standard deviation increases. In particular, this is noticeable when the standard
deviation of the log-normal distribution exceeds 3.8 dB, as the Erlang distribution obtained from
the fitting then reverts to the exponential distribution. Furthermore, when decreasing the standard
deviation of the log-normal distribution, the estimated value of k rapidly increases, so that the
model becomes less tractable.
C. Case III - Aloha vs CSMA
In this final case study, we will focus on the MAC protocol and how it affects performance
and tractability. To do this, we start from the rural intersection scenario, but replace the Aloha
MAC with a CSMA MAC. As the MAC affects not only the packet reception probability but
also the access probability, we will also consider throughput in this case study.
18
1) Packet reception probability: The packet reception probability for the CSMA case is given
in Proposition 7.
Proposition 7. Given a CSMA MAC with contention radius δ, exponential fading (i.e, S ∼
E(1, 1)) for each link, Euclidean loss function lE (·) with path loss exponent α = 2, and a
scenario as outlined in Section II, the success probability can be expressed as
P(β, xrx, xtx ) = e−
N β̃
P
LIH (β̃)LIV (β̃),
(22)
where β̃ = β/lE (xtx , xrx ), and
LIH (s) = exp −
LIV (s) = exp −
ˆ
ˆ
+∞
T
λMAC
[x,
0]
,
x
tx
H
2
−∞
+∞
1 + |xrx − x| /As
λMAC
V
T
[0, y] , xtx
dx
dy
2
[xrx , −y]T 2 /As
(23)
(24)
1+
T
T
MAC
where λMAC
[x,
0]
,
x
and
λ
[0,
y]
,
x
are given in (60) and (61), respectively.
tx
tx
H
V
−∞
Proof: See Appendix C.
As can be seen from Proposition 7, the expressions we obtain still involve an integral that
can be solved numerically easily and efficiently.
2) Numerical results: In order to evaluate the accuracy of the approximation introduced
in Section IV-A3, we start by comparing the analytically calculated outage probability to a
simulation with 50000 realizations of the fading parameters and the hard-core process induced
by the dependent thinning resulting from the CSMA scheme. This comparison can be seen in
Figure 4, which shows the analytical and simulated outage probability as a function of the distance between the receiver and the intersection for two different transmitter locations (xtx = [0, 0]
and xtx = [0, 150]), as well as two different CSMA contention radiuses δ ∈ {500 m, 10000 m}.
We observe good correspondence between simulation and analytical results, and an increase in
outage probability with increased distance to the intersection. We also note that when xtx = [0, 0],
it is possible to compare Figure 4 with Figure 2. We note that for δ = 10000 m, for a distance
of 100 m between Rx and intersection, CSMA has an outage probability of 0.003, while Aloha
is over 25 times worse, with an outage probability of 0.08.
19
lD
lD
lD
rS
lD
lD
lD
lD
rSlD
rSlD
100
outage probability, POut (β, xrx , xtx )
δ = 500 m
lDrS
rS
10−1
lD
rS
rS
lD
10
−2
lD
rSlD
lD
lD
lD
rS
lD
δ = 10000 m
xtx = [0, 0]
xtx = [0, 150]
rS
rS
10−3
100
101
102
103
distance between receiver and intersection, d [m]
Figure 4.
Comparison of analytical (blue lines) and simulated (red markers) outage probability POut (β, xrx , xtx ) versus
distance between receiver and intersection d, for different transmitter locations xtx as well as CSMA contention radiuses δ. The
receiver is located on the H-road, while the transmitter location is fixed to either xtx = [0, 0] (solid lines) or xtx = [0, 150]
(dashed lines). The different CSMA contention radiuses are δ = 500 m and δ = 10000 m, which in the region where the access
probability is constant, i.e., far away from the intersection, corresponds to pA = 0.1 and pA = 0.005, respectively.
To further study the performance gains achieved by using CSMA compared to Aloha, we
now look at both outage probability and throughput for a specific receiver and transmitter
configuration. The configuration that we consider is xrx = [0 0]T and xtx = [Rcomm 0]T . Note that
for the Aloha case this placement results in the worst possible throughput for a fixed lE (xtx , xrx ).
Figure 5 and Figure 6, show the outage probability as well as throughput as a function of the
access probability pA (xtx ), for two different values on Rcomm ∈ {100 m, 200 m}.
For Aloha (Figure 5), we see that with an increase in pA (xtx ), outage probability increases
due to the presence of more interferers. The throughput first increases (due to more active
transmitters) and then decreases (due to overwhelming amounts of interference), leading to an
optimal value of pA (xtx ). However, in order to guarantee a certain quality of service, one must
also consider a guarantee on the outage probability. For instance, if we want to guarantee an
outage probability of less than 10 % on the link when Rcomm = 100 m, the optimal value of
pA (xtx ) ≈ 0.006, leading to a throughput of around 0.0055 bits per unit time and bandwidth.
For CSMA (Figure 6), a low access probability (i.e., large contention region) reduces the outage
100
10−1
10−1
10−2
bC
throughput, T (β, xrx , xtx )
outage probability, POut (β, xrx , xtx )
20
bC
outage prob.
throughput
10−2
0
0.05
10−3
0.1
access probability, pA (xtx )
Figure 5. Outage probability POut (β, xrx , xtx ) and throughput T (β, xrx , xtx ) for the Aloha case as a function of the transmitter
access probability pA (xtx ). The receiver is located at xrx = [0, 0], and solid lines correspond to Rcomm = 100 m, while dashed
lines correspond to Rcomm = 200 m. The red circles indicate the maximum throughput that is possible to achieve while
guaranteeing that the outage probability is kept below the target value of 10 %.
probability. Similar to Aloha, the throughput first increases with increased access probability and
then decreases. To achieve an outage probability below 10 % when Rcomm = 100 m, the optimal
value of pA (xtx ) ≈ 0.023 (corresponding to a contention radius δ of about 1100 m), results in
a throughput of about 0.059 bits per unit time and bandwidth. Hence, in this scenario, using
CSMA instead of Aloha leads to more than a tenfold increase in the throughput for the same
communication range.
VI. C ONCLUSIONS
We have provided an overview of the dominant propagation properties of vehicular communication systems near intersections, for both rural and urban scenarios. Based on these properties,
we proposed a general procedure to analytically determine packet reception probabilities of individual transmissions as well throughput, mainly applicable to 802.11p communication. In contrast
to traditional cellular networks, the one-dimensional road geometry leads to non-homogeneous
packet reception probabilities and throughputs.
We have applied and validated this procedure to three case studies, relevant for vehicular
applications. The results indicate that the procedure is sufficiently general and flexible to deal
21
100
10−1
bC
10−2
10−1
throughput, T (β, xrx , xtx )
outage probability, POut (β, xrx , xtx )
bC
outage prob.
throughput
10−2
0
0.05
10−3
0.1
access probability, pA (xtx )
Figure 6.
Outage probability POut (β, xrx , xtx ) and throughput T (β, xrx , xtx ) for the CSMA case as a function of the
transmitter access probability pA (xtx ). The receiver is located at xrx = [−100, 0], and solid lines correspond to Rcomm = 100 m,
while dashed lines correspond to Rcomm = 200 m. The red circles indicate the maximum throughput that is possible to achieve
while guaranteeing that the outage probability is kept below the target outage probability of 10 %.
with a variety of scenarios, that its performance results match well with simulations, and that
it is able to capture the clustering of vehicles close to the intersection. The procedure can thus
serve as a useful design tool for communication system engineers, complementing simulations
and experiments.
Possible avenues for future research includes validation of the model agains actual measurements, adoption of advanced MAC schemes as well as 5G D2D features.
A PPENDIX A
P ROOF
OF
P ROPOSITION 4
In order to determine the packet reception probability when S ∼ E(1, 1), we follow the
general procedure from Section IV-C.
Step 1: The fading LT for the interfering links can be expressed as LSx (s) = 1/(1 + s).
Step 2: According to Section IV-A3 the intensity of the two interfering PPPs ΦMAC
and ΦMAC
H
V
MAC
T
MAC
T
are λH
[x 0] , xtx = pλH and λV
[0, y] , xtx = pλV , respectively.
Step 3: The LT of the interference for the two roads are derived as follows. Since the fading
22
of the interfering links is exponentially distributed, (10) simplifies to
ˆ +∞
λMAC
(x (z) , xtx )
R
dz .
LIR (s) = exp −
−∞ 1 + 1/(sθl(x (z) , xrx ))
(25)
Using (25) for the horizontal road with Euclidean path loss, and bearing in mind that x(z) =
[z 0]T , we can write
+∞
λMAC
(x(z), xtx )
H
dz
α
−∞ 1 + kxrx − x(z)k2 /As
ˆ +∞
1
(a)
= exp −pλH
dx
α
−∞ 1 + |xrx − x| /As
ˆ +∞
1
(b)
1/α
du
= exp −2pλH (As)
1 + uα
0
= exp −2pλH (As)1/α π/α csc (π/α)
ˆ
LIH (s) = exp −
(26)
(27)
(28)
(29)
where (a) uses the fact that the intensity is pλH on the H-road, and (b) involves a change of
variable u = |xrx − x| /(As)1/α . For the particular case of α = 2 the LT of the interference
further simplifies to
√
LIH (s) = exp −pλH π As .
(30)
For the V-road, using (25), we can in a similar way as for the H-road write
ˆ +∞
1
LIV (s) = exp −pλV
dy .
α
T
−∞ 1 + [xrx , −y]
/As
(31)
2
T
p
p
x2rx + y 2 = d2 + y 2 we can introduce ry =
=
Now using that the distance [xrx , −y]
2
p
d2 + y 2, with dry /dy = y/ry . Noting that a PPP remains PPP under a non-linear transformation
according to the mapping theorem [18, Theorem A.1], we have
!
ˆ +∞
ry
dry
p 2
(32)
LIV (s) = exp −2pλV
ry − d2 1 + ryα /As
d
ˆ +∞
1
1/α
√
= exp −pλV (As)
dω
(33)
ω − ω0 (1 + ω α/2 )
ω0
2
where we have carried out the following change of variable ω = ry / (As)1/α , and further in2
´ +∞ √
troduced ω0 = d/ (As)1/α . For α = 2, the integral can be computed as ω0 ( ω − ω0 (1 + ω))−1 dω=
√
π/ 1 + ω0 , which yields
−pλV πAs
(34)
LIV (s) = exp √
As + d2
23
Note that for d → 0, (34) reverts to (30), while for d → +∞, (34) tends to one.
Step 4: The fading on the useful link is characterized by its LT LS0 (s) = 1/(1 + s) and
CCDF F̄S0 (s) = exp (−s) .
Step 5: Using the LT of the interference from Step 3, and the CCDF of the fading from
Step 4 as a kernel, we can now determine P(β, xrx, xtx ) through (5). First using the CCDF, and
evaluating it in the desired point, we can write
F̄S0 t1 + t2 + Ñ β̃ = exp − t1 + t2 + Ñ β̃
(35)
As the interference from the H- and V-road is independent (i.e., ΦMAC
and ΦMAC
are independent)
H
V
we can now use (35) to express the transform in (5) as
P(β, xrx, xtx )
ˆ +∞
ˆ +∞
= exp −Ñ β̃
fIH (t1 ) exp −t1 β̃ dt1
fIV (t2 ) exp −t2 β̃ dt2
0
0
= e−Ñ β̃ LIH β̃ LIV β̃
(36)
(37)
Using the results from step 3, and the variable changes β̃ = β/lE (xtx , xrx ) and Ñ = N/P ,
finally allow us to express the packet reception probability as (18).
A PPENDIX B
P ROOF
OF
P ROPOSITION 6
We use the procedure from Section IV-C.
Step 1: The fading LTs for the interfering links from the H-road and the V-road can be
expressed as LSx (s) = 1/(1 + s) and LSx (s) = 1/(1 + sθV )kV , respectively.
Step 2: According to Section IV-A3 the intensity of the two PPPs ΦMAC
and ΦMAC
are pλH
H
V
and pλV , respectively.
Step 3: The LT of the interference for the two roads are derived in the following way. For the
, the fading LT as well as the loss function are the same as
H-road, with interferers x ∈ΦMAC
H
in the rural intersection case. Thus we can according to (29) expresse the LT of the interference
for a general α as
LIH (s) = exp −2pλH (As)1/α π/α csc (π/α) .
(38)
For the V-road we now have fading LT LSx (s) = 1/(1 + sθV )kV , intensity pλV , and Manhattan
loss function. Hence, using (10) we can write
24
LIV (s) = exp −
ˆ∞
−∞
= exp −pλV
(a)
= exp −pλV
λMAC
(x(z), xtx ) (1 − LSx (s lM (x(z), xrx ))) dz
V
ˆ∞
−∞
1 −
kX
V −1
q=0
kX
V −1
kV
q
1
(1 + sθV A [xrx , −y]T
ˆ∞
−∞
[xrx , −y]
T
( [xrx , −y]T
ˆ∞
αq
1
α
1
−α
1
)k V
kV −q
b
dy
dy
+ b)kV
(d + |y|)αq bkV −q
dy
α + b)kV
((d
+
|y|)
q=0
−∞
ˆ∞ αq kV −q
kX
V −1
u b
kV
(c)
= exp −pλV
du
α + b)kV
(u
q
q=0
(b)
= exp −pλV
kV
q
(39)
(40)
(41)
(42)
(43)
−∞
where (a) uses the Binomial Theorem and the variable change sθV A → b, (b) uses that for
points x ∈ ΦMAC
the distance kxrx − xk1 = |xrx | + |y| = d + |y|, and (c) uses the variable
V
change d + |y| → u. For q ≥ 0 , kV ≥ q + 1, b ≥ 0 and d > 0 the integral can be evaluated in
closed form, and for a general α we can express the LT of the interference as
LIV (s) =
(44)
kX
V −1
−q
As
1
1
kV
+q
Γ
exp −2pλV
αΓ
[k
]
θ
α
q
V
V
q=0
− α1 +q
!!
α
As
1
1
d
1
× −
,
Γ − + kV − q + d1+αq Γ [kV ] 2 F1 kV , + q, 1 + + q, −
θV
α
α
α
AsθV
where 2 F1 is the regularized hypergeometric function. Note that for α = 2 and kV = θV = 1
(i.e., exponential fading) this simplifies to
√
d
LIV (s) = exp −pλV As π − 2arctan √
,
As
and when d → 0 we get
√
LIV (s) = exp −pλV π As ,
(45)
(46)
25
i.e., it reverts to the same form as LIH (s) in (30).
Step 4: The fading on the useful link is characterized by its LT LS0 (s) = 1/(1 + sθ0 )k0 and
CCDF
−s/θ0
F̄S0 (s) = e
kX
0 −1
i=0
1 i
s
i!θ0i
(47)
Step 5: In the same manner as in Appendix A, we now use the LTs of the interference from
Step 3, and the CCDF of the fading from Step 4 to determine P(β, xrx, xtx ) through (5). First
using the CCDF, and evaluating it in the desired point, we can write
F̄S0 t1 + t2 + Ñ β̃
= e−β̃ (t1 +t2 +Ñ )/θ0
kX
0 −1
i=0
(a)
= e−ζ (t1 +t2 +Ñ )
kX
0 −1
i=0
(b)
−ζ Ñ
=e
kX
i
0 −1X
i
1 i
β̃
t
+t
+
Ñ
1
2
i!θ0i
i
ζi
t1 +t2 + Ñ
i!
i ζ i −ζt1
e (Ñ +t1 )j e−ζt2 ti−j
2 ,
j i!
i=0 j=0
(48)
(49)
(50)
(51)
where (a) involves the variable change ζ = β̃/θ0 and (b) uses the Binomial Theorem. Due to
the independence of the interference we can now use (51) to express the transform in (5) as
kX
i i
0 −1 X
i ζ (j) (i,j)
− ζN
C D ,
(52)
P(β, xrx, xtx ) = e P
j i!
i=0 j=0
where
C
(j)
=
ˆ
+∞
e−ζt1 (Ñ + t1 )j fIH (t1 )dt1
(53)
0
j
and
X j
Ñ j−n L[tn1 fIH (t1 )](ζ)
=
n
n=0
j
X
N j−n
j
dn
=
(−1)n
LI (ζ)
n
P
d(ζ)n H
n=0
D
(i,j)
=
ˆ
0
+∞
e−ζt2 ti−j
2 fIV (t2 )dt2
(54)
(55)
(56)
= L[ti−j
2 fIV (t2 )](ζ)
(57)
= (−1)i−j
(58)
di−j
LI (ζ)
d(ζ)i−j V
26
are obtained using the Laplace transform property tn f (t) ←→ (−1)n
dn
L [f
dζ n
(t)] (ζ). Note that
(52) and (55) use the variable change Ñ = N/P . Now using the results from step 4 express the
nth derivative of the LT of the interference from the H-road as
√ l 2−m+l−2n
n X
l
X
√
(−1)m −κ ζ
dn
2
−κ ζ −n
n
LI (ζ) = e
ζ
d(ζ)n H
m!
(−m
+
l)!
m=0
(59)
l=0
where (·)n is the Pochhammer symbol and κ = 2pλH (A)1/α π/α csc (π/α). For the V-road, there
is no general compact expression for the nth derivative of LIV (ζ), but an explicit expression can
in principle be calculated for any n, kV and θV .
A PPENDIX C
P ROOF
OF
P ROPOSITION 7
We use the procedure from Section IV-C.
Step 1: The fading LT for the interfering links can be expressed as LSx (s) = 1/(1 + s).
Step 2: According to Section IV-A3, the intensity of the two PPPs ΦMAC
and ΦMAC
are for
H
V
this case also a function of the transmitter location xtx . Using (13) we can express the intensity
for the H-road as
1−exp(−2δλH )
x ∈ R1
2δ
√
δ2 −x2 λV )λH
(60)
[x, 0]T , xtx = 1−exp(−2δλH −2
λMAC
√
x ∈ R2
H
2
2δλH +2 δ −x2 λV
0
else
p
p
2
2
in which R1 = {x| |x| > δ and (x − xtx )2 + ytx
> δ} and R2 = {x| |x| ≤ δ and (x − xtx )2 + ytx
>
δ}. Similarly for the V-road,
1−exp(−2δλV )
y ∈ R3
2δ
1−exp−2δλ −2√δ2 −y2 λ λ
V
H
V
√
(61)
λMAC
[0, y]T , xtx =
y ∈ R4
V
2
2
2δλV +2 δ −y λH
0
else
p
p
in which R3 = {y| |y| > δ and (y − ytx )2 + x2tx > δ} and R4 = {y| |y| ≤ δ and (y − ytx )2 + x2tx >
δ}.
Step 3: Using (25) the LT of the interference for the H- and V- road can be expresed as
ˆ +∞ λMAC [x, 0]T , xtx
H
dx
(62)
LIH (s) = exp −
1 + |xrx − x|α /As
−∞
27
and
ˆ
LIV (s) = exp −
+∞
−∞
T
λMAC
[0,
y]
,
x
tx
V
T
1 + [xrx , −y]
α
2
/As
dy
(63)
Step 4: The fading fading on the useful link is characterized by its LT LS0 (s) = 1/(1 + s)
and CCDF F̄S0 (s) = exp (−s).
Step 5: By applying a location dependent thinning, we approximate the interference from
the H- and V-road as independent. Hence, as the fading on the useful link is exponential (i.e.,
S0 ∼ E(1, 1)), we can in the same way as in Appendix A, express the packet reception probability
as P(β, xrx , xtx ) = e−Ñ β̃ LIH β̃ LIV β̃ . Using the results from Step 3, as well as the variable
change Ñ = N/P , we can for the particular value of α = 2 finally obtain (22). Note that for
a general transmitter location xtx , we are not able to evaluate the integrals in (62) and (63) in
closed form, but have to resort to numerical evaluation.
R EFERENCES
[1] E. Steinmetz, M. Wildemeersch, T. Q.S. Quek, and H. Wymeersch, “A Stochastic Geometry Model for Vehicular
Communication near Intersections,” in IEEE Global Communications Conference (GLOBECOM), Dec. 2015.
[2] G. Karagiannis, O. Altintas, E. Ekici, G. Heijenk, B. Jarupan, K. Lin, and T. Weil, “Vehicular Networking: A Survey
and Tutorial on Requirements, Architectures, Challenges, Standards and Solutions,” IEEE Communications Surveys &
Tutorials, vol. 13, no. 4, pp. 584–616, 2011.
[3] P. Papadimitratos, A. La Fortelle, K. Evenssen, R. Brignolo, and S. Cosenza, “Vehicular Communication Systems: Enabling
Technologies, Applications, and Future Outlook on Intelligent Transportation,” IEEE Communications Magazine, vol. 47,
no. 11, pp. 84–95, Nov. 2009.
[4] H. Hartenstein and K. P. Laberteaux, “A tutorial survey on vehicular ad hoc networks,” IEEE Communications Magazine,
vol. 46, no. 6, pp. 164–171, Jun. 2008.
[5] K. Dar, M. Bakhouya, J. Gaber, M. Wack, and P. Lorenz, “Wireless Communication Technologies for ITS Applications
[Topics in Automotive Networking],” IEEE Communications Magazine, vol. 48, no. 5, pp. 156–162, May 2010.
[6] F. Anjum, S. Choi, V. Gligor, R. Herrtwich, J.-P. Hubaux, P. Kumar, R. Shorey, and C.-T. Lea, “Guest Editorial Vehicular
Networks,” IEEE Journal on Selected Areas in Communications, vol. 25, no. 8, pp. 1497–1500, Oct. 2007.
[7] M. Alsabaan, W. Alasmary, A. Albasir, and K. Naik, “Vehicular Networks for a Greener Environment: A Survey,” IEEE
Communications Surveys & Tutorials, vol. 15, no. 3, pp. 1372–1388, Jul. 2013.
[8] L. Cheng, B. Henty, D. Stancil, F. Bai, and P. Mudalige, “Mobile Vehicle-to-Vehicle Narrow-Band Channel Measurement
and Characterization of the 5.9 GHz Dedicated Short Range Communication (DSRC) Frequency Band,” IEEE Journal on
Selected Areas in Communications, vol. 25, no. 8, pp. 1501–1516, Oct. 2007.
[9] J. F. Monserrat, H. Droste, O. Bulakci, J. Eichinger, O. Queseth, M. Stamatelatos, H. Tullberg, V. Venkatkumar,
G. Zimmermann, U. Dotsch, and A. Osseiran, “Rethinking the mobile and wireless network architecture: The METIS
research into 5G,” in European Conference on Networks and Communications (EuCNC), Jun. 2014.
28
[10] S. Mumtaz, K. M. Saidul Huq, and J. Rodriguez, “Direct mobile-to-mobile communication: Paradigm for 5G,” IEEE
Wireless Communications, vol. 21, no. 5, pp. 14–23, Oct. 2014.
[11] O. Yilmaz, Z. Li, K. Valkealahti, M. Uusitalo, M. Moisio, L. Petteri, and W. Carl, “Smart Mobility Management for D2D
Communications in 5G Networks,” in IEEE Wireless Communications and Networking Conference Workshops (WCNCW),
Apr. 2014, pp. 219–223.
[12] A. Khelil and D. Soldani, “On the suitability of Device-to-Device communications for road traffic safety,” in IEEE World
Forum on Internet of Things (WF-IoT), Mar. 2014, pp. 224–229.
[13] G. Araniti, C. Campolo, M. Condoluci, A. Iera, and A. Molinaro, “LTE for vehicular networking: A survey,” IEEE
Communications Magazine, vol. 51, no. 5, pp. 148–157, May 2013.
[14] R. Johri, J. Rao, H. Yu, and H. Zhang, “A multi-scale spatiotemporal perspective of connected and automated vehicles:
Applications and wireless networking,” IEEE Intelligent Transportation Systems Magazine, vol. 8, no. 2, pp. 65–73, Summer
2016.
[15] J. Santa, R. Toledo-Moreo, M. A. Zamora-Izquierdo, B. Úbeda, and A. F. Gómez-Skarmeta, “An analysis of communication
and navigation issues in collision avoidance support systems,” Transportation Research Part C: Emerging Technologies,
vol. 18, no. 3, pp. 351–366, Jun. 2010.
[16] C. F. Mecklenbrauker, A. F. Molisch, J. Karedal, F. Tufvesson, A. Paier, L. Bernado, T. Zemen, O. Klemp, and N. Czink,
“Vehicular Channel Characterization and Its Implications for Wireless System Design and Performance,” Proceedings of
the IEEE, vol. 99, no. 7, pp. 1189–1212, Jul. 2011.
[17] K. Sjöberg, “Medium Access Control for Vehicular Ad Hoc Networks,” Ph.D. dissertation, Chalmers University of
Technology, 2013.
[18] M. Haenggi and R. K. Ganti, “Interference in Large Wireless Networks,” Foundations and Trends in Networking, vol. 3,
no. 2, pp. 127–248, 2008.
[19] R. K. Ganti and M. Haenggi, “Interference and Outage in Clustered Wireless Ad Hoc Networks,” IEEE Transactions on
Information Theory, vol. 55, no. 9, pp. 4067–4086, Sep. 2009.
[20] N. Deng, W. Zhou, and M. Haenggi, “The Ginibre Point Process as a Model for Wireless Networks With Repulsion,”
IEEE Transactions on Wireless Communications, vol. 14, no. 1, pp. 107–121, Jan. 2015.
[21] A. Hunter, J. Andrews, and S. Weber, “Transmission capacity of ad hoc networks with spatial diversity,” IEEE Transactions
on Wireless Communications, vol. 7, no. 12, pp. 5058–5071, Dec. 2008.
[22] R. W. Heath, M. Kountouris, and T. Bai, “Modeling Heterogeneous Network Interference Using Poisson Point Processes,”
IEEE Transactions on Signal Processing, vol. 61, no. 16, pp. 4114–4126, Aug. 2013.
[23] B. Błaszczyszyn and H. Keeler, “Equivalence and comparison of heterogeneous cellular networks,” in IEEE 24th
International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC Workshops), Sep. 2013, pp.
153–157.
[24] Y. Jeong, J. W. Chong, H. Shin, and M. Z. Win, “Intervehicle Communication: Cox-Fox Modeling,” IEEE Journal on
Selected Areas in Communications, vol. 31, no. 9, pp. 418–433, Sep. 2013.
[25] Z. Tong, H. Lu, M. Haenggi, and C. Poellabauer, “A Stochastic Geometry Approach to the Modeling of DSRC for Vehicular
Safety Communication,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 5, pp. 1448–1458, May
2016.
[26] B. Błaszczyszyn, P. Mühlethaler, and Y. Toor, “Performance of MAC protocols in linear VANETs under different attenuation
and fading conditions,” in IEEE Conference on Intelligent Transportation Systems, Oct. 2009, pp. 1–6.
29
[27] ——, “Stochastic analysis of Aloha in vehicular ad hoc networks,” Annales des Télécommunications, vol. 68, no. 1, pp.
95–106, Jun. 2013.
[28] B. Błaszczyszyn, P. Mühlethaler, and N. Achir, “Vehicular Ad-hoc Networks using slotted Aloha: Point-to-Point, Emergency
and Broadcast Communications,” in IFIP Wireless Days, Nov. 2012, pp. 1–6.
[29] T. V. Nguyen and F. Baccelli, “A Stochastic Geometry Model for Cognitive Radio Networks,” The Computer Journal,
vol. 55, no. 5, pp. 534–552, Jul. 2011.
[30] E. Steinmetz, R. Hult, G. R. de Campos, M. Wildemeersch, P. Falcone, and H. Wymeersch, “Communication analysis for
centralized intersection crossing coordination,” in International Symposium on Wireless Communications Systems (ISWCS),
Aug. 2014, pp. 813–818.
[31] T. Abbas, K. Sjöberg, J. Karedal, and F. Tufvesson, “A measurement based shadow fading model for vehicle-to-vehicle
network simulations,” International Journal of Antennas and Propagation, vol. 2015, 2015, Article ID 190607, 12 pages,
2015.
[32] J. Karedal, N. Czink, A. Paier, F. Tufvesson, and A. F. Molisch, “Path Loss Modeling for Vehicle-to-Vehicle Communications,” IEEE Transactions on Vehicular Technology, vol. 60, no. 1, pp. 323–328, Jan. 2011.
[33] T. Abbas, A. Thiel, T. Zemen, C. F. Mecklenbrauker, and F. Tufvesson, “Validation of a non-line-of-sight path-loss model
for V2V communications at street intersections,” in International Conference on ITS Telecommunications (ITST), Nov.
2013, pp. 198–203.
[34] T. Mangel, O. Klemp, and H. Hartenstein, “5.9 GHz inter-vehicle communication at intersections: a validated non-line-ofsight path-loss and fading model,” EURASIP Journal on Wireless Communications and Networking, vol. 2011, no. 1, pp.
1–11, Nov. 2011.
[35] FP6 IST Project WINNER II, Deliverable 1.1.2 V1.2, “WINNER II Channel Models,” Tech. Rep., Sep. 2007. [Online].
Available: http://www.cept.org/files/1050/documents/winner2%20-%20final%20report.pdf
[36] M. Wildemeersch, C. H. Slump, and A. Rabbachin, “Acquisition of GNSS signals in urban interference environment,”
IEEE Transaction on Aerospace and Electronic Systems, vol. 50, no. 2, pp. 1078–1091, April 2014.
[37] M. Haenggi, “Mean interference in hard-core wireless networks,” IEEE Communications Letters, vol. 15, no. 8, pp. 792–
794, Aug. 2011.
[38] C. Abou-Rjeily and M. Bkassiny, “On the achievable diversity orders over non-severely faded lognormal channels,” IEEE
Communications Letters, vol. 14, no. 8, pp. 695–697, Aug. 2010.
[39] S. R. Cho and W. Choi, “Energy-efficient repulsive cell activation for heterogeneous cellular networks,” IEEE Journal on
Selected Areas in Communications, vol. 31, no. 5, pp. 870–882, May 2013.
| 3cs.SY
|
Hyper-reduction over nonlinear manifolds for large nonlinear
mechanical systems
Shobhit Jain, Paolo Tiso∗
arXiv:1710.05160v1 [cs.CE] 14 Oct 2017
October 17, 2017
Institute for Mechanical Systems, ETH Zürich,
Leonhardstrasse 21, 8092 Zürich, Switzerland
Abstract
Common trends in model order reduction of large nonlinear finite-element-discretized systems involve the introduction of a linear mapping into a reduced set of unknowns, followed by
Galerkin projection of the governing equations onto a constant reduction basis. Though this
reduces the number of unknowns in the system, the computational cost for obtaining the solution could still be high due to the prohibitive computational costs involved in the evaluation
of nonlinear terms. Hyper-reduction methods are then seen as a fast way of approximating the
nonlinearity in the system of equations. In the finite element context, the energy conserving sampling and weighing (ECSW) method has emerged as a stability and structure-preserving method
for hyper-reduction. Classical hyper-reduction techniques, however, are applicable only in the
context of linear mappings into the reduction subspace. In this work, we extend the concept of
hyper-reduction using ECSW to general nonlinear mappings, while retaining its desirable stability and structure-preserving properties. As a proof of concept, the proposed hyper-reduction
technique is demonstrated over models of a flat plate and a realistic wing structure, whose dynamics has been shown to evolve over a nonlinear (quadratic) manifold. An online speed-up of
over one thousand times relative to the full system has been obtained for the wing structure
using the proposed method, which is higher than its linear counterpart using the ECSW.
Keywords: EECSW; nonlinear manifolds; ECSW; hyper-reduction; model order reduction;
structure-preserving; nonlinear dynamics; Galerkin projection
1
Introduction
Model reduction and hyper-reduction
In spite of the ever increasing computational power and digital storage capabilities, reduced-order
models (ROM) of Finite Element (FE) discretized systems are a must when complex simulations
are needed to support design and optimization activities. Current state-of-the-art ROM techniques
are based on a reduced basis approach, i.e., the solution is sought on a linear, low-dimensional
subspace spanned by a few, carefully selected modes. The resulting residual is then projected onto
the same or a different subspace of the same dimension to obtain a well-posed reduced system. When
∗ Corresponding
author. Email: [email protected]
1
the original system of equations–often referred to as high-fidelity model (HFM)–features symmetric
operators, the residual is projected on the same basis used for the approximation of the solution. This
procedure is known as the Bubnov-Galerkin projection. The dimensionality reduction is guaranteed
by the limited size of the reduction variables as compared to the number of degrees of freedom in
the HFM.
When applied to nonlinear systems, projection-based ROMs, while accurate, are not able to
deliver significant speedups. This is due to the fact that the computation of the reduced nonlinear
terms at each solver iteration scales with the size of the HFM rather than with the one of the
reduction subspace. In the case of polynomial nonlinearities, one might consider pre-computing the
reduced nonlinear term [18, 17, 16], leading to higher-order tensors. Typically, FE models may also
contain rotational DOFs, where the associated nonlinearity cannot be pre-computed and stored in
the form of tensors. Furthermore, for high polynomial degrees, the computations with higher-order
tensors can become intractable, even more so than the HFM simulations [16]. For such cases, in the
most general setting, the above-mentioned scalability issue is successfully alleviated by a range of
hyper-reduction techniques, which aim at significantly reducing the cost of reduced nonlinear term
evaluation by computing them only at few, selected elements or nodes of the FE models, and cheaply
approximate the missing information.
The Gappy POD method [19], first introduced for image reconstruction and later applied in
dynamical systems [21, 22], was one of the first instances of hyper-reduction. However, as observed
by Farhat et al. [5], the term hyper-reduction was introduced much later in [15], where a selection
of the integration point in a FE model is performed in a somehow a priori manner. The empirical
interpolation method (EIM) [4] approximates a non-affine function using a set of basis functions
over a continuous, bounded domain, and hence finds applications for reduced basis representations
of parameterized partial differential equations (PDEs). The discrete empirical interpolation method
(DEIM) [1] achieves this on a spatially discretized domain. For FE applications, the unassembled
DEIM or UDEIM [2] was shown to be more efficient, where the unassembled element-level nonlinear
force vector is used in the DEIM algorithm to return a small set of elements (instead of nodes) over
which the nonlinearity is evaluated. For conservative/port-Hamiltonian systems, however, the DEIM
(or UDEIM) leads to a loss of numerical stability during time integration. Though a symmetrized
version of DEIM has been proposed recently [3] to avoid this issue, its potential for FE-based applications remains unexplored. To this end, the energy-conserving sampling and weighting (ECSW)
hyper-reduction method [5] is remarkable. The ECSW method directly approximates the reduced
nonlinear term and is particularly well-suited for FE-based structural dynamics applications, as it
preserves the symmetry of the operators involves and leads to numerical stability.
The main idea of ECSW is simple yet powerful: only few elements are interrogated to compute
the nonlinear forces, and weights (obtained using training) are attached to them to match the virtual
work done by their full counterparts onto virtual displacements given by the linear reduction basis
vectors. Albeit being efficient and relevant for FE-based mechanical systems, the ECSW method,
by construction, is limited to perform hyper-reduction of systems over strictly linear subspaces. To
the best of the authors’ knowledge, this limitation is shared by all other hyper-reduction techniques
in the literature, as well.
Nonlinear manifolds in model reduction
It is quite natural, however, for the dynamics of general nonlinear systems to evolve on inherently
nonlinear manifolds, and not over linear subspaces (cf. [20] for a discussion). Often, a nonlinear
manifold relevant for system dynamics may not be embeddable in a low-dimensional linear subspace.
It is then easy to see that the linear-mapping-based ROM would either fail to capture the system
response, or require frequent basis-updates (online) to account for the nonlinearity, making the
reduction process cumbersome and possibly redundant due to high computational costs in setting
up such a ROM. Many methods allow for model reduction using attracting nonlinear manifolds in
nonlinear dynamical systems – from a theoretical, as well as a data-driven, statistical perspective.
2
The approximate inertial manifold (AIM) provides an approximation of the solution of a PDE
over sufficiently long time scales, using the theory of inertial manifolds. This is done by enslaving the
amplitudes of fast modes in the system as a smooth graph over the amplitudes of a few selected modes
while assuming that the corresponding inertial forces are not excited (see [25, 6], for instance). The
so called static condensation approach (cf. [17]) achieves the same objective without the projection
of the governing equations on to eigenfunctions. This is usually done for systems exhibiting a
dichotomy of unknowns, such as the axial (in-plane) and transverse (out-of-plane) unknowns in a
beam or a plate. From the perspective of nonlinear dynamics, a good overview of nonlinear model
reduction techniques can be found in [24] and the references therein.
Over the past decade or so, the use of nonlinear manifolds in model reduction has received a
significant boost by many data-driven techniques, broadly referred to as nonlinear dimensionality
reduction (NLDR) or manifold learning, see [14], for instance. These techniques aim at constructing
low dimensional manifolds from high-dimensional data snapshots. In a ROM perspective, these
snapshots may be extracted from simulations of the underlying HFM, and the chosen NLDR is
deemed to provide a nonlinear mapping with the required smoothness [9]. In essence, these methods
provide a nonlinear generalization of the Proper Orthogonal Decomposition (POD), a well-known
technique which delivers an optimal linear subspace for the dimensionality reduction.
The use of nonlinear manifolds in model reduction is expected to not only further reduce the
size of the ROM in contrast to its linear mapping counterpart, but also shed light on the underlying
system behaviour. This is, especially, true for the case of thin walled structures, which exhibit a
time scale separation between slow, out-of-plane; and fast, in-plane dynamics. The fast dynamics
in such systems can often be enslaved statically to the slow one, in a nonlinear fashion [8]. Recent
efforts [11, 12] have shown that the geometrically nonlinear dynamics of structures, characterized
by bending-stretching coupling (as for instance, thin walled components) can be effectively captured
using a quadratic manifold (QM), constructed using model properties rather then HFM snapshots.
Our contribution
The nonlinear-manifold-based ROMs are usually obtained by projecting the HFM on the tangent
space of the manifold, in a virtual work sense. This generates configuration-dependent inertial terms
for the ROM in addition to the already existing projected nonlinear terms. For general nonlinear
systems, the evaluation and projection of the nonlinearity remains a bottleneck for computations,
much like the linear mapping counterpart. As discussed above, the state-of-the-art hyper-reduction
techniques allow for fast approximation of the (projected) nonlinearity in the case of linear-mappingbased ROMs. This functionality, however, has been missing in the context of nonlinear-manifoldbased ROMs.
In this contribution, we generalize the application of the ECSW hyper-reduction method to
nonlinear-manifold-based ROMs. This work is essentially inspired from the ideas in the seminal
work of Farhat et al. [5, 7], where ECSW method was introduced in the context of linear mappings.
Following similar lines as in [7], we further show that our extension of the ECSW (named hereafter
as extended ECSW: EECSW) preserves the Lagrangian structure associated with the Hamilton’s
principle. A natural consequence of this property is the numerical stability during hyper-reduction,
as shown in [7]. The proposed method is tested, first on an academically simple example, and then
on a realistic model of a wing using the recently proposed QM for reduction. Our numerical results
indicate that better speed ups can be achieved, as compared to the ones obtained with a linear POD
subspace reduction of comparable accuracy equipped with ECSW. Remarkably, EECSW selects less
elements for the hyper-reduction as compared to its linear counterpart, and more than compensates
the extra computational costs associated to the configuration dependent inertial terms, which are
not present in the case of linear subspace reduction.
This paper is organized as follows. The concept of projection based ROM on nonlinear mappings
is briefly reviewed in Section 2. The extension of ECSW for nonlinear mappings is the tackled in
Section 3, where also the preservation of the Lagrangian structure of the thus-obtained ROM is
3
shown. Numerical results showing the effectiveness of the method are reported in Section 4, and
finally, the conclusions are given in Section 5.
2
Model reduction using nonlinear mappings
The partial differential equations (PDEs) for momentum balance in a structural continuum are first
FE discretized along the spatial dimensions to obtain a system of second-order ordinary differential
equations (ODEs). Along with the initial conditions for generalized displacements and velocities,
these ODEs govern the response of the underlying structure. More specifically, this response can be
described by the solution to an initial value problem (IVP) of the following form:
Mü(t) + Cu̇(t) + f (u(t)) = g(t) ,
(1)
u(t0 ) = u0 , u̇(t0 ) = v0 ,
where the solution u(t) ∈ Rn is a high-dimensional generalized displacement vector, M ∈ Rn×n is the
mass matrix; C ∈ Rn×n is the damping matrix; f : Rn 7→ Rn gives the nonlinear elastic internal force
as a function of of the displacement u of the structure; and g(t) ∈ Rn is the time dependent external
load vector. The system (1), referred to as the high-fidelity model (HFM), exhibits prohibitive
computational costs for large values of the dimension n of the system. The classical notion of model
reduction aims to reduce this dimensionality by introducing a linear mapping on to a suitable lowdimensional invariant subspace. However, as discussed in the Introduction, more recent advances
in model reduction allow for the detection of system dynamics, evolving over nonlinear manifolds.
The idea of a projection-based reduced-order model using nonlinear mappings was presented in [11],
using a nonlinear mapping such as
u(t) ≈ Γ(q(t)),
(2)
where Γ : Rm 7→ Rn is the nonlinear mapping function. In general the mapping (2), when substituted
into the governing equations (1), results in a non-zero residual r as
MΓ̈ (q(t)) + CΓ̇ (q(t)) + f (Γ (q(t))) − g(t) = r.
The virtual work principle for a kinematically admissible displacement, δu =
nonlinear manifold leads to
r · δu = 0
h
i
=⇒ MΓ̈ (q(t)) + CΓ̇ (q(t)) + f (Γ (q(t))) · δu = g(t) · δu .
∂Γ(q)
∂q δq,
over the
Physically, this means that the residual force is orthogonal to the tangent space of the manifold.
This gives the corresponding reduced-order model as
h
i
PTΓ MΓ̈ (q(t)) + CΓ̇ (q(t)) + f (Γ (q(t))) = PTΓ g(t),
(3)
where PΓ = ∂Γ(q)
∈ Rn×m represents the tangent space of the nonlinear manifold in which the
∂q
system dynamics is expected to evolve. The acceleration and velocity vectors over the manifold can
be evaluated using the chain rule as
d
∂Γ(q)
Γ (q(t)) =
· q̇,
dt
∂q
2
d
∂Γ(q)
∂ Γ(q)
Γ̈ (q(t)) = Γ̇ (q(t)) =
· q̈ +
· q̇ · q̇ .
dt
∂q
∂q∂q
Γ̇ (q(t)) =
Using the above relations, (3) can also be expressed as
Mr (q) q̈ + Cr (q, q̇) q̇ + fr (q) = PTΓ g(t), ,
4
(4)
where Mr (q) = PTΓ MPΓ , Cr (q, q̇) = PTΓ CPΓ + PTΓ M
∂ 2 Γ(q)
∂q∂q
· q̇ and fr (q) = PTΓ f (Γ(q)).
Note that for Γ being a linear function, such that Γ(q) = Vq (for some V ∈ Rn×m spanning
a suitable low dimensional subspace), the ROM in (3) reduces to the familiar Galerkin-projectionbased ROMs as
VT MVq̈(t) + VT CVq̇(t) + VT f (Vq(t)) = VT g(t).
(5)
During time integration - usually referred to as online phase - of (5), the computational cost associated to the evaluation of the linear terms in (5) scales only with the number of reduced variables
m, since the corresponding reduced operators can be pre-computed offline. However, this is not
the case for the computation of the reduced nonlinear term VT f (Vq(t)). Specifically, for FE-based
applications, this evaluation is usually carried out online, in the following manner:
f̂ (q) := VT f (Vq) =
ne
X
VeT fe (Ve q),
(6)
e=1
where fe (ue ) ∈ RNe is the contribution of the element e towards the vector f (u) (Ne being the
number of DOFs for the element e), Ve is the restriction of V to the rows indexed by the DOFs
corresponding to e, and ne is the total number of elements in the structure. It is then easy to see from
(6), that the cost associated to the computation of the reduced nonlinear force f̂ (q) scales linearly
with the total number of elements ne in the structure, which is potentially high for large systems.
Thus, despite the reduction in dimensionality achieved in (5), the evaluation of the reduced nonlinear
term f̂ (q) emerges as a new bottleneck for the fast prediction of system response using the ROM
(5). Similar conclusions regarding computational bottlenecks hold for the nonlinear-reduction-based
ROM (3), as well.
3
Hyper-reduction and Extended ECSW
Hyper-reduction techniques help mitigate the above-stated high computational costs by approximation of the reduced-nonlinear term in a computationally affordable manner. As discussed in the
Introduction, the ECSW [5] is a remarkable method for hyper-reduction of linear-mapping-based
ROMs such as that in (5). After briefly summarizing ECSW, we propose here an extension to
ECSW, which facilitates hyper-reduction of nonlinear-manifold-based ROM, such as in (3). Much
like ECSW, we further show in that the proposed extended ECSW (EECSW) is numerically stable
due to preservation of the Lagrangian structure associated to the underlying mechanical system.
Analogous to (6), the projected nonlinear force (possibly a combination of inertial, damping or
elastic forces) in the ROM (3) can be expressed as a summation of the element-level contributions
as
ne
X
T
(PΓ (q))e he Γe (q), Γ̇e (q) , Γ̈e (q) ,
(7)
hr (q, q̇, q̈) = PTΓ h Γ(q), Γ̇ (q) , Γ̈ (q) =
e=1
where hr (q, q̇, q̈) is the part of the ROM (3) which is potentially unaffordable to compute (cost
scales with n or ne ), h is its full counterpart, with the he having the usual meaning for the elementlevel contribution from the element e. One easily observable distinction of EECSW from its linear
mapping counterpart is that one can precompute the reduced linear operators in the hyper-reduced
model. Due to the nonlinear mapping and projection on the tangent space, we see that EECSW
doesn’t feature reduced-linear operators, even for the linear mass, stiffness and damping matrices in
the HFM. Thus, in principle, one could include all the terms on the left hand of (3) in hr , i.e.,
h
i
hr (q, q̇, q̈) = PTΓ MΓ̈ (q(t)) + CΓ̇ (q(t)) + f (Γ (q(t))) .
The summation in (7) can be approximated as
(EECSW) hr (q, q̇, q̈) ≈ h̃r (q, q̇, q̈) =
X
e∈E
5
T
ξe (PΓ (q))e he Γe (q), Γ̇e (q) , Γ̈e (q) ,
(8)
where ξe ∈ R+ is a positive weight given to every element e ∈ E. These weights would be smartly
chosen using training to ensure a good approximation of the original sum over all ne elements.
Clearly, the evaluation of h̃r in (8) comes at a fraction of the computational cost associated to (7),
for |E| ne .
It is easy to see that this idea is a straight-forward generalization of ECSW to nonlinear-manifoldbased ROMs with projection to the tangent space, as given in (3). Indeed, for Γ(q) = Vq, hyperreduction of ROM (5) results in the approximation of the projected internal force in (6) as
X
f̂ (q) ≈ f̃ (q) =
ξe VeT fe (Ve q),
(9)
e∈E
which is same as the expression one would obtain from the direct application of ECSW, as well.
3.1
Element weight selection
The elements and weights are determined to approximate virtualwork over chosen training sets
which generally come from full solution run(s). We assume that u(i) , u̇(i) , ü(i) , i ∈ {1, . . . , nt }
represents a set of available training vectors. If the solution to the HFM, indeed,
evolves over a
nonlinear manifold as given by (2), then we must be able to find reduced unknowns q(i) , q̇(i) , q̈(i) 1
such that
u(i) ≈ Γ q(i)
(i)
(i)
(i)
u̇ ≈ PΓ q
q̇
∀i ∈ {1, . . . , nt },
(10)
#
"
2
∂ Γ(q)
· q̇(i) · q̇(i)
ü(i) ≈ PTΓ q(i) q̈(i) +
∂q∂q
q=q(i)
As will be clear later in this section, the optimal elements and weights corresponding to a reduced
mesh needs the set of reduced vectors q(i) , q̇(i) , q̈(i) during the training stage. In case these
are available from a previous ROM simulation, these would be directly useful. However, if not,
the nonlinear system of algebraic equations (10) can be solved for the reduced training vectors
q(i) , q̇(i) , and q̈(i) in a sequential manner. This system is potentially under-constrained (nt < n
usually) and , hence, the sought q(i) can be obtained from the solution to a nonlinear least-squares
problem as
q(i) = arg minm Γ (q) − u(i)
q∈R
2
2
.
(11)
Once q(i) is obtained, the second equation in (10) can be solved in a least squares sense for q̇(i) as
h
i−1
q̇(i) = PTΓ q(i) PΓ q(i)
PTΓ q(i) u̇(i) .
Finally, one can obtain the corresponding reduced accelerations as
"
h
i−1
∂ 2 Γ(q)
q̈(i) = PTΓ q(i) PΓ q(i)
PTΓ q(i)
ü(i) −
∂q∂q
#
· q̇(i) · q̇(i)
!
.
q=q(i)
1 Note that the ROM (4) can be expressed as:
−1
−1
T
q̈ + M−1
r (q)Cr (q, q̇) q̇ + Mr (q)fr (q) = Mr (q)PΓ g(t),
and by doing so, the inertial terms in the ROM become linear and require no hyper-reduction. This would also
avoid the need for the acceleration training vectors ü(i) or q̈(i) , as is the case in ECSW. Nonetheless, we treat hyperreduction and training in the straight-forward and general setting of (3) where the inertial terms may be hyper-reduced
and acceleration training may also be performed.
6
Note that for Γ being a linear function, such as Γ(q) = Vq, the system (11) reduces to the familiar
least squares solution of (10), available in closed form as
q(i) = (VT V)−1 VT u(i) ,
q̇(i) = (VT V)−1 VT u̇(i) ,
q̈(i) = (VT V)−1 VT ü(i) .
For general nonlinear mappings, however, iterative methods would be needed to solve (11). Starting
(i)
with an initial guess q0 for the solution, one such simple iterative procedure is the gradient-descent
method (cf. [13] for a review of methods for nonlinear least squares problems), and can be formulated
as follows:
2
(i)
(i)
r qk := Γ qk − u(i) ,
(12)
2
(i)
(i)
(i)
qk+1 = qk − γk ∇r qk ,
where γk is the well-known descent rate, which should be chosen small enough to facilitate smooth
decay to the local minimum of r, and the ∇r is the gradient of r, obtained using the configurationdependent tangent space PTΓ of the nonlinear manifold as
(i)
(i)
(i)
∇r qk = PTΓ (qk ) Γ qk − u(i) .
Once the reduced training vectors q(i) , q̇(i) , q̈(i) are obtained, the element level contribution of
projected internal force for each of the training vectors can be assembled in a matrix G as follows:
b1
g11 . . . g1ne
.. ∈ Rmnt ×ne ,
..
b = ... ∈ Rmnt ,
G = ...
.
.
gie = PΓ
gnt 1 . . . gnt ne
T
q(i)
he Γe (q(i) ), Γ̇e q(i) , Γ̈e q(i) ,
e
bnt
bi = fr q
(i)
=
ne
X
(13)
gie
e=1
∀i ∈ {1, . . . , nt },
e ∈ {1, . . . , ne } .
Analogous to ECSW, the set of elements and weights is then obtained by a sparse solution to the
following non-negative least-squares (NNLS) problem
(P 1) : ζ = arg
min
ζ̃∈Rne ,ζ̃≥0
kGζ̃ − bk22 ,
(14)
A sparse solution to (P 1) returns a sparse vector ζ, the non-zero entries of which form the reduced
mesh Z used in (9) as
Z = {e : ζe > 0}.
An optimally sparse solution to (P 1) is NP-hard to obtain. However, a greedy-approach-based
algorithm [10], which finds a sub-optimal solution, has has been found to deliver an effective reduced
mesh Z [5]. For the sake of completeness, we have included its pseudo-code in Algorithm 1. More
recent work [23] proposes and discusses different alternatives to accelerate the sampling procedure
during hyper-reduction.
7
Algorithm 1 Training for EECSW
Input: {u(i) }, i ∈ {1, . . . , nt }, τ (sparse NNLS tolerance)
Output: ξ ∈ Rne sparse, E ⊂ {1, . . . , ne }
14:
15:
16:
17:
for i ← 1 to nt do
q(i) ← LSQNLIN u(i)
. LSQNLIN: solves the nonlinear least squares problem (11), e.g.,
using gradient descent method (12)
end for
[G, b] ← Construct_G_b {q(i) }n1 t
. Construct_G_b: assembles G, b, as given by (13)
E ← ∅, Z ← {1, . . . , ne }, ξ ← 0 ∈ Rne
while kGξ − bk2 > τ kbk2 do
. sparse NNLS solver
µ ← GT (b − Gξ)
[|ν|, e] ← max (µ) . max: returns maximum value in a vector followed by its location (index)
E ← E ∪ {e}, Z ← Z\{e}
while true do
ζ E ← G†E b
. † represents pseudo-inverse
ζZ ← 0
if ζ E > 0 then
ξ←ζ
break
end if
η ← min ξk /(ξk − ζk )
18:
19:
20:
21:
22:
ξ ← ξ + η(ζ − ξ)
Z ← {i|ξi = 0}
E ← {1, . . . , ne }\Z
end while
end while
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
k∈E
3.2
3.2.1
Preservation of Lagrangian structure and numerical stability
Conservative systems
Following similar steps as in [7], we show here that the hyper-reduced model preserves the Lagrangian
structure, leading to numerical stability of the hyper-reduced model. For this purpose, we restrict
our discussion to a conservative system, where the Hamiltonian principle represents the conservation
of total mechanical energy as follows:
E = T (u̇) + V(u) = constant,
(15)
where T and V are the functionals associated to the kinetic energy and the potential energy of the
system, respectively. The kinetic energy is given by
n
T (u̇) =
e
X
1 T
1
u̇ Mu̇ =
u̇e Me u̇e ,
2
2
{z }
e=1 |
(16)
Te (ue )
where Me ∈ RNe ×Ne is the contribution of the element e towards the mass matrix. Furthermore,
the potential energy can be written as
Z
Z
V(u) = − f (u) · du + g · du ,
(17)
|
{z
} | {z }
V int (u)
8
V ext (u)
such that
∂V int (u)
,
∂u
ext
∂V (u)
g(u) = −
,
∂u
f (u) = −
(18)
(19)
and V int and V ext are the potential energies associated to internal and external forces, respectively.
The conservative counterpart of the equations of motion (1) can then be expressed using Lagrange
equations as:
∂V int (u)
d ∂T (u̇)
+
= g.
(20)
dt
∂ u̇
∂u
As shown in , the reduced-order model, obtained by introduction of a nonlinear mapping Γ and
projection of the resulting equations to the tangent space PΓ of the manifold, is given by
"
#
!
d ∂T (Γ̇(q))
∂V int (Γ(q))
T
PΓ
+
= PTΓ g,
(21)
dt
∂Γ(q)
∂ Γ̇(q)
where Γ : Rn 7→ Rm describes a general nonlinear mapping which approximates the solution u using
a reduced set of unknowns q. It is easy to see that this equation is equivalent to ROM (3) in the
conservative setting, i.e.,
∂V int (Γ(q))
PTΓ MΓ̈(q) +
= PTΓ g
(22)
∂Γ(q)
As is shown below, the projection of the equations to the tangent space of the manifold is essential
for preserving the Lagrangian structure associated to the corresponding Hamiltonian principle. Note
that ROM obtained using a Galerkin projection is a special case of the general setting in (22) when
a linear mapping Γ(q) = Vq is used. It is well known (as also shown in [7]) that such a Galerkin
projection is structure-preserving. We show here that this is also true for ROM (22), obtained by
projection on to the tangent space of general nonlinear manifolds.
The Hamiltonian, associated to (22), is given by
Er (q, q̇) = T (Γ̇(q)) + V(Γ(q)),
=⇒ Ėr (q, q̇, q̈) = Ṫ (Γ̇(q)) + V̇(Γ(q))
∂V ext (Γ(q))
∂T (Γ̇(q))
∂V int (Γ(q))
Γ̇(q) −
Γ̇(q)
Γ̈(q) +
∂Γ(q)
∂Γ(q)
∂ Γ̇(q)
∂V int (Γ(q))
= (Γ̇(q))T MΓ̈(q) + (Γ̇(q))T
− (Γ̇(q))T g
∂Γ(q)
∂V int (Γ(q))
T
T
−g
(∵ Γ̇(q) = PΓ q̇)
= q̇
PΓ MΓ̈(q) +
∂Γ(q)
=
=0
(Using (22))
(23)
Thus, Er = constant, and the total mechanical energy is conserved for ROM in (22).
Having established that tangent space projection leads to structure-preserving ROMs, we now
show that the same is true for hyper-reduced model obtained using EECSW, where the reduced
internal force from (22) is approximated (cf. (8)) as follows:
f̂ (q) = PTΓ f (Γ(q)) = PTΓ
ne
X
∂Veint (Γe (q))
∂Veint (Γe (q))
∂V int (Γ(q)) X
=
≈
ξe PTΓ e
.
PTΓ e
∂Γ(q)
∂Γe (q)
∂Γe (q)
e=1
e∈E
(24)
9
Furthermore, the inertial force can also be approximated as
PTΓ MΓ̈(u) =
ne
X
PTΓ
e
Me Γ̈e (q) ≈
e=1
X
ξe PTΓ
e
Me Γ̈e (q) =
e∈E
X
ξe PTΓ
e∈E
d ∂Te (Γ̇e (q))
.
e dt
∂ Γ̇e (q)
(25)
Thus, it follows from (24) that the potential energy V int (Γ(q)) associated to the reduced internal
force can also be approximated as
X
Vrint (q) := V int (Γ(q)) ≈ Ṽrint (q) =
ξe Veint (Γe (q)) ,
(26)
e∈E
and from (25) that the kinetic energy, T (Γ̇(q)), can be approximated as
X
Tr (q̇, q) := T Γ̇(q) ≈ T̃r (q̇, q) =
ξe Te (Γ̇e (q)) ,
(27)
e∈E
Here, Ṽrint (q) and T̃r (q̇, q) are the potential and the kinetic energy associated to the hyper-reduced
system of equations given as:
X
ξe PTΓ
X
Me Γ̈e (q) +
e
e∈E
ξe PTΓ
e∈E
∂Veint (Γe (q))
= PTΓ g.
e
∂Γe (q)
(28)
Analogous to the previous discussion, the Hamiltonian associated to the hyper-reduced system becomes
int
Ẽr (q, q̇) = T̃r (q̇, q) + V˜r (q) + V ext (Γ(q))
=⇒ Ẽ˙r (q, q̇, q̈) = T̃˙r (q̇, q) + Ṽ˙ rint (q) + V̇ ext (Γ(q))
= q̇
T
X
ξe PTΓ e
Me Γ̈e (q) +
e∈E
=0
X
e∈E
∂Veint (Γe (q))
− PTΓ g
∂Γe (q)
!
ξe PTΓ e
(Using (28)).
This shows that EECSW for ROMs given by (3) preserves the Lagrangian structure and conserves
energy according to the Hamilton’s principle for a conservative mechanical system. The numerical
stability of the hyper-reduced model is a direct consequence of this structure-preserving property as
noted in [7], as well.
3.2.2
Dissipative, non-conservative systems
To include dissipation effects which, typically, generate forces opposite to the generalized velocity
vector u̇ in the system, a dissipation functional D can be defined, such that the corresponding forces
are given by
∂D(u̇)
f dis (u̇) = −
.
∂ u̇
Assuming that the functional D has the same degree of homogeneity, m, in the generalized velocities
(u̇) for all elements in the mesh, we have
u̇T
ne
ne
X
∂D(u̇) X
∂De (u̇e )
=
=m
u̇Te
De (u̇e ),
∂ u̇
∂ u̇e
e=1
e=1
(29)
where De is the usual notation for the contribution of element e towards the total dissipation D,
which would be dependent on the nodal velocities u̇e only. For the special case of viscous damping,
i.e., m = 2, we have
1
D(u̇) = u̇T Cu̇.
2
10
Other values of m lead to different types of dissipation models, e.g., m = 1 leads to dry Coloumb
friction and m = 3 leads to a viscous drag in fluids.
The Hamilton’s principle in this non-conservative setting leads to
ne
X
∂D(u̇)
nc
T nc
De (u̇e ),
Ṫ + V̇ = −
+f
· u̇ = u̇ f − m
∂ u̇
e=1
(30)
where f nc is a non-conservative, time-dependent external force. The corresponding Lagrange equations of motion can then be expressed, equivalently, as
Mü +
∂D(u̇) ∂V int (u)
+
= g + f nc .
∂ u̇
∂u
(31)
Following the same procedure as shown above for conservative systems, it is easy to show that
the Lagrangian structure is preserved using EECSW for non-conservative systems, as well.
The nonlinear ROM associated to (31) is given by
#
"
∂D(Γ̇(q)) ∂V int (Γ(q))
T
+
= PTΓ [g + f nc ],
(32)
PΓ MΓ̈(q) +
∂Γ(q)
∂ Γ̇(q)
The approximation of the additional nonlinear term in (32) using EECSW is given by
ne
∂D
Γ̇
(q)
∂D
Γ̇
(q)
X
X
e
e
e
e
∂D(Γ̇(q))
PTΓ
=
PTΓ e
≈
PTΓ e
.
∂ Γ̇(q)
∂
Γ̇
(q)
∂ Γ̇e (q)
e
e=1
e∈E
(33)
It follows that the dissipation can be approximated as
Dr (q̇, q) := D(Γ̇(q)) ≈ D̃r (q̇, q) =
X
ξe De (Γ̇e (q)),
(34)
e∈E
where D̃r can be thought of as the dissipation associated to the hyper-reduced model given by
∂De Γ̇e (q)
int
X
∂V
(Γ
(q))
e
e
= PTΓ [g + f nc ].
ξe PTΓ e Me Γ̈e (q) +
+
(35)
∂Γe (q)
∂
Γ̇
(q)
e
e∈E
The rate of change of the total energy Ẽr associated to the hyper-reduced model (35) is given by
d
(Γ(q) · f nc )
Ẽ˙r (q, q̇, q̈) = T̃˙r (q̇, q) + Ṽ˙ rint (q) + V̇ ext (Γ(q)) −
dt
!
int
X
∂V
(Γ
(q))
e
e
= q̇T
ξe PTΓ e Me Γ̈e (q) +
− PTΓ [g + f nc ] − Γ(q) · ḟ nc
∂Γe (q)
e∈E
∂De Γ̇e (q)
X
=−
ξe q̇T PTΓ e
− Γ(q) · ḟ nc
(Using (35))
∂
Γ̇
(q)
e
e∈E
T ∂De Γ̇e (q)
X
=−
ξe Γ̇e (q)
− Γ(q) · ḟ nc
∂
Γ̇
(q)
e
e∈E
= −mD̃r (q̇, q) − Γ(q) · ḟ nc
(Using (29), (34))
T
=⇒ T̃˙r (q̇, q) + Ṽ˙ rint (q) + V̇ ext (Γ(q)) = −mD̃r (q̇, q) + Γ̇(q) f nc .
11
(36)
When compared with (30) and (34), Equation (36) proves that EECSW preserves the Lagrangian
structure associated to the Hamiltonian principle for a non-conservative system, as well.
The conditional/unconditional numerical stability properties of any chosen time integration
scheme are carried over to the hyper-reduced model as a direct consequence of the energy-conserving/
structure-preserving properties of EECSW. We refer the reader to Section 4.3 in [7] for the discussion
on how these properties lead to numerical stability.
4
4.1
Numerical results
Setup
To test and illustrate the performance of the EECSW method, proposed here, we require a highdimensional system which can be reduced using a suitable nonlinear mapping. To this end, we use
the quadratic manifold, introduced in [11], for model reduction. In the quadratic manifold approach,
the full set of unknowns u are mapped to a lower dimensional set of variables q using a quadratic
mapping as
1
u ≈ Γ(q) := Φq + Ω : (q ⊗ q) ,
2
(37)
where the columns of Φ ∈ Rn×M contains M significant vibration modes (VMs) extracted at equilibrium u = 0, and Ω ∈ Rn×M ×M is a third-order tensor containing the corresponding static modal
derivatives (SMDs). The symbol ⊗ denotes the dyatic product, while : indicates contraction over
the last two indexes. 2 .
For consistency, we use the same examples used in [11]: Model-I, a flat plate, simply-supported
on two opposite sides, and Model-II, a NACA-airfoil-based wing model. The details of both models
are shown in Figures 1 and 3. Time integration is performed using the implicit Newmark scheme
with a time step constant across all techniques for a fair comparison. Both structures are discretized
using flat, triangular-shell elements with 6 DOFs per node, i.e., 18 DOFs per element. For both
models, the accuracy of the results is compared to the corresponding full nonlinear solutions using
a mass-normalized global relative error (GRE) measure, defined as
rP
(u(t) − ũ(t))T M(u(t) − ũ(t))
GREM =
t∈S
rP
× 100,
u(t)T Mu(t)
t∈S
where u(t) ∈ Rn is the vector of generalized displacements at the time t, obtained from the HFM
solution, ũ(t) ∈ Rn is the solution based on the (hyper)reduced model, and S is the set of time
instants at which the error is recorded. The mass matrix M provides a relevant normalization for
the generalized displacements, which could be a combination of physical displacements and rotations,
as is the case in the shell models shown here.
The computations were performed on the Euler cluster of ETH Zürich. Since the success of
reduction techniques is often reported in terms of savings in simulation time, we define an online
speedup S ? computed according to the following simple formula:
S? =
Tf ull
? ,
Tsim
?
where Tf ull and Tsim
represent the CPU time taken during the time integration of full and (hyper)reduced solution respectively. The superscript ? denotes the reduction technique being used.
2 The mapping (37) can be written using the Einstein summation convention in the indicial notation as: Γ =
I
ΦIi qj + 21 ΩIij qi qj I ∈ {1, . . . , n}, i, j ∈ {1, . . . , M }
12
We compare the hyper-reduction using ECSW (where the ROMs are obtained from linear mappings), with the EECSW method. For both ECSW and EECSW, the convergence tolerance τ for
finding a sparse solution to the NNLS problem (14), as required in Algorithm 1, is chosen to be
0.01. This is within the practically recommended range as reported in [5]. A total of nt = 200
training vectors, chosen uniformly from the solution time-span, are used in sparse NNLS routine.
The following ROMs are considered in our comparison:
• Linear-mapping-based (Hyper-)Reduction:
POD: Here, the first most significant m Proper Orthogonal Decomposition (POD) modes
obtained from the HFM run are used to construct V. For a fair comparison, m is taken
equal as in LM.
ECSW-POD: Hyper-reduction using ECSW is performed, for the ROM obtained in the POD
approach. This is the original ECSW implementation as proposed in [5].
• Quadratic-manifold-based (Hyper-)Reduction:
QM: Here, the quadratic mapping (37) is used to construct the ROM as given by (3).
EECSW-QM: The ROM obtained in QM approach using the quadratic manifold is hyperreduced using the EECSW.
4.2
Flat Structure
Flat plate mesh
D
B
z, w
1
0
-1
40
35
30
y, v
25
20
x, u
15
A
C
20
15
10
10
5
5
0
(a)
0
(b)
Figure 1: Model - I: (a) Flat plate example - simply supported on sides A and D. The plate is
L = 40 mm long, H = 20 mm wide, t = 0.8 mm thick. The Young Modulus is E = 70 GPa, the
Poisson’s ratio is ν = 0.33, and the density is ρ = 2700 Kg/m3 . A uniform pressure is applied on
the plate, according to the time history p(t) = P sin(ωt), where P = 1 N/mm2 and ω = 2.097 × 104
rad/s. (b) The curved plate is discretized using flat, triangular shell elements. The FE mesh contains
400 elements and 1386 DOFs. The two opposite sides parallel to the y-axis are simply supported, a
uniform pressure is applied on the curved surface, according to the time history.
We consider the rectangular flat plate structure shown in Figure 1. A uniform pressure distribution
is chosen to act normal to the plate surface as the external load. A time varying amplitude (load
function) is used given by
g(t) = p(t)l,
p(t) = sin(ωt),
13
(38)
where l is a constant load vector corresponding to a uniform pressure distribution of 1 N/mm2 . Here,
p(t), termed as the dynamic load function, determines the time-dependency of the external load.
The results are shown for a periodic choice for p(t) as given by (38), where ω is chosen to be the first
natural frequency of the linearized system. Due to the resonant nature of the dynamic, it is easy to
excite geometrically nonlinear behavior in the system, as shown by Figure 2. Furthermore, Figure
2 shows the that the full nonlinear solution is reduced using the QM technique and, subsequently,
hyper-reduced using EECSW with good accuracy.
(a)
1.8
Full NL
Linear
QM
EECSW-QM
POD-2
POD-5
ECSW-POD-5
1.6
Displacement (mm)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
1.7
1.8
1.9
2
time(s)
2.1
2.2
2.3
×10−3
(b)
Figure 2: The linear solution (dashed-blue) grows quickly to high amplitudes for the choice of
resonant loading; the full solution (thick black) shows much lower amplitudes due to the geometrical
nonlinearities; the ROM obtained using quadratic manifold (yellow) and hyper-reduced solution for
this ROM using EECSW (dashed-red) capture the full solution accurately. Subfigure (b) highlights
the inaccuracy of the POD ROM with m = 2.
It can also be confirmed from the global relative error GREM and speed-up S ? , shown in Table
1, that the QM based nonlinear ROM reproduces the results of Model-I with good accuracy but,
as expected, does not deliver a good speed-up. The EECSW-based hyper-reduction, on the other
hand, gives results with a GREM of 1.37% but delivers a much higher speed-up of about 100. Note
14
also that a linear-mapping-based ROM with POD of the same size (m = 2) as the adopted nonlinear
mapping does not deliver a good accuracy.
Reduction
Technique
Basis
size (m)
2
5
5
2
2
POD
POD
ECSW-POD
QM
EECSW-QM
#
elements
400
400
12
400
6
GREM (%)
S?
5.62
1.81
2.52
1.32
1.37
2.96
2.84
77.7
3.03
99.93
Table 1: Comparison of global relative error GREM and speed-ups in Model-I for linear-mappingbased ROMs and their hyper-reduced counterparts on one hand, and for quadratic-manifold-based
ROMs and their hyper-reduced counterparts using EECSW on the other hand. Time for a full
solution run Tf ull = 74.4 s.
4.3
Wing Structure
Dynamic Load function p(t)
10000
8000
6000
4000
2000
0
0
0.05
0.1
0.15
0.2
0.25
0.3
time(s)
(a)
(b)
Figure 3: Model - II: (a) A wing structure with NACA 0012 airfoil (length(L) = 5 m, Width(W )
≈ 0.9 m, Height(H) = 0.1 m ) stiffened with ribs along the longitudinal and lateral direction. The
Young Modulus is E = 70 GPa, the Poisson’s ratio is ν = 0.33, and the density is ρ = 2700 Kg/m3 .
The wing is cantilevered at one end. Uniform pressure is applied on the highlighted area, with a
pulse load as given by (39); (b) The structure is meshed with triangular flat shell elements with
6 DOFs per node and each with a thickness of 1.5 mm. The mesh contains n = 135770 DOFs,
ne = 49968 elements. For illustration purposes, the skin panels are removed, and mesh is shown.
For the application of the proposed simulation-free hyper-reduction methods to more realistic models,
we consider the model of a NACA-airfoil-based wing structure, introduced in [11]. This model,
referred to as Model-II, contains truly high number of DOFs (n = 135770), thereby, allowing for
the appreciation of obtained accuracy and computational speed-ups. We simulate the response of
15
the structure to a low frequency pulse load, applied as a spatially uniform pressure load on the
highlighted area on the structure skin (cf. Figure 3). The dynamic load function, shown in Figure
4a, is given as
h
π
i
p(t) = A sin2 (ωt) H(t) − H
−t ,
(39)
ω
where H(t) is the Heaviside function and ω chosen as the mean of the first and second natural
frequency of vibration. We refer the reader to [11] for the detailed linearized, full nonlinear and
reduced nonlinear response plots of this model.
Figure 4b and Table 2 show that the hyper-reduction using EECSW for the quadratic-manifold
based ROM reproduce the full system response with a very good accuracy (GREM ≈ 1.71%) and
is more than a thousand times faster than the full response computation. As expected, we obtain
a much higher speed-up during hyper-reduction in comparison with Model-I. The speed-up without
hyper-reduction, however, are fairly similar between the two models, despite the striking difference
in the number of DOFs. Thus, the numerical results confirm the need and effectiveness of EECSW
for hyper-reduction of large nonlinear systems over nonlinear manifolds.
6
×10−3
0.15
Full
Linear
QM
EECSW-QM
POD-5
POD-10
ECSW-POD-10
5
4
0.1
y displacement (mm)
3
x displacement (mm)
Full
Linear
QM
EECSW-QM
POD-5
POD-10
ECSW-POD-10
2
1
0
-1
-2
0.05
0
-0.05
-0.1
-3
-4
0
0.05
0.1
0.15
0.2
0.25
-0.15
0.3
0
0.05
0.1
0.15
time (s)
time (s)
(a)
(b)
2
0.25
0.3
×10−3
Full
Linear
QM
EECSW-QM
POD-5
POD-10
ECSW-POD-10
1.5
1
z displacement (mm)
0.2
0.5
0
-0.5
-1
-1.5
-2
-2.5
0
0.05
0.1
0.15
0.2
0.25
0.3
time (s)
(c)
Figure 4: (a), (b), (c): Comparison of full, linear and various reduced responses at tip node 2 in x, y
and z directions. Note that the QM-based reduced solution (with 5 reduced unknowns) accurately
captures the full system response, where as the POD-based ROM with 5 modes fails to reproduce
the nonlinear system response (POD-5).
16
Figure 4 and Table 2 compare the response of Model-II to the applied load. Figure 4 further
shows that the response of the full system is captured more effectively using the quadratic manifold
as compared to the POD with the same number of unknowns (POD-5). This underscores the
importance of nonlinear manifolds in model reduction of nonlinear systems. In terms of speed as
well, we see in Table 2 that the EECSW method performs better than its linear counterpart.
Reduction
Technique
POD
POD
ECSW-POD
QM
EECSW-QM
Basis
size (m)
5
10
10
#
elements
49968
49968
152
GREM (%)
S?
5.01
0.98
1.05
3.07
2.77
794.6
5
5
49968
44
1.65
1.71
2.92
1098
Table 2: Starting with a linear modal superposition with M = 5 modes (first 5 VMs), reduction
techniques as described in Section 4.1 are formulated. Global Relative Error and speed-ups in ModelII for these reduction techniques are tabulated. A total of nt = 200 training vectors, chosen uniformly
from the solution time-span, are used to setup hyper-reduction in ECSW-POD and EECSW-QM.
Time for a full solution run Tf ull = 3.808 × 104 s.
It is interesting to see that EECSW-QM ends up sampling a much smaller set of elements as
compared to its linear counterpart EECSW-POD. The corresponding speed-up for EECSW-QM,
however, doesn’t show a proportionate increase. This can be explained on two accounts:
1. The additional convective terms in the nonlinear-mapping-based ROM (4) adds to the computational burden in comparison with its linear counterpart (5)
2. The evaluation of the nonlinear mapping (quadratic in this case), is expected to be more expensive that a linear mapping, which boils down to a standard matrix-vector product. Though
MATLAB is extremely efficient at performing such matrix-vector products, and hence, in the
evaluation of linear mappings, the same cannot be said about our sub-optimal implementation
for evaluation of the quadratic mapping at hand.
Nonetheless, in our numerical tests, EECSW more than compensates for these extra costs by outperforming ECSW-POD in terms of speed.
5
Conclusion
We have extended the concept of hyper-reduction for model reduction over nonlinear manifolds of
nonlinear, FE-based mechanical systems. In particular, the recently developed ECSW technique
has been extended to allow for hyper-reduction using smooth nonlinear mappings into the reduced
system of unknowns. In doing so, we showed that the ROM obtained by projection onto the tangent
space of the manifold, leads to the preservation of structure and stability properties of the ECSW
method.
We demonstrated the EECSW method on two structural dynamics examples by hyper-reduction
of a ROM based on the quadratic manifold, whereby better speed and accuracy were obtained using
the EECSW method in comparison with ECSW-POD based hyper-reduced models. Since these models were based on von-Karman nonlinearities, the degree of nonlinearity in the examples is expected
to be much lower than the nominal thin-walled structures featuring geometrically nonlinear behavior. A quadratic manifold was, therefore, effective in capturing the nonlinear behavior effectively
for these structures. However, nonlinear manifolds – possibly constructed from solution databases
– would play a crucial role in the effective reduction of models with higher degree of nonlinearities.
17
EECSW in such cases is expected to play a vital role during the hyper-reduction stage by providing
the necessary computational speed-ups. Testing the EECSW on ROMs based on nonlinear manifolds
constructed using data-driven techniques would be a part of our future endaevours.
Acknowledgements The authors acknowledge the support of the Air Force Office of Scientific
Research, Air Force Material Command, USAF under Award No.FA9550-16-1-0096. We also thank
Ludovic Renson for pointing us to the MATLAB-based tensor product toolbox TPROD.
References
[1] Chaturantabut, S., Sorensen, D. C. Nonlinear model reduction via discrete empirical interpolation, SIAM Journal on Scientic Computing (2010) 32(5): 2737-2764. DOI: 10.1137/090766498
[2] Tiso, P., Rixen, D. J. (2013) Discrete Empirical Interpolation Method for Finite Element Structural Dynamics. In: Kerschen G., Adams D., Carrella A. (eds) Topics in Nonlinear Dynamics,
Volume 1. Conference Proceedings of the Society for Experimental Mechanics Series, vol 35.
Springer, New York, NY. DOI: 10.1007/978-1-4614-6570-6_18
[3] Chaturantabut, S., Beattie, C., Gugercin, S. Structure-preserving Model Reduction for Nonlinear Port-Hamiltonian Systems arXiv:1601.00527 (2016)
[4] Barrault, M., Maday, Y., Nguyen, N. C., and Patera, A. T. An “empirical interpolation” method:
Application to efficient reduced-basis discretization of partial differential equations, C. R. Math.
Acad. Sci. Paris (2004) 339: 667–672. DOI: 10.1016/j.crma.2004.08.006
[5] Farhat, C., Avery, P., Chapman, T., Cortial, J. Dimensional reduction of nonlinear finite element dynamic models with finite rotations and energy-based mesh sampling and weighting for
computational efficiency, International Journal for Numerical Methods in Engineering (2014)
98(9): 625-662. DOI: 10.1002/nme.4668
[6] Matthies, H.G., Meyer, M., Nonlinear Galerkin methods for the model reduction of nonlinear
dynamical systems, Computers & Structures (2003) 81(9): 1277-1286.
[7] Farhat, C., Chapman, T., Avery, P. Structure-preserving, stability, and accuracy properties
of the energy-conserving sampling and weighting method for the hyper reduction of nonlinear
finite element dynamic models, International Journal for Numerical Methods in Engineering
(2015) 102(5): 1077-1110. DOI: 10.1002/nme.4820
[8] Jain, S., Tiso, P., Haller, G., Exact Nonlinear Model Reduction for a von Karman beam: SlowFast Decomposition and Spectral Submanifolds, submitted. arXiv:1703.03001v1 (2017)
[9] Daniel Millán, D., Arroyo, M. Nonlinear manifold learning for model reduction in finite elastodynamics, Computer Methods in Applied Mechanics and Engineering (2013) 261: 118-131
[10] Peharz, R., Pernkopf, F. Sparse nonnegative matrix factorization with `0 -constraints. Neurocomputing (2012); 80:38–46. DOI: 10.1016/j.neucom.2011.09.024
[11] Jain, S., Tiso, P., Rixen, D.J., Rutzmoser, J.B. A Quadratic Manifold for Model Order Reduction of Nonlinear Structural Dynamics, Computers & Structures (2017), 188: 80-94. DOI:
10.1016/j.compstruc.2017.04.005
[12] Rutzmoser, J.B., Rixen, D.J., Tiso, P., Jain, S. Generalization of Quadratic Manifolds for
Reduced Order Modeling of Nonlinear Structural Dynamics,Computers & Structures (2017),
192: 196-209. DOI: 10.1016/j.compstruc.2017.06.00
18
[13] Lines, L.R., Treitel, S. A review of least-squares inversion and its application to geophysical problems, Geophysical Prospecting (1984) 32(2): 159–186. DOI: 10.1111/j.13652478.1984.tb00726.x
[14] van der Maaten, L., Postma, E., van den Herik, J. Dimensionality Reduction: A Comparative
Review, TiCC TR 2009?005(2009), Tilburg University
[15] Ryckelynck D., A priori hyperreduction method: an adaptive approach, J. Comput. Phys.
(2005) 201(1): 346–366. DOI: 10.1016/j.jcp.2004.07.015
[16] Jain, S., Model Order Reduction for Non-linear Structural Dynamics, Master Thesis (2015),
Delft University of Technology. UUID: cb1d7058-2cfa-439a-bb2f-22a6b0e5bb2a
[17] Mignolet, M.P., Przekop, A., Rizzi, S.A., Spottswood, S.M. A review of indirect/non-intrusive
reduced order modeling of nonlinear geometric structures. J Sound Vib (2013) 332(10): 24372460. DOI: 10.1016/j.jsv.2012.10.017
[18] Barbič, J., James, D. L. Real-Time Subspace Integration for St. Venant-Kirchhoff Deformable
Models. ACM Trans. Graph. (2005) 24(3): 982–990. DOI: 10.1145/1073204.1073300
[19] Everson, R., Sirovich L. Karhunen-Loeve procedure for gappy data. Journal of the Optical
Society of America A (1995) 12(8):1657?1664. DOI: 10.1364/JOSAA.12.001657
[20] Haller, G. and Ponsioen, S. Exact model reduction by a slow?fast decomposition of nonlinear
mechanical systems, Nonlinear Dyn. (2017) 90: 617-647. DOI: 10.1007/s11071-017-3685-9
[21] Carlberg, K. Bou-Mosleh and Farhat, C. Efficient non-linear model reduction via a least-squares
Petrov?Galerkin projection and compressive tensor approximations, Int. J. Numer. Meth. Engng
(2011); 86: 155-181. DOI: 10.1002/nme.3050
[22] Carlberg, K., Farhat, C., Cortial, J. and Amsallem D. The GNAT method for nonlinear model
reduction: Effective implementation and application to computational fluid dynamics and turbulent flows, J. Comput. Phys. (2013), 242:623?647. DOI: 10.1016/j.jcp.2013.02.028
[23] Chapman, T. , Avery, P., Collins, P. and Farhat, C. Accelerated mesh sampling for the hyper reduction of nonlinear computational models, Int. J. Numer. Meth. Engng (2017), 109:1623–1654.
DOI: 10.1002/nme.5332
[24] Rega, G. and Troger, H. Dimension Reduction of Dynamical Systems: Methods, Models, Applications, Nonlinear Dyn. (2005), 41: 1?15. DOI: 10.1007/s11071-005-2790-3
[25] Ding, Q. and Zhang, K. Order reduction and nonlinear behaviors of a continuous rotor system,
Nonlinear Dyn. (2012), 67: 251?262. DOI: 10.1007/s11071-011-9975-8
19
| 5cs.CE
|
1
Student’s t Distribution based Estimation of
Distribution Algorithms for Derivative-free Global
Optimization
arXiv:1608.03757v2 [cs.NE] 25 Nov 2016
∗ Bin
Liu Member, IEEE, Shi Cheng Member, IEEE, Yuhui Shi Fellow, IEEE
Abstract
In this paper, we are concerned with a branch of evolutionary algorithms termed estimation of distribution (EDA), which
has been successfully used to tackle derivative-free global optimization problems. For existent EDA algorithms, it is a common
practice to use a Gaussian distribution or a mixture of Gaussian components to represent the statistical property of available
promising solutions found so far. Observing that the Student’s t distribution has heavier and longer tails than the Gaussian, which
may be beneficial for exploring the solution space, we propose a novel EDA algorithm termed ESTDA, in which the Student’s t
distribution, rather than Gaussian, is employed. To address hard multimodal and deceptive problems, we extend ESTDA further
by substituting a single Student’s t distribution with a mixture of Student’s t distributions. The resulting algorithm is named
as estimation of mixture of Student’s t distribution algorithm (EMSTDA). Both ESTDA and EMSTDA are evaluated through
extensive and in-depth numerical experiments using over a dozen of benchmark objective functions. Empirical results demonstrate
that the proposed algorithms provide remarkably better performance than their Gaussian counterparts.
Index Terms
derivative-free global optimization, EDA, estimation of distribution, Expectation-Maximization, mixture model, Student’s t
distribution
I. I NTRODUCTION
The estimation of distribution algorithm (EDA) is an evolutionary computation (EC) paradigm for tackling derivative-free
global optimization problems [1]. EDAs have gained great success and attracted increasing attentions during the last decade
[2]. Different from most of EC algorithms which are normally equipped with meta-heuristics inspired selection and variation
operations, the EDA paradigm is characterized by its unique variation operation which uses probabilistic models to lead the
search towards promising areas of the solution space. The basic assumption is that probabilistic models can be used for learning
useful information of the search space from a set of solutions that have already been inspected from the problem structure and
this information can be used to conduct more effective search. The EDA paradigm provides a mechanism to take advantage of
the correlation structure to drive the search more efficiently. Henceforth, EDAs are able to provide insights about the structure of
B. Liu is with School of Computer Science and Technology, Nanjing University of Posts and Telecommunications, Nanjing, Jiangsu, 210023 China. S. Cheng
is with School of Computer Science, Shaanxi Normal University, Xi’an, 710062 China. Y. Shi is with Department of Computer Science and Engineering,
Southern University of Science and Technology, Shenzhen, 518055 China.
∗ Correspondence author. E-mail: [email protected]
Manuscript received XX. X, 2016; revised XX X, 2016.
2
the search space. For most existent EC algorithms, their notions of variation and selection usually arise from the perspective of
Darwinian evolution. These notions offer a direct challenge for quantitative analysis by the existing methods of computer science.
In contrast with such meta-heuristics based EC algorithms, EDAs are more theoretically appealing, since a wealth of tools and
theories arising from statistics or machine learning can be used for EDA analysis. Convergence results on a class of EDAs have
been given in [3].
All existent EDAs fall within the paradigm of model based or learning guided optimization [2]. In this regard, different EDAs
can be distinguished by the class of probabilistic models used. An investigation on the boundaries of effectiveness of EDAs shows
that the limits of EDAs are mainly imposed by the probabilistic model they rely on. The more complex the model, the greater
ability it offers to capture possible interactions among the variables of the problem. However, increasing the complexity of the
model usually leads to more computational cost. The EDA paradigm could admit any type of probabilistic model, among which
the most popular model class is the Gaussian model. The simplest EDA algorithm just employs univariate Gaussian models,
which regard all design variables to be independent with each other [4, 5]. The simplicity of models makes such algorithms
easy to implement, while their effectiveness halts when the design variables have strong interdependencies. To get around of this
limitation, several multivariate Gaussian based EDAs (Gaussian-EDAs) have been proposed [4, 6, 7]. To represent variable linkages
elaborately, Bayesian networks (BNs) are usually adopted in the framework of multivariate Gaussian-EDA, while the learning of
the BN structure and parameters can be very time consuming [8, 9]. To handle hard multimodal and deceptive problems in a more
appropriate way, Gaussian mixture model (GMM) based EDAs (GMM-EDAs) have also been proposed [10, 11]. It is worthwhile
to mention that the more complex the model used, the more likely it encounters model overfitting, which may mislead the search
procedure [12].
The Student’s t distribution is symmetric and bell-shaped, like the Gaussian distribution, but has heavier and longer tails,
meaning that it is more prone to producing values that fall far from its mean. Its heavier tail property renders the Student’s t
distribution a desirable choice to design Bayesian simulation techniques such as the adaptive importance sampling (AIS) algorithms
[13–15]. Despite the purpose of AIS is totally different as compared with EDAs, AIS also falls within the iterative learning and
sampling paradigm like EDAs. The Student’s t distribution has also been widely used for robust statistical modeling [16]. Despite
of great success achieved in fields of Bayesian simulation and robust modeling, the Student’s t distribution has not yet gained any
attention from the EC community.
In this paper, we are concerned with whether the heavier tail property of the Student’s t distribution is beneficial for developing
more efficient EDAs. To this end, we derive two Student’s t distribution based algorithms within the paradigm of EDA. They
are entitled as estimation of Student’s t distribution algorithm (ESTDA) and estimation of mixture of Student’s t distribution
algorithm (EMSTDA), implying the usages of the Student’s t distribution and the mixture of Student’s t distributions, respectively.
We evaluate both algorithms using over a dozen of benchmark objective functions. The empirical results demonstrate that they
indeed have better performance than their Gaussian counterparts. In literature, the closest to our work here are the posterior
exploration based Sequential Monte Carlo (PE-SMC) algorithm [13] and the annealed adaptive mixture importance sampling
algorithm [17], while both were developed within the framework of Sequential Monte Carlo (SMC), other than EDA. In addition,
the faster Evolutionary programming (FEP) algorithm proposed in [18] is similar in the spirit of replacing the Gaussian distribution
with an alternative that has heavier tails. The major difference between FEP and our work here lies in that the former falls within
the framework of evolutionary strategy while the latter is developed within the EDA paradigm. In addition, the former employs
the Cauchy distribution, other than the student’s t or mixture of student’s t distributions here, to generate new individuals. To the
best of our knowledge, we make the first attempt here to bring Student’s t probabilistic models into the field of population based
EC.
The remainder of this paper is organized as follows. In Section II we revisit the Gaussian-EDAs. Besides the single Gaussian
distribution model based EDA, the GMM-EDA is also presented. In Section III we describe the proposed Student’s t model based
3
EDAs in detail. Like in Section II, both the single distribution and the mixture distribution model based EDAs are presented. In
Section IV, we present the performance evaluation results and finally in Section V, we conclude the paper.
II. G AUSSIAN
DISTRIBUTION BASED
EDA
In this section, we present a brief overview on the EDA algorithmic paradigm, the Gaussian-EDA and the GMM-EDA.
In this paper, we are concerned with the following derivative-free continuous global optimization problem
minf (x)
x∈χ
(1)
where χ denotes the nonempty solution space defined in Rn , and f : χ → R is a continuous real-valued function. The basic
assumption here is that f is bounded on χ, which means ∃fl > −∞, fu < ∞ such that fl ≤ f (x) ≤ fu , ∀x ∈ χ. We denote
the minimal function value as f ∗ , i.e., there exists an x∗ such that f (x) ≥ f (x∗ ) = f ∗ , ∀x ∈ χ.
A. Generic EDA scheme
A generic EDA scheme to solve the above optimization problem is presented in Algorithm 1. In this paper, we only focus on
the probabilistic modeling part in the EDA scheme, not intend to discuss the selection operations. In literature, the tournament
or truncation selection are commonly used with the EDA algorithm. Here we adopt truncation selection for all algorithms under
consideration. Specifically, we select, from {xi }N
i=1 , M individuals which produce the minimum objective function values. The
stopping criteria of Algorithm 1 is specified by the user based on the available budget, corresponding to the acceptable maximum
number of iterations here.
Algorithm 1: A main scheme of EDA
Input: the probabilistic model p(x|θ) with parameter θ, the population size N , the selection size M , which satisfies
M < N , and θ1 , the initialized value of θ.
Output: x̂ and f (x̂), which represent estimates of x∗ and f (x∗ ), respectively.
1 Initialization: Set θ = θ1 and then draw a random sample x̂ from p(·|θ = θ1 ). Set the iteration index k = 1;
2 while the stopping criterion is not met do
3
Sample N individuals from p(x|θk ), denote them by x1 , x2 , . . . , xN ;
4
Calculate the objective function values of the individualsyi = f (xi ), i = 1, 2, . . . , N ;
5
Find the minimum of {y1 , . . . , yN }, denote it by ymin and then find from {xi }N
i=1 the corresponding xmin which
produces ymin ;
6
if ymin < f (x̂) then
7
Let x̂ = xmin and f (x̂) = ymin ;
8
9
N
Select M individuals from {xi }N
i=1 based on the information {xi , yi }i=1 ;
Update θk−1 , based on the selected individuals, to get θk . See Subsections II-B, II-C, III-B and III-C for details on
specific operations with respect to the Gaussian model, GMM, the Student’s t and the mixture of Student’s t
models, respectively. Set k = k + 1.
B. Gaussian-EDA
In Gaussian-EDAs, the probabilistic model p(x|θ) in Algorithm 1 takes the form of Gaussian, namely p(x|θ) = N (x|µ, Σ),
where θ , {µ, Σ}, µ and Σ denote the mean and covariance of this Gaussian distribution, respectively.
As mentioned before, we employ a truncation selection strategy, which indicates that the selected individuals are those which
produce minimum objective function values. Given the selected individuals {xi , yi }M
j=1 , a maximum likelihood (ML) estimate of
4
the parameters of the Gaussian distribution is shown to be:
µ
Σ
where
T
=
=
PM
j=1
M
PM
xj
(2)
j=1 (xj
− µ)(xj − µ)
M −1
T
,
(3)
denotes transposition and all vectors involved are assumed to be column vectors both here and hereafter.
C. GMM-EDA
As its name indicates, the GMM-EDA uses a GMM to represent p(x|θ), namely p(x|θ) =
θ,
PL
{wl , µl , Σl }L
l=1 ,
l=1
PL
l=1
wl N (x|µl , Σl ), where
l is the index of the mixing components, w the probability mass of the mixing component which satisfies
wl = 1, and L denotes the total number of mixing components included in the mixture and is preset as a constant.
Now we focus on how to update θk−1 to get θk based on the selected individuals {xj }M
j=1 . We use an expectation-maximization
(EM) procedure, as shown in Algorithm 2, to update the parameters of the mixture model. For more details on EM based parameter
estimation for GMM, readers can refer to [19–22]. The stopping criterion can be determined by checking the changing rate of the
GMM parameters. If θnew has not been changed a lot for a fixed number of successive iterations, then we terminate the iterative
process of the EM procedure. Compared with a benchmark EM procedure, here we add a components deletion operation at the
end of the Expectation step. We delete components with extremely small mixing weights, because their roles in the mixture are
negligible, while at the Maximization step, they will occupy the same computing burden as the rest of components. This deletion
operation is also beneficial for avoiding disturbing numerical issues such as singularity of covariance matrix. For the influence of
this operation on the the number of survival mixing components, see Section IV-B.
III. S TUDENT ’ S T
DISTRIBUTION BASED
EDA
In this section, we describe the proposed ESTDA and EMSTDA in detail. Both of them are derived based on the application
of the Student’s t distribution in the EDA framework as presented in Algorithm 1. To begin with, we give a brief introduction to
the Student’s t distribution.
A. Student’s t distribution
Suppose that x is a d dimensional random variable that follows the multivariate Student’s t distribution, denoted by S(·|µ, Σ, v),
where µ denotes the mean, Σ a positive definite inner product matrix and v ∈ (0, ∞] is the degrees of freedom (DoF). Then the
density function of x is:
S(x|µ, Σ, v) =
)|Σ|−0.5
Γ( v+d
2
,
+ Md (x, µ, Σ)/v}0.5(v+d)
(πv)0.5d Γ( v2 ){1
(8)
where
Md (x, µ, Σ) = (x − µ)T Σ−1 (x − µ)
(9)
denotes the Mahalanobis squared distance from x to µ with respect to Σ, A−1 denotes the inverse of A and Γ(·) denotes the
gamma function. For ease of understanding, we give a graphical description of univariate Student’s t pdfs corresponding to
different DoFs in comparison with a standard Gaussian pdf in Fig.1. It is shown that the Student’s t pdfs have heavier tails than
the Gaussian pdf. The smaller the DoF v, the heavier the corresponding pdf’s tails. From Fig.1, we see that the Student’s t pdf
is more likely to generate an individual further away from its mean than the Gaussian model due to its long flat tails. It leads to
a higher probability of escaping from a local optimum or moving away from a plateau. We can also see that heavier tails lead
to a smaller hill around the mean, and vice versa. A smaller hill around the mean indicates that the corresponding model tuning
5
Algorithm 2: EM procedure to estimate parameters of the GMM within the EDA paradigm as presented in Algorithm 1
1
2
3
L
Input: the individuals {xj }M
j=1 , the model’s current parameter value θk−1 = {wk−1,l , µk−1,l , Σk−1,l }l=1 and the
admittable smallest weight W of a mixing component (we set W = 0.02 as a default value for use in our
experiments)
Output: θk
Initialization: Set θold = θk−1 . Set the iteration index i = 1;
while the stopping criterion is not met do
The Expectation step:
M
X
ǫ(l|xj )/M, l = 1, . . . , L,
wnew,l =
(4)
j=1
where ǫ(l|xj ) denotes the probability of the event individual xj belonging to the lth mixing component, which is
calculated as follows
wold,l N (xj |µold,l , Σold,l )
ǫ(l|xj ) = PL
.
(5)
l=1 wold,l N (xj |µold,l , Σold,l )
4
Delete mixing components whose mixing weights are smaller than W . Then update the value of L and increase the
mixing weights of the remaining mixing components proportionally to guarantee that their summation equals 1;
The Maximization step:
µnew,l
=
M
X
ǫ(l|xj )xj /M,
l = 1, . . . , L,
(6)
j=1
Σnew,l
=
M
X
ǫ(l|xj )(xj − µnew,l )(xj − µnew,l )T /M,
l = 1, . . . , L.
(7)
j=1
5
6
Set θnew = {wnew,l , µnew,l , Σnew,l }L
l=1 ;
Set θold = θnew and let k = k + 1.
Set θk = θnew ;
operation spends less time in exploiting the local neighborhood and thus has a weaker fine-tuning ability. Hence heavier tails
do not always bring advantages. This has been demonstrated by our numerical tests. See Fig.2d, which shows that for several
two-dimensional (2D) benchmark test problems, lighter tails associated with v = 50 provide better convergence result than heavier
tails corresponding to v = 5 and 10.
B. ESTDA
In ESTDA, the probabilistic model p(x|θ) in Algorithm 1 is specified to be a student’s t distribution, namely p(x|θ) =
S(x|µ, Σ, v), where θ , {µ, Σ}, and v is specified beforehand as a constant.
First, let’s figure out, given the parameter value θ, how to sample an individual x from p(·|θ). It consists of two steps that
is simulating a random draw τ from the gamma distribution, whose shape and scale parameters are set identically to v/2, and
sampling x from the Gaussian distribution N (·|µ, Σ/τ ).
Now let’s focus on, given a set of selected individuals {xj }M
j=1 , how to update the parameters of the student’s t distribution
in a optimal manner in terms of maximizing likelihood. As mentioned above, in generation of {xj }M
j=1 , a corresponding set of
gamma distributed variables {τj }M
j=1 is used. We can record these gamma variables and then use them to easily derive an ML
estimate for parameters of the Student’s t distribution [23, 24]:
µ
=
Σ
=
PM
j=1
PM
τj x j
i=1 τi
PM
j=1 τj (xj
− µ)(xj − µ)T
.
PM
i=1 τi
(10)
(11)
6
0.4
Gaussian
Student (v=1)
Student (v=3)
Student (v=10)
0.35
0.3
pdf
0.25
0.2
0.15
0.1
0.05
0
−6
−4
−2
0
x
2
4
6
Fig. 1: Comparison between the univariate Gaussian and Student’s t distributions (represented by “Student” in the figure). All
distributions have zero mean with a fixed variance 1. v denotes the DoF.
C. EMSTDA
The EMSTDA employs a mixture of student’s t distributions to play the role of the probabilistic model p(x|θ) in Algorithm
PL
PL
L
1. Now we have p(x|θ) =
l=1 wl S(x|µl , Σl , v), where θ , {wl , µl , Σl }l=1 ,
l=1 wl = 1, and v is specified beforehand
as a constant. Given the selected individuals {xj }M
j=1 , we resort to the EM method to update parameter values of the mixture
of Student’s t distributions. The specific operations are presented in Algorithm 3. EM methods based parameter estimation for
the mixture of Student’s t model can also be found in [13–15], where the mixture learning procedure is performed based on
a weighted sample set. The EM operations presented in Algorithm 3 can be regarded as an application of the EM methods
described in [13–15] to an equally weighted sample set. Similarly as in Algorithm 2, we add a components deletion operation
at the end of Expectation step. This deletion operation were used in the adaptive annealed importance sampling [14] algorithm
and the PE-SMC algorithm [13], which also adopt the EM procedure to do parameter estimation for a Student’s t mixture model.
By deleting components with extremely small mixing weights, we can avoid consuming computing burdens to update parameter
values for negligible components. As presented before in Subsection II-C, this deletion operation is also useful to avoid common
numerical issues such as singularity of covariance matrix. For the influence of this operation on the the number of survival mixing
components, see Subsection IV-B.
IV. P ERFORMANCE
EVALUATION
In this section, we evaluate the presented ESTDA and EMSTDA using a number of benchmark objective functions, which are
designed and commonly used in literature for testing and comparing optimization algorithms. See the Appendix Section for details
on the involved functions. The Gaussian-EDA and the GMM-EDA, which are described in Sections II-B and II-C, respectively,
are involved for comparison purpose.
Although scaling up EDAs to large scale problems has become one of the biggest challenges of the field [25], it is not our
7
Algorithm 3: EM procedure to estimate parameters of a mixture of student’s t model within the EDA paradigm as presented
in Algorithm 1
1
2
3
L
Input: the individuals {xj }M
j=1 , the model’s current parameter value θk−1 = {wk−1,l , µk−1,l , Σk−1,l }l=1 and the
admittable smallest weight W of a mixing component (we set W = 0.02 as a default value for use in our
experiments)
Output: θk
Initialization: Set θold = θk−1 . Set the iteration index i = 1;
while the stopping criterion is not met do
The Expectation step:
M
X
ǫ(l|xj )/M, l = 1, . . . , L,
wnew,l =
(12)
j=1
where ǫ(l|xj ) denotes the probability of the event individual xj belonging to the lth mixing component, which is
calculated as follows
wold,l S(xj |µold,l , Σold,l , v)
ǫ(l|xj ) = PL
.
(13)
l=1 wold,l S(xj |µold,l , Σold,l , v)
4
5
6
Delete mixing components whose mixing weights are smaller than W . Then update the value of L and increase the
mixing weights of the remaining mixing components proportionally to guarantee that their summation equals 1;
The Maximization step:
PM
j=1 ǫ(l|xj )xj (v + d)/ (v + Md (xj , µold,l , Σold,l ))
, l = 1, . . . , L,
(14)
µnew,l =
PM
j=1 ǫ(l|xj )(v + d)/ (v + Md (xj , µold,l , Σold,l ))
PM
T
j=1 ǫ(l|xj )(xj − µnew,l )(xj − µnew,l ) (v + d)/ (v + Md (xj , µold,l , Σold,l ))
, l = 1, . . . , L. (15)
Σnew,l =
PM
j=1 ǫ(l|xj )
Set θnew = {wnew,l , µnew,l , Σnew,l }L
l=1 ;
Set θold = θnew and let k = k + 1.
Set θk = θnew ;
intention here to discuss and investigate the applications of ESTDA and EMSTDA in high dimensional problems. The goal of
this paper is to demonstrate the benefits resulted from the heavier tails of Student’s t distribution in exploring the solution space
and finding the global optimum through the EDA mechanism. Therefore we only consider cases with d ≤ 10 here for ease of
performance comparison.
A. Experimental study for ESTDA
As described in Subsection III-A, the degree of difference between the shapes of a Student’s t and a Gaussian distribution
mainly depends on the value of DoF v. This Subsection is dedicated to investigating the influence of v on the performance of
ESTDA through empirical studies.
We will test on six objective functions, including the Ackley, De Jong Function N.5, Easom, Rastrigin, Michalewicz and the
Lèvy N.13 functions. We consider the 2D cases (i.e., d=2) here. These functions have diversing characteristics in their shapes
and thus give representativeness to the results we will report. See the Appendix for mathematical definitions of these functions.
We consider four ESTDAs, corresponding to different DoFs v = 5, 10, 50, 100, 500, respectively. The sample size and selection
size are fixed to be N = 1000 and M = 200 for all ESTDAs under consideration. A basic Gaussian-EDA is involved, acting as
a benchmark algorithm for comparison. We run each algorithm 30 times independently and record the best solutions it obtained
at the end of each iteration. We plot the means of the best fitness function values obtained over these 30 runs in Fig.2. It is
shown that the ESTDA (v = 5), performs strikingly better than the other candidate algorithms for the Ackley, De Jong Function
N.5 and Easom Functions. The ESTDA (v = 50) and ESTDA (v = 500) outperform the others significantly in handling the
8
Rastrigin function; and, for the remaining Michalewicz and Lèvy N.13 functions, all the algorithms converge to the optima at a
very similar rate. To summarize, the ESTDA outperforms the Gaussian-EDA markedly for four problems and performs similarly
as the Gaussian-EDA for the other two problems considered here. This result demonstrate that, the heavier tails of the Student’s t
distribution can facilitate the process of exploring the solution space and thus constitute a beneficial factor for the EDA algorithms
to find the global optimum faster.
2
10
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
−300
0
−10
1.3
10
10
−2
−250
mean of f(x)
−6
10
−8
10
−10
1.29
10
−200
−10
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
1.28
10
−10
10
1.27
10
mean of f(x)
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
−4
10
mean of f(x)
10
−150
−10
−100
−10
−12
10
−50
−10
1.26
−14
10
10
−16
10
0
0
5
10
15
20
25
30
iteration index
35
40
45
50
0
5
10
15
−10
0
5
10
iteration index
(a) Ackley Function
(b) De Jong Function N.5
15
iteration index
20
25
30
(c) Easom Function
2
0
10
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
0
10
−2
10
10
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
0.23
−10
ESTDA (v=5)
ESTDA (v=10)
ESTDA (v=50)
ESTDA (v=100)
ESTDA (v=500)
Gaussian−EDA
−2
10
−4
10
−4
−6
10
0.24
mean of f(x)
mean of f(x)
mean of f(x)
10
−10
−6
10
−8
10
−8
10
−10
10
−10
10
0.25
−10
−12
10
−12
10
−14
10
−14
0
5
10
15
20
25
30
iteration index
35
(d) Rastrigin Function
40
45
50
0
5
10
15
10
0
2
4
iteration index
(e) Michalewicz Function
6
8
10
iteration index
12
14
16
18
(f) Lèvy N.13 Function
Fig. 2: Convergence behaviors of ESTDAs corresponding to different DoFs and a Gaussian-EDA. The Y coordinate denotes
the averaged objective function values of the best solutions obtained from 30 independent runs of each algorithm. All test
functions involved here are 2D
B. A more thorough performance evaluation for ESTDA and EMSTDA
In this subsection, we perform a more thorough performance evaluation for ESTDA and EMSTDA. Both Gassuian-EDA and
GMM-EDA are also involved for comparison purpose. Including GMM-EDA we hope to further demonstrate that the heavier tails
of the Student’s t distribution are beneficial for exploration of solution space via the EDA mechanism. By contrasting EMSTDA
with ESTDA, we can inspect how much a mixture model can contribute to search the solution space via a Student’s t based EDA
mechanism.
To make a representative evaluation, we will consider in total 17 objective functions listed in Table I. For all but the Rastrigin
and Michalewicz functions, we consider 2D cases. For the Rastrigin and Michalewicz functions, we also consider 5D and 10D
cases. The population size per iteration N is set to be 1e3, 1e4, and 1e5 for 2D, 5D, 10D cases, respectively. The selection size
M is fixed to be 0.2 × N for all cases involved. We didn’t consider higher dimensional cases, for which a reliable probabilistic
model building requires far more samples, while it is impractical to carry out so large numbers of samples in reasonable time
scales. Definitions of all the involved objective functions are presented in the Appendix of the paper.
We fix the value of the DoF v to be 5 for all test functions, except that, for problems involving the Rastrigin function, v is
set to be 50, since it gives better convergence behavior as reported in subsection IV-A. For every algorithm, the iterative EDA
procedure terminates when the iteration number k in Algorithm 1 is bigger than 50. For EMSTDA and GMM-EDA, the iterative
EM procedure terminates when the iteration number i in Algorithms 2 and 3 is bigger than 2. Each algorithm is run 30 times
independently. Then we calculate the average and standard error of the converged fitness values. The results are given in Table I.
For each test function, the best solution is marked with bold font. Then we count for each algorithm how many times it output
9
TABLE I: Convergence results yielded from 30 independent runs of each algorithm on each test problem. a ± b in the table
denotes that the average and standard error of the best fitness values obtained from 30 independent runs of the corresponding
algorithm are a and b, respectively. The best solution for each problem is marked with bold font.
Test Problems
Ackley
2D
Dejong N.5
2D
Easom
2D
Rastrigin
2D
5D
10D
Michalewicz 2D
5D
10D
Lèvy N.13
2D
Cross-in-tray 2D
Drop-wave
2D
Eggholder
2D
Griewank
2D
Holder table
2D
Lèvy
2D
Schaffer
2D
Schwefel
2D
Shubert
2D
Perm
2D
Rosenbrock
2D
Goal:f (x⋆ )
0
1
-1
0
0
0
-1.8013
-4.6877
-9.66015
0
-2.0626
-1
-959.6407
0
-19.2085
0
0
0
-186.7309
0
0
ESTDA
0±0
18.1207±1.8130
-0.9330±0.2536
0±0
0±2.13 × 10−12
0.0383±0.0447
-1.8013±0
-4.6877±9.36 × 10−9
-9.5384±0.0475
0±0
-2.0626±0
-0.9884±0.0129
-588.9196±75.2446
19.5764±3.7905
-19.0835±0.1918
0±0
0±1.7064 × 10−6
368.3134±75.4837
-186.7309±4.05 × 10−13
0±0
0.0420±0.0418
EMSTDA
0.0128±0.0701
3.2370±2.6410
-0.9587±0.1862
0.0050±0.0202
0.6562±0.7985
0.5383±0.5980
-1.8013±0
-4.6404±0.0561
-9.4226±0.2107
0.0014±0.0053
-2.0626±0
-0.9909±0.0125
-731.8013±145.5383
1.5197±5.5576
-19.2085±0
0±0
0.0001±4.5280 × 10−4
184.2835±118.4596
-186.7309±1.64 × 10−13
0±0
0.0036±0.0129
Gaussian-EDA
3.1815±1.3004
19.6536±1.1683
0.2262±0.4186
0±0
0±1.70 × 10−11
0.0396±0.0272
-1.8013±0
-4.6500±0.0099
-9.1047±0.1353
0±0
-2.0626±0
-0.9990±0.0010
-560.7251±4.2080
30.4232±1.0396
-19.1860±0.0821
0±0
0±0
436.8232±2.5521
-186.7309±1.38 × 10−4
0±0
0.0477±0.0558
GMM-EDA
1.5332±1.8186
6.4677±6.3530
-0.3153±0.4589
0.0182±0.0643
0.3575±0.6293
0.5341±0.8317
-1.8013±0
-4.6459±0.0175
-9.1426±0.1415
0.0043±0.0179
-2.0626±0
-0.9938±0.0097
-686.5236±147.5427
15.7949±16.3001
-19.2085±0
0±1.0605 × 10−4
0±2.0208 × 10−4
247.0598±123.2987
-186.7309±2.48 × 10−5
0±4.0029 × 10−6
0.0151±0.0382
a better solution as compared with all the rest of algorithms and put the result into Table II. We see that the EMSTDA get the
highest score 7, followed by score 5 obtained by ESTDA, while their Gaussian counterparts only obtain relatively lower scores
0 and 2, respectively. This result further coincides with our argument that the Student’s t distribution is more preferable than the
Gaussian for use in designing EDAs. We also see that, between ESTDA and EMSTDA, the latter performs better than the former.
This demonstrates that using mixture models within a Student’s t based EDA is likely to be able to bring additional advantages.
What slightly surprises us in Table II is that the GMM-EDA seems to perform worse than the Gaussian-EDA, while after a
careful inspection on Table I, we see that GMM-EDA beats Gaussian-EDA strikingly in cases involving the Ackley, Dejong N.5,
Easom, Michalewicz 10D, Eggholder, Griewank, Holder table, Schwefel and Rosenbrock functions, while for cases involving
the Michalewicz 2D, Lèvy N.13, Cross-in-tray, Drop-wave, Lèvy, Schaffer, Shubert and Perm functions, GMM-EDA performs
identically or similarly as the Gaussian-EDA. Hence, in summarize, GMM-EDA actually performs better than Gaussian-EDA in
this experiment.
To investigate the effect of the component deletion operator used in Algorithms 3 and 2, we record the number of survival
mixing components per iteration of the ESTDA and EMSTDA, and calculate its mean averaged over 30 independent runs of each
algorithm. The result for six typical objective functions is depicted in Fig. 3. We see that, for both algorithms, the mean number
of survival components decreases as the algorithm iterates on. This phenomenon suggests that the mixture model mainly takes
effect during the early exploration stage, while, when the algorithm goes into the latter exploitation phase, there usually remains
much less components in the mixture model.
TABLE II: A simple statistics made for the results presented in Table I. Each figure in this table represents the number of times
the corresponding algorithm have outputted a better solution than all the rest of competitor algorithms under consideration.
Note that in some cases, e.g, the Cross-in-tray function case, more than one algorithms provide an identical best solution. For
such cases, no algorithm will get a score in the statistics made here.
ESTDA
5
EMSTDA
7
Gaussian-EDA
2
GMM-EDA
0
10
3.5
3.5
4
EMSTDA
GMM−EDA
EMSTDA
GMM−EDA
EMSTDA
GMM−EDA
3.5
3
3
2.5
2.5
mean of L
mean of L
mean of L
3
2
2
1.5
1.5
2.5
2
1.5
1
1
0
5
10
15
20
25
iteration index
30
35
40
45
0
5
10
15
20
1
25
iteration index
(a) Ackley Function
0
5
(b) De Jong Function N.5
4
4
EMSTDA
GMM−EDA
15
20
iteration index
25
30
35
(c) Easom Function
4
EMSTDA
GMM−EDA
3.9
3.5
10
EMSTDA
GMM−EDA
3.5
3.8
3
2.5
3.7
mean of L
mean of L
mean of L
3
3.6
2
2.5
2
3.5
1.5
1
1.5
3.4
0
5
10
15
20
iteration index
25
(d) Rastrigin Function
30
35
3.3
0
5
10
15
20
25
1
0
5
10
iteration index
(e) Michalewicz Function
15
20
iteration index
25
30
35
(f) Lèvy N.13 Function
Fig. 3: Mean number of survival mixing components in the mixture model.
V. C ONCLUDING
REMARKS
In this paper we introduced the Student’s t distribution, for the first time, into a generic paradigm of EDAs. Specifically,
we developed two Student’s t based EDAs, which we term as ESTDA and EMSTDA here. The justification for the proposed
algorithms lies in the assumption that the heavier tails of the Student’s t distribution are beneficial for an EDA type algorithm
to explore the solution space. Experimental results demonstrate that for most cases under consideration, the Student’s t based
EDAs indeed performed remarkably better than their Gaussian counterparts implemented in the same algorithm framework. As
a byproduct of our experimental study, we also demonstrated that mixture models can bring in additional advantages, for both
Student’s t and Gaussian based EDAs, in tackling global optimization problems.
All algorithms developed here are based on density models that use full covariance matrices. The reason is two-fold. One is
for the ease of presentation of the basic idea, namely, replacing the commonly used Gaussian models by heavier tailed Student’s
t models in EDAs. The other is for conducting a completely fair performance comparison, since, if we combine the differing
state-of-art techniques into the algorithms used for comparison, it may become difficult for us to judge whether a performance gain
or loss is caused by replacing Gaussian with the Student’s t model or the other state-of-art techniques. Since the full covariance
model based EDA framework is so basic that many state-of-art techniques can be taken as special cases of it, we argue that the
ESTDA and EMSTDA algorithms presented here can be generalized by combining other techniques, such as factorizations of the
covariance matrices, dimension reduction by separating out weakly dependent dimensions or random embedding, other divide and
conquer or problem decomposition approaches, to handle specific optimization problems.
APPENDIX: B ENCHMARK
OBJECTIVE FUNCTIONS USED IN THIS PAPER
Here we introduce the objective functions used in Section IV for testing algorithms designed to search global minimum. Note
that in this section, we use xi to denote the ith dimension of x for ease of exposition.
11
A. Ackley Function
This function is defined to be
v
u d
u1 X
f (x) = −a exp −bt
x2 − exp
d i=1 i
d
1X
cos(cxi )
d i=1
!
+ a + exp(1) ,
(16)
where a = 20, b = 0.2 and c = 2π. This function is characterized by a nearly flat outer region, and a large peak at the centre,
thus it poses a risk for optimization algorithms to be trapped in one of its many local minima. The global minima is f (x⋆ ) = 0,
at x⋆ = (0, . . . , 0).
B. De Jong Function N.5
The fifth function of De Jong is multimodal, with very sharp drops on a mainly flat surface. The function is evaluated on the
square xi ∈ [−65.536, 65.536], for all i = 1, 2, as follows
f (x) =
0.002 +
2
X
5
i=1
1
i + (x1 − a1i )6 + (x2 − a2i )6
!−1
,
(17)
where
a
=
−32
−16
0
16
32
−32
9
16
32
−32
−32
−32
−32
−32
−16
32
32
32
The global minimum f (x⋆ ) ≈ 1 [18].
.
C. Easom Function
This function has several local minimum. It is unimodal, and the global minimum has a small area relative to the search
space. It is usually evaluated on the square xi ∈ [−100, 100], for all i = 1, 2, as follows
f (x) = cos(x1 ) cos(x2 ) exp −(x1 − π)2 − (x2 − π)2 .
(18)
The global minimum f (x⋆ ) = −1 is located at x⋆ = (π, π).
D. Rastrigin Function
This function is evaluated on the hypercube xi ∈ [−5.12, 5.12], for all i = 1, . . . , d, with the form
f (x) =
10d +
d
X
i=1
[x2i
!
− 10 cos(2πxi )] .
(19)
Its global minimum f (x⋆ ) = 0 is located at x⋆ = (0, . . . , 0). This function is highly multimodal, but locations of the local
minima are regularly distributed.
E. Michalewicz Function
This function has d! local minima, and it is multimodal. The parameter m defines the steepness of the valleys and ridges;
a larger m leads to a more difficult search. The recommended value of m is m = 10. It is usually evaluated on the hypercube
xi ∈ [0, π], for all i = 1, . . . , d, as follows
f (x) =
d
X
i=1
sin(xi ) sin2m
ix2i
π
.
(20)
12
The global minimum in 2D case is f (x⋆ ) = −1.8013 located at x⋆ = (2.20, 1.57), in 5D case is f (x⋆ ) = −4.687658 and in
10D case is f (x⋆ ) = −9.66015.
F. Lèvy Function N.13
This function is evaluated on the hypercube xi ∈ [−10, 10], for i = 1, 2, as below
f (x) = sin2 (3πx1 ) + (x1 − 1)2 [1 + sin2 (3πx2 )] + (x2 − 1)2 [1 + sin2 (2πx2 )] ,
(21)
with global minimum f (x⋆ ) = 0, at x⋆ = (1, . . . , 1).
G. Cross-in-tray Function
This function is defined to be
f (x) = −0.0001 (|g(x1 , x2 )| + 1)0.1 ,
where
g(x1 , x2 ) = sin(x1 ) sin(x2 ) exp
100 −
p
x21 + x22
π
(22)
!
.
It is evaluated on the square xi ∈ [−10, 10], for i = 1, 2, with f (x⋆ ) = −2.06261. This function has multiple global minima
located at x⋆ = (1.3491, −1.3491), (1.3491, 1.3491), (−1.3491, 1.3491) and (−1.3491, −1.3491).
H. Drop-wave Function
This is multimodal and highly complex function. It is evaluated on the square xi ∈ [−5.12, 5.12], for i = 1, 2, as follows
p
1 + cos(12 x21 + x22 )
f (x) = −
.
0.5(x21 + x22 ) + 2
(23)
The global minimum f (x⋆ ) = −1 is located at x⋆ = (0, 0).
I. Eggholder Function
This is a difficult function to optimize, because of a large number of local minima. It is evaluated on the square xi ∈
[−512, 512], for i = 1, 2, as follows
r
p
x1
+ 47 − x1 sin
|x1 − (x2 + 47)| .
f (x) = −(x2 + 47) sin
x2 +
2
(24)
The global minimum f (x⋆ ) = −959.6407 is located at x⋆ = (512, 404.2319).
J. Griewank Function
This function has many widespread local minima, which are regularly distributed. It is usually evaluated on the hypercube
xi ∈ [−600, 600], for all i = 1, . . . , d, as follows
f (x) =
!
d
d
Y
X
xi
x2i
−
cos √ + 1 .
4000 i=1
i
i=1
The global minimum f (x⋆ ) = 0 is located at x⋆ = (0, . . . , 0).
(25)
13
K. Holder table Function
This function has many local minimum, with four global one. It is evaluated on the square xi ∈ [−10, 10], for i = 1, 2, as
follows
f (x) = − sin(x1 ) cos(x2 ) exp
100 −
p
x21 + x22
π
!
.
(26)
Its global minimum f (x⋆ ) = −19.2085 is located at x⋆ = (8.05502, 9.66459), (8.05502, −9.66459), (−8.05502, 9.66459) and
(−8.05502, −9.66459).
L. Lèvy Function
This function is evaluated on the hypercube xi ∈ [−10, 10], for all i = 1, . . . , d, as follows
f (x) =
2
sin (πw1 ) +
d−1
X
i=1
where wi = 1 +
xi −1
,
4
2
2
2
2
!
(wi − 1) [1 + 10 sin (πwi + 1)] + (wd − 1) [1 + sin (2πwd )] .
(27)
for all i = 1, . . . , d. The global minimum f (x⋆ ) = 0 is located at x⋆ = (1, . . . , 1).
M. The second Schaffer Function
This function is usually evaluated on the square xi ∈ [−100, 100], for all i = 1, 2. It has a form as follows
f (x) =
0.5 +
sin2 (x21 − x22 ) − 0.5
[1 + 0.001(x21 + x22 )]2
.
(28)
Its global minimum f (x⋆ ) = 0 is located at x⋆ = (0, 0).
N. Schwefel Function
This function is also complex, with many local minima. It is evaluated on the hypercube xi ∈ [−500, 500], for all i = 1, . . . , d,
as follows
f (x) =
418.9829d −
d
X
xi sin(
i=1
p
!
|xi |) .
(29)
Its global minimum f (x⋆ ) = 0 is located at x⋆ = (420.9687, . . . , 420.9687).
O. Shubert Function
This function has several local minima and many global minima. It is usually evaluated on the square xi ∈ [−10, 10], for all
i = 1, 2.
f (x) =
5
X
i cos((i + 1)x1 + i)
i=1
!
5
X
!
i cos((i + 1)x2 + i) .
i=1
(30)
Its global minimum is f (x⋆ ) = −186.7309.
P. Perm Function 0,d,β
This function is evaluated on the hypercube xi ∈ [−d, d], for all i = 1, . . . , d, as follows
!2
d
d
X
X
1
.
(j + β) xij − i
f (x) =
j
j=1
i=1
Its global minimum f (x⋆ ) = 0 is located at x⋆ = (1, 21 , . . . , d1 ).
(31)
14
Q. Rosenbrock Function
This function is also referred to as the Valley or Banana function. The global minimum lies in a narrow, parabolic spike.
However, even though this spike is easy to find, convergence to the minimum is difficult [26]. It has the following form
f (x) =
d−1
X
i=1
[100(xi+1 − x2i )2 + (xi − 1)2 ],
(32)
and is evaluated on the hypercube xi ∈ [−5, 10], for all i = 1, . . . , d. The global minimum f (x⋆ ) = 0 is located at x⋆ =
(1, . . . , 1).
R EFERENCES
[1] M. Pelikan, D. E. Goldberg, and F. G. Lobo, “A survey of optimization by building and using probabilistic models,” Computational
optimization and applications, vol. 21, no. 1, pp. 5–20, 2002.
[2] M. Hauschild and M. Pelikan, “An introduction and survey of estimation of distribution algorithms,” Swarm and Evolutionary Computation,
vol. 1, no. 3, pp. 111–128, 2011.
[3] Q. Zhang and H. Muhlenbein, “On the convergence of a class of estimation of distribution algorithms,” IEEE Trans. on evolutionary
computation, vol. 8, no. 2, pp. 127–136, 2004.
[4] P. Larranaga and J. A. Lozano, Estimation of distribution algorithms: A new tool for evolutionary computation. Springer Science & Business
Media, 2002, vol. 2.
[5] M. Sebag and A. Ducoulombier, “Extending population-based incremental learning to continuous search spaces,” in International Conf. on
Parallel Problem Solving from Nature.
Springer, 1998, pp. 418–427.
[6] P. Larrañaga, R. Etxeberria, J. A. Lozano, and J. M. Peña, “Optimization in continuous domains by learning and simulation of gaussian
networks,” 2000.
[7] P. A. Bosman and D. Thierens, “Expanding from discrete to continuous estimation of distribution algorithms: The IDEA,” in International
Conference on Parallel Problem Solving from Nature.
Springer, 2000, pp. 767–776.
[8] W. Dong, T. Chen, P. Tiňo, and X. Yao, “Scaling up estimation of distribution algorithms for continuous optimization,” IEEE Trans. on
Evolutionary Computation, vol. 17, no. 6, pp. 797–822, 2013.
[9] Q. Zhang, J. Sun, E. Tsang, and J. Ford, “Hybrid estimation of distribution algorithm for global optimization,” Engineering computations,
vol. 21, no. 1, pp. 91–107, 2004.
[10] Q. Lu and X. Yao, “Clustering and learning gaussian distribution for continuous optimization,” IEEE Trans. on Systems, Man, and Cybernetics,
Part C (Applications and Reviews), vol. 35, no. 2, pp. 195–204, 2005.
[11] C. W. Ahn and R. S. Ramakrishna, “On the scalability of real-coded bayesian optimization algorithm,” IEEE Trans. on Evolutionary
Computation, vol. 12, no. 3, pp. 307–322, 2008.
[12] H. Wu and J. L. Shapiro, “Does overfitting affect performance in estimation of distribution algorithms,” in Proc. of the 8th annual conf. on
Genetic and evolutionary computation.
ACM, 2006, pp. 433–434.
[13] B. Liu, “Posterior exploration based sequential monte carlo for global optimization,” arXiv preprint arXiv:1509.08870, 2015.
[14] ——, “Adaptive annealed importance sampling for multimodal posterior exploration and model selection with application to extrasolar planet
detection,” The Astrophysical Journal Supplement Series, vol. 213, no. 1, p. 14, 2014.
[15] O. Cappé, R. Douc, A. Guillin, J.-M. Marin, and C. P. Robert, “Adaptive importance sampling in general mixture classes,” Statistics and
Computing, vol. 18, no. 4, pp. 447–459, 2008.
[16] K. L. Lange, R. J. Little, and J. M. Taylor, “Robust statistical modeling using the t distribution,” Journal of the American Statistical
Association, vol. 84, no. 408, pp. 881–896, 1989.
[17] B. Liu and C. Ji, “A general algorithm scheme mixing computational intelligence with bayesian simulation,” in Proc. of the 6th Int’l Conf.
on Advanced Computational Intelligence (ICACI).
IEEE, 2013, pp. 1–6.
[18] X. Yao, Y. Liu, and G. Lin, “Evolutionary programming made faster,” IEEE Trans. on Evolutionary Computation, vol. 3, no. 2, pp. 82–102,
1999.
15
[19] J. A. Bilmes et al., “A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaussian mixture and hidden
markov models,” International Computer Science Institute, vol. 4, no. 510, p. 126, 1998.
[20] L. Xu and M. I. Jordan, “On convergence properties of the EM algorithm for Gaussian mixtures,” Neural computation, vol. 8, no. 1, pp.
129–151, 1996.
[21] C. Biernacki, G. Celeux, and G. Govaert, “Choosing starting values for the EM algorithm for getting the highest likelihood in multivariate
Gaussian mixture models,” Computational Statistics & Data Analysis, vol. 41, no. 3, pp. 561–575, 2003.
[22] N. Vlassis and A. Likas, “A greedy EM algorithm for Gaussian mixture learning,” Neural processing letters, vol. 15, no. 1, pp. 77–87, 2002.
[23] C. Liu, “ML estimation of the multivariatetdistribution and the EM algorithm,” Journal of Multivariate Analysis, vol. 63, no. 2, pp. 296–312,
1997.
[24] C. Liu and D. B. Rubin, “ML estimation of the t distribution using EM and its extensions, ECM and ECME,” Statistica Sinica, pp. 19–39,
1995.
[25] K. Ata, J. Bootkrajang, and R. J. Durrant, “Toward large-scale continuous EDA: A random matrix theory perspective,” Evolutionary
computation, vol. 24, no. 2, pp. 255–291, 2015.
[26] V. Picheny, T. Wanger, and D. Ginsbourger, “A benchmark of kriging-based infill criteria for noisy optimization,” Struct Multidiscip Optim,
vol. 10, p. 1007, 2013.
| 9cs.NE
|
arXiv:1512.04315v1 [math.AC] 14 Dec 2015
RELATIVE HILBERT CO-EFFICIENTS
AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
Abstract. Let (A, m) be a Cohen-Macaulay local ring of dimension d and
let I ⊆ J be two m-primary ideals with I a reduction of J. For i = 0, . . . , d
I
th Hilbert coefficient of J (I) respectively. We
let eJ
i (A) (ei (A)) be the i
I
th relative Hilbert coefficient
call the number ci (I, J) = eJ
i (A) − ei (A) the i
of J with respect to I. If GI (A) is Cohen-Macaulay then ci (I, J) satisfy
various constraints. We also show that vanishing of some ci (I, J) has strong
implications on depth GJ n (A) for n ≫ 0.
Introduction
Let (A, m) be a Cohen-Macaulay local ring of dimension d and let J be an
m−primary ideal. The Hilbert-Samuel function of A with respect to J is HJ (n) =
λ(A/J n+1 ), (here λ(−) denotes the length). It is well known that HJ is of polynomial type i.e. there exists PJ (X) ∈ Q[X] such that HJ (n) = PJ (n) for all n ≫ 0.
We write
X+d
X +d−1
PJ (X) = eJ0 (A)
− eJ1 (A)
+ · · · + (−1)d eJd (A).
d
d−1
Then the numbers eJi (A) for i = 0, 1, · · · , d are the Hilbert coefficients of A with
respect to J. The number eJ0 (A) is called the multiplicity of A with respect to J.
Now assume for convenience A has infinite residue field. Then J has
La minimal reduction q generated by a system of parameters of A. Let GrJ (A) = n≥0 J n /J n+1
be the associated graded ring of A with respect to J. There has been a lot of research regarding properties of J and q and the depth properties of GrJ (A). For
example if J 2 = qJ then we say J has minimal multiplicity and in this case GrJ (A)
is Cohen-Macaulay (see, [14, 2.7]).
In the context of this paper we consider minimal reduction to be an absolute
reduction of J. The main new idea of this paper is that it is convenient to consider
reduction I of J not necessarily minimal but having the crucial property that
GrI (A) is Cohen-Macaulay. We note that if q is a minimal reduction of J then
it is generated by system of parameters of A and so necessarily Grq (A) is CohenMacaulay.
As I is a reduction of J then necessarily eI0 (A) = eJ0 (A). Let
ci (I, J) = eJi (A) − eIi (A)
for i ≥ 1.
Date: March 1, 2018.
1991 Mathematics Subject Classification. Primary 13A30; Secondary 13D40, 13D07.
Key words and phrases. associated graded rings, Cohen-Macaulay local rings, reductions, integral closure, normal ideals.
1
2
AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
Then we say ci (I, J) to be the ith relative Hilbert coefficient of J with respect to
I. We note that if q is a minimal reduction of J then eqi (A) = 0 for i ≥ 1 and so
ci (q, J) = eJi (A) for i ≥ 1.
Let us recall the classical Northcott’s inequality
eJ1 (A) ≥ eJ0 (A) − λ(A/J)
(see, [10]). But eJ0 (A) = λ(A/q) where q is a minimal reduction of J. So Northcott’s
inequality can be rewritten as
eJ1 (A) ≥ λ(J/q).
Furthermore if equality hods then by Huneke (see, [4]) and Ooishi (see, [11]) GrJ (A)
is Cohen-Macaulay.
Our first result which easily follows from a deep result of Huckaba and Marley
[5, Theorem 4.7] is the following.
Theorem 1. Let (A, m) be a Cohen-Macaulay local ring and let I ⊂ J be m−
primary ideals with I a reduction of J. Assume GrI (A) is Cohen-Macaulay. Then
c1 (I, J) ≥ λ(J/I).
If equality holds then GrJ (A) is also Cohen-Macaulay.
We give a different proof of Theorem 1. Although it’s longer than the proof
using Huckaba and Marley result, it has the advantage that it’s techniques can be
generalized to prove other results.
In [9] Narita proved that eJ2 (A) ≥ 0. Furthermore if dim A = 2 then eJ2 (A) = 0
if and only if reduction number of J n is 1 for n ≫ 0. In particular GrJ n (A) is
Cohen-Macaulay for n ≫ 0.
Our genelization of Narita’s result is:
Theorem 2. Let (A, m) be a Cohen-Macaulay local ring and let I ⊂ J be m−
primary ideals with I a reduction of J. Assume GrI (A) is Cohen-Macaulay. Then
c2 (I, J) ≥ 0.
If dim A = 2 and c2 (I, J) = 0 then GrJ n (A) is Cohen-Macaulay for n ≫ 0.
If we assume J is integrally closed then we have the following result:
Theorem 3. Let (A, m) be a Cohen-Macaulay local ring of dimension 2. Let I ⊂ J
be m− primary ideals with I a reduction of J. Assume J is integrally closed and
GrI (A) is Cohen-Macaulay. If c1 (I, J) = λ(J/I) + 1 then
f2 /I 2 ) ≤ 2λ(J/I) + 1.
2λ(J/I) ≤ λ(J
f2 /I 2 ) = 2λ(J/I) + 1 then GrJ n (A) is Cohen-Macaulay for all n ≫ 0.
If λ(J
f2 denotes the Ratliff-Rush closure of J 2 (see,[12]).
In Theorem 3, J
Narita gave an example which shows that eJ3 (A) of an m−primary ideal J can
be negative. Recall that an ideal J is said to be normal if J n is integrally closed
for all n ≥ 1. In [6] Itoh proved that if dim A ≥ 3 and J is a normal ideal then
eJ3 (A) ≥ 0. We prove:
RELATIVE HILBERT CO-EFFICIENTS
3
Theorem 4. Let (A, m) be a Cohen-Macaulay local ring of dimension d ≥ 3. Let
I ⊂ J be m− primary ideals with I a reduction of J. Assume J is normal and
GrI (A) is Cohen-Macaulay. Then
c3 (I, J) ≥ 0.
If d = 3 and c3 (I, J) = 0 then GrJ n (A) is Cohen-Macaulay for all n ≫ 0.
In the main body of the paper we consider a more general situation (A, m) →
(B, n) is a finite map with dim A = dim B, I an m-primary A-ideal, J an n-primary
ideal with IB a reduction of J. We now describe in brief the contents of this paper.
In section one we discuss a few prelimary results that we need. In section two we
prove Theorem 1. In section three we prove Theorems 2,3. We prove Theorem 4 in
section 4. Finally in section 5 we give a few examples which illustrates our results.
1. Preliminaries
Throughout this paper we follow the following hypothesis unless otherwise stated.
ψ
Hypothesis: Let (A, m) −→ (B, n) be a local homomorphism of Cohen-Macaulay
local rings such that
(1) B is finite as an A−module and dim A = dim B.
(2) I ⊂ A and J ⊂ B are ideals such that ψ(I)B is a reduction of J.
Remark 1.1. ψ(I)B is not necessarily a minimal reduction of J.
Remark 1.2. Note that B/n is a finite extension of A/m. Set δ = dimA/m B/n.
Then for any B−module N of finite length we have λA (N ) = δλB (N ).
The following result gives a necessary and sufficient condition for ψ(I)B to be a
reduction of J.
Lemma 1.3. Let ψ : (A, m) −→ (B, n) be a local homomorphism of CohenMacaulay local rings such that B is a finite A−module and dim A = dim B. Let I
ideal in A and let J ideal in B with ψ(I) ⊂ J. Then GrJ (B) is finitely generated
as a GrI (A)−module if and only if ψ(I)B is a reduction of J.
Proof. Suppose ψ(I)B is a reduction of J. Let c ≥ 1 be such that ψ(I)J n = J n+1
for all n ≥ c. Then J n+1 /J n+2 = ψ(I)(J n /J n+1 ) for all n ≥ c. Therefore GrJ (B)
is a finite GrI (A)−module.
Conversely suppose that GrJ (B) is a finite GrI (A)−module. Then there exists
n0 such that GrI (A)1 GrJ (B)n = GrJ (B)n+1 for all n ≥ n0 . Thus for n ≥ n0
I Jn
J n+1
= 2 . n+1
n+2
J
I J
ψ(I)J n + J n+2
.
=
J n+2
So J n+1 = ψ(I)J n + J n+2 . Therefore by Nakayama’s lemma J n+1 = ψ(I)J n for all
n ≥ n0 .
L
Remark 1.4. Let R(I, A) = A[It] = n≥0 I n tn be the Rees ring of A with respect
L
to I. If M is a finite A module then set R(I, M ) = n≥0 I n M tn the Rees module
of M with respect to I. It can also be easily shown ψ(I)B is a reduction of J if and
only if R(J, B) is a finite R(I, A)−module.
4
AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
Remark 1.5. Set W =
L
n≥0
J n+1 /I n+1 B. Then we have
0 −→ R(I, B) −→ R(J, B) −→ W (−1) −→ 0
an exact sequence of R(I, A) modules . So W (−1) is a finite R(I, A) module. Hence
W is a finite R(I, A)−module.
The following is our main object to study associated graded modules and Hilbert
coefficients.
L
Definition 1.6. Let M be an A−module. Set LI (M ) = n≥0 M/I n+1 M. Then
the A−module LI (M ) can be given an R(I, A)−module structure as follows. The
Rees ring R(I, A) is a subring of A[t] and so A[t] is an R(I, A)−module. Therefore
M [t] = M ⊗A A[t] is an R(I, A)−module. The exact sequence
0 −→ R(I, M ) −→ M [t] −→ LI (M )(−1) −→ 0
defines an R(I, A)−module structure on LI (M )(−1) and so on LI (M ). Notice
LI (M ) is not a finitely generated R(I, A)-module.
Remark 1.7. Let x be M superficial with respect to I and set u = xt ∈ R(I, A)1 .
Notice that LI (M )/uLI (M ) = LI (M/xM ).
By [13, Proposition 5.2] we have the following:
Remark 1.8. Let x ∈ I\I 2 . Then x∗ ∈ GrI (A)1 is GrI (M )−regular if and only if
u = xt ∈ R(I, A)1 is LI (M )−regular.
1.9. Let ψ : A −→ B as before and
is m−primary and
L dim A = d. Assume that I L
J is n−primary. Define LI (B) = n≥0 B/I n+1 B and LJ (B) = n≥0 B/J n+1 . As
LJ (B) is a R(J, B)− module and so as a R(I, A)−module. For each n ≥ 0 we have
J n+1
B
B
−→ n+1 −→ n+1 −→ 0
I n+1 B
I
B
J
an exact sequence of A− modules. So we get
0 −→
0 −→ W −→ LI (B) −→ LJ (B) −→ 0
an exact sequence of R(I, A) modules. Therefore
n+1
X
J
hI (z) − δhJB (z)
zn = B
λA
.
n+1
I
B
(1 − z)d+1
Note that hIB (1) − δhJB (1) = 0. So we can write
δhJB (z) = hIB (z) + (z − 1)r(z).
Therefore we have
1) δeJ0 (B) = eI0 (B).
2) δeJi (B) = eIi (B) + (r(i−1) (1)/(i − 1)!) for i ≥ 1.
Remark 1.10. If δeJ1 (B) 6= eI1 (B) then dim W = d.
We need the following technical result.
Lemma 1.11. Let ψ : A −→ B as before. Assume the residue field of A is infinite.
Then there exists x ∈ I such that
(1) x is A superficial with respect to I.
(2) ψ(x) is B superficial with respect to J.
RELATIVE HILBERT CO-EFFICIENTS
5
Proof. Note that GrJ (B) is a finite GrI (A)−module. Also ψ induces a natural map
ψ̂ : GrI (A) −→ GrJ (B). Let z ∈ GrI (A)1 be a GrI (A) ⊕ GrJ (B) filter regular.
Then note that ψ̂(z) is GrJ (B) filter regular. Let x ∈ I be such that x∗ = z. Then
clearly x is A superficial with respect to I. Also note that ψ(x)∗ = ψ̂(z). So ψ(x)
is B superficial with respect to J.
The following result easily follows by induction on the dimension of the ring.
Corollary 1.12. Let ψ : A −→ B as before. Assume that the residue field of A is
infinite. Let dim A = d ≥ 1. Then there exist x = x1 , · · · , xd ∈ I such that
(1) x is A superficial sequence with respect to I.
(2) ψ(x) is B superficial sequence with respect to J.
Proof. Follows easily by induction on d and using Lemma 1.11.
If ψ(I)B ( J then W 6= 0. More over we have the following result:
Lemma 1.13. Let ψ : A −→ B as before. Assume that GrI (B) is Cohen-Macaulay.
Then the following are equivalent:
(1) δeJ1 (B) = eI1 (B).
(2) GrI (B) = GrJ (B).
Proof. We first prove that GrJ (B) is Cohen-Macaulay. By Sally descent and
Lemma 1.11 we may assume that dim B = 1. Now consider the exact sequence
0 −→ W −→ LI (B) −→ LJ (B) −→ 0,
L n+1 n+1
where W =
J
/I
B. As δeJ1 (B) = eI1 (B) we get λ(W ) < ∞. Let M be the
i
unique homogeneous maximal ideal of R(I, A) and H i (−) = HM
(−). As GrI (B)
0
I
is Cohen-Macaulay by Remark 1.8 we get H (L (B)) = 0. So we get W = 0.
Therefore LI (B) = LJ (B). So H 0 (LJ (B)) = 0. Thus GrJ (B) is Cohen-Macaulay.
Now assume dim B ≥ 2. We prove LI (B) = LJ (B). This will prove the result.
Note that depth W ≥ 1. Set u := xt where x is A−superficial with respect to I and
ψ(x) is B−superficial with respect to J. Then we have an exact sequence
0 −→
LI (B)
LJ (B)
W
−→
−→
−→ 0.
uW
uLI (B)
uLJ (B)
By induction and Remark 1.7 we get LI (B)/uLI (B) = LJ (B)/uLJ (B). So W =
uW . By graded Nakayama’s Lemma W = 0. Hence LI (B) = LJ (B).
2. Extension of Northcott’s inequality.
The following result easily follows from a result due to Huckaba and Marley (see,
[5, Theorem 4.7]). However our techniques to prove this extends to prove our other
results.
Theorem 2.1. Let ψ : A −→ B as before. Assume that I is m−primary and J
is n−primary. If GrI (B) is Cohen-Macaulay then δeJ1 (B) ≥ eI1 (B) + λ(J/IB). If
equality holds then GrJ (B) is also Cohen-Macaulay.
Proof. By L
Sally machine we may assume that dim A = 1. Set Wi = J i+1 /I i+1 B
and W =
i≥0 Wi . Note that λ(Wi ) = e0 (W ) for i ≫ 0. Let (x) be a minimal
reduction of I. Let u = xt ∈ R(I, A)1 . Then we have
6
AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
0 −−−−→
Wi −−−−→
u
y
LI (B)i
u
y
−−−−→
LJ (B)i
u
y
−−−−→ 0
0 −−−−→ Wi+1 −−−−→ LI (B)i+1 −−−−→ LJ (B)i+1 −−−−→ 0
a commutative diagram with exact rows. As GrI (B) is Cohen-Macaulay so x∗ is
u
GrI (B) regular. So by Remark 1.8, u is LI (B) regular. Thus LI (B)i −→ LI (B)i+1
u
is injective. As W ⊂ LI (B) we also get Wi −→ Wi+1 is injective. Therefore
λ(W0 ) ≤ λ(W1 ) ≤ · · · ≤ λ(Wi ) = e0 (W ) for i ≫ 0. So
δeJ1 (B) = eI1 (B) + e0 (W )
≥ eI1 (B) + λ(J/IB).
Now suppose δeJ1 (B) = eI1 (B) + λ(J/IB). Then e0 (W ) = λ(J/IB). Thus λ(Wn ) =
u
λ(W0 ) for all n ≥ 0. So we get Wi −→ Wi+1 is an isomorphism. By Snake lemma
we get
u
LJ (B)i −→ LJ (B)i+1
is injective. Therefore u is LJ (B) regular. By Remark 1.8 x∗ is GrJ (B)−regular.
Hence GrJ (B) is Cohen-Macaulay.
Now we give an example where Theorem 2.1 holds.
Example 2.2. Let A = Q[|X, Y, Z, W |]/(XY − Y Z, XZ + Y 3 − Z 2 ). Let x, y, z, w
be the images of X, Y, Z, W in A respectively. Set m = (x, y, z, w). Then (A, m)
is a two dimensional Cohen-Macaulay local ring. Let I = (x, y, w). Note that I is
m−primary and z is integral over I. So I is a reduction of m. It is proved in [8,
example 3.6] GrI (A) is Cohen-Macaulay. Using CoCoA (see, [1]) we have computed
m
I
eI1 (A) = 6, em
1 (A) = 7 and λ(A/I) = 2. So e1 (A) = e1 (A) + λ(m/I). Hence by
Theorem 2.1 we get Grm (A) is Cohen-Macaulay.
The following example shows that the condition GrI (B) is Cohen-Macaulay is
essential.
Example 2.3. Let A = Q[|X, Y, Z, U, V, W |]/(Z 2, ZU, ZV, U V, Y Z − U 3 , XZ −
V 3 ), with X, Y, Z, U, V, W inderterminates. Let x, y, z, u, v, w be the images of
X, Y, Z, U, V, W in A. Set m = (x, y, z, u, v, w). Then (A, m) is a three dimensional
Cohen-Macaulay local ring. Let I = (x, y, u, w). Note that v 4 = vxz = 0 and z 2 = 0
in A. Thus z, v are integral over I. So I is a reduction of m. Let J = (x, y, w).
Then J is a minimal reduction I. Using CoCoA (see, [1]) we have checked that
PI (t) =
4 + t2 + t3
(1 − t)3
and Pm (t) =
1 + 3t + 3t3 − t4
.
(1 − t)3
We also checked λ(I/J) = 2, λ(I 2 /I 2 ∩ J) = 1 and λ(I 3 /I 3 ∩ J) = 0. Therefore by
[5, Theorem 4.7] we get depth GrI (A) < 3. Hence GrI (A) is not Cohen-Macaulay.
Also note that h−polynomial of Grm (A) has negative coefficient. So Grm (A) is also
I
not Cohen-Macaulay. It is easy to see that em
1 (A) = e1 (A) + λ(m/I).
3. The case of dimension two
Let a be an ideal in a Notherian ring S and M a finite S module. Then for
n+k
n M := ∪
n ≥ 1, a]
M :M ak ) is called the Ratliff-Rush closure of an M .
k≥1 (a
RELATIVE HILBERT CO-EFFICIENTS
7
In general if a ⊂ b be two ideals in a ring S then it need not imply that e
a⊂e
b.
However for reduction of ideals we have the following:
Proposition 3.1. Let S be a Notherian ring and let a ⊂ b be a reduction of b.
n ⊂b
n for all n ≥ 1.
f
Then af
n . So xak ⊂ an+k for some k. Thus xak br ⊂ an+k br for all r ≥ 0.
Proof. Let x ∈ af
As a ⊂ b is a reduction so abs = bs+1 for s ≫ 0. Choose r ≥ s. Then ak br = bk+r .
n.
Therefore xbk+r ⊂ bn+k+r . Thus x ∈ bf
L
n+1 M . Then L
^
e I (M ) is
e I (M ) =
3.2. Let M be an A−module. Define L
n≥0 M/I
e A)−module so R(I, A)−module. Set
a R(I,
M
M
n+1 /I n+1 B.
n+1 and W
e J (B) =
f=
J]
L
B/J]
n≥0
Then we have
n≥0
f −→ LI (B) −→ L
e J (B) −→ 0
0 −→ W
an exact sequence of R(I, A) modules. Note that hIB (1) = δhJB (1) = δe
hJB (1).
Therefore we can write
re(z)
r (z) and HW
δe
hJB (z) = hIB (z) + (z − 1)e
.
f (z) =
(1 − z)d
Therefore
1) δe
eJ0 (B) = eI0 (B).
2) δe
eJi (B) = eIi (B) + re(i−1) (1)/(i − 1)!.
Now we extend a famous result of Narita concerning second Hilbert coefficient
(see, [9]).
Proposition 3.3. Let ψ : A −→ B as before and dim A ≥ 2. Let I be m−primary
and J be n−primary. Assume that GrI (B) is Cohen-Macaulay. Then
δeJ2 (B) ≥ eI2 (B).
Proof. We may assume that dim A = 2. Let M be the unique homogeneous maximal ideal of R(I, A). Let H i (−) := HM (−) be the ith local cohomology module.
As GrI (B) is Cohen-Macaulay so H i (GrI (B)) = 0 for i = 0, 1. Also note that
f J (B)) = 0. By Remark 1.8 H 0 (L
e J (B)) = 0 and H i (LI (B)) = 0 for i = 0, 1.
H 0 (Gr
As we have
f −→ LI (B) −→ L
e J (B) −→ 0
0 −→ W
an exact sequence of R(I, A) modules. By considering long exact sequence in local
f ) = 0 for i = 0, 1. Hence W
f is Cohen-Macaulay. So
cohomology we get H i (W
re(1) (1) ≥ 0. Also note that eI2 (B) ≥ 0.
Now
δe
eJ2 (B) = eI2 (B) + re(1) (1).
≥ eI2 (B).
f is CohenRemark 3.4. From the proof of Proposition 3.3 one can see that W
Macaulay if dim A = 2.
By analysing the case of equality in the above Theorem we prove:
8
AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
Theorem 3.5. Let ψ : A −→ B as before and dim A = 2. Let I be m−primary
and J be n−primary. Assume that GrI (B) is Cohen-Macaulay. Suppose δeJ2 (B) =
f J (B) is Cohen-Macaulay. Consequently GrJ n (B) is Cohen-Macaula
eI2 (B). Then Gr
y for n ≫ 0.
f is CohenProof. We have δeJ2 (B) = eI2 (B) + re(1) (1). By remark 3.4 we get W
(1)
J
I
(1)
Macaulay. So re (1) ≥ 0. By hypothesis δe2 (B) = e2 (B). So re (1) = 0. Therefore
re(z) = c(constant) and hence
c
.
HW
f (z) =
(1 − z)2
Let x, y be a I superficial sequence. Set u = xt, v = yt. Then u, v ∈ R(I, A)1 . Now
f=W
f /uW
f , LI (B) = LI (B)/uLI (B) and L
e J (B) = L
e J (B)/uL
e J (B). As u is
set W
f ⊕ LI (B) ⊕ L
e J (B)−regular we get
W
f −→ LI (B) −→ L
e J (B) −→ 0
0 −→ W
an exact sequence. So we get a commutative diagram
0 −−−−→
fi −−−−→
W
v
y
LI (B)i
v
y
−−−−→
e J (B)i
L
v
y
−−−−→ 0
fi+1 −−−−→ LI (B)i+1 −−−−→ L
e J (B)i+1 −−−−→ 0
0 −−−−→ W
f are Cohen-Macaulay we get that v is LI (B) ⊕
with exact rows. As GrI (B) and W
v f
v
f −regular. So LI (B)i −→
fi −→
fi ) =
W
Wi+1 are injective. As λ(W
LI (B)i+1 and W
v
fi+1 ) for all i. We get W
fi −→
fi+1 is an isomorphism. By Snake Lemma
λ(W
W
v
e J (B)i −→
e J (B)i+1 is injective. So depth L
e J (B) ≥ 2. Hence
L
it follows that L
f
f
by remark 1.8 depth Gr J (B) ≥ 2. So Gr J (B) is Cohen-Macaulay. In particular
GrJ n (B) is Cohen-Macaulay for n ≫ 0.
Now we give an example where Theorem 3.5 holds.
Example 3.6. Let A = Q[|X, Y, Z, U, V |]/(Z 2 , ZU, ZV, U V, Y 2 Z − U 3 , X 2 Z − V 3 ),
with X, Y, Z, U, V inderterminates. Let x, y, z, u, v be the images of X, Y, Z, U, V
in A. Set m = (x, y, z, u, v). Then (A, m) is a two dimensional Cohen-Macaulay
local ring. Let I = (x, y, z, u) and J = (x, y, z, u, v 2 ). Note that v 4 − vx2 z = 0 in
A. Thus v is integral over I. So I is a reduction of J. Let q = (x, y). Then q is
a minimal reduction of I. Using CoCoA (see, [1]), we have computed eI1 (A) = 4
and eI2 (A) = eJ2 (A) = 1. We also checked that λ(I/q) = 3, λ(I 2 /I 2 ∩ q) = 1. By
[5, Theorem 4.7] we get GrI (A) is Cohen-Macaulay. Hence by Theorem 3.5 we get
f J (A) Cohen-Macaulay. Hence GrJ n (A) is Cohen-Macaulay for n ≫ 0.
Gr
For integrally closed ideals we prove:
Theorem 3.7. Let ψ : A −→ B as before and dim A = 2. Let I be m−primary and
J be n−primary. Assume that GrI (B) is Cohen-Macaulay. Suppose J is integrally
closed and
δeJ1 (B) = eI1 (B) + λ(J/IB) + 1.
Then
f2 /I 2 B) ≤ 2λ(J/IB) + 1.
a) 2λ(J/IB) ≤ λ(J
RELATIVE HILBERT CO-EFFICIENTS
9
f2 /I 2 B) = 2λ(J/IB) + 1. Then Gr
f J (B) is Cohen-Macaulay. Consequently
b) If λ(J
GrJ n (B) is Cohen-Macaulay for n ≫ 0.
f is Cohen-Macaulay. Let
Proof. By remark 3.4 W
X
fn )z n =
HW
λA (W
f (z) =
n≥0
re(z)
(1 − z)2
f . Note that all the co-efficients of re(z) are positive. Write
be the Hilbert series of W
re(z) = r0 + r1 z + · · · + rs z s .
f2 /I 2 B) = 2r0 + r1 . We have
Then we have r0 = λ(Je/IB) = λ(J/IB) and λ(J
δeJ1 (B) = eI1 (B) + re(1).
= eI1 (B) + λ(J/IB) + 1.
Therefore re(1) = λ(J/IB) + 1. Hence r1 + · · · + rs = 1. So (a) follows.
Suppose (b) holds. i.e.
f2 /I 2 B) = 2λ(J/IB) + 1.
λ(J
Then r1 = 1 and rj = 0 for j ≥ 2. Let x, y be a I superficial sequence. Set u = xt.
f /uW
f . Then we have
Then u ∈ R(I, A)1 . Also set U = W
0 −→ U −→
e J (B)
LI (B)
L
−→
−→ 0
e J (B)
uLI (B)
uL
an exact sequence. Note that λ(Un ) = λ(U2 ) for n ≥ 2. Also note that v = yt in
R(I, A) is LI (B)/uLI (B) regular. So v is U regular. So we have
U
−→ 0
vU
an exact
sequence. By Hilbert series (U/vU )j =
0 for j ≥ 2. Now by setting K =
v eJ
J
J
J
e
e
e
ker (L (B)/uL (B))(−1) −→ L (B)/uL (B) we get by Snake Lemma Kj = 0
0 −→ U (−1) −→ U −→
for j ≥ 2. Also note that K0 = 0.
Claim: K1 = 0. To prove the claim set F = {Jfn }. Then F is a filtration on B.
Then F = {Jfn + (x)/(x)} is the quotient filtration on B/xB. Put q = J/(x) = F¯1 .
We may assume that q is integrally closed. As Jfn = J n for n ≫ 0. So we get
F¯n = qn for n ≫ 0.
We prove that F2 : y = F¯1 = q. Let a ∈ F2 : y. So ya ∈ F2 . So y n+1 a ∈ y n F2 ⊂
F2+n = qn+2 for n ≫ 0. This implies a ∈ qe = q̄ = q. It follows that K1 = 0. Thus
e J (B) has depth ≥ 2. This implies Gr
f J (B) is Cohen-Macaulay. So
K is zero. So L
GrJ n (B) is Cohen-Macaulay for n ≫ 0.
Here we give an example where our Theorem 3.7 holds:
Example 3.8. Let A = Q[|X, Y, Z, W |]/(X 2 − Y 2 Z, XY 4 − Z 2 ). Let x, y, z, w
denotes the images of X, Y, Z, W in A respectively. Let I = (x, y, w) and m =
(x, y, z, w). Then (A, m) is a two dimensional Cohen-Macaulay local ring. Using
CoCoA (see, [1]) we have computed
PI (t) =
2 + 2t
(1 − t)2
and Pm (t) =
1 + 2t + t2
.
(1 − t)2
10 AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
I
f2 2
We have em
1 (A) = e1 (A) + λ(m/I) + 1. We have also checked that λ(m /I ) = 3 =
f m (A) is Cohen-Macaulay.
2λ(m/I) + 1. Hence by Theorem 3.7(b) we get Gr
4. The case of third Hilbert coefficient
In this section we deal with third Hilbert coefficients and generalize a remarkable
result of Itoh for normal ideals(see, [6]).
Theorem 4.1. Let A = B and ψ = idA . Let dim A ≥ 3. Let I, J be m−primary
ideals and I is a reduction of J. Suppose GrI (A) is Cohen-Macaulay and J is
asymptotically normal. Then
eJ3 (A) ≥ eI3 (A).
Proof. By standard argument it suffices to consider dim A = 3. As J is asymptotically normal by [3, Theorem 3.1] there exists n0 such that depth GrJ n (A) ≥ 2 for
all n ≥ n0 .
L
Now set T = I n , K = J n and W = n≥0 K n+1 /T n+1 for n ≥ n0 . Then we get
an exact sequence
0 −→ W −→ LT (A) −→ LK (A) −→ 0
of R(T, A)−modules. Note that depth LK (A) ≥ 2. So we get W is Cohen-Macaulay
of dimension 3. Thus ei (W ) ≥ 0 for 0 ≤ i ≤ 3. Hence
T
eK
3 (A) = e3 (A) + e2 (W ).
≥ eT3 (A).
As eJ3 (A) = eJ3 (A) and eI3 (A) = eI3 (A) for all n ≥ 1. Therefore eJ3 (A) ≥ eI3 (A).
n
n
By analysing the case of equality we prove the following:
Theorem 4.2. Let A = B and ψ = idA . Assume dim A = 3. Let I, J be two
m−primary ideals and I is a reduction of J. Suppose GrI (A) is Cohen-Macaulay
and J is asymptotically normal. If eJ3 (A) = eI3 (A) then GrJ n (A) is Cohen-Macaulay
for all n ≫ 0.
Proof. As J is asymptotically normal by [3, Theorem 3.1] there exists n0 such
that for all n ≥ n0 , depth GrJ n (A) ≥ 2. Now set T = I n , K = J n and W =
L
n+1
/T n+1 for n ≥ n0 . From the proof of Theorem 4.1 we see that W is
n≥0 K
Cohen-Macaulay of dimension 3 and eJ3 (A) = eI3 (A) + e2 (W ).
Suppose eI3 (A) = eJ3 (A). Then e2 (W ) = 0. So the Hilbert series of W is given
by
r0 + r1 s
.
HW (s) =
(1 − s)3
Let x, y, z be a K ⊕ T superficial sequence. Set u = xt, v = yt and w = zt. Note
that u, v, w ∈ R(T, A)1 . Also set U = W/(u, v)W, LK (A) = LK (A)/(u, v)LK (A)
and LT (A) = LT (A)/(u, v)LT (A). Then we get an exact sequence
0 −→ U −→ LT (A) −→ LK (A) −→ 0.
Now consider the commutative diagram
RELATIVE HILBERT CO-EFFICIENTS
11
0 −−−−→ U (−1) −−−−→ LT (A)(−1) −−−−→ LK (A)(−1) −−−−→ 0
w
w
w
y
y
y
0 −−−−→
U
−−−−→
LT (A)
−−−−→
LK (A)
−−−−→ 0
Also note that Hilbert series of U/wU is given by
HU/wU (s) = r0 + r1 s.
w
Therefore (U/wU )j = 0 for j ≥ 2. Now set E = ker LK (A)(−1) −→ LK (A) .
Note that Ej = 0 for all j ≥ 2 by Snake Lemma. Also E0 = 0.
Claim E1 = 0: To prove this set F = {K m }. Then F is a filtration on A. Also F =
{K m + (x, y)/(x, y)} is the quotient filtration on A/(x, y)A. Put q = K/(x, y) = F¯1 .
We may assume that q is integrally closed. Note F¯m = qm for m ≥ 1.
We prove that F2 : z = F¯1 = q. Let a ∈ F2 : z. So za ∈ F2 = q2 . This implies
a ∈ qe ⊂ q̄ = q. It follows that E1 = 0. Thus E is zero. So depth LK (A) ≥ 1.
Thus depth LK (A) ≥ 3. Therefore depth GrK (A) ≥ 3. Hence GrK (A) is CohenMacaulay.
5. Examples.
In this section we show that there are plenty of examples where Theorem 2.1
holds.
Example 5.1. Let (R, m) be a regular local ring. Let (B, n) = (R/a, m/a) be a
Cohen-Macaulay local ring. Suppose dim R = t and dim B = d. Then ht(a) = t − d.
Set g = t − d. Then there exists a regular sequence u = u1 , · · · , ug of length g. Set
A = R/(u). Then we get a surjective ring homomorphism
ψ
A ։ B.
Let q be a minimal reduction of mA . Set I = (q :A mA ). Clearly q ⊂ I ⊂ mA . By
[2, Theorem 2.1] we get
I 2 = qI.
So ψ(I 2 ) = ψ(q)ψ(I). Thus GrI (B) has minimal multiplicity. As ψ(q)B is a
minimal reduction ψ(mA ) = n so we get ψ(I)B is a reduction of ψ(mA ) = n. Hence
by Theorem 2.1 we get
en1 (B) ≥ eI1 (B) + λ(n/I).
Example 5.2. Suppose (A, m) be a Gorenstein local ring which not regular. Let
J be any m−primary ideal. Set I := (q :A m) where q is a minimal reduction of J.
Then
eJ1 (A) ≥ eI1 (A) + λ(J/I).
Proof. It is enough to prove that I has minimal multiplicity. By [2, Theorem 2.1]
we have q ⊂ I ⊂ J and I 2 = qI. Thus I has minimal multiplicity.
ψ
Example 5.3. Let (A, m) → (B, n) be a local homomorphism of Cohen-Macaulay
local rings with dim A = dim B. Let I be an m primary ideal in A. If A is regular
and GrI (A) is Cohen-Macaulay. Then GrI (B) is Cohen-Macaulay. If ψ(I)B is a
reduction J in B then
eJ1 (B) ≥ eI1 (B) + λ(J/I).
12 AMIR MAFI, TONY J. PUTHENPURAKAL, RAKESH B. T. REDDY AND HERO SAREMI
Proof. By Auslander-Buchsbaum formula we get B is free as an A−module. As
GrI (A) Cohen-Macaulay, so GrI (B) is Cohen-Macaulay. Hence by Theorem 2.1 we
get the inequality.
Example 5.4. Let (A, m) be Cohen-Macaulay local ring and I be an m−primary
ideal. Let I ⊂ J ⊂ I¯ (integral closure of I). If GrI (A) is Cohen-Macalay then by
Theorem 2.1 we get
eJ1 (A) ≥ eI1 (A) + λ(J/I).
If equality holds above then GrJ (A) is Cohen-Macaulay.
Acknowledgements
The first and last authors would like to thank IIT Bombay especially Department
of Mathematics for it’s hospitality during the preparation of this work.
References
[1] Neisi, G., Capani, A.,Robbiano and L.: CoCoA, a system for doing computations in commutative algebra. available via anonymous ftp from cocoa.dima.unige.it 1995.
[2] A. Corso, C. Polini, and W. V. Vasconcelos, links of prime ideals, Math. Proc. Cambridge
Phil. Soc. 115 (1994), 431-436.
[3] S. Huckaba and C. Huneke, Normal ideals in regular rings, J. Reine Angew. Math 510 (1999)
63-83.
[4] C. Huneke, Hilbert functions and symbolic powers, Michigan Math. J. 34 (1987), 293-318.
[5] S. Huckaba and T. Marley, Hilbert coefficients and depth of associated graded rings, J.London
Math.Soc.,56(2) (1997), 64-76.
[6] S. Itoh, Coefficients of normal Hilbert polynomials, J. Algebra 150 (1992), 101-117.
[7] S. Itoh, Hilbert coefficients of integrally closed ideals, J. Algebra 176 (1995), 638-652.
[8] A. V. Jayanthan and Ramakrishna. N On the depth of fiber cones of stretched m−primary
ideals, To appear in Indian J. Pure Appl. Math.
[9] M. Narita, A note on the coefficients of Hilbert characteristic functions in semi-regular local
rings, Proc. Camb. Phil. Soc. 59 (1963), 269-275.
[10] D. G. Northcott, A note on the coefficients of the abstract Hilbert function , J. London Math.
Soc. 35 (1960), 209-214.
[11] , Ooishi, △-genera and sectional genera of commutative rings , Hiroshima Math. J. 17 (1987),
145-158.
[12] L.J. Ratliff Jr and D.E. Rush, Two notes on reduction of ideals, Indiana Univ. Math. J. 27
(6) (1978) 929-934.
[13] T. J. Puthenpurakal, Ratliff-Rush Filtration, Regularity and Depth of Higher Associated
Graded Modules Part I,J.Pure Applied Algebra, 208, 159-176(2007).
[14] P. Valabrega, G. Valla, Form rings and regular sequences, Nogoya Math. J. 72(1978) 93-101.
Department of Mathematics, University of Kurdistan, P.O. Box: 416, Sanandaj, Iran.
E-mail address: a [email protected]
Department of Mathematics, IIT Bombay, Powai, Mumbai 400 076
E-mail address: [email protected]
E-mail address: [email protected]
Department of Mathematics, Sanandaj Branch Islamic Azad University, Sanandaj,
Iran.
E-mail address: [email protected]
| 0math.AC
|
1
On Binary Distributed Hypothesis Testing
Eli Haim and Yuval Kochman
arXiv:1801.00310v1 [cs.IT] 31 Dec 2017
Abstract
We consider the problem of distributed binary hypothesis testing of two sequences that are generated by an i.i.d.
doubly-binary symmetric source. Each sequence is observed by a different terminal. The two hypotheses correspond
to different levels of correlation between the two source components, i.e., the crossover probability between the two.
The terminals communicate with a decision function via rate-limited noiseless links. We analyze the tradeoff between
the exponential decay of the two error probabilities associated with the hypothesis test and the communication rates.
We first consider the side-information setting where one encoder is allowed to send the full sequence. For this setting,
previous work exploits the fact that a decoding error of the source does not necessarily lead to an erroneous decision
upon the hypothesis. We provide improved achievability results by carrying out a tighter analysis of the effect of
binning error; the results are also more complete as they cover the full exponent tradeoff and all possible correlations.
We then turn to the setting of symmetric rates for which we utilize Körner-Marton coding to generalize the results,
with little degradation with respect to the performance with a one-sided constraint (side-information setting).
I. I NTRODUCTION
We consider the distributed hypothesis testing (DHT) problem, where there are two distributed sources, X and
Y , and the hypotheses are given by
(0)
H0 : (X, Y ) ∼ PX,Y
(1)
H1 : (X, Y ) ∼ PX,Y ,
(0)
(1a)
(1b)
(1)
where PX,Y and PX,Y are different joint distributions of X and Y . The test is performed based on information
sent from two distributed terminals (over noiseless links), each observing n i.i.d. realizations of a different source,
where the rate of the information sent from each terminal is constrained. This setup, introduced in [1, 2], introduces
a tradeoff between the information rates and the probabilities of the two types of error events. In this work we
focus on the exponents of these error probabilities, with respect to the number of observations n.
When at least one of the marginal distributions depends on the hypothesis, a test can be constructed based only on
the type of the corresponding sequence. Although this test may not be optimal, it results in non-trivial performance
(positive error exponents) with zero rate. In contrast, when the marginal distributions are the same under both
hypotheses, a positive exponent cannot be achieved using a zero-rate scheme, see [3].
One may achieve positive exponents while maintaining low rates, by effectively compressing the sources and then
basing the decision upon their compressed versions. Indeed, many of the works that have considered the distributed
hypothesis testing problem bear close relation to the distributed compression problem.
January 3, 2018
DRAFT
2
Ahlswede and Csiszár [4] have suggested a scheme based on compression without taking advantage of the
correlation between the sources; Han [5] proposed an improved scheme along the same lines. Correlation between
the sources is exploited by Shimokawa et al. [6, 7] to further reduce the coding rate, incorporating random binning
following the Slepian-Wolf [8] and Wyner-Ziv [9] schemes. Rahman and Wagner [10] generalized this setting and
also derived an outer bound. They also give a “quantize and bin” interpretation to the results of [6]. Other related
works include [11–15]. See [16, 10] for further references.
We note that in spite of considerable efforts over the years, the problem remains open. In many cases, the gap
between the achievability results and the few known outer bounds is still large. Specifically, some of the stronger
results are specific to testing against independence (i.e., under one of the hypotheses X and Y are independent),
or specific to the case where one of the error exponents is zero (“Stein’s-Lemma” setting). The present work
significantly goes beyond previous works, extending and improving the achievability bounds. Nonetheless, the
refined analysis comes at a price. Namely, in order to facilitate analysis, we choose to restrict attention to a simple
source model.
To that end, we consider the case where (X, Y ) is a doubly symmetric binary source (DSBS). That is, X and Y
△
are each binary and symmetric. Let Z = Y ⊖ X be the modulo-two difference between the sources.1 We consider
the following two hypotheses:
H0 : Z ∼ Ber (p0 )
(2a)
H1 : Z ∼ Ber (p1 ) ,
(2b)
where we assume throughout that p0 ≤ p1 ≤ 1/2. Note that a sufficient statistic for hypothesis testing in this case
is the weight (which is equivalent to the type) of the noise sequence Z. Under communication rate constraints,
a plausible approach would be to use a distributed compression scheme that allows lossy reconstruction of the
sequence Z, and then base the decision upon that sequence.
We first consider a one-sided rate constraint. That is, the Y -encoder is allocated the full rate of one bit per source
sample, so that the Y sequence is available as side information at the decision function. In this case, compression
of Z amounts to compression of X; a random binning scheme is optimal for this task of compression, lossless or
lossy.2 Indeed, in this case, the best known achievability result is due to [6], which basically employs a random
binning scheme.3
A natural question that arises when using binning as part of the distributed hypothesis testing scheme is the effect
of a “bin decoding error” on the decision error between the hypotheses. The connection between these two errors
is non-trivial as a bin decoding error inherently results in a “large” noise reconstruction error, much in common
with errors in channel coding (in the context of syndrome decoding). Specifically, when a binning error occurs,
1 Notice
2 More
that in this binary case, the uniform marginals mean that Z is necessarily independent of X.
precisely, it gives the optimal coding rates, as well as the best known error exponents when the rate is not too high.
3 Interestingly,
when p1 = 1/2 (testing against independence), the simple scheme of [4] which ignores the side-information altogether is
optimal.
January 3, 2018
DRAFT
3
the reconstruction of the noise sequence Z is roughly consistent with an i.i.d. Bernoulli 1/2 distribution. Thus, if
one feeds the weight of this reconstructed sequence to a simple threshold test, it would typically result in deciding
that the noise was distributed according to p1 , regardless of whether that is the true distribution or not. This effect
causes an asymmetry between the two error probabilities associated with the hypothesis test. Indeed, as the Stein
exponent corresponds to highly asymmetric error probabilities, the exponent derived in [6] may be interpreted as
taking advantage of this effect.4
The contribution of the present work is twofold. First we extend and strengthen the results of [6]. By explicitly
considering and leveraging the properties of good codes, we bound the probability that the sequence Z happens
to be such that Y ⊖ Z is very close to some wrong yet “legitimate” X, much like an undetected error event in
erasure decoding [17]. This allows us to derive achievability results for the full tradeoff region, namely the tradeoff
between the error exponents corresponding to the two types of hypothesis testing errors.
The second contribution is in considering a symmetric-rate constraint. For this case, the optimal distributed
compression scheme for Z is the Körner-Marton scheme [18], which requires each of the users to communicate at
a rate H (Z); hence, the sum-rate is strictly smaller than the one of Slepian-Wolf, unless Z is symmetric. Thus,
the Körner-Marton scheme is a natural candidate for this setting. Indeed, it was observed in [4, 16] that a standard
information-theoretic solution such as Slepian-Wolf coding may not always be the way to go, and [16] mentions the
the Körner-Marton scheme in this respect. Further, Shimokawa and Amari [19] point out the possible application
of the Körner-Marton scheme to distributed parameter estimation in a similar setting and a similar observation is
made in [20]. However, to the best of our knowledge, the present work is the first to propose an actual KörnerMarton-based scheme for distributed hypothesis testing and to analyze its performance. Notably, the performance
tradeoff obtained recovers the achievable tradeoff derived for a one-sided constraint.
The rest of this paper is organized as follows. In Section II we formally state the problem, define notations and
present some basic results. Section III and IV provide necessary background: the first surveys known results for
the case of a one-sided rate constraint while the latter provides definitions and properties of good linear codes. In
Section V we present the derivation of a new achievable exponents tradeoff region. Then, in Section VI we present
our results for a symmetric-rate constraint. Numerical results and comparisons appear in Section VII. Finally,
Section VIII concludes the paper.
II. P ROBLEM S TATEMENT AND N OTATIONS
A. Problem Statement
Consider the setup depicted in Figure 1. X and Y are random vectors of blocklength n, drawn from the (finite)
source alphabets X and Y, respectively. Recalling the hypothesis testing problem (1), we have two possible i.i.d.
distributions. In the sequel we will take a less standard notational approach, and define the hypotheses by random
variable H which takes the values 0, 1, and assume a probability distribution function PX,Y |H ; Therefore H = i
4 Another interesting direction, not pursued in this work, is to change the problem formulation to allow declaring an “erasure” when the
probability of a bin decoding error exceeds a certain threshold.
January 3, 2018
DRAFT
4
φX
X
iX ∈ MX
ψ
φY
Y
Fig. 1.
Ĥ
iY ∈ MY
Problem setup.
(i)
refers to Hi of (1) and (2).5 We still use for the distribution PX,Y |H=i (for i = 0, 1) the shortened notation PX,Y .
Namely, for any x ∈ X n and y ∈ Y n , and for i ∈ {0, 1},
P (X = x, Y = y|H = i) =
n
Y
(i)
PX,Y (xj , yj ).
j=1
A scheme for the problem is defined as follows.
△
Definition 1: A scheme Υ = (φX , φY , ψ) consists of encoders φX and φY which are mappings from the set of
length-n source vectors to the messages sets MX and MY :
φX : X n 7→ MX
(3a)
φY : Y n 7→ MY .
(3b)
and a decision function, which is a mapping from the set of possible message pairs to one of the hypotheses:
ψ : MX × MY 7→ {0, 1}.
(4)
Definition 2: For a given scheme Υ, denote the decision given the pair (X, Y) by
△
Ĥ = ψ (φX (X), φY (Y)) .
(5)
The decision error probabilities of Υ are given by
△
ǫi = P Ĥ 6= H H = i ,
i = 0, 1.
(6a)
Definition 3: For any E0 > 0 and E1 > 0, the exponent pair (E0 , E1 ) is said to be achievable at rates (RX , RY )
if there exists a sequence of schemes
△
(n)
(n)
Υ(n) = φX , φY , ψ (n) ,
5 We
n = 1, 2, . . .
(7)
do not assume any given distribution over H, as we are always interested in probabilities given the hypotheses.
January 3, 2018
DRAFT
5
(n)
(n)
with corresponding sequences of message sets MX and MY
(n)
and error probabilities ǫi , i ∈ {0, 1}, such that6
1
(n)
log MX ≤ RX
n
1
(n)
lim sup log MY ≤ RY ,
n→∞ n
lim sup
(8a)
n→∞
(8b)
and
lim inf −
n→∞
1
(n)
log ǫi ≥ Ei ,
n
i = 0, 1.
(8c)
The achievable exponent region C (RX , RY ) is the closure of the set of all achievable exponent pairs.7
The case where only one of the error probabilities decays exponentially is of special interest; we call the resulting
(n)
quantity the Stein exponent after Stein’s Lemma (see, e.g., [21, Chapter 12]). When ǫ1
is exponential, the Stein
exponent is defined as:
△
σ1 (RX , RY ) = sup {E1 : ∃(E0 , E1 ) ∈ C (RX , RY )} .
(9)
E0 >0
σ0 (RX , RY ) is defined similarly.
We will concentrate on this work on two special cases of rate constraints, where for simplicity we can make the
notation more concise.
1) One-sided constraint where RY = ∞. We shall denote the achievable region and Stein exponents as CX (RX ),
σX,0 (RX ) and σX,1 (RX ).
2) Symmetric constraint where RX = RY = R. We shall denote the achievable region and Stein exponents as
C (R), σ0 (R) and σ1 (R).
Note that for any R we have that C (R) ⊆ CX (R).
Whenever considering a specific source distribution, we will take (X, Y ) to be a DSBS. Recalling (2), that means
△
that X and Y are binary symmetric, and the “noise” Z = Y ⊖ X satisfies:
P (Z = 1|H = i) = pi ,
i = 0, 1
(10)
for some parameters 0 ≤ p0 ≤ p1 ≤ 1/2 (note that there is loss of generality in assuming that both probabilities
are on the same side of 1/2).
B. Further Notations
The following notations of probability distribution functions are demonstrated for random variables X, Y and Z
over alphabets X , Y and Z, respectively. The probability distribution function of a random variable X is denoted
by PX , and the conditional probability distribution function of a random variable Y given a random variable X is
6 All
logarithms are taken to the base 2, and all rates are in units of bits per sample.
7 For
simplicity of the notation we omit here and in subsequent definitions the explicit dependence on the distributions (P (0) , P (1) ).
January 3, 2018
DRAFT
6
denoted by PY |X . A composition PX and PY |X is denoted by PX PY |X , leading to the following joint probability
distribution function PX,Y of X and Y :
△
PX PY |X (x, y) = PX (x)PY |X (y|x),
(11)
for any pair x ∈ X and y ∈ Y.
The Shannon entropy of a random variable X is denoted by H (PX ), and the Kullback-Leibler divergence of
a pair of probability distribution functions (P, Q) is denoted by D (P kQ). The mutual information of a pair of
random variables (X, Y ) is denoted by I PX , PY |X . The similar conditional functionals of the entropy, divergence
and mutual information are defined by an expectation over the a-priori distribution: the conditional entropy of a
random variable X given a random variable Z is denoted by
X
△ X
H PX|Z PZ =
PX (x)
PY |X (y|x) log
x∈X
y∈Y
1
.
PY |X (y|x)
(12)
The divergence of a pair of conditional probability distribution functions PX|Z and PY |Z is denoted by
D PX|Z kPY |Z PZ .
The conditional mutual information of a pair of random variables (X, Y ) given a random variable Z is denoted by
I PX|Z , PY |X,Z PZ ,
and notice that it is equal to
H PX|Z PZ − H PX|Y,Z PZ PX|Z .
If there is a Markov chain Z ↔ X ↔ Y , then we can omit the Z from PY |X,Z and the expression becomes
I PX|Z , PY |X PZ .
Since we concentrate on a binary case, we need the following. Denote the binary divergence of a pair (p, q),
where p, q ∈ (0, 1), by
△
Db (pkq) = p log
1−p
p
+ (1 − p) log
,
q
1−q
(13)
which is the Kullback-Leibler divergence of the pair of probability distributions ((p, 1 − p), (q, 1 − q)). Denote the
binary entropy of p ∈ (0, 1) by
△
Hb (p) = p log
1
1
+ (1 − p) log
,
p
1−p
(14)
which is the entropy function of the probability distribution (p, 1 − p). Denote the Gilbert-Varshamov relative
January 3, 2018
DRAFT
7
distance of a code of rate R, δGV : [0, 1] 7→ [0, 1/2] by
△
δGV (R) = Hb−1 (1 − R).
(15)
The operator ⊕ denotes addition over the binary field. The operator ⊖ is equivalent to the ⊕ operator over the
binary field, but nevertheless, we keep them for the sake of consistency.
The Hamming weight of a vector u = (u1 , . . . , un ) ∈ {0, 1}n is denoted by
wH (u) =
n
X
1{ui =1} ,
(16)
k=1
where 1{·} denotes the indicator function, and the sum is over the reals. The normalized Hamming weight of this
vector is denoted by
δH (u) =
1
wH (u).
n
(17)
Denote the n dimensional Hamming ball with center c and normalized radius r ∈ [0, 1] by
△
Bn (c, r) = {x ∈ {0, 1}n|δH (x ⊖ c) ≤ r} ,
(18)
The binary convolution of p, q ∈ [0, 1] is defined by
△
p ∗ q = (1 − p)q + p(1 − q).
(19)
Definition 4 (Bernoulli Noise): A Bernoulli random variable Z with P (Z = 1) = p is denoted by Z ∼ Ber (p).
An n dimensional random vector Z with i.i.d. entries Zi ∼ Ber (p) for i = 1, . . . , n is called a Bernoulli noise, and
denoted by
Z ∼ BerV (n, p)
(20)
Definition 5 (Fixed-Type Noise): Denote the set of vectors with type a ∈ [0, 1] by
△
Tn (a) = {x ∈ {0, 1}n : δH (x) = a}.
(21a)
N ∼ Uniform (Tn (a))
(21b)
A noise
is called an n-dimensional fixed-type noise of type a ∈ [0, 1].
.
−1
∞
log(an /bn ) = 0. We write
For any two sequences, {an }∞
n=1 and {bn }n=1 , we write an = bn if limn→∞ n
an ≤ bn if limn→∞ n−1 log(an /bn ) ≤ 0.
For any two sequences of random vectors Xn , Yn ∈ X n (n = 1, 2, . . .), we write
Xn = Yn
D
January 3, 2018
(22)
DRAFT
8
if
.
P (Xn = xn ) = P (Y = xn )
(23)
uniformly over xn ∈ X n , that is,
lim
n→∞
1
PX (xn )
log n
=0
n
PYn (xn )
(24)
PX (xn )
1
log n
≤0
n
PYn (xn )
(25)
uniformly over xn ∈ X n . We write Xn 6 Yn if
D
lim
n→∞
uniformly over xn ∈ X n .
The set of non-negative integers are denoted by Z+ , and the set of natural numbers, i.e., 1, 2, . . ., by N.
C. Some Basic Results
When the rate is not constrained, the decision function has access to the full source sequences. The optimal
tradeoff of the two types of errors is given by the following decision function, depending on the parameter T ≥ 0
(Neyman-Pearson [22]),8
0, P (0) (x, y) ≥ T · P (1) (x, y)
X,Y
X,Y
ϕ(x, y) =
1, otherwise.
(26)
Proposition 1 (Unconstrained Case): Consider the hypothesis testing problem as defined in Section II-A, where
there is no rate constraint, i.e. RX = RY = ∞, then (E0 , E1 ) ∈ C (∞) if and only if there exists a distribution
(∗)
function PX,Y over the pair (X , Y) such that
(i)
(∗)
Ei ≤ D PX,Y PX,Y , for i = 0, 1.
(27)
For proof, see e.g. [21]. Note that in fact rates equal to the logarithms of the alphabet sizes suffice.
For the DSBS, the Neyman-Pearson decision function is a threshold on the weight of the noise sequence. We
denote it (with some abuse of notations) by
△
ϕt (x, y) = ϕt (δH (x ⊕ y)) ,
where ϕt : R 7→ {0, 1} is a threshold test,
0, w ≤ t
ϕt (w) =
1, w > t.
(28)
8 In order to achieve the full Neyman-Pearson tradeoff, special treatment of the case of equality is needed. As this issue has no effect on error
exponents, we ignore it.
January 3, 2018
DRAFT
9
It leads to the following performance:
Corollary 1 (Unconstrained Case, DSBS): For the DSBS, CX (1) = C (1), and they consist of all pairs (E0 , E1 )
satisfying that for some t ∈ (p0 , p1 ),
Ei ≤ Db (tkpi ), for i = 0, 1.
(29)
We now note a time-sharing result, which is general to any given achievable set.
Proposition 2 (time-sharing): Suppose that (E0 , E1 ) ∈ C (RX , RY ). Then ∀α ∈ [0, 1] :
(αE0 , αE1 ) ∈ C (αRX , αRY ).
(30)
The proof is standard, by noting that any scheme may be applied to an α-portion of the source blocks, ignoring
the additional samples. Applying this technique to Corollary 1, we have a simple scheme where each encoder sends
only a fraction of its observed vector.
Corollary 2: Consider the DSBS hypothesis testing problem as defined in Section II-A. For any rate constraint
R ∈ [0, 1], for any t ∈ (p0 , p1 )
(R · Db (tkp0 ), R · Db (tkp1 )) ∈ C (R)
(31)
Specializing to Stein’s exponents, we have:
σ0 (R) ≥ R · Db (p1 kp0 )
(32a)
σ1 (R) ≥ R · Db (p0 kp1 ),
(32b)
Of course we may apply the same result to the one-sided constrained case, i.e., CX and the corresponding Stein
exponents.
III. O NE -S IDED C ONSTRAINT: P REVIOUS R ESULTS
In this section we review previous results for the one-sided constraint case RY = ∞. We first present them for
(0)
(1)
general distributions PX,Y , PX,Y and then specialize to the DSBS.
A. General Sources
Ahlswede and Csiszár have established the following achievable Stein’s exponent.
Proposition 3 ([4, Theorem 5]): For any RX > 0,
(0)
(1)
σX,1 (R) ≥ D PX PX
+
max
PV |X :
(0)
I PX ,PV |X ≤RX
January 3, 2018
(0)
(∗)
D PV,Y PV,Y ,
(33)
DRAFT
10
(∗)
(0)
where PV,Y and PV,Y are the marginals of
(0)
△
(0)
(0)
(∗)
△
(0)
(1)
PV,X,Y = PV |X PX PY |X
and
PV,X,Y = PV |X PX PY |X ,
respectively.
The first term of (33) reflects the contribution of the type of X (which can be conveyed with zero rate), while the
second reflects the contribution of the lossy version of X sent with rate RX . Interestingly, this exponent is optimal
(1)
(1)
for case PY |X = PY , known as test against independence.
Han has improved upon this exponent by conveying the joint type of the source sequence X and its quantized
version (represented by V ) to the decision function.9
Proposition 4 ([5, Theorems 2,3]): For any RX ≥ 0,
(1)
(0)
σX,1 (RX ) ≥ D PX PX +
max
PV |X :
(0)
I PX ,PV |X ≤RX ,
σHAN (V ),
(34a)
|V|≤|X |+1
where
△
σHAN (V ) =
(0)
(1)
(∗)
D PY |X,V kPY |X PX PV |X
min
(∗)
PY |V,X :
(∗)
(34b)
(0)
PV,Y =PV,Y
(0)
(∗)
and where PV,Y and PV,Y are the marginals of
(0)
△
(0)
(0)
PV,X,Y = PV |X PX PY |X
(35a)
and
(∗)
△
(0)
(∗)
PV,X,Y = PV |X PX PY |V,X ,
(35b)
respectively.
The following result by Shimokawa et al., gives a tighter achievable bound by using the side information Y when
encoding X.
Proposition 5 ([6, Corollary III.2],[16, Theorem 4.3]): Define
△
(0)
(0)
σSHA (V ) = −I PX|Y , PV |X PY
9 Han’s
result also extends to any rate pair (RX , RY ); however, we only state it for the single-sided constraint.
January 3, 2018
DRAFT
11
+
(0)
(1)
(∗)
D PY |X,V kPY |X PX PV |X ,
min
(∗)
PY |V,X :
(∗)
H
(0)
(∗)
PV |Y
(36a)
(0)
PY =PY ,
(0)
(0)
PỸ ≥H PV |Y PY
(∗)
where PV,Y and PV,Y are the marginals of the distributions defined in (35a) and (35b), respectively. Then, for any
RX > 0,
(0)
(1)
σX,1 (RX ) ≥ D PX PX
+
max
PV |X :
(0)
(0)
I PX|Y ,PV |X PY
≤RX ,
min {σHAN (V ), RX + σSHA (V )} .
(36b)
|V|≤|X |+1
Notice that for PV |X
(0)
such that I PX , PV |X ≤ RX , the bound of the last proposition will be not greater than
the bound of Proposition 4. Therefore the overall bound yields by taking the maximal one.
It is worth pointing out that for distributed rate-distortion problem, the bound in Proposition 5 is in general
suboptimal [23].
A non-trivial outer bound derived by Rahman and Wagner [10] using an additional information at the decoder,
which does not exist in the original problem.
Proposition 6 ([10, Corollary 5]): Suppose that
(1)
(0)
PX = PX .
(37a)
(1)
(0)
Consider a pair of conditional distributions PZ|X,Y and PZ|X,Y such that
(1)
(0)
PZ|X = PZ|X
(37b)
and such that X ↔ Z ↔ Y under the distribution
(1)
△
(1)
(1)
PX,Y,Z = PX,Z PZ|X,Y .
(37c)
Then, for any RX > 0,
(1)
(0)
σX,1 (RX ) ≤ D PY |Z kPY |Z PZ
(0)
(0)
I PY |Z , PX|Y,Z PV |X PZ .
+
max
PV |X :
(0)
(0)
≤RX ,
I PX|Y ,PV |X PY
(37d)
|V|≤|X |+1
B. Specializing to the DSBS
We now specialize the results of Section III-A to the DSBS. Throughout, we choose the auxiliary variable V to
be connected to X by a binary symmetric channel with crossover probability a; with some abuse of notation, we
write e.g. σ(a) for the specialization of σ(V ). Due to symmetry, we conjecture that this choice of V is optimal,
January 3, 2018
DRAFT
12
up to time sharing that can be applied according to Proposition 2; we do not explicitly write the time-sharing
expressions.
The connection between the general and DSBS-specific results can be shown;.However, we follow a direction
that is more relevant to this work, providing for each result a direct interpretation, explaining how it can be obtained
for the DSBS; in doing that, we follow the interpretations of Rahman and Wagner [10].
The Ahlswede-Csiszár scheme of Proposition 3 amounts to quantization of the source X, without using Y as side
information.
Corollary 3 (Proposition 3, DSBS with symmetric auxiliary): For any RX > 0,
σX,1 (RX ) ≥ σAC (δGV (RX )),
(38a)
where
△
σAC (a) = Db (a ∗ p0 ka ∗ p1 ).
(38b)
This performance can be obtained as follows. The encoder quantizes X using a code that is rate-distortion optimal under the Hamming distortion measure; specifically, averaging over the random quantizer, the source and
reconstruction are jointly distributed according to the RDF-achieving test channel, that is, the reconstruction X̂ is
obtained from the source X by a BSC with crossover probability a that satisfies the RDF, namely a = δGV (RX ).
The decision function is φt (X̂, Y) which can be seen as two-stage: first the source difference sequence is estimated
as Ẑ = Y ⊖ X̂, and then a threshold is applied to the weight of that sequence, as if it were the true noise. Notice
that given H = i, Ẑ ∼ BerV (n, a ∗ pi ); the exponents are thus the probabilities of such a vector to fall inside
or outside a Hamming sphere of radius nt around the origin. As Proposition 3 relates to a Stein exponent, the
threshold t is set arbitrarily close to a ∗ p0 , resulting in the following; one can easily generalize to an achievable
exponent region.
The Han scheme of Proposition 4 amounts (for the DSBS) to a similar approach, using a more favorable
quantization scheme. In order to express its performance, we use the following exponent, which is explicitly
evaluated in Appendix A. While it is a bit more general than what we need at this point, this definition will allow
us to present later results in a unified manner.
Definition 6: Fix some parameters p, a, t, w ∈ [0, 1]. Let cn ∈ {0, 1}n, n = 1, 2, . . . be a sequence of vectors
such that limn→∞ δH (cn ) = w. Let Z ∼ BerV (n, p) and let U ∼ Uniform (Tn (a)). Then:
EBT (p, a, w, t) , lim −
n→∞
1
log P (Z ⊕ U ∈ Bn (cn , t)) .
n
(39)
Corollary 4 (Proposition 4, DSBS with symmetric auxiliary): For any RX > 0,
σX,1 (RX ) ≥ σHAN (δGV (RX )),
January 3, 2018
(40a)
DRAFT
13
where
△
σHAN (a) = EBT (p1 , a, 0, a ∗ p0 ) .
(40b)
One can show that σHAN (a) ≥ σAC (a), where the inequality is strict for all p1 < 1/2 (recall that for p1 = 1/2,
“testing against independence”, the Alswhede-Csiszár scheme is already optimal). The improvement comes from
having quantization error that is fixed-type a (recall Definition 5) rather than Bernoulli. Thus, Ẑ is “mixed” uniformBernoulli; the probability of that noise to enter a ball around the origin is reduced with respect to that of the Bernoulli
Ẑ of Corollary 3.
The Shimokawa et al. scheme of Proposition 5 is similar in the DSBS case, except that the compression of X
now uses side-information. Namely, Wyner-Ziv style binning is used. When the bin is not correctly decoded, a
decision error may occur. The resulting performance is given in the following.
Corollary 5 (Proposition 5, DSBS with symmetric auxiliary): For any RX > 0,
σX,1 (RX ) ≥
max
0≤a≤δGV (RX )
min {σHAN (a), σSHA (RX , a)} ,
(41a)
where
△
σSHA (R, a) = R − Hb (a ∗ p0 ) + Hb (a)
(41b)
This exponent can be thought of as follows. The encoder performs fixed-type quantization as in Han’s scheme,
except that the quantization type a is now smaller than δGV (RX ). The indices thus have rate 1 − Hb (a). Now these
indices are distributed to bins; as the rate of the bin index is RX , each bin is of rate 1 − Hb (a) − RX . The decision
function decodes the bin index using the side information Y, and then proceeds as in Han’s scheme.
The two terms in the minimization (41a) represent the sum of the events of decision error combined with bindecoding success and error, respectively. The first is as before, hence the use of σSHA . For the second, it can be
shown that as a worst-case assumption, X̂ resulting from a decoding error is uniformly distributed over all binary
sequences. By considering volumes, the exponent of the probability of the reconstruction to fall inside an nt-sphere
is thus at most 1 − Hb (t); a union bound over the bin gives σSHA .
Remark 1: It may be better not to use binning altogether (thus avoiding binning errors), i.e., the exponent of
Corollary 5 is not always higher than that of Corollary 4.
Remark 2: An important special case of this scheme is when lossless compression is used, and Wyner-Ziv
coding reduces to a side-information case of Slepian-Wolf coding. This amounts to forcing a = 0. If no binning
error occurred, we are in the same situation as in the unconstrained case. Thus, we have the exponent:
min (Db (p0 kp1 ), σSHA (RX )) ,
(42a)
where
△
σSHA (R) = σSHA (R, 0) = R − Hb (p0 ).
January 3, 2018
(42b)
DRAFT
14
Coding component
Oblivious to Y
Using side-information Y
Lossless
TS
Bin + TS
TABLE I
S UMMARY OF POSSIBLE SCHEMES . TS STANDS FOR TIME - SHARING , Q
Lossy
Q + TS [4, 5]
Q + Bin + TS [6]
STANDS FOR QUANTIZATION , B IN STANDS FOR BINNING .
We have seen thus that various combinations of quantization and binning; Table III-B summarizes the different
possible schemes.
An upper bound is obtained by specializing the Rahman-Wagner outer bound of Proposition 6 to the DSBS.
Corollary 6 (Proposition 6, DSBS with symmetric additional information):
△
σRW (R, ζ, a) = min min max
0≤ζ≤p0 0≤b1 ≤1 0≤a≤1/2
b1 · ζ · (1 − a) + b0 · (1 − ζ) · a
Hb (γ) − ζ ∗ aHb
ζ ∗a
b1 · ζ · a + b0 · (1 − ζ) · (1 − a)
+ (1 − ζ ∗ a)Hb
1−ζ ∗a
p0 − a
Db b0 · (1 − a) + b1 · a
,
1 − 2a
(43)
△ p1 −a·(1−b1 )
.
1−a
where b0 =
We note that it seems plausible that the exponent for p1 = 1/2, given by Corollary 4, is an upper to general p1 ,
i.e.,
σX,1 (RX ) ≤ 1 − Hb (p0 ∗ δGV (R)).
(44)
Next we compare the performance of these coding schemes in order to understand the effect of each of the
different components of the coding schemes on the performance.
IV. BACKGROUND : L INEAR C ODES
AND
E RROR E XPONENTS
In this section we define code ensembles that will be used in the sequel, and present their properties. Although
the specific properties of linear codes are not required until Section VI, we put an emphasis on such codes already;
this simplifies the proofs of some properties we need to show, and also helps to present the different results in a
more unified manner.
A. Linear Codes
Definition 7 (Linear Code): We define a linear code via a k × n generating matrix G over the binary field. This
induces the linear codebook:
C = {c : c = uG, u ∈ {0, 1}k },
January 3, 2018
(45)
DRAFT
15
where u ∈ {0, 1}k is a row vector.
Assuming that all rows of G are linearly independent, there are 2k codewords in C, so the code rate is
R=
k
.
n
(46)
Clearly, for any rate (up to 1), there exists a linear code of this rate asymptotically as n → ∞.
A linear code is also called a parity-check code, and may be specified by a (n − k) × n (binary) parity-check
matrix H. The code C contains all the n-length binary row vectors c whose syndrome
△
s = cHT
(47)
is equal to the n − k all zero row vector, i.e.,
△
C = c ∈ {0, 1}n : cHT = 0 .
(48)
Given some general syndrome s ∈ {0, 1}n−k , denote the coset of s by
△
Cs = {x ∈ {0, 1}n : xHT = s}.
(49)
The minimum Hamming distance quantizer of a vector x ∈ {0, 1}n with respect to a code C ⊆ {0, 1}n is given
by
△
QC (x) = arg min δH (x ⊖ c).
c∈C
(50)
For any syndrome s with respect to the code C, the decoding function fC (s) : {0, 1}n−k 7→ {0, 1}n gives the
coset leader, the minimum Hamming weight vector within the coset of s:
△
fC (s) = arg min δH (z).
z∈Cs
(51)
Maximum-likelihood decoding of a parity-check code, over a BSC Y = X ⊕ Z, amounts to syndrome decoding
x̂ = y ⊖ fC (y) [24, Theorem 6.1.1]. The basic “Voronoi” set is given by
△
Ω0 = z : z ⊖ fC (zHT ) = 0 .
(52)
The ML decision region of any codeword c ∈ C is equal to a translate of Ω0 , i.e.,
△
Ωc = y : y ⊖ fC (yHT ) = c
= Ω0 + c.
(53)
(54)
B. Properties of Linear Codes
Definition 8 (Normalized Distance Distribution): The normalized distance (or weight) distribution of a linear
code C for a parameter 0 ≤ w ≤ 1 is defined to be the fraction of codewords c 6= 0, with normalized weight at
January 3, 2018
DRAFT
16
most w, i.e.,
△
ΓC (w) =
1
|C|
X
1{δH (c)≤w} ,
(55)
c∈C\{0}
where 1{·} is the indicator function.
Definition 9 (Normalized Minimum Distance): The normalized minimum distance of a linear code C is defined
as
△
δmin (C) = min δH (c)
c∈C\{0}
(56)
Definition 10 (Normalized Covering Radius): The normalized covering radius of a code C ∈ {0, 1}n is the
smallest integer such that every vector x ∈ {0, 1}n is covered by a Hamming ball with radius r and center at some
c ∈ C, normalized by the blocklength, i.e.:
△
ρcover (C) = max n min δH (x ⊖ c).
x∈{0,1}
c∈C
(57)
Definition 11 (Normalized Packing Radius): The normalized packing radius of a linear code C is defined to be
half the normalized minimal distance of its codewords, i.e.,
△
ρpack (C) =
1
δmin (C).
2
(58)
C. Good Linear Codes
We need two notions of goodness of codes, as follows.
Definition 12 (Spectrum-Good Codes): A sequence of codes C (n) ⊆ {0, 1}n, n = 1, 2, . . . with rate R is said to
be spectrum-good if for any w ≥ 0,
. (n)
ΓC (n) (w) = ΓR (w)
where
(n)
ΓR (w)
2−nDb (wk1/2) , w > δ (R)
GV
=
.
0,
otherwise
(59)
Definition 13 (Covering-Good): A sequence of codes C (n) ⊆ {0, 1}n, n = 1, 2, . . . with rate R is said to be
covering-good if
ρcover (C (n) ) −→ δGV (R)
n→∞
(60)
The existence of linear codes satisfying these properties is well known. Specifically, consider the ensemble of
constructed by random generating matrices, where each entry of the matrix is drawn uniformly and statistically
independent with all other entries, then almost all members have a spectrum close to (59), see e.g. [24, Chapter
5.6-5.7]; in addition, for almost all members, a process of appending rows to the generating matrix (with vanishing
January 3, 2018
DRAFT
17
rate) results in a normalized covering radius close to δGV [25, Theorem 12.3.5]. These existence arguments are
detailed in Appendix C. We need the following properties of good codes.
Spectrum-good codes obtain the best known error exponent for the BSC. Namely, for a BSC with crossover
probability p, they achieve E BSC (p, R), given by
△
E BSC (p, R) = max {Er (p, R), Eex (p, R)} ,
(61a)
1
1
△
Er (p, R) = max ρ − (1 + ρ) log p 1+ρ + (1 − p) 1+ρ − ρR
(61b)
where
ρ∈[0,1]
is the random-coding exponent, and
△
Eex (p, R) = max −ρ log
ρ≥1
i ρ1
1 1h p
− ρR.
2 p(1 − p)
+
2 2
(61c)
is the expurgated exponent. Notice that as the achievability depends only upon the spectrum, it is universal in p.
As for covering-good codes, we need the following result, shown that the quantization noise induced by coveringgood codes is no worse than a noise that is uniform over an nδGV -Hamming ball.
Lemma 1: Consider a covering-good sequence of codes C (n) ⊆ {0, 1}n, n = 1, 2, . . . of rate R. Then,
X ⊖ QC (n) (X) 6 N,
(62a)
X ∼ Uniform ({0, 1}n)
N∼ Uniform Bn 0, ρcover (C (n) ) .
(62b)
D
for
(62c)
Furthermore, the same holds when adding any random vector to both sides, i.e., for any Z:
X ⊖ QC (n) (X) ⊕ Z 6 N ⊕ Z,
(63)
D
The proof appears in Appendix B.
D. Nested Linear Codes
We briefly recall some basic definitions and properties related to nested linear codes. The reader is referred to
[26] for further details.
Definition 14 (Nested Linear Code): A nested linear code with rate pair (R1 , R2 ) is a pair of linear codes (C1 , C2 )
with these rates, satisfying
C2 ⊆ C1 ,
January 3, 2018
(64)
DRAFT
18
i.e., each codeword of C2 is also a codeword of C1 (see [26]). We call C1 and C2 fine code and coarse code,
respectively.
If a pair {(n, k1 ), (n, k2 )} of parity-check codes, k1 ≥ k2 , satisfies condition (64), then the corresponding paritycheck matrices H1 and H2 are interrelated as
HT2
|{z}
=[
(n−k2 )×n
HT1 ,
|{z}
∆HT ],
| {z }
(65)
(n−k1 )×n (k1 −k2 )×n
where H1 is an (n − k1 ) × n matrix, H2 is an (n − k2 ) × n matrix, and ∆H is a (k1 − k2 ) × n matrix. This implies
that the syndromes s1 = xHT1 and s2 = xHT2 associated with some n-vector x are related as s2 = [s1 , ∆s], where
the length of ∆s is k1 − k2 bits. In particular, if x ∈ C1 , then s2 = [0, . . . , 0, ∆s]. We may, therefore, partition C1
into 2k1 −k2 cosets of C2 by setting s1 = 0, and varying ∆s, i.e.,
[
C1 =
C2,s2 ,
where s2 = [0, ∆s].
(66)
∆s∈{0,1}k1 −k2
Finally, for a given pair of nested codes, the “syndrome increment” ∆s is given by the function
∆s = x · ∆HT .
(67)
Proposition 7: Let the syndrome increment ∆s be computed for c ∈ C1 . Then, the coset leader corresponding to
the syndrome of c with respect to C2 is given by
fC2 (cHT1 ) = fC2 ([0, ∆s]),
(68)
where 0 is a zero row vector of length n − k1 .
For a proof, see, e.g., [26].
Definition 15 (Good Nested Linear Code): A sequence of nested linear codes with rate pair (R1 , R2 ) is said to
be good if the induced sequences of fine and coarse codes are covering-good and spectrum-good, respectively.
The existence of good nested linear codes follows naturally from the procedures used for constructing spectrumgood and covering-good codes; see Appendix C for a proof.
We need the following property of good nested codes.
(n)
(n)
Corollary 7: Consider a sequence of good nested codes (C1 , C2 ), n = 1, 2, . . . with a rate pair (R1 , R2 ). Let
△
X ∼ Uniform ({0, 1}n) and Z ∼ BerV (n, p) be statistically independent. Denote U = QC (n) (X), where quantization
1
with respect to a code is defined in (50). Then,
P QC (n) (X ⊕ Z) 6= U ≤ P QC (n) U ⊕ N′ ⊕ Z 6= U
2,S
(69)
2,S
△
where N′ ∼ Uniform (Bn (0, δGV (R1 ))) is statistically independent of (U, Z), where S = UHT2 , and where H2 is a
(n)
parity check matrix of the coarse code C2 .
January 3, 2018
DRAFT
19
The proof, which relies upon Lemma 1, is given in Appendix B.
E. Connection to Distributed Source Coding
As the elements used for the schemes presented (quantization and binning) are closely related to distributed
compression, we present here some material regarding the connection of the ensembles presented to such problems.
A covering-good code achieves the rate-distortion function of a binary symmetric source with respect to the
Hamming distortion measure, which amounts to a distortion of δGV (R). Furthermore, it does so with a zero error
probability.
A spectrum-good code is directly applicable to the Slepian-Wolf (SW) problem [8], where the source is DSBS.
Specifically, a partition of all the binary sequences into bins of rate Rbin can be performed by a code of rate
R = 1 − Rbin (which can be alternatively be seen as a nested code with R + Rbin = 1), and the SW decoder can
be seen as a channel decoder where the input codebook is the collection of sequences in the bin and the channel
output is Y n , see [27]. Thus, it achieves the exponent
E BSC (p, Rbin ) = E BSC (p, 1 − R)
As in channel coding, this error exponent is achievable universally in p.
The achievability of the random-coding and expurgated exponents for the general discrete SW problem was
established by Csiszár et al. [28] and [29],10 Indeed, the connection between channel coding and SW coding is
fundamental (as already noted in [27]), and the optimal error exponents (if they exist) are related, see [31–33].
Nested codes are directly applicable to the Wyner-Ziv (WZ) problem [9], where the source is DSBS and under
the Hamming distortion measure, see [26]. When a good ensemble is used, the exponent of a binning error event is
at least E BSC (p, Rbin ). As the end goal of the scheme is to achieve low distortion with high probability, the designer
has the freedom to choose Rbin that strikes a good balance between binning errors and other excess-distortion events,
see [34].
V. O NE -S IDED C ONSTRAINT: N EW R ESULT
In this section we present new achievable exponent tradeoffs for the same one-sided case considered in the
previous section. To that end, we will employ the same binning strategy of Corollary 5. However, our analysis
technique allows to improve the exponent, and to extend it from the Stein setting to the full tradeoff.
For our exponent region, we need the following exponent. It is a variation upon EBT (Definition 6), where the
fixed-type noise is replaced by a noise uniform over a Hamming ball.
10 Indeed, Csiszár has already established the expurgated exponent for a class of additive source-pairs which includes the DSBS in [29].
However, as the derivation was for general rate pairs rather than for the side-information case, it faced inherent difficulty in expurgation in a
distributed setting. This was solved by using linear codes; see [30] for a detailed account in a channel-coding setting.
January 3, 2018
DRAFT
20
Definition 16: Fix some parameters p, a, t, w ∈ [0, 1/2]. Let cn ∈ {0, 1}n, n = 1, 2, . . . be a sequence of vectors
such that limn→∞ δH (cn ) = w. Let Z ∼ BerV (n, p) and let N∼ Uniform (Bn (0, a)). Define
EBB (p, a, w, t) , lim −
n→∞
1
log P (N ⊕ Z ∈ Bn (cn , t)) .
n
(70)
The following can be shown using standard type considerations.
Lemma 2:
EBB (p, a, w, t) = −Hb (a) + min [Hb (r) + EBT (p, r, w, t)]
0≤r≤a
(71)
We are now ready to state the main result of this section.
Theorem 1 (Binary Hypothesis Testing with One-Sided Constraint): Consider the hypothesis testing problem as
defined in Section II-A for the DSBS, with a rate constraint RX ∈ [0, 1]. For any parameters a ∈ [0, δGV (RX )] and
t ∈ [a ∗ p0 , a ∗ p1 ],
(SI)
(a,
t;
p
,
R
),
E
(a,
t;
p
,
R
)
∈ CX (RX ),
E (SI)
0
X
1
X
0
1
(72)
where
(a, t; p0 , RX ) = min EBB (p0 , a, 1, 1 − t) , E BSC (a ∗ p0 , Rbin )
△
(a,
t;
p
,
R
)
=
min
E
(p
,
a,
0,
t)
,
E
(p
,
a,
t,
R
)
,
E (SI)
1
X
BB
1
c
1
bin
1
△
E (SI)
0
(73a)
(73b)
where
△
Ec (p1 , a, t, Rbin ) = max
n
− Rbin +
min
δGV (Rbin )<w≤1
o
Db (wk1/2) + EBB (p1 , a, w, t) , E BSC (a ∗ p1 , Rbin ) ,
(74)
and where
△
Rbin = 1 − Hb (a) − RX
(75)
and EBB (p, a, w, t) is defined in Definition 16.
We prove this theorem using a quantize-and-bin strategy similar to that of Corollary 5, implemented with good
nested codes as defined in Section IV-D. In each of the two minimizations in (73), the first term is a bound on
the exponent of a decision error resulting from a bin-decoding success, while the second is associated with a bindecoding error, similar to the minimization in (41a). Using the properties of good codes, we provide a tighter and
more general (not only a Stein exponent) bound; the key part is the derivation of Ec , the error exponent given H1 ,
and given a bin-decoding error: we replace the worst-case assumption that the “channel output” is uniform over
all binary sequences by the true distribution, centered around the origin; in particular, it means that given an error,
points close to the decision region of the correct codeword, thus not very close to any other codeword, are more
likely.
January 3, 2018
DRAFT
21
After the proof, we remark on the tightness of this result.
Proof: For a chosen a, denote
△
RQ = δGV −1 (a)
(76a)
= 1 − Hb (a)
(76b)
= RX + Rbin .
(76c)
(n)
(n)
Consider a sequence of linear nested codes (C1 , C2 ), n = 1, 2, . . . with rate pair (RQ , Rbin ), which is good
in the sense of Definition 15. For convenience, the superscript of the code index in the sequence will be omitted.
The scheme we consider uses structured quantization and binning. Specifically, given a vector X, we denote its
quantization by
△
U = QC1 (X).
(77)
Note that we can decompose Y as follows:
Y= X⊕Z
= U ⊕ N ⊕ Z,
(78a)
(78b)
where the quantization noise N = X ⊖ U is independent of U (and of course also of Z) since X is uniformly
distributed.
For sake of facilitating the analysis of the scheme, we also define
Y′ = U ⊕ N′ ⊕ Z,
(79a)
N′ ∼ Uniform (Bn (0, a))
(79b)
where
is independent of the pair (U, Z). That is, Y′ satisfies the same relations with U as Y, except that the quantization
noise N is replaced by a spherical noise with the same circumradius.
The encoded message is the syndrome increment (recall (67)) of U, i.e.,
φX (X) = ∆S
= U · ∆H T .
(80a)
(80b)
Since the rates of C1 and C2 are RQ and Rbin , respectively, the encoding rate is indeed RX .
Let
△
S = UH2T .
January 3, 2018
DRAFT
22
By Proposition 7, since U ∈ C1 , the decoder can recover from ∆S the coset C2,S of syndrome S.
The reconstructed vector at the decoder is given by
△
Û = QC2,S (Y),
(81)
△
Ŵ = δH Y ⊖ Û .
(82)
Denote
After computing Ŵ , given a threshold t ∈ [a ∗ p0 , a ∗ p1 ], the decision function is given by
△
ψ(φX (X), Y) = ϕt (Ŵ ),
(83a)
where ϕt (w) is the threshold function (28).
Denote
△
W = δH (Y ⊖ U).
(84)
△
Denote the decoding error event by EC = {Û 6= U} and the complementary event by Ec . Using elementary
probability laws, we can bound the error events given the two hypotheses as follows:
ǫ0 = P Ŵ > t H = 0
= P Ec , Ŵ > t H = 0 + P E c , Ŵ > t H = 0
≤ P Ec , Ŵ > t H = 0 + P (W > t|H = 0)
≤ P (Ec |H = 0) + P (W ≥ t|H = 0) .
(85a)
(85b)
(85c)
(85d)
And similarly,
ǫ1 = P Ŵ ≤ t H = 1
= P Ec , Ŵ ≤ t H = 1 + P E c , Ŵ ≤ t H = 1
≤ P Ec , Ŵ ≤ t H = 1 + P (W ≤ t|H = 1) .
(86a)
(86b)
(86c)
Comparing with the required exponents (73), it suffices to show the following four exponential inequalities.
(87a)
(87b)
(87c)
P (W ≥ t|H = 0) ≤ EBB (p0 , a, 1, 1 − t)
P (Ec |H = 0) ≤ E BSC (a ∗ p0 , Rbin )
P (W ≤ t|H = 1) ≤ EBB (p1 , a, 0, t)
P Ec , Ŵ ≤ t H = 1 ≤ Ec (p1 , a, t, Rbin )
January 3, 2018
(87d)
DRAFT
23
In the rest of the proof we show these. For (87a), we have:
P (W ≥ t|H = 0)
(88a)
= P (δH (Y ⊖ U) ≥ t|H = 0)
(88b)
= P (N ⊕ Z ∈
/ Bn (0, t)|H = 0)
(88c)
= P (N ⊕ Z ∈ Bn (1, 1 − t)|H = 0)
(88d)
≤ P N′ ⊕ Z ∈ Bn (1, 1 − t) H = 0
.
= 2−nEBB(p0 ,a,1,1−t) ,
(88e)
(88f)
where 1 is the all-ones vector, (88c) follows by substituting (77), the transition (88e) is due to Lemma 1 and the
last asymptotic equality is due to Definition 16. The proof of (87c) is very similar and is thus omitted.
For (87b), we have:
P (EC |H = 0)
= P Û 6= U H = 0
(89a)
(89b)
= P QC2,S (X ⊕ Z) 6= QC1 (X) H = 0
≤ P QC2,S QC1 (X) ⊕ N′ ⊕ Z 6= QC1 (X) H = 0
= P QC2,S U ⊕ N′ ⊕ Z 6= U H = 0
= P QC2,S (Y′ ) 6= U H = 0
≤ 2−nE BSC (a∗p0 ,Rbin ) ,
(89c)
(89d)
(89e)
(89f)
(89g)
where (89d) is due to Corollary 7, the last inequality follows from the spectrum-goodness of the coarse code C2
and Y′ was defined in (79a). Notice that the channel exponent is with respect to an i.i.d. noise, but it is easy to
show that the exponent of a mixed noise can only be better.
Lastly, For (87d) we have:
P EC , Ŵ ≤ t H = 1
= P Û 6= U, Ŵ ≤ t H = 1
= P Û 6= U, Y ∈ Bn Û, t H = 1
o
n
[
= P
Û = c, Y ∈ Bn (c, t) H = 1
(90a)
(90b)
(90c)
(90d)
c∈C2,S \{U}
= P Û 6= U, Y ∈
January 3, 2018
[
c∈C2,S \{U}
Bn (c, t) H = 1
(90e)
DRAFT
24
≤ max
P (EC |H = 1) , P Y ∈
Bn (c, t) H = 1 .
[
c∈C2,S \{U}
(90f)
Due to the spectrum-goodness of the coarse code C2 , the first term in the maximization is exponentially upperbounded by
2−nE BSC (a∗p1 ,Rbin ) .
For the second term, we proceed as follows.
[
Bn (c, t) H = 1
P Y ∈
(91a)
c∈C2,S \{U}
≤ P Y′ ∈
[
c∈C2,S \{U}
= P Y ⊖ U ∈
= P Y′ ⊖ U ∈
= P N′ ⊕ Z ∈
≤
X
Bn (c, t) H = 1
[
′
c∈C2,S \{U}
[
c∈C2 \{0}
[
c∈C2 \{0}
Bn (c ⊖ U, t) H = 1
Bn (c, t) H = 1
(91d)
Bn (c, t) H = 1
(91e)
P N′ ⊕ Z ∈ Bn (c, t) H = 1
X
X
P N′ ⊕ Z ∈ Bn (c, t) H = 1
X
X
2−nEBB(p1 ,a,j/n,t)
nδGV (Rbin )≤j≤n c∈C2 :nδH (c)=j
.
=
(91c)
c∈C2 \{0}
=
(91b)
(91f)
(91g)
(91h)
nδGV (Rbin )≤j≤n c∈C2 :nδH (c)=j
=
X
|C2 | · ΓC0 (w) · 2−nEBB (p1 ,a,j/n,t)
(91i)
X
2nRbin · 2−nDb (j/nk1/2) · 2−nEBB (p1 ,a,j/n,t)
(91j)
nδGV (Rbin )≤j≤n
≤
nδGV (Rbin )≤j≤n
.
= 2−n[−Rbin +minδGV (Rbin )<w≤1 Db (wk1/2)+EBB (p1 ,a,w,t)]
(91k)
= 2−nEc (p1 ,a,t,Rbin ) ,
(91l)
where (91b) is due to Lemma 1, (91f) is due to the union bound, the lower limit in the outer summation in (91g) is
valid since spectrum-good codes have no non-zero codewords of lower weight, in (91j) we substituted the spectrum
of a spectrum-good code, and in (91k) we substitute w = j/n and use Laplace’s method.
At this point we remark on the tightness of the analysis above.
January 3, 2018
DRAFT
25
Remark 3: There are two points where our error-probability analysis can be improved.
1) For the exponent of the probability of bin-decoding error (under both hypotheses) we used E BSC (a ∗ pi , Rbin ).
However, one may use the fact the quantization noise is not Bernoulli but rater bounded by a sphere to derive
a larger exponent.
2) In (90e) we have the probability of a Bernoulli noise to fall within a Hamming ball around some non-zero
codeword, and also outside the basic Voronoi cell. In the transition to (90f) we bound this by the maximum
between the probabilities of being in the Hamming balls and being outside the basic Voronoi cell.
While solving the first point is straightforward (though cumbersome), the second point (exponentially tight evaluation
of (90e)) is an interesting open problem, currently under investigation. We conjecture that except for these two points,
our analysis of this specific scheme is exponentially tight.
Remark 4: In order to see that the encoder can be improved, consider the Stein-exponent bound, obtained by
setting t = a ∗ p0 in (73b):
σX,1 (RX ) ≥ min EBB (p1 , a, 0, a ∗ p0 ) , Ec (p1 , a, a ∗ p0 , RX ) ,
(92)
cf. the corresponding expression of the scheme by Shimokawa et al. (41a),
min EBT (p1 , a, 0, a ∗ p0 ) , σSHA (RX , a) .
Now, one can show that we have an improvement of the second term. However, clearly by (71), EBB (p1 , a, 0, a ∗ p0 ) ≤
EBT (p1 , a, 0, a ∗ p0 ). That is, quite counterintuitively, a quantization noise that has always weight a is better than
one that may be smaller. The reason is that a “too good” quantization noise may be confused by the decision
function with a low crossover probability between X and Y , favoring Ĥ = 0. In the Stein case, where we do not
care at all about the exponent of ǫ0 , this is a negative effect. We can amend the situation by a simple tweak: the
encoder will be the same, except that when it detects a quantization error that is not around a it will send a special
symbol forcing Ĥ = 1. It is not difficult to verify that this will yield the more favorable bound
σX,1 (RX ) ≥ min EBT (p1 , a, 0, a ∗ p0 ) , Ec (p1 , a, a ∗ p0 , RX ) .
(93)
A similar process, where if the quantization noise is below some threshold Ĥ = 1 is declared, may also somewhat
extend the exponent region in the regime “close” to Stein (low E0 ), but we do not pursue this direction.
Remark 5: Of course, the two-stage decision process where Ĥ is a function of Ŵ is sub-optimal. It differs from
the Neyman-Pearson test that takes into account the probability of all possible values of W .
Remark 6: Using time sharing on top of the scheme, we can obtain improved performance according to Proposition 2.
Remark 7: In the special case a = 0 the scheme amounts to binning without quantization, and the nested code
maybe replaced by a single spectrum-good code. In this case the expressions simplify considerably, and we have
January 3, 2018
DRAFT
26
the pair:
(0,
t;
R
)
=
min
D
(tkp
),
E
(p
,
R
)
E (SI)
X
b
0
0
bin
BSC
0
△
(SI)
E 1 (0, t; RX ) = min Db (tkp1 ), Ec (p1 , 0, t, Rbin ) ,
(94a)
(94b)
where
Ec (p1 , 0, t, Rbin ) = max
n
− Rbin +
min
δGV (RX )<w≤1
o
Db (wk1/2) + EBB (p1 , 0, w, t) , E BSC (p1 , Rbin ) ,
(95)
and where EBB (p1 , 0, w, t) is defined in (70).
VI. S YMMETRIC C ONSTRAINT
In this section we proceed to a symmetric rate constraint RX = RY = R. In this part our analysis specifically
hinges on the linearity of codes, and specifically builds on the Körner-Marton coding scheme [18]. Adding this
ingredient to the analysis, we get an achievable exponent region for the symmetric constraint in the same spirit of
the achievable region in Theorem 1, where the only loss due to constraining RY is an additional spherical noise
component. Before stating our result, we give some background on the new ingredient.
A. Körner-Marton Compression
The Körner-Marton problem has the same structure as our DHT problem for the DSBS, except that the crossover
probability is known (say p), and the decision function is replaced by a decoder, whose goal is to reproduce the
difference sequence Z = Y ⊖ X with high probability. By considering the two corresponding one-sided constrained
problems, which amount to SI versions of the SW problem, it is clear that any rate R < Hb (p) is not achievable.
The Körner-Marton scheme allows to achieve any rate R > Hb (p) in the following manner.
Assume that the two encoders use the same linear codebook, with parity-check matrix H. Further, both of them
send the syndrome of their observed sequence:
φX (X) = XHT
(96a)
φY (Y) = YHT .
(96b)
In the first stage of the decoder, the two encoded vectors are summed up, leading to
φY (Y) ⊖ φX (X) = YHT ⊖ XHT
= ZHT ,
(97a)
(97b)
which is but the syndrome of the difference sequence. This is indistinguishable from the situation of a decoder for
the SI SW problem,
The decoder is now in the exact same situation as an optimal decoder for a BSC with crossover probability p,
and code rate 1 − R. By the fact that linear codes allow to approach the capacity 1 − Hb (p), the optimal rate
January 3, 2018
DRAFT
27
follows. Further, if spectrum-good codes are used, the exponent E BSC (p, 1 − R) is achievable, i.e., there is no loss
in the exponent w.r.t. the corresponding side-information SW problem.
B. A New Bound
We now present an achievability result that relies upon a very simple principle:as in the Körner-Marton decoder,
after performing the XOR (97) the situation is indistinguishable from that of the input to a SW decoder, also in
DHT we can
Theorem 2 (Binary Hypothesis Testing with Symmetric Constraint): Consider the hypothesis testing problem as
defined in Section II-A for the DSBS, with a symmetric rate constraint R ∈ [0, 1]. For any parameter t ∈ [p0 , p1 ],
(KM)
(t;
R),
E
(t;
R)
∈ C (R),
E (KM)
0
1
(98)
where
E (KM)
(t; R) = E (SI)
0
0 (0, t; R)
(99a)
(t; R) = E (SI)
E (KM)
1
1 (0, t; R),
(99b)
where the one-sided constraint exponents with a = 0 are given in (94).
Proof: Let the codebook be taken from a spectrum-good sequence. Let the encoders be the Körner-Marton
encoders (96). The decision function first obtains ZH T as in the Körner-Marton decoder (97), and then evaluates
Ẑ = QC (ZH T )
and applies the threshold function to
△
W ′ = δH Ẑ .
Noticing that W ′ is equal in distribution to Ŵ in the proof of Theorem 1 when a = 0, and that all error events
only functions of that variable, the proof is completed.
It is natural to ask, why we restrict ourselves under a symmetric rate constraint to a binning-only scheme. Indeed,
lossy versions of the Körner-Marton problem have been studied in [35, 36]. One can construct a scheme based on
nested codes, obtain a lossy reconstruction of the noise sequence Z and then test its weight. However, unlike the
reconstruction in the single-sided case which includes a Bernoulli component (Z) and a quantization noise bounded
by a Hamming ball (N), in the symmetric-rate case we will have a combination of a Bernoulli component with two
quantization noises. This makes the analysis considerably more involved. An idea that comes to mind, is to obtain
a bound by replacing at least one of the quantization noises with a Bernoulli one; however, we do not see a clear
way to do that. Thus, improving Theorem 2 by introducing quantization is left for future research.
VII. P ERFORMANCE C OMPARISON
In this section we provide a numerical comparison of the different bounds for the DSBS.
January 3, 2018
DRAFT
28
We start with the Stein setting, where we can compare with the previously-known results. Our achievable performance for the one-sided constrained case is given by (93) (which coincides with (92) for the parameters we checked).
We compare it against the unconstrained performance, and against the previously best-known achievable exponent,
namely the maximum between Corollaries 4 and 5.11 To both, we also apply time sharing as in Proposition 2. It can
be seen that the new exponent is at least as good, with slight improvement for some parameters. As reference, we
show the unconstrained performance, given by Corollary 1. Also shown on the plots, is the performance obtained
under a symmetric rate constraint, found by constraining the quantization parameter to a = 0. It can be seen that
for low p1 the symmetric constraint yields no loss with respect to the one-sided constraint.
0.4
0.14
0.35
0.12
0.3
0.1
0.25
0.08
0.2
0.06
0.15
0.04
0.1
0.02
0.05
0
0
0.02
0.04
0.06
0.08
0.1
p0
0.12
0.14
(a) p1 = 0.25
0.16
0.18
0.2
0
0
0.01
0.02
0.03
0.04
0.05
p
0.06
0.07
0.08
0.09
0
(b) p1 = 0.1. The three top curves coincide.
Fig. 2. Stein exponent comparison. The rate is 0.3 bits, the X axis is p0 . From top to bottom: unconstrained performance, new exponent,
previously known exponent, new exponent without quantization (achievable with symmetric rate).
Beyond the Stein setting, we plot the full exponent tradeoff, achievable by Theorems 1 and 2. In this case we are
not aware of any previous results we can compare to. We thus only add the unconstrained tradeoff of Corollary 1,
and the simple strategy of Corollary 2. Also here it can be seen that the symmetric constraint imposes a loss for
high p1 , but not for a lower one.
VIII. C ONCLUSIONS
In this work we introduced new achievable error exponents for binary distributed hypothesis testing for binary
symmetric i.i.d sources. One may wonder, naturally, regarding the extension beyond the binary symmetric case.
In that respect, a distinction should be made between two parts of the work. Under a one-sided rate constraint,
linear codes were used merely for concreteness and for convenience of extension to the symmetric constraint; the
same results should hold for a random code ensemble. Thus, there is no fundamental problem in extending our
analysis to any discrete memoryless model.
11 In [7], the performance is evaluated using an asymmetric choice of the auxiliary variable. We have verified that the symmetric choice we
use performs better.
January 3, 2018
DRAFT
29
0.35
0.1
0.09
0.3
0.08
0.25
0.07
0.06
E1
E1
0.2
0.05
0.15
0.04
0.03
0.1
0.02
0.05
0.01
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0
0.8
0
0.02
0.04
0.06
0.08
E0
0.1
E0
0.12
0.14
0.16
0.18
0.2
(b) p1 = 0.1. The new tradeoff curves with and without quantization
coincide.
(a) p1 = 0.25
Fig. 3. Exponent tradeoff comparison. The rate is 0.3 bits, p0 = 0.01. From top to bottom: unconstrained tradeoff, new tradeoff, new tradeoff
without quantization (achievable with symmetric rate), time-sharing only tradeoff.
In contrast, in the setting of a symmetric rate constraint, we explicitly use the “matching” between the closedness
under addition of linear codes, and the additivity of the relation between the sources. Thus, our approach which
has almost no loss with respect to the single-sided constraint, cannot be extended beyond (nearly) additive cases.
The question whether a different approach can achieve that, remains open.
Finally, we stress again the lack of tight outer bounds for the problem, except for cases where the communication
constraints do not limit the exponents.
ACKNOWLEDGMENT
The authors thank Uri Erez for sharing insights throughout the work. They also thank Vincent Y. F. Tan for
introducing them the distributed hypothesis testing problem, and Nir Weinberger for helpful discussions.
A PPENDIX A
E XPONENT OF
A
H AMMING BALL
In this appendix we evaluate the exponent of the event of a mixed noise entering a Hamming ball, namely
EBT (p, a, w, t) of Definition 6.
Lemma 3:
EBT (p, a, w, t) =
min
Hb (a) − wHb
γ∈[max(0,a+w−1),min(w,a)]
γ
w
− (1 − w)Hb
a−γ
1−w
+ Ew (p, 1 − (w + a − 2γ), w + a − 2γ, t − (w + a − 2γ))
(100)
where
△
Ew (p, α, β, t) =
January 3, 2018
min
x∈[max(0,t),min(α,β+t)]
αDb
x
α
p + βDb
x−t
p
β
(101)
DRAFT
30
The proof follows from the lemmas below.
Lemma 4 (Difference of weights): Let Z1 and Z2 be two random vectors. Assume that Zi ∼ BerV (ki , p) and
let Wi = wH (Zi ) for i = 1, 2. If ki grow with n such that
k1
= α,
n
k2
= β,
lim
n→∞ n
lim
n→∞
and further let a sequence t grow with n such that
lim
n→∞
t
=τ
n
where τ ∈ (−β, α). Then,
lim −
n→∞
1
log P (W1 − W2 = t) = Ebb (p, α, β, τ ),
n
where Ew (p, α, β, τ ) is given by (101).
Proof:
P (W1 − W2 = t) =
k1
X
P (W1 = w) P (W2 = w − t)
(102a)
w=0
min(k1 ,k2 +t)
X
=
P (W1 = w) P (W2 = w − t)
(102b)
w=max(0,t)
.
=
min(k1 ,k2 +t)
X
2
−k1 Db
p −k2 Db w−t
p
k
w
k1
2
(102c)
2
w=max(0,t)
.
=
.
=
max
w∈[max(0,t),...,min(k1 ,k2 +t)]
2
−k1 Db
x∈[max(0,τ ),min(α,β+τ )]
w
k1
p −k2 Db w−t
p
k
2
2−n[αDb ( α kp)+βDb (
x
max
x−τ
β
2
(102d)
kp)] ,
(102e)
where (102c) follows by the exponent of the probability of a type class.
The following lemma will assist in proving Lemma 6 which follows.
Lemma 5 (Mixed noise, fixed dimension): Let n ∈ N, and consider a noise that is a mixture of a noise U ∼
Uniform (Tn (a)) (uniform over a fixed type) and a Bernoulli vector Z ∼ BerV (n, p), where a ∈ 1/n · {0, . . . , n}
and p ∈ [0, 1]. Further let c ∈ {0, 1}n where w = δH (c) ∈ 1/n · {0, . . . , n}, be the center of a “distant” sphere.
Then, for a sphere radius with normalized radius t ∈ 1/n · {0, . . . , n},
n·min(w,a)
P (wH (c ⊕ U ⊕ Z) = nt) =
X
m=n·max(0,a+w−1)
nw
m
n−nw
na−m
n
na
P (W1,m − W2,m = nt − nw − na + 2m)
(103)
where W1,m ∼ Binomial (n − (nw + na − 2m), p) and W2,m ∼ Binomial (nw + na − 2m, p).
January 3, 2018
DRAFT
31
Proof: Define the following sets of indices I, M1 , M2 :
△
I = {i : ci = 1}
(104a)
△
M1 = {i : UI,i = 1}
n
o
△
M2 = i : UI,i = 1 .
(104b)
(104c)
In words, I is the set of indices where c contains ones; M1 is the subset (within I) where U contains ones and
M2 is defined similarly over the complement of I.
Then,
P (wH (c ⊕ U ⊕ Z = nt)) = P (wH (cI ⊕ UI ⊕ ZI ) + wH (cI ⊕ UI ⊕ ZI ) = nt)
(105a)
= P (nw − wH (UI ⊕ ZI ) + wH (UI ⊕ ZI ) = nt)
h
i
= P nw − wH (UI,M1 ⊕ ZI,M1 ) + wH UI,M1 ⊕ ZI,M1
i
h
+ wH UI,M2 ⊕ ZI,M2 + wH UI,M2 ⊕ ZI,M2 = nt
h
i h
i
= P nw − M1 − wH (ZI,M1 ) + wH ZI,M1 + M2 − wH ZI,M2 + wH ZI,M2 = nt
= P wH (ZI,M1 ) − wH ZI,M1 − wH ZI,M2 + wH ZI,M2 = n(t − w) + M1 − M2
= P wH (ZI,M1 ) − wH ZI,M1 − wH ZI,M2 + wH ZI,M2 = n(t − w − a) + 2M1
(105b)
(105c)
(105d)
(105e)
(105f)
n·min(w,a)
X
=
P (M1 = m)
m=n·max(0,a+w−1)
· P wH (ZI,M1 ) − wH ZI,M1 − wH ZI,M2 + wH ZI,M2 = n(t − w − a) + 2m M1 = m
n·min(w,a)
nw n−nw
X
m
na−m
P (W1,m − W2,m = n(t − w − a) + 2m)
=
n
(105g)
(105h)
na
m=n·max(0,a+w−1)
△
△
where (105f) follows since |I| = nw, by denoting M1 = |M1 |, M2 = |M2 | and noting that M1 + M2 = na;
equality (105g) follows since M1 ≤ na and M1 ≤ nw, and since M2 ≤ na and M2 ≤ n(1 − w) (therefore
M1 ≥ n(a + w − 1)); and equality (105h) follows by denoting W1,m ∼ Binomial (n − (nw + na − 2m), p) and
W2,m ∼ Binomial (nw + na − 2m, p).
Lemma 6 (Mixed noise, asymptotic dimension): Consider a sequence of problems as in Lemma 5 indexed by the
blocklebgth n, with parameters an → a, wn → w and tn → t. Then:
1
lim − P (wH (cn ⊕ Un ⊕ Zn ) = ntn ) = EBT (p, a, w, t) .
n
n→∞
(106)
Proof: A straightforward calculation shows that
P (wH (cn ⊕ Un ⊕ Zn ) = ntn )
January 3, 2018
(107a)
DRAFT
32
.
=
n·min(wn ,an )
X
2nwn Hb (
m/n
wn
−m/n
) 2n(1−wn )Hb ( an1−w
) 2−nHb (an )
n
m=n·max(0,an +wn −1)
· 2−nEw (p,1−(wn +an −2m/n),wn +an −2m/n,tn −(wn +an −2m/n))
.
=
max
a−m
2−n[−wHb ( w )−(1−w)Hb ( 1−w )+Hb (a)] · 2−nEw (p,1−(w+a−2m),w+a−2m,t−(w+a−2m)).
m
m∈[max(0,a+w−1),min(w,a)]
(107b)
(107c)
The proof of Lemma 3 now follows:
P (wH (cn ⊕ Un ⊕ Zn ) ≤ tn ) =
tn
X
P (wH (cn ⊕ Un ⊕ Zn ) = nτ )
(108a)
τ =0
ntn
. X −nEBT (p,a,w,τ /n)
=
2
(108b)
.
= max 2−nEBT (p,a,w,τ )
(108c)
= 2−n minτ ∈[0,t] EBT (p,a,w,τ )
(108d)
τ =0
τ ∈[0,t]
A PPENDIX B
Q UANTIZATION -N OISE P ROPERTIES
OF
G OOD C ODES
In this section we prove Lemma 1 and Corollary 7, which contain the properties of good codes that we need for
deriving our achievable exponents.
First we define the covering efficiency of a code C as:
△
η(C) =
|Bn (0, ρcover (C))|
.
|Ω0 |
(109)
Lemma 7: Consider a covering-good sequence of codes C (n) ⊆ {0, 1}n, n = 1, 2, . . . of rate R. Then,
.
η(C (n) ) = 1
(110)
Proof: Since for all c ∈ C (n) we have that |Ωc | = |Ω0 |, it follows that
.
|Ω0 | = 2n(1−R) .
(111)
Therefore,
.
η(C (n) ) =
.
=
2−n(1−R)
Bn 0, ρcover (C (n) )
2−n(1−R)
(n)
2−nHb (ρcover (C ))
= 2n[Hb (ρcover (C
January 3, 2018
−1
(n)
))−(1−R)]
(112a)
(112b)
(112c)
DRAFT
33
= 2n[Hb (ρcover (C
(n)
))−Hb (δGV (R))]
.
= 1,
(112d)
(112e)
where the last asymptotic equality is due to (60) and due to the continuity of the entropy.
Proof of Lemma 1: Since for any v ∈
/ Ω0
P (X ⊖ QC (n) (X) = v) = 0
(113a)
≤ P (N = v) ,
(113b)
it is left to consider points v ∈ Ω0 . To that end, since the code is linear, due to symmetry
X ⊖ QC (n) (X) ∼ Uniform (Ω0 ) .
(114)
Thus, for v ∈ Ω0 ,
P (X ⊖ QC (n) (X) = v)
P (N = v)
=
|Ω0 |
(115a)
−1
Bn 0, ρcover (C (n) )
−1
(115b)
= η(C (n) )
(115c)
.
= 1,
(115d)
where the last (asymptotic) equality is due to Lemma 7. This completes the last part (Note that the rate of convergence
is independent of v, and therefore the convergence is uniform over v ∈ Ω0 ). The second part follows by the linearity
of convolution: For any v ∈ {0, 1}n,
P (X ⊖ QC (n) (X) ⊕ Z = v)
X
PZ (z)P (X ⊖ QC (n) (X) ⊕ z = v)
=
(116a)
(116b)
z∈{0,1}n
=
X
PZ (z)P (X ⊖ QC (n) (X) = v ⊖ z)
(116c)
X
PZ (z)P (N = v ⊖ z)
(116d)
z∈{0,1}n
≤
z∈{0,1}n
= P (N ⊕ Z = v) ,
(116e)
where the (asymptotic) inequality is due to the first part of the lemma.
Proof of Corollary 7: Let N = X ⊖ U. Note that since the code is linear and X is uniform over {0, 1}n, it
follows that N is independent of the pair (U, Z). Thus,
P QC (n) (X ⊕ Z) 6= U
(117a)
2,S
January 3, 2018
DRAFT
34
= P QC (n) (U ⊕ [X ⊖ U] ⊕ Z) 6= U
2,S
= P QC (n) (U ⊕ N ⊕ Z) 6= U .
(117b)
(117c)
2,S
Recalling the definition of N′ in (79b), we have
P QC (n) (U ⊕ N ⊕ Z) 6= U
2,S
= P QC (n) (U ⊕ N ⊕ Z) 6= U
2,S
= EU P QC (n) (u ⊕ N ⊕ Z) 6= u U = u
2,S
≤ EU P QC (n) u ⊕ N′ ⊕ Z 6= u U = u
2,S
= P QC (n) U ⊕ N′ ⊕ Z 6= U ,
(118a)
(118b)
(118c)
(118d)
(118e)
2,S
where (118d) follows by applying Lemma 1 for each u.
A PPENDIX C
E XISTENCE
OF
G OOD N ESTED C ODES
In this appendix we prove the existence of a sequence of good nested codes, as defined in Definition 15. To that
end, we first state known results on the existence of spectrum-good and covering-good codes.
By [24], random linear codes are spectrum-good with high probability. That is, let C (n) be a linear code of
blocklength n and rate R, with a generating matrix G(n) drawn i.i.d. Bernoulli-1/2. Then there exist some sequences
(n)
ǫS
(n)
and δS
approaching zero, such that for all w,
(n)
(n)
(n)
P ΓC (n) (w) > (1 + δS )ΓR (w) ≤ ǫS .
(119)
As for covering-good codes, a construction based upon random linear codes is given in [25]. For a generating
(n)
matrix G(n) at blocklength n, a procedure is given to generate a new matrix GI
(n)
(n) △
= GI (G(n) ), with kI
= ⌈log n⌉
rows. The matrices are combined in the following way:
(n)
(n)
G
×
n
k
I
I
k ′(n) × n
G′(n) =
(n)
(n)
k
×
n
G
(120)
Clearly, adding GI does not effect the rate of a sequence of codes. Let C (n) be constructed by this procedure, with
(n)
G(n) drawn i.i.,d. Bernoulli-1/2. It is shown in [25, Theorem 12.3.5], that there exists sequences ǫC
(n)
and δC
approaching zero, such that
(n)
(n)
P ρcover (C (n) ) > (1 + δC )δGV (R) ≤ ǫC .
January 3, 2018
(121)
DRAFT
35
A nested code of blocklength n and rates (R, Rbin ) has a generating matrix
(n)
(n)
G̃
k̃
×
n
G(n) =
k (n) × n
(n)
(n)
kbin × n
Gbin
(122)
(n)
That is, G(n) and Gbin are the generating matrices of the fine and coarse codes, respectively.
We can now construct good nested codes in the following way. We start with random nested codes of rate
(n)
(R, Rbin ). We interpret the random matrix G(n) as consisting of matrices G̃(n) and Gbin as above, both i.i.d.
(n)
Bernoulli-1/2. We now add GI
(n)
= GI (G(n) ) as in the procedure of [25], to obtain the following generating
matrix:
(n)
GI
(n)
G′(n) = G̃
(n)
Gbin
(n)
kI
×n
k̃ (n) × n
k ′(n) × n
(n)
kbin × n
(123)
(n)
Now we construct the fine and coarse codes using the matrices G′(n) and Gbin , respectively; the rate pair does not
change due to the added matrices. By construction, the fine and coarse codes satisfy (121) and (119), respectively.
(n)
Thus, by the union bound, the covering property is satisfied with δC
(n)
(n)
and the spectrum property with δS ,
(n)
simultaneously, with probability 1 − ǫC − ǫS . We can thus construct a sequence of good nested codes as desired.
R EFERENCES
[1] T. Berger, “Decentralized estimation and decision theory,” in The IEEE 7th Spring Workshop Inf. Theory, Mt.
Kisco, NY, Sep. 1979.
[2] R. Ahlswede and I. Csiszár, “To get a bit of information may be as hard as to get full information,” IEEE
Trans. Information Theory, vol. 27, no. 4, pp. 398–408, July 1981.
[3] H. M. H. Shalaby and A. Papamarcou, “Multiterminal detection with zero-rate data compression,” IEEE Trans.
Information Theory, vol. 38, no. 2, pp. 254–267, Mar. 1992.
[4] R. Ahlswede and I. Csiszár, “Hypothesis testing with communication constraints,” IEEE Trans. Information
Theory, vol. 32, no. 4, pp. 533–542, July 1986.
[5] T. S. Han, “Hypothesis testing with multiterminal data compression,” IEEE Trans. Information Theory, vol. 33,
no. 6, pp. 759–772, Nov. 1987.
[6] H. Shimokawa, T. S. Han, and S. Amari, “Error bound of hypothesis testing with data compression,” in Proc.
Int. Symp. Info. Theory (ISIT), June 1994, p. 114.
[7] H. Shimokawa, “Hypothesis testing with multiterminal data compression,” Master’s thesis, Feb. 1994.
[8] D. Slepian and J. K. Wolf, “Noiseless coding of correlated information sources,” IEEE Trans. Information
Theory, vol. 19, pp. 471–480, July 1973.
[9] A. D. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the decoder,”
IEEE Trans. Information Theory, vol. 22, pp. 1–10, Jan. 1976.
January 3, 2018
DRAFT
36
[10] M. Rahman and A. Wagner, “On the optimality of binning for distributed hypothesis testing,” IEEE Trans.
Information Theory, vol. 58, no. 10, pp. 6282–6303, Oct. 2012.
[11] T. S. Han and K. Kobayashi, “Exponential-type error probabilities for multiterminal hypothesis testing,” IEEE
Trans. Information Theory, vol. 35, no. 1, pp. 2–14, Jan. 1989.
[12] S. Amari, “On optimal data compression in multiterminal statistical inference,” IEEE Trans. Information
Theory, vol. 57, no. 9, pp. 5577–5587, Sep. 2011.
[13] Y. Polyanskiy, “Hypothesis testing via a comparator,” in Proc. Int. Symp. Info. Theory (ISIT), July 2012, pp.
2206–2210.
[14] G. Katz, P. Piantanida, and M. Debbah, “Distributed binary detection with lossy data compression,” IEEE
Trans. Information Theory, vol. 63, no. 8, pp. 5207–5227, Aug 2017.
[15] ——, “A new approach to distributed hypothesis testing,” in 2016 50th Asilomar Conference on Signals,
Systems and Computers, Nov. 2016, pp. 1365–1369.
[16] T. S. Han and S. Amari, “Statistical inference under multiterminal data compression,” IEEE Trans. Information
Theory, vol. 44, no. 6, pp. 2300–2324, Oct. 1998.
[17] G. D. Forney, Jr., “Exponential error bounds for erasure, list, and detection feedback schemes,” IEEE Trans.
Information Theory, vol. 14, pp. 206–220, Mar. 1968.
[18] J. Körner and K. Marton, “How to encode the modulo-two sum of binary sources,” IEEE Trans. Information
Theory, vol. 25, pp. 219–221, Mar. 1979.
[19] H. Shimokawa and S. Amari, “Multiterminal estimation theory with binary symmetric source,” in Proc. Int.
Symp. Info. Theory (ISIT), Sep. 1995, p. 447.
[20] M. El Gamal and L. Lai, “Are Slepian-Wolf rates necessary for distributed parameter estimation?” CoRR,
vol. abs/1508.02765, 2015. [Online]. Available: http://arxiv.org/abs/1508.02765
[21] T. M. Cover and J. A. Thomas, Elements of Information Theory. New York: Wiley, 1991.
[22] J. Neyman and E. S. Pearson, “The testing of statistical hypotheses in relation to probabilities a priori,”
Mathematical Proceedings of the Cambridge Philosophical Society, vol. 29, pp. 492–510, 10 1933. [Online].
Available: http://journals.cambridge.org/article S030500410001152X
[23] A. B. Wagner, B. G. Kelly, and Y. Altuğ, “Distributed rate-distortion with common components,” IEEE Trans.
Information Theory, vol. 57, no. 7, pp. 4035–4057, July 2011.
[24] R. G. Gallager, Information Theory and Reliable Communication. John Wiley & Sons, 1968.
[25] G. Cohen, I. Honkala, S. Litsyn, and A. Lobstein, Covering Codes. Elsevier (North Holland Publishing Co.),
1997.
[26] R. Zamir, S. Shamai, and U. Erez, “Nested linear/lattice codes for structured multiterminal binning,” IEEE
Trans. Information Theory, vol. 48, pp. 1250–1276, June 2002.
[27] A. D. Wyner, “Recent results in the Shannon theory,” IEEE Trans. Information Theory, vol. 40, no. 1, pp.
2–10, Jan. 1974.
[28] I. Csiszár and J.Körner, “Towards a general theory of source networks,” IEEE Trans. Information Theory,
January 3, 2018
DRAFT
37
vol. 26, no. 2, pp. 155–165, Mar. 1980.
[29] I. Csiszár, “Linear codes for sources and source networks: Error exponents, universal coding,” IEEE Trans.
Information Theory, vol. 28, no. 4, pp. 585–592, July 1982.
[30] E. Haim, Y. Kochman, and U. Erez, “Distributed structure: Joint expurgation for the multiple-access channel,”
IEEE Trans. Information Theory, vol. 63, no. 1, pp. 5–20, Jan. 2017.
[31] J. Chen, D.-K. He, A. Jagmohan, and L. A. Lastras-Montaño, “On the reliability function
of
variable-rate
Slepian-Wolf
coding,”
Entropy,
vol.
19,
no.
8,
2017.
[Online].
Available:
http://www.mdpi.com/1099-4300/19/8/389
[32] ——, “On the reliability function of variable-rate Slepian-Wolf coding,” in Proceedings of the 45th annual
Allerton Conference on Communication, Control and Computing, Sep. 2007.
[33] N. Weinberger and N. Merhav, “Optimum trade-offs between error exponent and excess-rate exponent of
Slepian-Wolf coding,” in Proc. Int. Symp. Info. Theory (ISIT), June 2015, pp. 1565–1569.
[34] B. G. Kelly and A. B. Wagner, “Reliability in source coding with side information,” IEEE Trans. Information
Theory, vol. 58, no. 8, pp. 5086–5111, Aug. 2012.
[35] D. Krithivasan and S. S. Pradhan, “Distributed source coding using Abelian group codes: A new achievable
rate-distortion region,” IEEE Trans. Information Theory, vol. 57, no. 3, pp. 1495–1519, Mar. 2011.
[36] A. B. Wagner, “On distributed compression of linear functions,” IEEE Trans. Information Theory, vol. 57,
no. 1, pp. 79–94, Jan. 2011.
January 3, 2018
DRAFT
| 7cs.IT
|
In-place Graph Rewriting with Interaction Nets
Ian Mackie
LIX, CNRS UMR 7161, École Polytechnique, 91128 Palaiseau Cedex, France
Shinya Sato
University Education Center, Ibaraki University, 2-1-1 Bunkyo, Mito, Ibaraki 310-8512, Japan
An algorithm is in-place, or runs in-situ, when it does not need any additional memory to execute
beyond a small constant amount. There are many algorithms that are efficient because of this feature,
therefore it is an important aspect of an algorithm. In most programming languages, it is not obvious when an algorithm can run in-place, and moreover it is often not clear that the implementation
respects that idea. In this paper we study interaction nets as a formalism where we can see directly,
visually, that an algorithm is in-place, and moreover the implementation will respect that it is inplace. Not all algorithms can run in-place however. We can nevertheless still use the same language,
but now we can annotate parts of the algorithm that can run in-place. We suggest an annotation for
rules, and give an algorithm to find this automatically through analysis of the interaction rules.
1
Introduction
An algorithm runs in-place, or in-situ, if it needs a constant amount of extra space to run. For an algorithm
to be in-place, the input is usually overwritten, so mutable data-structures need to be supported by the
programming language. There are many well-known in-place algorithms, in particular from the domain
of sorting. One example is bubble sort, that we can write in Java for instance:
static void bubble() {
int t;
for (int i = n-1; i >= 0; --i)
for (int j = 1; j <= i; ++j)
if (a[j-1] > a[j])
{ t = a[j-1]; a[j-1] = a[j]; a[j] = t; }
}
With some knowledge of what the above instructions do, and tracing a few steps of the execution, we
can soon realise that it runs in-place: one additional memory location (t) is all that is needed to sort the
array a of integers. In many programs however, it is not obvious that an algorithm can run in-place, and
moreover it is often not clear that the underlying implementation respects that idea. This issue becomes
more pertinent when we examine different programming paradigms and different programming styles,
especially when we have dynamic data-structures.
In Figure 1 we give four fragments of programs for inserting an element into a sorted list (so part
of the insertion sort algorithm). These programs are written in Prolog, Haskell and Java, with the latter
written using two different programming styles. Insertion can be written so that it runs in-place, but
it is not easy to see which of these examples are (or can be) in-place unless we start to examine the
compiler, the run-time system, and the definition of functions like cons in the case of Java. Declarative
languages (functional and logic based in this example) are designed to be referentially transparent, so the
data-structures are updated in a non-destructive way. Moreover, it is not the programmer who decides
how memory is allocated and organised in these languages. On the other-hand, languages like C and Java
A. Corradini & H. Zantema (Eds.)
Computing with Terms and Graphs (TERMGRAPH 2016)
EPTCS 225, 2016, pp. 15–24, doi:10.4204/EPTCS.225.4
c I. Mackie & S. Sato
This work is licensed under the
Creative Commons Attribution License.
In-place Graph Rewriting with Interaction Nets
16
(the imperative fragment) the programmer does the memory allocation (and de-allocation also in some
languages) explicitly, and therefore has a better idea of resource usage. These examples illustrate some
of the difficulties in knowing if the program will run in-place or not.
Program 1: Prolog
Program 2: Haskell
insert([Y|Ys], X, [Y|Zs]):Y < X, !, insert(Ys, X, Zs).
insert(Ys, X, [X|Ys]).
insert e [] = [e]
insert e (x:xs) = if e < x then e:x:xs
else x:(insert e xs)
Program 3: Java (functional style)
Program 4: Java (destructive style)
static List insert(int x, List l) {
if (isEmpty(l) || x < l.head)
return cons(x, l);
else
return cons(l.head, insert(x, l.tail));
}
static List insert(int x, List l) {
if (isEmpty(l) || x < l.head)
return cons(x, l);
else {
l.tail = insert(x, l.tail);
return l;
}
}
Figure 1: Example programs
In-place algorithms are important because they can lead to more efficient algorithms, and even change
time complexity. If the data-structure supports it, concatenation of two lists can be constant time if done
in-place, but linear if not. Memory allocation is also expensive, so minimising it also makes it more
efficient. Although less important in some ways, there are devices that have limited space (embedded
systems, hand-held devises, etc.), so limiting the space usage if there is no run-time impact is always
advantageous.
In this paper we use a formalism where we can see directly, in fact visualise, that an algorithm is
in-place, and moreover the implementation respects that it is in-place. We use the graphical rewriting
system of interaction nets [5] as our programming paradigm. This visual language has many similarities
with term rewriting systems in that they are user-defined systems. For this reason, they can be considered
as specification languages. However, they are also a model of computation that requires all aspects of
the computation to be explained, including copying and garbage collection. For this reason they are like
an implementation model, or low-level language. It is the mixture of these features that allows us to see
directly how the program can be implemented, and thus see how the memory is allocated.
Not all algorithms can run in-place however. The formalism will still be of use though, and we
identify three different uses of the information we can ascertain from interaction rules:
1. If the rewrite rules have a particular property then the algorithm is in-place (and will be implemented in-place).
2. If the rules can be applied in a given way, so a strategy is needed, then the algorithm can be
implemented in-place.
3. If neither of the above hold, then we can still make use of the formalism by re-using as much data
as possible in the computation.
For the final point, we can either ask the programmer to annotate the rules, or develop an algorithm to
do this automatically. In this paper, we give examples to motivate the first and last points—more details
including the second point will be given in a longer version of this paper.
I. Mackie & S. Sato
17
The space usage of algorithms, as well as the time complexity, are fundamental in algorithm design
and analysis, and well documented in many textbooks. There have also been a number of works that
give a bound on the space usage through type systems, for example [3], and [4]. Our approach is more
syntactical, and uses properties of the underlying run-time system.
Overview. The rest of this paper is structured as follows. In the next section we recall the background,
and give some examples to motivate the ideas. We then give some case studies of examples that are
in-place in Section 3. In Section 4 we introduce an annotation for the rules which allows for node reuse.
We then go on in Section 5 to show an algorithm to annotate a rule automatically in the case of using a
fixed-size node representation for nodes. After a brief discussion on how we can use this information in
a compiler in Section 6, we conclude in Section 7.
2
Background
In the graphical rewriting system of interaction nets [5], we have a set Σ of symbols, which are names
of the nodes. Each symbol has an arity ar that determines the number of auxiliary ports that the node
has. If ar(α ) = n for α ∈ Σ, then α has n + 1 ports: n auxiliary ports and a distinguished one called the
principal port labelled with an arrow. Nodes are drawn as follows:
...
x
x
ë
A net built on Σ is an undirected graph with nodes as the vertices. The edges of the net connect nodes
together at the ports such that there is only one edge at every port. A port which is not connected is called
a free port. Two nodes (α , β ) ∈ Σ × Σ connected via their principal ports form an active pair, which is
the interaction net analogue of a redex. A rule ((α , β ) ⇒ N) replaces the pair (α , β ) by the net N. All
the free ports are preserved during reduction, and there is at most one rule for each pair of nodes. The
following diagram illustrates the idea, where N is any net built from Σ.
x
y
...
y
y
x
N
...
ì
x
⇒
...
...
x
ë
y
We refer to the rule ((α , β ) ⇒ N) as α ⊲⊳ β . The most powerful property of this graph rewriting system
is that it is one-step confluent—all reduction sequences are permutation equivalent. We use an extension
of these pure interaction nets: values can be stored in the nodes, and rules can test these values. This is
done is such a way as to preserve the one-step confluence property. We use this extension in the insertion
sort example below.
It is possible to reason about the graphical representation of nets, but it is convenient to have a textual
calculus for compact representation. There are several calculi in the literature, and here we review one
calculus [2], which is a refined version of [1].
Agents: Let Σ be a set of symbols, ranged over by α , β , . . ., each with a given arity ar : Σ → IN. An
occurrence of a symbol is called an agent, and the arity is the number of auxiliary ports.
Names: Let N be a set of names, ranged over by x, y, z, etc. N and Σ are assumed disjoint. Names
correspond to wires in the graph system.
In-place Graph Rewriting with Interaction Nets
18
Terms: A term is built on Σ and N by the grammar: t ::= x | α (t1 , . . . ,tn ) | $t, where x ∈ N , α ∈ Σ,
ar(α ) = n and t1 , . . . ,tn ,t are terms, with the restriction that each name can appear at most twice.
If n = 0, then we omit the parentheses. If a name occurs twice in a term, we say that it is bound,
otherwise it is free. We write s,t, u to range over terms, and ~s,~t,~u to range over sequences of
terms. A term of the form α (t1 , . . . ,tn ) can be seen as a tree with the principal port of α at the
root, and the terms t1 , . . . ,tn are the subtrees connected to the auxiliary ports of α . The term $t
represents an indirection node which is created by reduction, and is not normally part of an initial
term. Intuitively, $t corresponds to a variable bounded with t (or a state such that an environment
captures t).
Equations: If t, u are terms, then the unordered pair t = u is an equation. Θ will be used to range over
sequences of equations.
Rules: Rules are pairs of terms written: α (x1 , . . . , xn ) = β (y1 , . . . , ym ) ⇒ Θ, where (α , β ) ∈ Σ × Σ is the
active pair, and Θ is the right-hand side of the rule. We will abbreviate in the following the leftand right-hand sides of the rule by LHS and RHS respectively. All names occur exactly twice in a
rule, and there is at most one rule for each pair of agents.
3
In-place algorithms: case studies
Here we give some example interaction net systems that demonstrate the ideas we have discussed previously. The first one is unary numbers with addition. We represent the following term rewriting system:
add(Z,y)=y, add(S(x),y)=add(x,S(y)) as a system of nets with nodes Z, S, Add, and two rewrite
rules:
Z
S
Add
Add
y
x
x
⇒
S
S
Add
x
Add
x
y
Z
⇒
x
x
x
x
The following is an example of add(S(Z),Z):
Add
S
Add
S
→
Z
r
S
→
r
Z
Z
Z
r
Z
Next, we introduce an example for the Ackermann function defined by:
ack 0 n = n+1, ack m 0 = ack (m-1) 1, ack m n = ack (m-1) (ack m (n-1)).
We can build the interaction net system on the unary natural numbers that corresponds to the term rewriting system as follows:
A
y
Z
⇒
r
A2
A
S
y
Z
⇒
y
r
A
r
Pred
⇒
S
x
A2
r
r
x
S
S
x
A2
S
⇒
Dup
Pred
x
r
x
r
y
y
x
A
Z
y
A
r
I. Mackie & S. Sato
19
where the node Dup duplicates S and Z nodes, and the node Pred erases the S node:
Dup
S
⇒
Dup
S
Dup
⇒
Z
Z
Z
Pred
S
⇒
S
The following is an example of rewriting:
A
S
→
A2
S
→
Dup
S
S
Z
Z
S
Z
S
Dup ⊲⊳ Z
→
S
Z
S
S
Z
A
S
S
Z
Z
A
A
Pred
A
S
ack 1 2
Pred ⊲⊳ S
→
Z
Pred
A
S
S
A
Z
Z
Z
ack 0 (ack 1 1)
Z
Observation for in-place running. Interaction nets are quite unique as a programming paradigm because we are basically writing programs using the internal data-structure. We characterise three kinds of
rewriting rule:
• Case 1: there are two nodes in the right-hand side (RHS). The two nodes of the active pair can be
reused. Thus, no matter which way we evaluate, the algorithm for these rules can run in constant
space. The rules Add ⊲⊳ S, A ⊲⊳ S and Dup ⊲⊳ Z are classified in this case.
• Case 2: there are less than two nodes in the RHS. The active pair nodes can be reused as nodes
that occur in the RHS, so in terms of the memory space, it can run in constant space as well. For
instance, the rules Add ⊲⊳ Z, A ⊲⊳ Z and Pred ⊲⊳ S are classified in this case.
• Case 3: there are more than two nodes in the RHS. Here, active pair nodes can be reused as nodes
that occur in the RHS, but additional memory space is required for other nodes. We divide this
into two very different categories:
1. An active pair creates another active pair that is Case 2 above. These two reductions together
make the algorithm in-place. For instance, in the last two-step reductions of ack 1 2 to
ack 0 (ack 1 1) in the example, we can save memory space for two nodes when we take
Pred ⊲⊳ S and Dup ⊲⊳ Z in this order, in comparison with the order Dup ⊲⊳ Z and Pred ⊲⊳ S.
2. For instance, the rules A2 ⊲⊳ Z, A2 ⊲⊳ S and Dup ⊲⊳ S are classified in this case. These rules
are not in-place, but the total cost can be reduced by choosing the two reused nodes well. For
instance, in the rule A2 ⊲⊳ S, it is better to reuse the A2 in the LHS as the right side A (not the
left side) because the information of the right auxiliary port (denoted as “r”) can be reused.
Thus we try to reuse the memory in the best way possible.
Insertion sort is a well-known in-place algorithm. The first three interaction rules below encode
insertion of an item into a sorted list, and the final two rules encode the insertion sort algorithm.
I(x)
nil
⇒
x
nil
In-place Graph Rewriting with Interaction Nets
20
I(x)
y
x≤y
⇒
x
y
I(x)
y
x>y
⇒
y
I(x)
IS
nil
⇒
nil
IS
x
⇒
I(x)
IS
These five rules encode the whole program—there is nothing else. Our real point however, is that in
this case a trace of the execution (an animation of this algorithm) is showing no more and no less than
what is needed to explain this algorithm. It is in-place because to begin with we need to put an additional
IS node, and the final rule for IS erases this. We invite the reader to trace the following example net:
IS
2
4
1
3
nil
This example is the full system of interaction nets for the insertion sort algorithm, and it runs inplace. There are other examples, for instance reversing a list. In this case, we need to start with adding a
rev node, and the final rule deletes it. The following two rules implement reversing any list:
rev
nil
⇒
h
rev
⇒
rev
h
The starting configuration is shown in the following example, which will reverse the list in-place
with five interactions:
1
rev
4
2
3
4
nil
nil
Annotating Rules
In the previous section we saw that there are interaction systems that can run in-place. To ensure that
this is actually achieved at the implementation level, we need to make sure that for these examples the
nodes in the rule are reused when building the new net. This idea can be also used even when there are
more than two nodes in the right-hand side of the rule, and in this case there is a choice of how to reuse
the nodes. We next introduce an annotation to show which nodes in the RHS of the rule are reused. This
helps the compiler to analyse information so that it can improve on the in-place execution of parts of the
program.
Annotation: ∗L and ∗R. We introduce ∗L and ∗R to denote where the left-hand side and the right-hand
side nodes in the LHS of a rule are used for in-place computation, respectively. For instance, the rule
Add ⊲⊳ S is written as follows:
I. Mackie & S. Sato
21
Add
S
*R
y
x
x
⇒
*L
x
y
x
The advantage is that the compiler is easily able to know, traversing the net, not only where the
active pair nodes are used, but also which information about the connections should be preserved. For
instance, in the above example, the information denoted as “x1 ” and “y1 ” in the LHS of the rule should
be preserved, say as “_x1 ” and “_y1 ”, because these are overwritten in the RHS of the rule, and the “x1 ”
and “y1 ” in the RHS should be replaced by the “_x1 ” and “_y1 ”.
Annotation for changing of the node name. To change the node name, we introduce a name cast,
such as the type cast of the C programming language, with the ∗L and ∗R. For instance, the rule A2 ⊲⊳ S
is written as follows:
A2
S
⇒
Dup
Pred
(*R)
x
x
r
y
A
(*L)
A
y
r
In this example, the compiler can also know that the information denoted as “x” and “y” in the LHS
of the rule should be preserved, by checking the connection of ∗L and ∗R.
Advantage of using a fixed-size node representation for nodes. We represent nodes as a fixed-size
node, thus fixed-size auxiliary ports. For this we need to use more space than necessary, but we can
manage and reuse nodes in a simpler way [6]. Here, we assume that auxiliary ports are assigned by the
order from the left-hand side to the right-hand side, the rule A ⊲⊳ Z is written as follows:
A
y
Z
r
⇒
S
(*L)
y
r
So we can reuse not only the ∗L node, but also the pointer information (denoted as “y”).
5
Deriving Annotations
In the previous section we showed how rules can be annotated with information about the reuse of nodes.
Here, in the case of using a fixed-size node representation, we define a function to calculate how a given
term is similar to others. We first introduce some notation. For strings we write double quotes (“ and ”).
We use the notation {x} in a string as the result of replacing the occurrence {x} with its actual value. For
instance, if x = “abc” and y = 89 then “1{x}2{y}” = “1abc289”. We use + as an infix binary operation
to concatenate strings. We also use a symbol − to show the empty sequence. In order to show where in
a term is in a sequence of equations, we use the following term path notation: nth (L | R) : arg1 arg2 . . ..
For instance, the term t in x = y, α (β (s,t), z)= w is denoted as 2L:12 because the t occurs in α (β (s,t), z),
which is the left-hand side term of the second equation, occurs in β (s,t), which is the first argument of
the term, and in the second argument of the term β (s,t).
Using this, we can now give an important definition for this paper:
In-place Graph Rewriting with Interaction Nets
22
Definition 1 (Node matching) The function Match below takes a term and a sequence of equations,
and returns a list of a pair (score, term path) where the score contains the number of the matched agent
and matched arguments. We use Standard ML notation for lists, thus, in the following the operator “@”
is list concatenation.
Match(t, (e1 , . . . , en )) = Matche (t, e1 , 1)@ · · · @Matche (t, en , n)
Matche (t, s = u, pos) = Matcht (t, s, “{pos}L : ”)@Matcht (t, u, “{pos}R : ”)
= [((0, 0), tpath)]
Matcht (α (~x), x, tpath)
| Matcht (α (~x), β (t1 , . . . ,tn ), tpath) = [((agentPts, namePts), tpath)]
@Matcht (α (~x), t1 , “{tpath}1”)@ · · ·
@Matcht (α (~x), tn , “{tpath}n”)
where agentPts = if α = β then 1 else 0
namePts = Matchns (~x, (t1 , . . . ,tn ))
Matchns (−,~t)
= 0
| Matchns (~x, −)
= 0
~
| Matchns ((x,~x), (t, t)) = (if x = t then 1 else 0) + Matchns (~x,~t)
Thus Match will return a list of the matching metrics for each node, and give the location in the net
for each. We can then easily extract the best one from this information. The following examples illustrate
how this information is calculated.
Example 2 The rule Add ⊲⊳ S given previously is written textually as
Add(x1 , x2 ) = S(y1 ) ⇒ Add(S(x1 , x2 )) = y1
First we take the left hand side term Add(x1 , x2 ) of the active pair.
Match(Add(x1 , x2 ), Add(S(x1 ), x2 ) = y1 )
returns the following list:
[((1,1),1L:), ((0,1),1L:1), ((0,0),1L:11), ((0,0),1L:2), ((0,0),1R:)]
The first element of this result ((1, 1), 1L :) shows that the highest score (1, 1) is obtained when we
annotate the term 1L :, which is Add(S(x1 ), x2 ), such as (∗L)(S(x1 ), x2 ). Next we take the right hand side
term S(y1 ). The following is the result of the function Match for S(y1 ):
[((0,0),1L:), ((1,0),1L:1), ((0,0),1L:11), ((0,0),1L:2), ((0,0),1R:)]
This result shows that there is one agent term 1L : 1 (thus S(x1 )) that has the same id with S(y1 ), and
no agent terms that have the same occurrence of the argument. Therefore, S(x1 ) should be annotated as
(∗R)(x1 ). These annotations correspond to the first graph in Section 4.
Example 3 The rule A2 ⊲⊳ S given previously is written textually as
A2(x, r) = S(y) ⇒ x = Dup(A(y, w), Pred(A(w, r)))
The result of Match(A2(x, r), x = Dup(A(y, w), Pred(A(w, r)))) is as follows:
[((0,0),1L:), ((0,0),1R:), ((0,0),1R:1), ((0,0),1R:11), ((0,0),1R:12),
((0,0),1R:2), ((0,1),1R:21)), ((0,0),1R:211), ((0,0),1R:212)].
I. Mackie & S. Sato
23
This shows that the term 1R : 21, which is A(w, r), should be annotated by (∗L) because it has the highest
score (0, 1).
In the case of S(y), the result of the Match is as follows:
[((0,0),1L:), ((0,0),1R:), ((0,1),1R:1), ((0,0),1R:11), ((0,0),1R:12),
((0,0),1R:2), ((0,0),1R:21)), ((0,0),1R:211), ((0,0),1R:212)].
Thus, taking account of using a fixed-size node representation, we find that we should annotate the term
1R : 1, which is A(y, w), by (∗R), though we annotated Pred in the graph in Section 4. Of course, the
evaluation of the score depends on the implementation method. However, in these cases, (0, 1) must be
highest because the others are (0, 0).
6
Discussion
The algorithm given above can be put to use in the compilation of rules. There is some choice for the
compiler if several nodes get the same score. We briefly summarise here some details about the lowlevel language, and a description of how data-structures are used. Our contribution here is to extend
those ideas with reuse. In the longer version of this paper we will give the algorithm for compilation in
detail and focus on the data-structures that are used. An important result that we get about compilation
is that we do the least amount of work in implementing the interaction rules.
The concrete representation of an interaction net can be summarised by the following diagram, where
Γ represents a net, EQ a stack of equations, and I an interface of the net:
EQ
...
Γ
...
I
The following is the representation of the net Add(Z, r) = S(w), Add(Z, w) = S(Z), where N are nodes
representing variables:
Z
N
N
S
Add
Z
Add
Z
EQ
I
The interface of the net is where we collect all the free edges of a net. We have a set of instructions
that can manipulate this data-structure, and the compilation of a rewrite rule needs to generate code
to manipulate this structure. What we have achieved in this paper is an annotation of a rule so that
we can generate code to manipulate data-structures of the kind shown above with the least amount of
work. Specifically, we limit the allocation of memory on the heap, and also avoid unnecessary garbage
collection. We also find the best way to build the RHS of the rule, by re-using the memory cells and the
existing pointers.
Returning to the ideas presented in the introduction, we have obtained the following:
• Case 1: if there are two nodes in the RHS of the rule, then the algorithm for executing these rules
will run in-place, and moreover we will perform the fewest updates of pointers to build the RHS.
This is because the algorithm above will find the best way of reusing both nodes and connections.
In-place Graph Rewriting with Interaction Nets
24
• Cases 2 and 3: Even when the algorithm is not in-place because we have more than two nodes or
fewer than two nodes in the RHS, the ideas of this paper still gives the optimal implementation of
the rule. We note however that if we change the data-structure, there may be other solutions to this
problem, so the result we obtain is with respect to the chosen data-structure.
One aspect of in-place interaction nets reduction that we have not examined in the current paper is
pairing up interaction rules by using a specific reduction strategy. For example, if an interaction system
has a rule with one node in the RHS, and another rule with three, then we might be able to schedule these
two rules to be performed together, this maintaining the algorithm in-place. There are a number of ways
we can do this at the implementation level, and we will include details of this aspect in the long version
of the paper.
7
Conclusion
We have introduced a notion to rewrite graphs, specifically interaction nets, in-place. We have identified
a number of in-place algorithms and this lead us to an annotation to facilitate the implementation of
in-place rewriting by re-using nodes of active pairs. The main feature of this annotation is that is will
give the best possible reuse, so will allow the least amount of work to be done when rewriting the graph.
We are currently working to incorporate the ideas of this paper into an implementation, using the model
previously developed in [2]. We hope to present details of that, and benchmark results at a future time.
References
[1] Maribel Fernández & Ian Mackie (1999): A Calculus for Interaction Nets. In G. Nadathur, editor: Proceedings
of the International Conference on Principles and Practice of Declarative Programming (PPDP’99), Lecture
Notes in Computer Science 1702, Springer-Verlag, pp. 170–187, doi:10.1007/10704567 10.
[2] Abubakar Hassan, Ian Mackie & Shinya Sato (2015): An Implementation Model for Interaction Nets. In:
Proceedings 8th International Workshop on Computing with Terms and Graphs, TERMGRAPH 2014, Vienna,
Austria, July 13, 2014., pp. 66–80, doi:10.4204/EPTCS.183.5.
[3] Martin Hofmann (2000): A Type System for Bounded Space and Functional In-Place Update–Extended
Abstract. In Gert Smolka, editor: Programming Languages and Systems, 9th European Symposium on
Programming, ESOP 2000, European Joint Conferences on the Theory and Practice of Software, ETAPS
2000, Berlin, Germany, Proceedings, Lecture Notes in Computer Science 1782, Springer, pp. 165–179,
doi:10.1007/3-540-46425-5 11.
[4] John Hughes & Lars Pareto (1999): Recursion and Dynamic Data-structures in Bounded Space: Towards
Embedded ML Programming. In Didier Rémi & Peter Lee, editors: Proceedings of the fourth ACM SIGPLAN International Conference on Functional Programming (ICFP ’99), Paris, 1999, ACM, pp. 70–81,
doi:10.1145/317636.317785.
[5] Yves Lafont (1990): Interaction Nets. In: Proceedings of the 17th ACM Symposium on Principles of Programming Languages (POPL’90), ACM Press, pp. 95–108, doi:10.1145/96709.96718.
[6] Simon L. Peyton Jones (1987): The Implementation of Functional Programming Languages. Prentice Hall
International.
| 6cs.PL
|
Constraint Exploration and Envelope of Simulation
Trajectories
Oswaldo Terán1, Bruce Edmonds and Steve Wallis
{o.teran, b.edmonds, s.wallis}@mmu.ac.uk
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building, Aytoun Street, Manchester, M1 3GH, UK.
http://www.cpm.mmu.ac.uk/
Abstract
The implicit theory that a simulation represents is precisely not in the individual choices but
rather in the ‘envelope’ of possible trajectories – what is important is the shape of the whole
envelope. Typically a huge amount of computation is required when experimenting with
factors bearing on the dynamics of a simulation to tease out what affects the shape of this
envelope. In this paper we present a methodology aimed at systematically exploring this
envelope. We propose a method for searching for tendencies and proving their necessity
relative to a range of parameterisations of the model and agents’ choices, and to the logic of
the simulation language. The exploration consists of a forward chaining generation of the
trajectories associated to and constrained by such a range of parameterisations and choices.
Additionally, we propose a computational procedure that helps implement this exploration by
translating a Multi Agent System simulation into a constraint-based search over possible
trajectories by ‘compiling’ the simulation rules into a more specific form, namely by
partitioning the simulation rules using appropriate modularity in the simulation. An example
of this procedure is exhibited.
KEYWORDS: Constraint Search, Constraint Logic Programming, Proof, Emergence,
Tendencies
1. Constraint Declarative Programming and Exploration of Simulation
Trajectories
Several approaches have turned up as an answer to the need for declarative programming
with a more flexible manipulation of the semantic than traditional Logic Programming
(LP) and forward chaining systems do. The call has been for techniques allowing a
semantic driven search (Frühwirth et al., 1992).
The first answer came from Logic Programming, Constraint Logic Programming,
commonly using Prolog, both as a platform and as the programming style (Frühwirth et
al., 1992). It is based in backward chaining inference. A second answer came from Rule
Based Forward Chaining systems. Examples are Constraint Handling Rules (CHR and its
improved version CHRv; see Abdennadher et al., 1999) Constraint Rule Base
Programming (CRP) (Liu et al.), Satchmo and CPUHR-tableux calculus (Addennadher,
1995; Idem, 1997). Among the advantages of these over CLP are allowing alternative
logical extensions via split and backtracking (e. g. Satchmo, CHRv, CRP), introduction of
1 Also a member of the Centro de Simulación y Modelos (CESIMO: Centre for Simulation and
Modelling) and the Department of Operations Research of the University of Los Andes, Venezuela
(http://cesimo.ing.ula.ve/).
user defined constraints (e.g. CHR) and Meta and Higher-Order reasoning via re-writing
of rules (e. g. Satchmo).
Until now Constraint Logic Programming’s aim has been to search for a solution or
satisfaction of a goal. In CLP the aim is to look for a proof (like in LP) while in constraint
forward chaining systems to find a model satisfying certain conditions has been the
purpose. So, in the first approach the conclusion is based (in some sense) in a whole
exploration while in the second only one among the possible solutions is searched (one
extension).
A similar situation to this in LP has appeared in other areas of research, for example,
in simulation – a model oriented search approach. There usually constraints are given via
fixed parameters of the model and choices representing simulation optional trajectories.
For example, in Multi-Agent Systems (MAS) choices might be due to agents’ decisions.
In traditional simulation, the dynamics or the program is generally analysed using
Scenario Analysis (Domingo et al., 1996) and Monte Carlo techniques (Zeigler, 1976). In
the first the dynamics of each trajectory is considered while in the second a sample of
them is studied through statistical techniques. Due to the fact that a lot of computational
resources are required in the former and several not always desirable assumptions and
simplifications are made in the later, there is still a need for more appropriate techniques
for analysing simulation dynamics. For example, in social simulation it is important to
search for tendencies being non-contingent, that is, being common to several paths, and
non-expected from the simulation design (Edmonds, 1999). These are called emergent
tendencies. The theoretical value of the analysis of the dynamics of a simulation can be
seen, for example, in alignment of models (Axtell et al., 1996).
On the other hand, declarative programming seems suitable for simulation due to its
flexibility, expressiveness and the correspondence of the simulation extension to certain
logic extensions allowing formalisations and a promise for formal proofs. In this sense,
declarative programs open new ways for exploring simulation dynamics, apart from the
named traditional methods. However, in applications like social simulation proving
usually is almost impossible because of the huge amount of required computation – too
many simulation trajectories appear.
Constraint logic programming seems promising for alleviating these drawbacks. A
systematic - controlled constraint search for alternative trajectories in a simulation will
allow bring in stronger conclusions and more assertive and fruitful theoretical
exploitation of the experience than when using traditional tools. Constraint search is not
something new in declarative programming. Both backward and forward chaining
inference based tools have been developed for constraint reasoning. However, these
approaches could not be straightforwardly appropriated for simulation applications.
Our main goal in this presentation is to propose a framework for a constraint search of
tendencies in simulation trajectories and a technique for implementing it in a declarative
simulation language. Constraints in simulation are due to parameters of the model and
choices, where each ‘choice’ means that the simulation takes one of the possible
‘trajectories’. We are particularly interested in searching for emergent tendencies. The
search will be in a subspace of trajectories defined by the range of allowed constraints
and parameters, and the logic of the simulation language. Additional advantages to make
the exploration efficient can be taken from this semantic driven search -it will be the case
2
in the technique to be proposed. We used a declarative simulation language (SDML:
Strictly Declarative Modelling Language; Moss et al., 1998) suited for constraint searches
because of its facilities for backward and forward simulation, backtracking, re-writing of
rules and an internal assumption manager allowing certain predefined manipulations of
constraints.
We will begin in section 2, by outlining the main features of SDML for a constraint
search. The implementational concerns of the technique, i.e. the proposed architecture
for doing the constraint-based model search in a “hunt” of tendencies is described in
section 3. Following this (section 4), we will give an example of a technique applying this
architecture. Then in section 5, we will present a case where this methodology using the
technique previously described is used. In section 6, we briefly position this approach
with respect to general theorem proving, proving in Multi-Agent Systems and constraint
logic programming. Finally, some conclusions are made.
2. Towards the implementation of a suitable platform for a constraint
envelope of trajectories using SDML
SDML (Strictly Declarative Modelling Language; Moss et. al, 1998) is the declarative
Multi-Agent System in which we have developed our experiments. As a source of
comparisons and ideas, we have also programmed our model in the Theorem Prover
OTTER (McCune, 1995; Wos et. al, 1965; Wos, 1988; Chiang et al., 1973).
2.1 Relevant characteristics and features SDML offers are
• Good underlying logical properties of the system. SDML’s underlying logic is the
Strongly Grounded Autoepistemic Logic (SGAL) described by Kurt Konolige
(Konolige, 1995).
• Its backtracking procedure facilitates the exploration of alternative trajectories via
the splitting of simulation paths according to agents’ choices and model’s
parameters.
• The assumptions manager in SDML tracks the use of assumptions. Assumptions
result from choices.
• A good collection of useful primitives relevant to, for example, social simulation.
• The type meta-agent. A meta-agent (meta, for our purposes) is an agent
“attached” to another agent as a controller; it is able to program that agent. This is
used here not as an agent per se but as a module used to ‘compile’ rules into an
efficient form as well as to monitor and control the overall search process and
goals.
2.2 Internal Manipulation of Constraints in SDML
A
partition
is
a
grouping of rules according to their dependencies SDML
does.
Dependencies among rules in different partitions give dependencies among
partitions. Rules in a partition do not have dependencies on the subsequent partitions.
Assumptions are made for each partition in accordance to choices in such a partition.
The SDML assumption manipulator is a sort of Truth Maintenance System (TMS) for
each partition (see figure 1). For example, when certain value for a variable has been
deduced under two different assumptions, then a disjunction of the two original
assumptions is placed for this datum in the database. This reasoning is helped by the
introduction of choices, when defining data and applying rules, and backtracking. Once a
3
Rulebase (user application)
Meta Module
Partition 1
Inference Engine
Partition n
Checking for firing and
partitioning of rules
Backtracking
Database
Assumptions
Manager
Assumptions
Figure 1. Overview of SDML’s framework.
contradiction (e.g. the predicate false in the consequent of a rule) is found the system
backtracks and a new choice is made in that partition. When all choices have been
unsuccessfully tested in a partition the system backtracks to a previous partition to make a
different choice there.1
Meta is a module to write rules in the system at the beginning of the simulation. It
permits manipulations and reasoning in a higher level. Its semantic manipulation
obviously increases flexibility, e.g. to write rules referring explicitly to instances of a type
rather than to types. Meta will be one of the corners stones of the technique to be
presented. Such a manipulation will be very useful both, for driving the search and for
making it efficient in terms of computational time.
In SDML, choices are introduced via built in predicates. For example, the primitive
randomChoice allows choosing randomly from several alternative values. Each choice
will define a different simulation path labelled with an assumption. Another example is
notInferred, one of the primitives allowing non-monotonic reasoning, which allows
generation of data when a certain fact is not present in the database, e.g. notInferred b,
implies c, will put c in the database if b is not in the database. If this value, b, is written
later during the search in the database, then the assumption becomes false. In case the rule
written b is in a different partition from the one where the assumption was made, then c
and any other data supported by the assumption is withdrawn via backtracking to the
partition where the assumption was made.
3. Constraint Model Generation and Envelope of Trajectories
3.1 Constrained exploration of trajectories
We propose the use of an exhaustive constraint-based search over a range of possible
trajectories in order to establish the necessity of postulated emergent tendencies. Thus a
subset of the possible simulation parameterisations and agent choices is specified; the
target emergent tendencies are introduced in the form of negative constraints; and an
automatic search over the possible trajectories is performed. (See figure 2).
4
(QYHORSHRIDOOWHQGHQFLHV
$VXEVSDFHRIWHQGHQFLHV
$OWHUQDWLYHVHWWLQJ
RISDUDPHWHUV
6WUXFWXUHRIWKHPRGHOIRUFHUWDLQ
FRPELQDWLRQRISDUDPHWHUV
7UDMHFWRULHVIRUDVWUXFWXUH%UDQFKHV
DUHGXHWRDJHQWV¶FKRLFHV
0DSSLQJRIFHUWDLQ
WHQGHQF\ IURP WKH
WUDMHFWRULHV
Figure 2.A constraint-based exploration of possible simulation dynamics
3.2 Characterising the envelope of tendencies
In order to distinguish between the exceptional and the representative in a simulation, we
will formally describe the envelope of certain tendencies in a simulation. This might be
done by:
• Certain properties satisfied by the observed tendency.
• A mathematical description of a subspace of the tendencies or of a subspace given
a bound of the tendencies.
• Representative or typical instances of such a tendency.
• A mapping from the setting of trajectories, as given by the alternative arrangement
of parameters and agents’ choices, to certain knowledge (maybe properties) about
the tendency: (parameters X choices)Å (know. of the tend.)
3.3 Proving the necessity of a tendency
We want to be able to generalise about tendencies going from observation of individual
trajectories to observation of a group of trajectories generated for certain parameters and
choices. Actually, we want to know if a particular tendency is a necessary consequence of
the system or a contingent one. For doing this we propose to translate the original MAS
along with the range of parameterisations and agents’ choices into a platform (described
in the next section) where the alternative trajectories can be unfolded. Each trajectory will
correspond to a possible trajectory in the original MAS. Once one trajectory is shown to
satisfy the postulated tendency another set of parameters and agents’ choices is selected
and the new trajectory is similarly checked. If all possible trajectories are successfully
tested, the tendency is proved to be necessary relative to the logic of the simulation
language, the range of parameterisations and agents’ choices.
The idea is to translate the MAS into a constraint-based platform in an automatic or
near automatic way without changing the meaning of the rules that make it up in order to
perform this automatic testing. In this way a user can program the system using the
agent-based paradigm with all its advantages; inspect single runs of the system to gain an
intuitive understanding of the system and then check the generality of this understanding
for fragments of the system via this translation into a constraint-based architecture.
In the example shown below, all trajectories are explored for one combination of
parameters, eight agents’ choices per iteration and seven iterations. A simple tendency
was observed characterised by a mathematical description of its boundaries. This
5
characterisation was handled as a theorem. The theorem was proved to be necessary
following a procedure similar to the one described in the previous paragraphs.
3.4 What is new in this model-constrained methodological approach
It is our goal in this paper to propose an alternative approach for exploring and analysing
simulation trajectories. It will allow the entire exploration and subsequent analysis of a
subspace of the whole space of simulation trajectories. We are suggesting the generation
of trajectories in a semantically constrained way. Constrictions will be context-dependent
(over the semantics of the trajectory itself) and will be driven via the introduction of a
controller or meta-module.
4. Implementing a suitable constraint-based programming platform for
the envelope of trajectories
The main goal of the programming strategy to be described is to increase the efficiency in
terms of simulation time, thus making an efficient constraint-based search possible. The
improvements will be achieved by making the rules and states more context-specific. This
enables the language’s inference engine to exploit more information about the logical
dependencies between rules and thus increase the efficiency.
4.1 An Overview of the system.
We implemented the proposed architecture in three modules; let us call them model,
prover and meta. The following diagram illustrates this:
PRGHO
6XEPRGXOHV
,QLWLDOLVDWLRQ
6SHFLI\SDUDPHWHUV
6SHFLI\&KRLFHV
&DOFXODWLRQVDQG
GHFLVLRQV
$JHQW JLYHQ WKH JHQHUDO
HQYLURQPHQW RI WKH SURRI
SURYHU
6XEPRGXOHV
,QVWDQFHVRI7UDQVLWLRQ5XOHV
,QVWDQFHRIUXOHIRUWKHRUHP
FKHFNLQJ
PHWD
6XEPRGXOHV IRU ZULWLQJ
UXOHVIRU
6WDWH7UDQVLWLRQ
&KHFNLQJWKHRUHP.
$JHQW H[HFXWLQJ WKH WUDQVLWLRQ UXOHV
DQGFRQVHTXHQWO\WKHSURRI
$JHQWUHVSRQVLEOHIRU
FRQWUROOLQJWKHSURRI
)LJXUH3K\VLFDOYLHZRIWKHV\VWHP
4.2 Description of System Modules
We have found it convenient to distinguish and model as distinct entities three basic
elements of a simulation: the static structure of the model, the dynamics of the simulation
and the way this dynamics is “managed” by certain meta-rules or by a controller. Each of
those entities is programmed in a different module:
1. model, sets up the structure of the model, that is, it gives the environment of the
simulation: range of parameters, initialisations, alternative choices and basic
(backward chaining) rules for calculations.
2. prover, generates the dynamics of the simulation. This is a sub-module of model
(i.e. it is contained in model). This will basically contain the transition rules,
auxiliary rules for generating pre-processing required data and the conditions to
6
test the necessity of the theorem. All of them are rules to be executed while the
simulation is going on.
3. meta, is responsible for controlling the dynamics of the simulation. Its meta-rules
write the transition rules and the theorem in (as well as others required by) the
module prover. A picture of the system is given in Figure 3.
4.3 Program dynamics
Modules’ rules are executed in the following sequence:
1. model: initialising the environment for the proof (setting parameters, etc..)
2. meta: creating and placing the transition rules in prover.
3. prover: carrying on the simulation using the transition rules and backtracking
when a contradiction is found.
The program backtracks from a path once the conditions for the theorem are verified,
then a new path with different choices and/or parameters is picked up. Next figure
describes a transition step.
Meta-prover, writes on Prover the rules for
period 1, if it is the beginning of the
simulation; or for period STIi+1 and using
data from periods 1, 2,… STIi, in other cases.
Prover just after the model has settled
the parameters of the simulation and
the initialisations (i = 0), or after
certain STI, let’s say STIi.
meta-prover
transition
prover
prover
STIi
STIi+1
)LJXUH6WDWHWUDQVLWLRQIURP67, WR67,L
L
4.4 Split of the rules: a source of efficiency.
In forward chaining simulation the antecedent retrieves instance data from the past in
order to generate data for the present (and maybe the future):
past facts Å present and future facts
Traditionally, the set of transition rules are implemented to be general for the whole
simulation. A unique set of transition rules is used at any STI (Simulation Time Instant or
iteration).
As the simulation evolves, the size of the database increases and the antecedents have
to discriminate among a growing amount of data. At STIi, there would be data from (i-1)
alternative days matching the antecedent. As the simulation evolves it becomes slower
because of the discrimination the program has to carry out among this (linearly) growing
amount of data.
Using the proposed technique, we would write a transition rule for each simulation
time. The specific data in the antecedent as well as in the consequent could be instanced.
Where possible, a rule for each datum, the original rule will generate, would be written.
This will be illustrated in the example of the next section.
7
2ULJLQDOWUDQVLWLRQUXOHOHWXVVD\WRZULWHWKHYDULDEOH9
One rule per each:
time period number ««L«Q
SDUDPHWHU M«P
MP «M«P
6SOLWUXOHIRUGD\L DQGSDUDPHWHUS
WKHDQWHFHGHQWFRQWDLQV
H[SOLFLW UHIHUHQFH WR GDWD JLYHQ SHU WUDQVLWLRQ UXOHV LQ
WKLVRULQSUHYLRXVLWHUDWLRQV «L
H[SOLFLW UHIHUHQFH WR SDUDPHWHUV JLYHQ LQ LQLWLDOLVDWLRQ
RUGHWHUPLQHGLQWKLVRULQSUHYLRXVLWHUDWLRQV DQG
FDOOUXOHVIRUFDOFXODWLRQVDQGGHFLVLRQV
WKHFRQVHTXHQWJLYHV
9DWWLPHSHULRGL
YDOXHVRIWKHYDULDEOH
Figure 5. Splitting of rules by time period and a combination of parameters.
This technique represents a step forward in improving the efficiency of declarative
programs. One could, in addition, make use of partitions and time levels to introduce
further modularity – this would further speed up the search process and increase the
memory that is needed. Partitions permit the system to order the rules to fire in an
efficient way according to their dependencies. Time levels let us discriminate among data
lasting different amounts of time. The splitting of rules lets us discriminate among the
transition rules for different simulation times given a more specific instancing of data at
any STI.
4.5 Measuring the efficiency of the technique
Comparing the two programs, the original MAS simulation and the constraint-based
translation we obtain a speed up by a factor of O(NM), where N is the number of agents
and M is the number of STIs. SDML already has facilities for discriminating among STIs,
but their use is not convenient for the sort of simulation we are doing (exploring scenarios
and/or proving) because of the difficulties for accessing data from any time step at any
time. If we had used this facility still the simulation would have been speeded up by N.
Notice that all these values are only estimations because a program stops trying to fire a
rule as soon as it finds out that one of its clauses is false.
It is clear that the greater the number of entities in the simulation or the number of
STIs, the larger the benefits from the technique. We must notice that the speeding up of
the simulation is only one dimension of the efficiency given by the technique.
4.6 Translating a traditional MAS architecture into a model-exploration MAS
architecture.
Before splitting the rules the original MAS is reduced in a sort of unencapsulation of
the hierarchy of agents into the architecture shown in figure 3. Additional variables must
8
be added into predicates and functions in order to keep explicit the reference to the
"owner" agent of any instance of a variable. This will facilitate the check for tendencies,
the testing of the theorem and any other data manipulation. It is as if the agent where
replaced by its rulebase, see figure 6.
In the original architecture, each agent has its own rulebase (RB) and database (DB).
The agent’s structure is given by its set <RB, DB> as well as by the structure of any
subagents.
MAS Architecture:
each square represents an agent (A) with its rulebase (RB) and database (DB)
UNIVERSE (or A): <RB, DB>
A2: <RB2, DB2>
A1 : <RB1,DB1>
A3 : <RB3,DB3>
A2.1 : <RB2.1, DB2.1>
A2.2: <RB2.2, DB2.2>
Logically partitioned: each square represents a partition
unencapsulation
partition 1
Main
Partition
partition 3
partition 2.1
partition 2
partition 2.2
)LJXUH8QHQFDSVXODWLQJD0$6¶DUFKLWHFWXUH
Using the technique, the initialisation of the static structure is accomplished by the
module model, as explained above. The transition rules (dynamic structure) will be set up
by the module meta into the module prover. There is still a hierarchy, both in the
structure of the model and in the dynamics of the simulation – it is given by the
precedence in the rulebase partition. Additional partitions will appear as the rules are
split, discriminating among rules according to the factors used for splitting (e.g. STIs,
consumers and producers in the example, see below).
5. An example.
A simple model of a producer-consumer system, which was previously built in SDML
and in the Theorem Prover OTTER, was rebuilt using the proposed modelling strategy. In
the new model the exploration of possibilities is speeded up by a factor of 14. Also, the
model built in OTTER, though faster than the original model in SDML, is several times
slower than the improved model built in SDML.
Some of the split transition rules were the ones for creating (at each STI) producers’
prices and sales, consumers’ demand and order, warehouses’ level and factories’
production. Among the rules for auxiliary data split were the ones for calculating: totalorder and total-sales (a sum of the orders for all producers), total-order and total-sales per
producer, and total-order and total-sales per consumer.
9
5.1 Example of a split rule: Rule for prices.
This rule calculates a new price for each producer at each STI (which we called day),
according to its own price and sales, and the price and sales of a chosen producer, at the
immediately previous STI.
The original rule in SDML was like this:
for all (producer)
for all (consumer)
for all (day)
(
price(producer,myPrice,day)
totalSales(totalSales,day)
sales(producer,mySales,day)
choiceAnotherProducer(anotherProducer)
price(anotherProducer,otherPrice, day)
calculateNewPrice(mySales,totalSales,
otherPrice,myPrice,newPrice)
implies
price(producer, newPrice, day + 1)
)
and
and
and
and
and
The new rule (in the efficient program) will be “broken” making explicit the
values of prices and sales per each day.
In the following, we show the rule per day-i and producer-j:
for all (consumer)
(
price(producer-j, myPrice, day-i)
totalSales(totalSales, day-i)
sales(producer, mySales, day-i)
choiceAnotherProducer(anotherProducer)
price(anotherProducer, otherPrice, day-i)
calculateNewPrice(mySales,totalSales,otherPrice,
myPrice,newPrice)
implies
price(producer-j, newPrice, (day-i) + 1)
)
and
and
and
and
and
If the name for the clauses price and sales (e. g.those clauses associated with the data
recalculated at each STI) are used to make explicit the day, the rule will have the
following form. It is important to observe that only one instance of newprice in the
consequent is associated with only one transition rule and vice verse:
for all (consumer)
(
price-i(producer-j, myPrice)
totalSales-i(totalSales)
sales-i(producer-j, mySales)
choiceAnotherProducer(anotherProducer)
price-i(anotherProducer, otherPrice)
calculateNewPrice(mySales,totalSales, otherPrice,
myPrice,newPrice)
implies
price-(i+1)(producer-j, newPrice)
)
10
and
and
and
and
and
5.2 Other facts about the example
There were eight “types” of rules (see table below) involved in the application of the
technique, those associated with generating the dynamics of the simulation. Excluding the
rules for testing the theorem and setting up producers’ choices, which did not suffer
additional split, the other six rules in the original model where split into 96 rules in the
new model. All of them were split by transition time step (six transitions); among these
six, two suffered additional split by producer (there were three producers), and one,
among the last two, was also split by consumer (there were three consumers). This gives:
(4 + (1 + (1 * 3)) * 3) * 6 = 96 rules in the new model replacing the referred six rules in
the old model.
'HVFULSWLRQRIWKHUXOH UXOHIRU
1RRIUXOHVLQWKH 5XOHVSOLWE\
1R RI UXOHV LQ
RULJLQDOPRGHO
WKHQHZPRGHO
&KHFNLQJWKHRUHP
3URGXFHUV¶ FKRLFH RI DQRWKHU FRQVXPHU IRU
DOUHDG\ VSOLW E\
FRPSDULQJFHUWDLQGDWDZKHQFKDQJLQJSULFH
67,
&DOFXODWLQJWRWDORUGHUVE\FRQVDQGSURG
67,
&DOFXODWLQJWRWDOSURGXFHUV¶VDOHV
67,
'HWHUPLQLQJ ' FRQVXPHUV¶ GHPDQG DQG
67,
RUGHU
'3URGXFHUV¶SULFH
67,
'3URGXFHUV¶SURGXFWLRQDQGOHYHOLQVWRUH
67,DQGSURGXFHU
'3URGXFHUV¶VDOH
67,SURGXFHUDQG
FRQVXPHU
7DEOH&RPSDULQJWKHQXPEHURIUXOHVLQWKHRULJLQDODQGLQWKHQHZLPSOHPHQWDWLRQ
Though splitting of the rules increases the necessary amount of memory for keeping
the rules, it together with partitioning of the rulebase according to rule dependencies,
which has already been implemented in SDML, allows a much faster instantiation of the
rules and consequently speeds up the simulation. This is due to the fact that rules refer to
data more explicitly so that when intending to fire a rule the data considered to check the
antecedent of the rule is limited to a smaller part of the database. The size of the searched
part of the database keeps stable as the simulation time progresses while without splitting
this search space, in SDML, might grow linearly as the simulation time goes forward.
Rules split by only STI might also have been “broken” by consumer and producer, but it
was not necessary as a model fast enough for our purpose of proving the necessity of
certain tendency was obtained with the described implementation. It is more in our
interest to use this experience more intensively in future modelling.
5.3 What the technique enables
In this example, the described technique was used to prove that the size of the interval of
prices (that is: biggest price - smaller price, each day) decreases over time during the first
six STIs over a range of one parameterisation and eight choices for the agents at each STI.
An exponential decrease of this interval was demonstrated in all the simulation paths. A
total of 32768 simulation trajectories were tested. It was not possible to simulate beyond
this number of days because of the limitations imposed by computer memory. The
complete search process took only 24 hours.
11
Though the tendency we have shown is simple and quantitative, it is obvious that the
technique is applicable in more interesting cases of emergent tendencies, even if they
have a qualitative nature.
This technique is useful not only because of the speeding up of the simulation but also
for its appropriateness when capturing and proving tendencies under the specified
constraints. In the example, the meta-module was used to write the rule with the
hypothesis (theorem) to be tested on prover-module at the beginning of the simulation. If
the meta-module were able to write rules on prover-module while the simulation is going
on, the theorem we wanted to prove could be adapted according to the results of the
simulation via relaxing constraints. For example, the technique could be implemented in
a way that we only give the program hints related to the sort of proof we are interested in.
Then the meta-module would “elaborate”, via adapting over time in a context dependent
manner, a set of hypotheses or theorems.
6. Other Approaches
6.1 Using OTTER (McCune 1995), a resolution-style first order Theorem Prover
In simulation, strategies like a Future Event List (FEL), in event-driven simulation, and
partition of the space of rules and a hierarchy of nested time levels, in declarative
simulation systems (e.g. some MAS), are used. The criteria for firing rules is well
understood, and procedures like weighting and subsumption usually are not necessary.
Additionally, redundant data for some purpose could be avoided in MAS with appropriate
compilation techniques.
The advantages given for the weighting procedure in OTTER are yielded in MAS
systems like SDML by procedures such as partitioning, where chaining of the rules
allows firing the rules in an efficient order according to their dependences.
The main difficulty for our simulation purposes when using OTTER was the lack of
facilities for accessing the database. Clearly the introduction of a meta-module and other
facilities for reasoning about predicates and rules in SDML brings a great improvement in
this sense.
6.2 DESIRE
Among other approaches for the practical proof of MAS properties, the more pertinent
might be the case conducted by people working in DESIRE (Engelfriet et. al., 1998).
They propose the hierarchical verification of MAS properties, and succeeded in doing this
for a system. However, their aim is the verification of a computational program – it is
proved that the program behaves in the intended way. It does not include the more
difficult task, which we try to address, of proving properties of the simulation dynamics.
6.3 Satchmo and other CLP constraint programs
Some of Satchmo’s and other constraint programming languages’ facilities are similar to
SDML’s ones, for example, backtracking and Satchmo’s false predicate. However, they
present certain built-in facilities for manipulation of constraints that SDML has not. For
instance, reasoning about terms in CLP(X) or consistence techniques to prune the range
of trajectories in other CLP (Frühwirth, 1992). Instead, SDML allows facilities to
introduce alternative values for the manipulated entities (e.g. predicates, clauses, integer
variables) which can be used as constraints (clauses for choosing, e.g. randomChoice) as
well as a meta module able to reason about terms or rules. A meta module can build rules
12
taking advantage of the simulation semantics. Permitting it to act while the simulation is
going on will allow it to adapt the search to the simulation results. Because of all this
SDML is able to control the manipulation of constraints flexibly and transparently for the
user.
On the other hand, in CLP the aim is to look for a proof while in constraint forward
chaining systems the purpose is to find a model satisfying certain conditions. In the first
group of languages the conclusion is based (in some sense) in a whole (implicit)
exploration of trajectories while in the second group only one among the possible
solutions is searched (based on one trajectory). In contrast we have proposed a
methodology for searching and proving the necessity of a tendency in a subspace of
trajectories, which seems more appropriate in many applications where a broaden proving
is prohibitive due to the huge amount of required computational resources.
It should be possible to rewrite the example in some of these programs, for instance in
CHRv. It is our purpose to enrich our methodological approach from the experience of
people working in constraint programming. It will be one of our tasks to rewrite future
models in some of these languages in order to bring in new ideas.
7. Conclusions and future work
We have argued and shown using an example the pertinence of a methodology for a
constrained exploration and envelope of trajectories as a complement to traditional
methods dealing with post-hoc analysis of the dynamics of simulations. We have
suggested a forward chaining semantically constrained generation of trajectories.
Like in Constraint Logic Programming the constraint of the generation of simulation
extensions respond to the need for a systematic and controlled exploration of the
simulation dynamics. Our aim is to analyse subspaces of trajectories rather than searching
for a model. When searching for a model, a “committed” or “non-determinist” (do not
care about determinism in Abdennadher words; Abdennadher, 1999) search, i.e. search
for an extension without backtracking, is done.
A platform to implement this methodology has been proposed. It consists of a
modular structure according to strategic parts of a simulation: a first module, model, sets
up the static structure of the simulation; then a second module, prover, generates the
dynamics of the simulation; and finally a meta-module is responsible for controlling the
dynamics of the simulation. The second characteristic of this platform is a partitioning of
the space of rules and splitting of transition rules by STI, parameters and choices.
The control of the search via a meta-module makes the manipulation of the
constraints more flexible, transparent and handy for the user (a controller in an even
higher level) than when it is introduced into the same level than the transition rules of the
simulation. Constrains are context-dependent (over the semantic of the trajectory itself) as
the meta-module is able to access the semantics of the simulation setting up in advance
one among the possible combination of agents’ choices and model parameters for each
simulation run.
In the offered example only a combination of parameters but all possible
combinations of agents’ choices were followed. The meta-module sat up a possible
combination of choices after the combination of parameters of the model was defined and
changed it after backtracking. Backtracking occurred as soon as certain conditions
characterizing certain tendency where identified in the trajectory. This allowed prove the
13
necessity of the tendency for that combination of parameters and the range of agents’
choices. Thought in the example the tendency was identified by the modeller, also a
monitoring module aiming to recognize relevant tendencies might have been introduced.
In this example the meta-module acts only at the beginning of the simulation. More
powerful and flexible programs are possible when the meta-module is able to act at any
STI. Transition rules can be evolving, in a sense that the meta-module builds the rules for
a STI once every fact for the previous STI are known. This allows adaptation (e.g.
relaxing) of the constraints defining the theorem according to the “discovered”
circumstances into the explored dynamics of the simulation.
The splitting of the transition rules represents a step forward in improving the
efficiency of declarative programs, one additional to the use of partitions and time levels.
It lets us discriminate among the transition rules for different simulation time steps given
a more specific instancing of data at any time step. Thus this alleviates some of the
drawbacks of declarative programming due to the necessary grasping and updating of all
state variables at any STI.
Our future work will be focused both in the search for a refining and improving of
the presented platform as well as in its application for proving more interesting cases of
emergent tendencies.
Acknowledgements
SDML has been developed in VisualWorks 2.5.2, the Smalltalk-80 environment produced by ObjectShare.
Free distribution of SDML for use in academic research is made possible by the sponsorship of ObjectShare
(UK) Ltd. The research reported here was funded by CONICIT (the Venezuelan Governmental
Organisation for promoting Science), by the University of Los Andes, and by the Faculty of Management
and Business, Manchester Metropolitan University.
References
Abdennadher S.(1999), “Constraint Handling Rules: Applications and Extensions”, Invited Talk, 2nd
International Workshop on Optimization and Simulation of Complex Industrial Systems. Extensions
and Applications of Constraint-Logic Programming and 7th International Workshop on Deductive
Databases and Logic Programming, Tokyo, Japan, in conjunction with the 12th International
Conference on Applications of Prolog, INAP'99.
Abdennadher. S. and H. Schütz(1997), “Model Generation with Existentially Quantified Variables and
Constraints”, Sixth International Conference on Algebraic and Logic Programming, Springer LNCS.
Abdennadher, S., F. Bry, N. Eisinger and T. Geisler (1995), “The Theorem Prover Satchmo: Strategies,
Heuristics, and Applications - System Description”, Journées Francophones de Programmation en
Logique, JFPL'95, Dijon.
Axtell, R., R. Axelrod, J. M. Epstein, and M. D. Cohen (1996), “Aligning Simulation Models: A Case of
Study and Results”, Computational Mathematical Organization Theory, 1(2), pp. 123-141.
Chiang, C.L., and R. C.T. Lee (1973), Symbolic Logic and Mechanical Theorem Proving, Academic
Press, London, UK
Domingo C., G. Tonella and O. Terán (1996), “Generating Scenarios by Structural Simulation”, in AI,
Simulation and Planning High Autonomy Systems, The Univ. of Arizona, pp 331-336.
Edmonds, B. (1999), “Modelling Bounded Rationality In Agent-Based Simulations using the Evolution of
Mental Models”. In Brenner, T. (Ed.), Computational Techniques for Modelling Learning in
Economics, Kluwer, 305-332.
Engelfriet, J., C. Jonker and J. Treur (1998) “Compositional Verification of Multi-Agent Systems in
Temporal Multi-Epistemic Logic”, AI Group, Vrije Universiteit Amsterdam, The Netherlands.
14
Frühwirth, T., A. Herold, V. Küchenhoff, T. Le Provost, P. Lim, E. Monfroy and M. Wallace (1992).
Constraint Logic Programming - An Informal Introduction”, Chapter in Logic Programming in Action
(G. Comyn et al., Eds.), Springer LNCS 636, September.
Liu, B., Jaffar J. and Yap R., "Constraint Rule-Based Programming", School of Computing, National
University of Singapore, Singapore. Accessible at:
http://www.comp.nus.edu.sg/~joxan/res.html.
Konolige, K. (1995), “Autoepistemic Logic”, in Handbook of Logic in Artificial Intelligence and Logic
Programming (4),Oxford Science Publications, pp. 217-295.
McCune, W. (1995), OTTER 3.0 Reference Manual Guide, Argonne National Laboratory, Argonne, IL.
Moss, S., H. Gaylard, S. Wallis, B. Edmonds (1998), “SDML: A Multi-Agent Language for
Organizational Modelling”, Computational Mathematical Organization Theory, 4(1), 43-69.
Moss, S., B. Edmonds, S. Wallis (1997), "Validation and Verification of Computational Models with
Multiple Cognitive Agents", Centre for Policy Modelling Report, CPM-97-25,
http://www.cpm.mmu.ac.uk/cpmrep25.html
Wos, L. (1988), Automated Reasoning: 33 Basis Research Problems, Prentice Hall, New Jersey, USA.
Wos, L., Robinson, G. A., and Carson, D. F. (1965), “Efficiency and Completeness of the Set of Support
Strategy in Theorem Proving”, J. ACM 14, N° 4, 698-709.
Zeigler, B. (1976), Theory of Modleling and Simulation, Robert E. Krieger Publishing Company, Malabar, Fl, USA.
15
| 6cs.PL
|
arXiv:0807.1475v1 [cs.CE] 9 Jul 2008
Simulations of Large-scale WiFi-based
Wireless Networks: Interdisciplinary
Challenges and Applications
Maziar Nekovee and a,b Radhika S. Saksena b
a Mobility
b Centre
Research Centre, British Telecom, Polaris 134, Adastral Park,
Martlesham, Suffolk IP5 3RE, UK
for Computational Science, University College London, 20Gordon Street,
London WC1H 0AJ, UK
Abstract
Wireless Fidelity (WiFi) is the fastest growing wireless technology to date. In addition to providing wire-free connectivity to the Internet WiFi technology also enables
mobile devices to connect directly to each other and form highly dynamic wireless
adhoc networks. Such distributed networks can be used to perform cooperative communication tasks such ad data routing and information dissemination in the absence
of a fixed infrastructure. Furthermore, adhoc grids composed of wirelessly networked
portable devices are emerging as a new paradigm in grid computing. In this paper
we review computational and algorithmic challenges of high-fidelity simulations of
such WiFi-based wireless communication and computing networks, including scalable topology maintenance, mobility modelling, parallelisation and synchronisation.
We explore similarities and differences between the simulations of these networks
and simulations of interacting many-particle systems, such as molecular dynamics (MD) simulations. We show how the cell linked-list algorithm which we have
adapted from our MD simulations can be used to greatly improve the computational performance of wireless network simulators in the presence of mobility, and
illustrate with an example from our simulation studies of worm attacks on mobile
wireless adhoc networks.
Key words: Wireless Computer Networks, Large-scale Simulations, Interacting
Many-Particle Systems, Wireless Grids
PACS:
Preprint submitted to Elsevier
9 July 2008
1
Introduction
Modern world has become increasingly mobile. As a result, traditional ways of
connecting users to the Internet (and to each other) via physical cables have
proved inadequate. Wireless communications [1], on the other hand, poses no
restrictions on the user’s mobility and allows a great deal of flexibility, both on
the part of users and service providers. Wireless connectivity for voice via mobile telephony made it possible for people to connect to each other regardless
of location. This has had a profound influence on the business of telecommunications, as well as the society as a whole [2]. New wireless technologies
targeted at computer networks promise to do the same for Internet access,
connecting wirelessly not only laptops and portable devices but also millions
of cars, sensors, consumer devices, etc to each other and to the global Internet.
The most successful, and fastest growing, example of such wireless technologies
is WiFi (Wireless Fidelity)[3]. Like cellular technology, WiFi uses a number of
base stations to connect user devices to an existing fixed network (these base
stations are called access points). However, unlike cellular systems which are
centralised, WiFi systems operate in a highly distributed fashion. Each WiFi
device is responsible for managing its own connectivity, mobility and access to
the radio spectrum. Furthermore, unlike cellular systems, nearby WiFi devices
can directly connect to each other and form self-organising wireless adhoc
networks [4,5]. Such networks are highly dynamic and flexible. They can be
created (and torn down) on the fly in order to route data packets between
participating devices, or to the closest Internet gateway. Adhoc technology can
also be used to connect together a collection of WiFi accesspoints which then
form a so-called mesh network [6]. This can help to greatly extend the range
of WiFi coverage without the need for connecting every single accesspoint to
the fixed network.
Initially WiFi technology was used to provide connectivity to “nomadic” users
in coffeshops, airports etc, and for wire-free Internet access in homes and offices. The last few years, however, have seen the emergence of much more ambitious applications of this technology. For example, it is expected that WiFibased wireless access will enable the coverage of entire cities, thus providing
citizens on the move with high-speed (11-54 megabits per second) connectivity. Other frontiers in WiFi technology include high-speed Internet access
to automobile users, and WiFi-based vehicular adhoc networks and vehicular
grids [7].
In addition to the above applications in telecommunications, wireless adhoc
grids based on WiFi and related technologies are emerging as a new paradigm
in grid computing [8,9]. Adhoc grid environments enable mobile users to join
together wirelessly and share computing resources, services and information
2
[9]. One example of such adhoc grids are wireless sensor grids for medical,
industrial and environmental monitoring. Another one are wireless computational grids where WiFi-enabled devices are networked together in order to
perform complex computing and data aggregation tasks, which are beyond
the capabilities of a single device [7].
The increasing complexity and the very large scale of such emerging WiFi
systems has created a need for scalable high-fidelity simulation platforms that
can help scientists, engineers and network planners accurately predict and
optimise their performance prior to large-scale deployment. The aim of this
paper is to review computational challenges that are involved in creating such
simulation platforms, including scalable topology construction and mobility,
parallel and distributed simulations on grid platforms, and synchronisation.
We also show that there are interesting similarities between the simulations
of WiFi-based wireless networks and molecular dynamics (MD) simulations of
interacting many-particle systems, and illustrate how these could be exploited
in practice.
The rest of this paper is organised as follows. In section 2 we give a brief description of the main ingredients for simulations of WiFi-based networks. This
is followed by an examination of the computational and algorithmic challenges
of such simulations, and how to address these. Section 4 describes, as an example, aspects of our simulation studies of computer worm attacks on mobile
wireless adhoc networks. We close the paper in section 5 with conclusions.
2
Modelling Ingredients
There has been significant previous research in modelling [12] and simulations
of wired communication networks [10,11]. However, modelling of wireless networks is very distinct from modelling of wired networks in that the physical
medium properties, i.e. radio propagation and interference, cannot be separated from the higher layer network protocols, because strong interactions
impact performance and drive engineering design decisions. Furthermore the
ability of users to (rapidly) change their physical location while maintaining
connectivity greatly increases the dynamism of these networks, in comparison
to fixed networks. In this section we shall focus on describing these distinctive
ingredients for the modelling of WiFi-based wireless networks. Other components in high-fidelity simulations of these networks, which are outside the scope
of the current paper include modelling of various packet routing mechanisms
[1,14] and transport protocols [13] in WiFi environments.
Building on the idea of creating parallels between these networks and interacting many-particle systems, we first consider the simplest building block of
3
such networks, namely a single pair of communicating WiFi devices. We then
consider many such pairs operating in the vicinity of each other, and discuss
the interaction topologies and operational rules of the resulting systems.
2.1 A single pair of WiFi devices
At present WiFi devices are constrained by regulators to operate in pre-defined
frequency bands. Depending on the IEEE standard used in the device these
frequency bands are either in the 2.4-2.5 GHz or in the 5.2-5.8 GHz range.
Each of these bands is divided into a number of frequency channels.
Consider two WiFi devices i and j which communicate with each other using
a common frequency channel, fi . The received radio signal strength at device
j resulting from a transmission by device i depends on a variety of effects.
These include free space attenuation of radio waves, the response of the environment and mobility; the latter will be neglected for clarity of presentation
but will be picked up in section 2.5. Effects due to the environment include
reflection at surfaces, diffraction due to obstacles, and transmissions through
walls. Phenomenologically, in the absence of detailed information on the environment, these effects can be described using the so-called pathloss model
[15], which states that signal power at a receiving device j is related to the
signal power of the transmitting node i via the following equation:
P ij =
Pi
.
cfi rijα
(1)
In the above equation rij is the distance between node i and node j, P i
and P ij are the transmit power and the received power, respectively, and cfi
is a frequency-dependent constant. For free space propagation α = 2, but
depending on the specific indoor/outdoor propagation scenario it is found
empirically that this exponent can vary typically between 2 and 5. A data
transmission by node i is correctly received at node j, i.e. i can establish a
communication link with j, provided that:
P i /cfi rijα
P ij
=
≥ βth .
ν
ν
(2)
In the above equation βth is a sensitivity threshold and ν is the noise level at
node j.
Condition (2) translates into a maximum transmission range for node i:
rti
=
Pi
cfi βth ν
4
!1/α
,
(3)
such that device i can establish a wireless link with device j only if j is within
a circle of radius rti .
2.2 A collection of WiFi devices: Interference effects
In the above we considered a stand-alone sender-receiver pair of WiFi devices.
In reality, however, many pairs of nearby wireless nodes may simultaneously attempt to establish links, either between themselves or to nearby access points.
Due to the broadcast nature of radio transmissions, a radio signal transmitted
towards a specific node can interfere with communications of many nearby
network nodes and contribute to their noise level. Consequently, a successful
transmission from node i to node j depends not only on the transmit power of
node i and its distance to node j but also on the activity of all other nearby
nodes. In particular, aggregate transmission of nearby devices may result in
a situation where a transmitter-receiver pair cannot establish a link despite
the fact that they are within the range of each other. Such interference effects
need to be accurately taken into account in modelling data communications
in wireless networks in order to obtain realistic results [16].
To model interference, for each signal transmitted from a sender i to a receiver
j the aggregate received power resulting from all other nearby sender devices
needs to be computed. Signal arrival at node j is then considered successful
only if the ratio of the received power from i to aggregate noise is above the
sensitivity threshold, βth . Computing the impact of interference on the transmissions of devices is one of the most computationally expensive components
in the simulations of wireless networks as it requires the computation of O(N 2 )
pairwise interactions. However, taking advantage of the fact that the interference effect decays as 1/r α with the distance between two devices, one usually
limits the computation of the interference terms to devices which are within a
so-called interference range of a given device. Conventionally the interference
range is chosen to be rij = 2rtj 1 .
2.3 Medium Access Control (MAC)
The interference problem is perhaps the most important issue in deployment
of high-density WiFi networks. WiFi technology attempts to mitigate this
problem using a distributed random access mechanisms called Medium Access Control (MAC). The MAC protocol used by WiFi-based wireless devices
1
The precise choice of the interference range depends on both the decay exponent
α and the density of nodes in the system and may have to be increased significantly
in order to obtain accurate results [17]
5
follows the IEEE 802.11 standard [3], which specifies a set of rules that enable
nearby devices to coordinate their transmissions in a distributed manner, in
such a way that devices whose radio transmissions may interfere with each
other do not get access to the same frequency channel at the same time.
The IEEE 802.11 MAC is a complex protocol and we do not attempt to
describe here the full model of this protocol. Instead we focus on the most
relevant aspect of this protocol, the so-called listen-before-talk (LBT) rule.
This rule dictates that each device should check the occupancy of the wireless
medium before starting a data transmission and refrain for a random time
from transmitting if it senses that the medium is busy. Roughly speaking, the
net effect of this mechanism is to create an interference-free “exclusion zone”
around each transmitting device (which is roughly of the size ∼ π(2rti)2 ),
thereby reducing (but not entirely eliminating) the possibility of packet collisions. The presence of the MAC introduces novel spatio-temporal correlations
in the dynamics of data communications which need to be taken into account
in realistic simulations of these networks [35].
2.4 Graph representation of interactions in WiFi Networks
From the above models of radio propagation, an abstract communication
graph for a collection of WiFi devices can be constructed. This is achieved
by creating an edge between node i and all other nodes in the plane that are
within the transmission range of i, and repeating this procedure for all nodes
in the network. In general wireless devices may use different transmit powers
such that the existence of a wireless link from i to j does not imply that a link
from j to i also exists. Consequently the resulting communication graph is
directed. Assuming, however, that all devices use the same transmit power P ,
and a corresponding transmission range rt , the topology of the resulting network can be described as a two dimensional random geometric graph (RGG)
[18]. Similarly, one constructs an interference graph for the network by creating an edge between any two nodes which are within a radius ri of each other.
Fig 1. shows, as an examples, the communication and interference graphs created by a collection of WiFi devices distributed randomly in a 1000 × 1000
m2 rectangular area.
Mathematically, a graph is represented by a corresponding adjacency matrix A,
where the element aij = 1 if two nodes are connected, and zero otherwise. Since
the adjacency matrices corresponding to the communication and interference
graphs of WiFi networks are usually sparse, they can be efficiently encoded in
the computer memory in the form of neighbour lists.
6
Fig. 1. The communication (left panel) and the corresponding interference graph
(right panel) of a wireless adhoc network created by WiFi devices uniformly distributed in a 1000 × 1000 m2 rectangular area.
2.5
Modelling User Mobility
The movement of users carrying WiFi devices can greatly influence the performance of WiFi systems. The impact of user mobility is twofold. First of all,
as a result of mobility, devices continuously move in and out of each others’
transmission and/or interference ranges and this will result in time-dependent
network topologies. Secondly, mobility causes spatio-temporal variations in the
user density, and consequently, the traffic load offered to each access point.
Accurately modelling user mobility is therefore of great importance in highfidelity simulations of WiFi system. Several mobility models have been presented in recent simulation studies of WiFi and Bluetooth-based mobile wireless networks. The most widely used of such models assume random and uncorrelated movements of individual devices. These include the random walk
model and variations thereof, such as the random-waypoint model [19]. However, such simple models are unable to reproduce important features of user
mobility patterns, which result from a combination of correlations [22], environmental constraints [22] and social interactions between users [20].
Fortunately there has been much previous research on agent-based realistic
modelling of both human and vehicular mobility [22,21], and such models can
be coupled to network simulators in order to examine in detail the impact of
user mobility. Very recently, for example, we used microscopic car-following
models to investigate various properties of vehicular adhoc networks operating
in realistic highway traffic scenarios [27].
A computational issue in coupling high-fidelity mobility simulators to wireless
network simulators is the large difference in the time scales of the two types of
7
simulations. For example, car-following models typically update the position
of vehicles every second. On the other hand, a typical timestep of wireless
network simulations is ∼ 1 µs. To ensure accurate results the combined simulations of mobility and wireless communication should be performed using
the smallest time-step in the problem. In the above example, this means a 106
increase in the update frequency of the vehicular traffic simulator [25].
3
Computational Challenges
3.1 Network topology construction and maintenance
One of the most computationally intensive portion of the simulations of WiFi
networks is the construction of the neighbour lists which encode the topology
of the network. In static networks these neighbour lists can be constructed
once and for all at the beginning of each simulation. In networks comprising
mobile nodes, the neighbour lists need to be updated with every update of
nodes’ positions. Consequently, optimisation of neighbour list construction
algorithm becomes critical to the performance of the simulation code.
A brute force implementation of the neighbour list construction involves checking the distance of each node from all the other nodes in the system in order
to determine its neighbours. This approach involves a double nested loop iteration over all the nodes in the system and scales as O(N 2 ), where, N is the
number of nodes in the network. The computation becomes very expensive as
we go to larger networks, higher node densities and to networks with highly
dynamic nodes (such as vehicular networks).
A similar issue is faced in simulations of interacting many-particle systems
such as molecular dynamics simulations, where, for each particle in the system, the interaction forces with the remaining particles need to be calculated
to simulate its dynamics. Analogous to the transmission/interference range
in WiFi networks, in MD simulations of liquids with short-range interaction
potentials, the interaction force computation is truncated at a cut-off radius,
rc . When rc is equal to or smaller than one-third of the linear dimension of the
simulation box, a cell-linked list method is often adopted which brings down
the force computations from an O(N 2 ) calculation to O(N) (in this case N is
the number of particles in the many-particle simulation system).
The cell-linked list approach [28] applied to the construction of network topology in WiFi systems works as follows. First the two-dimensional network
area/simulation cell is divided into sub-cells such that the linear dimension of
each sub-cell is equal to the transmission (interference) range. Any node in
8
Fig. 2. Figure illustrating the cell linked list method [28] applied to a simulation
of WiFi networks. The network area is sub-divided into sub-cells with a linear
dimension equal to the the transmission range, rt . For the central node in cell 5,
only the other occupants in its own sub-cell and those in the immediate neighbouring
sub cells (1,2,3,4,6,7,8,9) feature in its neighbour list.
a sub-cell can only interact with nodes in its own sub-cell and in the immediate neighbouring sub-cells and is invisible to nodes in all other sub-cells in
the network. This is illustrated in Figure 2, where, in order to determine the
neighbour list for the central node in cell 5, one only needs to consider nodes
within cell 5 and in the immediate neighbouring cells (1,2,3,4,6,7,8,9). The
brute force method of constructing neighbour lists with two nested loops over
all particles in the system is replaced by 1) a loop over all nodes to determine
which sub-cells they lie in - an O(N) operation and 2) a loop over all nodes
to find their neighbour nodes in the immediate neighbouring sub-cells - an
O(N × Nc ) algorithm, where Nc is the average number of nodes per sub-cell.
We note that the above decomposition of the simulation system into linked
cells is naturally suited to domain decomposition parallelisation of the simulations, which we shall discuss in the next section.
3.2 Parallel and distributed simulations on grid platforms
Due to the short-range nature of wireless communications, parallel simulations
of WiFi-based systems on massively parallel computers or tightly coupled grid
platforms can be performed most effectively using domain decomposition. The
area in which WiFi devices operate is divided into a number of regular subdomains with dimensions larger than the maximum interference radius of the
wireless devices that comprise the network. The entire communication stack
of all devices within each subdomain is then allocated to one processor and
inter-processor communications are only performed when nodes move from one
processor to another, or there are radio signal propagation across sub-domain
9
boundaries.
Several parallel simulators tools for WiFi-based mobile wireless networks which
exploit the domain decomposition strategy have been proposed and implemented in recent years [23,24,25,30]. However, there is very limited published
work in examining the performance of these simulators in the context of largescale parallel simulations. Benchmark studies performed on relatively small
number of processors (6 − 12 PEs) however, indicate that only sub-linear
parallel speedups can be achieved, presumably due to a combination of communications and synchronisation overheads.
In addition to domain-decomposition, a task farming strategy can be exploited
in simulations of large-scale wireless networks in order to perform Monte Carlo
runs over an ensemble of network realisations and/or to explore the performance for a range of system parameters. In this case multiple runs of the same
code are spawned on a set of slave processors and the results are collected and
further processed by the master processor at the end of computation. Unlike
domain decomposition, task farming requires no synchronisation and very limited interprocess communications. Therefore linear parallel speedups can be
achieved even for simulations performed on loosely coupled grid platforms.
3.3 Synchronisation
A WiFi network consists of a number of devices each having its own internal
set of states (e.g. the number of data packets in the incoming queue of a
laptop, or the random backoff time of the MAC protocol). These states change
stochastically (and therefore asynchronously) in response to events which are
generated either internally or due to interactions with other devices in the
system. For example, the arrival of a voice call will change the state of the
outgoing data queue of a WiFi-enabled mobile phone.
Parallel simulations of interacting systems with such asynchronous dynamics
(also known as Parallel Discrete Event Simulations) requires the use of a synchronisation protocol among the processing elements (PES) in order to ensure
that causality errors are not introduced in the simulation results. An overview
of the synchronisation of parallel discrete event simulation and a comprehensive discussion of commonly used synchronisation schemes can be found in
[29]. Conservative synchronisation schemes are conventionally used in parallel
simulations of wireless networks [17,31]. Each PE defines its own lookahead as
the minimum duration (measured using the simulation clock) for which it will
not send any message on its outgoing links. Periodically, a global minimum
of all PE’s simulation time plus their lookahead values is computed. Each PE
can then process all events inside its own domain that take place within this
10
time window safely without the need of additional synchronisation.
When the number of processing elements becomes large, conservative synchronisation schemes may result in large fluctuations in the rates at which
different PES progress, hence greatly reducing the computational scalability
of the parallel discrete event simulation [32]. It has been demonstrated recently
[32] that by changing the communication topology of PES from a regular grid
to a small-world-type topology, the above problem can be eliminated and high
parallel efficiency achieved. We are currently in the process of implementing
such schemes in the parallel version of our wireless network simulator and the
results will be reported elsewhere.
4
Case Study: Simulations of Worm Attacks on Mobile Wireless
Adhoc Networks
Worms are self-replicating computer viruses which can propagate through
computer networks without any human intervention [26]. With wireless networks becoming increasingly popular, many security experts predict that these
networks will soon be a main target of attacks by worms and other type of
malware [33,34]. In addition to individual devices, open resource sites in wireless information or computational grids could well be the next wave of targets
for such wireless worm attacks. A qualitative understanding of such attacks is
of great importance, both in assessing their risk and for the design of effective
detection and prevention strategies.
Worm and virus attacks on the Internet have been the subject of extensive
empirical, theoretical and simulation studies, and there have been a number of
studies on securing conventional wired grids against such attacks [37]. Investigation of virus spreading in wireless networks and wireless grids in general
and worms in particular is, however, at its infancy. In a recent study [35] we
used Monte Carlo simulations to investigate the spreading of worm epidemics
in static wireless adhoc networks. These studies point out to important differences between the propagation patterns of worms in wired and wireless
networks and highligh the importance of incorporating interference effects,
network topology and medium access control for realistic modelling of data
communications in these networks.
In this section we briefly describe aspects of these simulations and further
advance them by incorporating the impact of device mobility in the dynamics
of worm propagation.
11
4.1 Worm propagation model
Following [35] we assume that wireless worms primarily utilise multihop forwarding for their propagation in adhoc networks, a mechanism which does not
require any Internet connectivity. With respect to an attacking worm we use
the so-called susceptible-infected-removed (SIR) model from mathematical biology [36], adapted to the context of wireless communications. We assume that
nodes in the network to be in one of the following three states: vulnerable,
infected, or immune. Infected nodes try to broadcast the worm to their neighbours at every possible opportunity. Vulnerable nodes can become infected
with probability λ when they receive a transmission containing a copy of the
worm from an infected neighbour. Finally, infected nodes get patched and become immune to the worm with probability δ. We denote by S(t), I(t) and
R(t) the population of vulnerable, infected and immune nodes, respectively.
4.2 Simulation details
We simulated the propagation of worms in mobile wireless adhoc networks
comprising N devices in a L2 = 1000 × 1000 m2 area. At the start of each simulation the devices were spread randomly in the simulation area, after which
they were allowed to move following simple random walks (Periodic boundary
conditions were used at the edges of the simulation area). The worm spreading dynamics was simulated on top of the resulting time-dependent network
using Monte Carlo simulations. Each Monte Carlo run starts by infecting a
single randomly chosen node and proceeds for a certain number of simulation
timesteps, iupdate , after which the positions of the nodes are updated according
to the random walk model. We use this form of updating in order to mimic
the difference in the timescales between the spreading process and the user
mobility. Each simulation continues in this fashion until the epidemic dies out
(i.e. no infected node is left in the network). We typically average our results
over 500 Monte Carlo runs in order to obtain statistically significant data.
Furthermore, the results were also averaged over simulations starting from
different initial infected seeds.
4.3 Results
First we consider the improvement in computational performance gained from
using the cell linked-list algorithm for updating the network topology. A comparison of the performance of the cell-linked list and brute force method for
different movement update frequencies, ν = 1/iupdate , is shown in Figure 3. It
12
5000
brute force
cell linked-list
CPU Time (seconds)
4000
3000
2000
1000
0
0
10
5
15
20
iupdate
Fig. 3. Scaling of the computational cost vs. inverse update frequency for the simulation of an adhoc network composed of dynamic nodes whose positions are updated
periodically with frequency ν. The CPU times are compared for two simulations in
which, respectively, the brute force method and the cell linked list method were
used.
can be seen that using the cell-linked list algorithm greatly reduces the computational cost associated with updating network topology and, as expected,
this reduction becomes more significant as the update frequency increases.
Next we consider the impact of node mobility on the dynamics of worm propagation in the network. As an example, the time evolution of the population
of infected nodes, I(t), is plotted in Figure 4. The results were obtained for
mobile wireless adhoc network composed of 4000 nodes and using λ = 0.3 and
δ = 0.1. The nodes’ positions were updated every 1, 2, 5, 10 and 20 time-steps
during the worm propagation, i.e ν1 = 1, 2, 5, 10, 20. It can be seen that node
mobility has a significant impact on the spreading dynamics. In particular, as
mobility increases (i.e. the network is updated more frequently) the epidemic
peak (the maximum number of infected devices) becomes more pronounced
and also occurs at earlier times. These results indicate that dynamic adhoc
networks are more vulnerable to worm attacks than static networks.
Qualitatively, we can explain the above behaviour in the following way. In
a fixed wireless adhoc network the maximum number of nodes to which an
infected device can spread the worm in the course of its infection is limited
by the total number of devices which are within its transmission range. On
average this is given by nA, where n is device density and A = πrt2 is the total
area covered by device’s transmission. Switching on mobility enables infected
devices to sweep on a larger area than A, hence increasing the maximum number of nodes that each device can infect before getting patched. Consequently,
both the speed and the magnitude of the epidemic are increased with increased
13
1/υ = 1
1/υ = 2
1/υ = 5
1/υ = 10
1/υ = 20
0.5
I(t)
0.4
0.3
0.2
0.1
0
0
20
40
60
80
100
t
Fig. 4. Population of infected nodes vs. simulation time in an adhoc network composed of dynamic nodes whose positions are updated periodically with frequency
ν. The evolution of the infected node population for different update frequencies is
plotted.
device mobility.
5
Conclusions
WiFi and related technologies not only allow users to access the Internet on
the move, they are also enabling mobile devices to connect directly to each
other and form adhoc networks for distributed voice, video and data communication. Such adhoc networks also form a flexible communication backbone for
wireless adhoc grids, an emerging paradigm in distributed and grid computing.
With the proliferation of WiFi-enabled mobile devices such grids will enable
innovative applications based on sharing and federating computing and information resources of billions of wireless devices such as sensors, smartphones,
PDAS and laptops.
In this paper we described aspects of computational modelling of such WiFibased wireless networks, and examined some of the important computational
challenges which are involved in high-fidelity simulations of these networks.
We argued that viewing these networks in terms of interacting many-particle
systems provides a useful framework for understanding and addressing some
of these challenges. We demonstrated this point by using the cell-linked list
technique, which is widely used in simulations of such systems, for scalable updating of network topology in large-scale simulations of mobile wireless adhoc
networks. Furthermore, we discussed parallel simulations of WiFi-based wireless networks and showed that conventional parallelisation strategies used in
14
parallel simulations of interacting many-particle systems, such as domain decomposition, are readily applicable to such simulations. However, these strategies need to be complimented with scalable inter-processor synchronisation
schemes in order to deal with the asynchronous nature of interactions in wireless networks.
High-fidelity simulation platforms for WiFi-based wireless networks capable
of effectively utilising the power of computational grid and massively parallel
computing are currently at their infancy. Such platforms, however, will be
necessary in planning and optimising the highly complex next generation WiFi
networks. They are also important in realistically analysing the potentials and
challenges of future adhoc grid platforms [8,9], such as security, scalability, and
intermittent network connectivity resulting from mobility.
Our experience shows that in designing such platforms one can greatly benefit
from computational and algorithmic techniques developed in other branches
of computational science. In addition to scalable network topology construction and parallelisation, which were discussed in the current paper, another
example that comes to mind is the use of fast multipole expansion techniques
for O(N) interference computation in wireless networks with long-range radio
signal propagation [38]. At the same time, high-fidelity simulations of WiFibased wireless networks presents an array of new computational challenges
which are the subject of our ongoing and future research.
Acknowledgements
M.N. acknowledges support from the Royal Society through an Industry Fellowship. He thanks the Centre for Computational Science at UCL for hospitality and his colleagues at BT’s Mobility Research Centre for stimulating
discussions.
References
[1] A. Goldsmith Wireless Communications, Cambridge University Press, 2005.
[2] Nomads at last: A special report on mobile telecoms, The Economist, 1-18, April
12th, 2008.
[3] M. S. Gast 802.11 Wireless Networks, Second Edition, O’Reily, 2005; W.
Stalling, Wireless Communications Networks, Prentice Hall, 2005.
[4] R. Hekmat, Adhoc networks: Fundamental properties and network topologies,
Springer, 2006.
15
[5] D. Grigoras and M. Riordan, Future Generation Computer Systems 23, 990996, 2007.
[6] Y.
Zhang,
J.
Luo,
H.
Hou
(eds.),
Wireless Mesh Networking: Architectures,Protocols and Standards, Auerbach
Publications, NY, 2007.
[7] M. Nekovee, Proc. Workshop on Ubiquitious Computing and e-Research,
Edinburgh, UK, May 2005.
[8] L. W. McKnight, J. Howison, S. Bradner, Wireless grids: Distributed resource
sharing by mobile, nomadic, and fixed devices, IEEE Internet Computing,
Special Issue on Wireless grids, 24-31, July/August 2004.
[9] R. Moreno-Vozmediano, A hybrid mechanism for resource/service discovery in
ad-hoc grids, Future Generation Computer Systems, 2008 (article in press).
[10] The Network Simulator – ns-2, www.isi.edu/nsnam/ns.
[11] G. Riley, Proc. of the 2003 ACM Winter Simulation Conference, 2003.
[12] R. Pastor-Satorras and A. Vespignani, Evolution and Structure of the Internet:
A Statistical Physics Approach, Cambridge University Press, 2006.
[13] U. O. Ibom, Proc. of the IEEE Information Networking, 1-5, 2008.
[14] J. J. Garcia, Future Generation Computer Systems, 81-93,1988.
[15] T. Rappaport, Wireless Communications, Principle and Parctice, Prentice-Hall,
2000.
[16] M. Takai, J. Martin, and R. Bagrodia, Proc. of ACM MOBIHOC’01, 87-94,
2001.
[17] Z. Ji, J. Zhou, M. Takai, J. Martin, and R. Bagrodia, Proc. of the 18th Workshop
on Parallel and Distributed Simulation (PASDs’04).
[18] M. Penrose, Random Geometric Graphs, Oxford University Press, Oxford, 2003.
[19] See, e.g., T. Camp, J. Boleng, and V. Davis, Wireless Communication and
Mobile Computing, Special Issue on Mobile Adhoc Networking: Research,
Trends and Applications 2, 483-502, 2002.
[20] M. Musolesi and C. Mascolo, Mobile Computing and Communications Review
1, 1-12,2006.
[21] P. Ball, Critical Mass: How One Things Leads to Anorther, Arrow Books, 2006.
[22] D. Helbig, Rev. Mod. Phys. 73, 1067-1141 (2001).
[23] X. Zeng, R. Bagrodia, M. Gerla, Proc. 12th Workshop on Parallel and
Distributed Simulations (PADS 98), 154-161, 1998.
[24] www.scalable-networks.com
16
[25] L. Bononi, M. Di Felice, G. D. Angelo, M. Bracuto, L. Donatiello, Computer
Networks 52, 155-179, 2008.
[26] P. Szor, The Art of Computer Virus Research and Defense, Symantec Press,
2006.
[27] M. Nekovee, Proc. IEEE Vehicular Technology Conferenc (VTC07-Spring),
2486-2490, 2007.
[28] M. P. Allen and D.J. Tildesley, Computer Simulations of Liquids, Clarendon,
Oxford, 1987.
[29] R. Fujimoto, Parallel
Interscience,NY, 1999.
and Distributed Simulation
Systems
,
Wiley-
[30] Z. Ji, J. Zhou, M. Takai, J. Martin, R. Bogardia, Proc. 18th IEEE Workshop
on Parallel and Distributed Simulations (PADS’04), 2004.
[31] L. Bononi, M. Di Felice, M. Bertini, E. Croci, Proc. of MSWIM’06,
Torremolinos, Spain, 28-35, 2004.
[32] G. Korniss, M. A. Novotny, H. Guclu, Z. Toroczkai and P. A. Rikvold, Science
299, 677 (2003).
[33] M. Hypponen, Scientific American, 70–77, November 2006.
[34] J. Kleinberg, Nature 449, 287-288, 2007.
[35] M. Nekovee, New J. Phys. 9, 189, 2007.
[36] N. T. J. Bailey, The Mathematical Theory of Epidemics, Hafner Publishing,
1957.
[37] K. Hwang, Y-k Kwok, S. Song, M. C. Y. Chen, Y. Chen, R. Zhou and X. Lou,
GridSec: Trusted grid computing with security binding and self-defence against
network worms and DDoS attacks, Proc. International Conference on Grid
Computing Security and Resource Management (GSRM’05), Atlanta, USA,
may 2005.
[38] L. F. Perrone and D. M. Nicol, Proc. IEEE MASCOTS 2000.
17
| 5cs.CE
|
Parsimony, Model Adequacy and Periodic Correlation
arXiv:1611.01535v1 [stat.AP] 4 Nov 2016
in Time Series Forecasting
A. Ian McLeod
Department of Statistical and Actuarial Sciences
The University of Western Ontario
London, Ontario N6A 5B7
Canada
McLeod, A.I., (1993), Parsimony, model adequacy and periodic correlation
in forecasting time series, International Statistical Review 61/3, 387-393.
Stable URL: http://www.jstor.org.proxy1.lib.uwo.ca/stable/1403750
DOI: 10.2307/1403750
2
Parsimony and Model Adequacy in Forecasting
Summary
The merits of the modelling philosophy of Box & Jenkins (1970) are
illustrated with a summary of our recent work on seasonal river flow forecasting. Specifically, this work demonstrates that the principle of parsimony, which has been questioned by several authors recently, is helpful
in selecting the best model for forecasting seasonal river flow. Our work
also demonstrates the importance of model adequacy. An adequate model
for seasonal river flow must incorporate seasonal periodic correlation. The
usual autoregressive-moving average (ARMA) and seasonal ARMA models are not adequate in this respect for seasonal river flow time series. A
new diagnostic check, for detecting periodic correlation in fitted ARMA
models is developed in this paper. This diagnostic check is recommended
for routine use when fitting seasonal ARMA models. It is shown that this
diagnostic check indicates that many seasonal economic time series also
exhibit periodic correlation. Since the standard forecasting methods are
inadequate on this account, it can be concluded that in many cases, the
forecasts produced are sub-optimal. Finally, a limitation of the arbitrary
combination of forecasts is also illustrated. Combining forecasts from an
adequate parsimonious model with an inadequate model did not improve
the forecasts whereas combining the two forecasts of two inadequate models did yield an improvement in forecasting performance. These findings
also support the model building philosophy of Box & Jenkins. The nonintuitive findings of Newbold & Granger (1974) and Winkler & Makridakis
(1983) that the apparent arbitrary combination of forecasts from similar
models will lead to forecasting performance is not supported by our case
study with river flow forecasting.
Keywords: Combined Forecasts; Diagnostic Check for Periodic Correlation;
Forecasting Seasonal Time Series; Model Adequacy; Parameter Parsimony.
A.I. McLeod
1
3
Introduction
The main purpose of this paper is to discuss some general statistical
principles which are elucidated by our recent work in river flow forecasting
(McLeod et al., 1987; Noakes et al., 1985; Thompstone et al., 1985a). Also
based on these case studies, a new diagnostic check for periodic correlation in the residuals of fitted ARMA models is developed. This diagnostic
check is suitable for routine use when fitting seasonal ARMA models.
Briefly our experience with river flow time series suggests that the best
forecasting results are obtained by following the general model building
philosophy implicit in Box & Jenkins (1970) with suitable modifications
and improvements. In general terms, this approach is iterative and advocates choosing the most parsimonious adequate statistical model. Two basic principles of special relevance are:
Principle 1: Model Adequacy.
The model is considered adequate if it incorporates all relevant information and if when calibrated to the data, no important significant departures from the statistical assumptions made can be found.
Principle 2: Model Parsimony.
The principle of parsimony means that the simplest possible model
should be chosen.
One can view the problem of statistical modelling as choosing an adequate statistical model which is the most parsimonious. In mathematical programming terminology we could say that the problem of statistical
modelling has an objective function which is to minimize the model complexity (Model Parsimony) subject to the constraint of Model Adequacy.
4
Parsimony and Model Adequacy in Forecasting
In §2 the results of a case study of forecasting monthly river flow time
series is summarized. Here the importance of incorporating periodic correlation in the forecasting model is demonstrated. For a seasonal time series
denoted by zr,p where r denotes the year and p denotes the seasonal period, the periodic correlation coefficient is defined by
ρm (ℓ) = √
γm (ℓ)
,
(γm (0)γm−ℓ (0))
where
γm (ℓ) = Cov(zr,m , zr,m−ℓ ).
The concept of periodically correlated processes was introduced by
Gladyshev (1961). The first application of periodic time series models
seems to have been by hydrologists Thomas & Fiering (1962). Since that
time there have been very extensive developments in the theory and applications of periodically correlated time series. For a review of the probabilistic literature on periodically correlated processes, see Yaglom (1986,
§26.5; 1987). Miamee (1990) and Sakai (1991) have derived new theoretical results and conditions on the spectral density function of periodically
correlated time series. On the statistical methodology side, contributions
to periodically correlated time series modelling have been made by Jones &
Brelsford (1967), Moss & Bryson (1974), Pagano (1978), Cleveland & Tiao
(1979), Troutman (1979), Tiao & Grupe (1980), Sakai (1982), Dunsmuir,
W. (1983), Thompstone et al. (1985b), Vecchia (1985a, 1985b), Vecchia et
al. (1985), Li & Hui (1988), Jiménez et al. (1989), Hurd & Gerr (1991),
Osborn, D.R. (1991) and Vecchia, A.V. & Ballerini, R. (1991). Periodic
time series models are often used for modelling seasonal time series – especially environmetric series. However, several other interesting applications
include multiple spectral estimation (Newton, 1982) and multichannel signal processing (Sakai, 1990).
In some situations, as in the case study in §3, a comprehensive mod-
elling approach which satisfies both adequacy and parsimony principles
A.I. McLeod
5
may not be practical either for reasons of expediency or because a suitable
model cannot be found with available methodology. In this case, we have
found combined forecasts to be useful. On the other hand, if a good model
can be found, our experience suggests that the forecast cannot be significantly improved by combining it with forecasts from models which are less
parsimonious or less adequate. This latter result is at variance with the results reported by Winkler & Makridakis (1983) and Newbold & Granger
(1974). Perhaps this is due to the fact that the river flow time series used
in our studies are generally longer and more homogeneous than the economic series used by the aforementioned authors. The skill of the modeller
in developing an adequate model could also be a factor.
In order to make the hydrological data sets used in the case studies
referred to in this paper readily accessible to other researchers, all data is
available in the StatLib archive riverflow.
6
2
Parsimony and Model Adequacy in Forecasting
Monthly River Flow Case Study
The data in this case study (Noakes et al., 1985) consisted of thirty
mean monthly river flows for periods of from 37 to 64 years. Various models and model selection procedures were used to calibrate a model to each
data set omitting the last three years of data. The one-step ahead forecasts were then compared for the last three years (36 values). The best
forecasts as judged by the root mean-square error and other criteria were
obtained with the family of periodic autoregressive models.
The periodic autoregression model equation may be written
φm (B)(Zr,m − µm ) = ar,m
(1)
where Zr,m denotes the logarithmic flow for the r th year and mth month,
µm denotes the corresponding monthly mean, ar,m , r = 1, 2, . . . , m =
1, 2, . . . , 12 are a sequence independent normal random numbers with mean
2
zero and variance, σm
, and
φm (B) = 1 − φm,1 B − . . . − φm,pm B pm
(2)
where B is the backshift operator on t, where t = 12(r − 1) + m. Several
model selection techniques were used to select pm (m = 1, 2, . . . , 12). It
was found that a periodic autoregression which was determined by choosing pm as small as possible to achieve an adequate fit gave the best forecasts. This was accomplished by initially determining pm based on a plot
of the periodic partial autocorrelation function and then checking the adequacy of the fitted model. Our approach is thus a natural extension of
that of Box & Jenkins (1970).
On the other hand, a subset periodic autoregression approach was
found to produce comparatively very poor forecasts. In this approach, for
each period all possible autoregressions with some parameters constrained
to zero and with pm = 12 were examined (212 possibilities) and the best
model was selected with the Akaike Information Criterion (Akaike, 1974)
A.I. McLeod
7
as well as the Bayes Information Criterion (Schwarz, 1978). It was also noticed that the resulting models were always less parsimonious than that
selected by the first approach.
The seasonal ARMA model developed by Box & Jenkins (1970, Ch. 9)
did not perform very well either. In this case, the diagnostic check, developed in the next §4, indicates that this is due to model inadequacy.
The periodic autoregression and seasonal ARMA represent quite dif-
ferent families of time series models. Not only do the models differ in the
correlation structure but the specification of seasonality is purely stochastic in the seasonal ARMA model and purely deterministic in the case of
the periodic autoregression. Moreover neither specification is likely to be
absolutely correct. Thus although the periodic autoregression model forecasted best and was considered to represent a more valid statistical model,
it might be thought from the experience reported by Newbold & Granger
(1974) and Winkler & Makridakis (1983) that combining the periodic autoregression and seasonal ARMA forecasts would be helpful. As shown in
McLeod et al. (1987) this is not the case. In particular with method 1 of
Winkler & Makridakis (1983, p. 152) the periodic autoregression forecast
had a smaller mean square error at least 17 times out of 30. Thus combined forecasts cannot be recommended in this situation.
8
3
Parsimony and Model Adequacy in Forecasting
Quarter-Monthly River Flow Case Study
The object of this study (Thompstone et al., 1985a) was to obtain one-
step-ahead forecasts of the quarter-monthly, i.e. almost weekly, inflows to
the Lac St. Jean reservoir system operated by Alcan Limited. Complete
time series on past quarter-month inflows, precipitation and snowmelt
in the river basin were available for 30 years. A Box-Jenkins multiple
transfer-function noise model with precipitation and snowmelt as inputs
was found to provide an adequate fit to the deseasonalized data in many
respects except that it did not account for the periodic correlation effect.
A periodic autoregression model was also fit but this model did not take
into account the covariates precipitation and snowmelt. It could be suggested that at this stage a periodic-transfer-function noise model should be
developed to take into account both factors. However such a model could
easily involve too many parameters and, in any case, it was not possible
to calibrate it with our existing computer software. Perhaps future work
will result in a suitable model. Finally, a third model which was a semitheoretical hydrological model which incorporates various hydrological and
meteorological information in a conceptual model of river flow. The conceptual modelling approach has been strongly advocated by certain hydrologists who feel that time series methods are too empirical.
All three models were calibrated on data for 27 years and then used
to produce one-step-ahead forecasts over the next three years (144 periods). The root mean square error for transfer-function noise, periodic autoregression and conceptual model for forecasting logarithmic flows were
respectively 0.2790, 0.3009 and 0.3894. When the forecasts were combined
by simple averaging the root mean square error dropped to 0.1355. More
sophisticated combination techniques were found to lead to even further
improvements.
It is interesting to note that the empirical time series approach outperformed the more theoretical conceptual approach which has been strongly
A.I. McLeod
advocated by some hydrologists. A similar phenomenon with macroeconomic time series forecasting as previously been found (Naylor et al.,
1972).
9
10
4
Parsimony and Model Adequacy in Forecasting
A New Diagnostic Check For Periodic Autocorrelation
The seasonal ARMA model of order (p, d, q)(ps, ds , qs )s may be written
Φ(B s )φ(B)∇ds s ∇d Zt = Θ(B s )θ(B)at ,
(3)
where Zt is the observation at time t and at is a sequence of independent
normal random variables with mean zero and variance σ 2 . For monthly
time series s = 12 and t = 12(r − 1) + m, where r and m represent the year
and month respectively. The polynomials Φ(B s ), φ(B), Θ(B s ) and θ(B) of
degrees ps , p, qs and q specify the autoregressive and moving average components of the model. The terms ∇s = 1 − B s and ∇ = 1 − B represent the
seasonal and non-seasonal differencing operators. Using standard model
selection techniques (Box & Jenkins, 1970; Hipel et al., 1977) it was found
that most monthly river flow time series could be tentatively modelled as
a seasonal ARMA model of order (p, 0, 1)(0, 1, 1)12, where p = 0, 1 or 2.
The diagnostic check described below can be used to check for model inadequacy due to periodic correlation in the residuals of such fitted models.
The residual periodic autocorrelation at lag k ≥ 1 may be written
P
âr,m âr,m−k
r̂m (k) = qP r
,
P 2
2
â
â
r r,m−k
r r,m
(4)
where âr,m denotes the seasonal ARMA model residual for period t =
12(r − 1) + m (r = 1, . . . , N ; m = 1, . . . , 12). If the seasonal ARMA
model is adequate then using the methodology in McLeod (1978) it can be
√
√
shown for any fixed M ≥ 1, Nr̂(m) = N (r̂m (1), . . . , r̂m (M )) is asymptotically normal with mean zero and covariance matrix (1M − Q/12)/N ,
where 1M is the identity matrix of order M and Q = XI −1 X T , where
√
X and I are given in eq. (44) of McLeod (1978). Moreover, N r(m) and
√ (m′ )
Nr
are asymptotically independent when m 6= m′ . Since the diag-
onal elements of Q are all less than one, it follows that to a good approximation, r̂m (1), m = 1, . . . , 12 are jointly normally distributed with mean
A.I. McLeod
11
vector zero, diagonal covariance matrix and Var(r̂m (1)) = N −1 . A diagnostic check for detecting periodic autocorrelation in seasonal ARMA model
residuals is given by
S=N
12
X
2
r̂m
(1)
(5)
m=1
which should be approximately χ2 -distributed on 12 df.
As a check on the asymptotic approximation involved, a brief simulation experiment was performed. A (1, 0, 0)(0, 0, 0)12 model with φ1 =
−0.9, −0.6, −0.3, 0.3, 0.6 and 0.9 was simulated. Table 1 summarizes the
results on S for one thousand simulations with N = 17. The empirical significance level of a nominal 5% test was estimated by counting the number
of times that S exceeded 21.0261. From Table 1, the approximation is seen
to be adequate for practical purposes. In further experiments with N = 34
and 68, the approximation was seen to improve although the empirical significance level was still slightly less than 0.05 in all cases. This suggests
that in general the significance will be slightly overestimated. For example,
if the observed value of S indicates significant periodic correlation at the
5% level, the true significance level will be slightly less than 5%.
[Table 1 here]
The data on the Saugeen River (1915–1976) is illustrative of the usefulness of this new diagnostic check. A (1, 0, 1) (0,1,1)12 model was fit to the
logarithmic flows and passed all diagnostic checks given in Box & Jenkins
(1970). However, it was found that S = 59.6 indicating very significant
residual periodic correlation. As indicated in the next section, it appears
that many seasonal economic time series also exhibit such periodic residual
correlations.
12
Parsimony and Model Adequacy in Forecasting
Table 1
Empirical mean, variance and significance level of S
with N = 17 in 1000 simulations using a nominal 5% test.
φ1
Mean
Variance
Significance
level
–0.9
11.9
19.3
0.032
–0.6
11.5
18.6
0.025
–0.3
11.3
19.0
0.023
0.0
10.9
16.6
0.016
0.3
11.1
19.7
0.027
0.6
11.5
18.6
0.026
0.9
11.7
19.7
0.030
A.I. McLeod
5
13
Application to Forecasting Economic Time Series
Many seasonal economic time series may exhibit periodic correlation
which most of the standard approaches do not take into account. The diagnostic check of §4 may be applied routinely when fitting seasonal ARMA
models. Table 2 shows the results of testing the seasonal ARMA models
fitted by Miller & Wichern (1977, p.432) to four Wisconsin series. It is
seen that in two out of the four series there is very significant periodic correlation. In these cases, models which take this correlation into account
may be expected to produce improved forecasts. A comprehensive new approach to the modelling and forecasting of such series is given by McLeod
(1992).
[Table 2 here]
Acknowledgements
This research was supported by NSERC.
14
Parsimony and Model Adequacy in Forecasting
Table 2
Diagnostic Test For Residual Periodic Correlation
For Four Wisconsin Series From Miller & Wichern
Category
S
d.f.
Significance
level
Food Products
25.36
12
0.013
Fabricated Metals
36.8
12
0.0002
Transportation Equipment
11.6
12
0.478
Trade
6.98
12
0.859
A.I. McLeod
15
References
Akaike, H. (1974). A new look at the statistical model identification.
IEEE Trans. Autom. Control 19, 716–723.
Box, G. E. P. & Jenkins, G. M. (1970). Time Series Analysis Forecasting
and Control. San Francisco: Holden-Day.
Cleveland, W.P. & Tiao, G.C. (1979). Modeling seasonal time series.
Économie Appliquée 32, 107–129.
Dunsmuir, W. (1984). Time series regression with periodically correlated
errors and missing data. In Time Series Analysis of Irregularly Observed Data, Ed. E. Parzen. Springer-Verlag: New York.
Hipel, K. W., McLeod, A. I. & Lennox, W. C. (1977). Advances in BoxJenkins modelling. Water Resources Res. 13, 567–586.
Hurd, H.L. and Gerr, N.L. (1991). Graphical methods for determining the
presence of periodic correlation. J. Time Ser. Anal. 12, 337–350.
Gladyšev, E.G. (1961). Periodically correlated random sequences. Soviet
Math. Dokl. 2, 385–388.
Jiménez, C., McLeod, A.I. & Hipel, K.W. (1989). Kalman filter estimation
for periodic autoregressive-moving average models. Stochastic Hydrology and Hydraulics 3, 227-240.
Jones, R.H. & Brelsford, W. (1967). Time series with periodic structure.
Biometrika 54, 403–408.
Li, W.K. & Hui, Y.V. (1988). An algorithm for the exact likelihood of periodic autoregressive moving average models. Commun. Statist. Simulation Comput. 14, 1483–1494.
16
Parsimony and Model Adequacy in Forecasting
Miamee, A.G. (1990). Periodically correlated processes and their stationary dilations. SIAM, J. Appl. Math. 50, 1194–1199.
McLeod, A. I. (1978). On the distribution of residual autocorrelations in
Box-Jenkins models. J. R. Statist. Soc. B 40, 296–302.
McLeod, A. I., Noakes, D. J., Hipel, K. W. & Thompstone, R. M. (1987).
Combining hydrological forecasts. J. Water Resourc. Planning & Manage. Div. Proc. ASCE 113, 29–41.
McLeod, A. I. (1992, to appear). An extension of Box-Jenkins seasonal
models.
Miller, R. B. & Wichern, D. W. (1977). Intermediate Business Statistics.
New York: Holt, Reinhart and Winston.
Moss, M.E. and Bryson, M.C. (1974). Autocorrelation structure of
monthly streamflows. Water Resources Res. 10, 733–744.
Naylor, T. II, Seaks, T.G. & Wichern, D.W. (1972). Box-Jenkins methods:
An alternative to econometric models. Int. Statist Rev. 40, 123–137.
Newbold, P. & Granger, C. W. J. (1974). Experience with forecasting univariate time series and the combination of forecasts. J. R. Statist. Soc.
A 137, 131–165.
Newton, H.J. (1982). Using periodic autoregressions for multiple spectral
estimation. Technometrics 24, 109–116.
Noakes, D. J., McLeod, A. I. & Hipel, K. W. (1985). Forecasting monthly
riverflow time series. Int. J. Forecast. 1, 179–190.
Osborn, D.R. (1991). The implications of periodically varying coefficients
for seasonal time series processes. J. Econometrics 48, 373–384.
Pagano, M. (1978). On periodic and multiple autoregressions. Ann.
Statist. 6, 1310–1317.
A.I. McLeod
17
Sakai, H. (1982). Circular lattice filtering using Pagano’s method. IEEE
Trans. Acoust. Speech Signal Process. 30, 279–287.
Sakai, H. (1990). Cirular lattice filtering for recursive least squares and
ARMA modeling. In Linear Circuits, Systems and Signal Processing,
Ed. N. Nagai, New York: Dekker.
Sakai, H. (1991). On the spectral density matrix of a periodic ARMA process. J. Time Ser. Anal. 12, 73–82.
Schwarz, G. (1978). Estimating the dimension of a model. Ann. Statist.
6, 461–464.
Thomas, H. A. & Fiering, M. B. (1962). Mathematical synthesis of stream
flow sequences for the analysis of river basins by simulation. In Design
of Water Resources, Ed. Maass, A., Hufschmidt, M. M., Dorfman, R.,
Thomas, H. A., Marglin, S. A. & Fair, G. M. Harvard University Press.
Tiao, G.C. & Grupe, M.R. (1980). Hidden periodic autoregressive-moving
average models in time series data. Biometrika 67, 365–73.
Thompstone, R. M., Hipel, K. W. & McLeod, A. I. (1985a). Forecasting
quarter-monthly riverflow. Water Resources Bull. 21, 731–741.
Thompstone, R. M., Hipel, K. W. & McLeod, A. I. (1985b). Simulation
of monthly hydrological time series. In Stochastic Hydrology. Ed. A .I.
McLeod, Dordrecht: Reidel.
Troutman, B.M. (1979). Some results in periodic autoregression. Biometrika 66, 219–228.
Vecchia, A.V. (1985a). Maximum likelihood estimation for periodic
autoregressive-moving average models. Technometrics 27, 375–384.
18
Parsimony and Model Adequacy in Forecasting
Vecchia, A.V. (1985b). Periodic autoregressive-moving average modeling
with applications to water resources. Water Resources Bull. 21, 721730.
Vecchia, A.V., Obeysekera, J.T., Salas, J. D. & Boes, D. C. (1985). Aggregation and estimation for low-order periodic ARMA models. Water
Resources Res. 19, 1297–1306.
Vecchia, A.V. & Ballerini, R. (1991). Testing for periodic autocorrelations
in seasonal time series data. Biometrika 78, 53–63.
Winkler, R. L. & Makridakis, S. (1983). The combination of forecasts.
J. R. Statist. Soc. A 146, 150–157.
Yaglom, A.M. (1986). Correlation Theory of Stationary and Related Random Functions I. Basic Results. New York: Springer-Verlag.
Yaglom, A.M. (1987). Correlation Theory of Stationary and Related Random Functions II. Supplementary Notes and References. New York:
Springer-Verlag.
A.I. McLeod
19
Résumé
Les mérites de la philosophie de la modélisation de Box et Jenkins
(1970) sont illustrées par un résumé de nos recherches récentes sur la
prévision de saisonniers débits en rivière. En particulier, nos résultats
démontrent que le principe de la parcimonie, que plusieurs auteurs ont
mis en question, est utile à la sélection du meilleur modèle pour prévoir
les saisonniers débits en rivière. Nos recherches démontrent l’importance
de la compétence d’un modèle. Un modèle adéquat de saisonniers débits
en rivière doit incorporer la corrélation périodique et saisonnière. Les
modèles autorégressifs à moyenne mobile (ARMA) habituels et les modèles
ARMA saisonniers ne sont pas adéquats à cet égard pour les séries de
saisonniers débits en rivière. Dans cet article, on développe une nouvelle
méthode pour déceler la corrélation périodique dans les modèles ARMA
ajustés. Cette méthode est à utiliser habituellement dans l’ajustement
des modèles ARMA saisonniers. Cette méthode indique que beaucoup de
séries chronologiques économiques font preuve de la corrélation périodique
aussi. Puisque les méthodes ordinaires de prévision ne sont pas adéquats,
on peut conclure que dans beaucoup de cas les prévisions produites sont
moins qu’optimales. En dernier lieu, une limitation à la combinaison arbitraire des prévisions est illustrée aussi. La combinaison des prévisions
d’un modèle parcimonieux et adéquat avec celles d’un modèle inadéquat
n’améliora pas les prévisions. Cependant, le fait de combiner les deux
prévisions de deux modèles inadéquats produisit une amélioration de la
performance de la prévision. Ces résultats appuient aussi la philosophie
de la modélisation de Box et Jenkins. Les résultats non-intuitifs de Newbold et Granger (1974) et de Winkler et Makridakis (1983) indiquent
que la combinaison apparente et arbitraire de prévisions de modèles semblables menèra à la performance des prévisions. Cette conclusion n’est
pas soutenue par notre étude de cas portant sur la prévision des débits en
rivière.
| 10math.ST
|
arXiv:1709.03138v1 [cs.CV] 10 Sep 2017
Fully Convolutional Neural Networks
for Dynamic Object Detection
in Grid Maps
Master’s Thesis of
Florian Pierre Joseph Piewak
Department of Computer Science
Institute for Anthropomatics
and
FZI Research Center for Information Technology
Reviewer:
Prof. Dr.–Ing. J. M. Zöllner
Second reviewer:
Prof. Dr.–Ing. R. Dillmann
Advisors:
Dipl.–Inform. Michael Weber
Dipl.–Ing. Timo Rehfeld
Research Period: December 13th , 2015 –
KIT – University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association
June 12th , 2016
www.kit.edu
Fully Convolutional Neural Networks
for Dynamic Object Detection
in Grid Maps
by
Florian Pierre Joseph Piewak
Master’s Thesis
June 2016
Master’s Thesis, FZI
Department of Computer Science, 2016
Reviewer: Prof. Dr.–Ing. J. M. Zöllner, Prof. Dr.–Ing. R. Dillmann
Department Technical Cognitive Systems
FZI Research Center for Information Technology
Affirmation
Ich versichere wahrheitsgemäß, die Arbeit selbstständig verfasst, alle benutzten Hilfsmittel vollständig und genau angegeben und alles kenntlich gemacht zu haben, was aus Arbeiten anderer
unverändert oder mit Abänderungen entnommen wurde sowie die Satzung des KIT zur Sicherung
guter wissenschaftlicher Praxis in der jeweils gültigen Fassung beachtet zu haben.
Sunnyvale,
Florian Pierre Joseph Piewak
June 2016
iv
Abstract
One of the most important parts of environment perception is the detection of obstacles in the
surrounding of the vehicle. To achieve that, several sensors like radars, LiDARs and cameras are
installed in autonomous vehicles. The produced sensor data is fused to a general representation
of the surrounding. In this thesis the dynamic occupancy grid map approach of Nuss et al. [37] is
used while three goals are achieved. First, the approach of Nuss et al. [36] to distinguish between
moving and non-moving obstacles is improved by using Fully Convolutional Neural Networks to
create a class prediction for each grid cell. For this purpose, the network is initialized with public
pre-trained network models and the training is executed with a semi-automatic generated dataset.
The pre-trained Convolutional Neural Networks are trained on colored images, which is another
kind of input as used in this thesis. For this reason, several parameters of the Fully Convolutional
Neural Networks like the network structure, different combination of inputs and the weights of labels are optimized. The second goal is to provide orientation information for each detected moving
obstacle. This could improve tracking algorithms, which are based on the dynamic occupancy grid
map. The orientation extraction based on the Convolutional Neural Network shows a better performance in comparison to an orientation extraction directly over the velocity information of the
dynamic occupancy grid map. A general problem of developing machine learning approaches like
Neural Networks is the number of labeled data, which can always be increased. For this reason,
the last goal is to evaluate a semi-supervised learning algorithm, to generate automatically more
labeled data. The result of this evaluation shows that the automated labeled data does not improve
the performance of the Convolutional Neural Network. All in all, the best results are combined
to compare the detection against the approach of Nuss et al. [36] and a relative improvement of
34.8% is reached.
v
Zusammenfassung
Ein elementarer Bestandteil der Umgebungswahrnehmung eines Fahrzeuges, ist die Erkennung
von Verkehrsteilnehmern. Dies wird mittels unterschiedlicher Sensoren, wie zum Beispiel Radar,
LiDAR oder Kamera realisiert, wobei die aufgenommen Daten zu einer allgemeinen Umgebungsrepräsentation fusioniert werden. In dieser Arbeit wird die Umgebungsrepräsentation von Nuss
et al. [37] genutzt, welche eine Implementierung einer dynamischen Belegungsrasterkarte vorstellen. Die vorliegende Arbeit besteht aus drei Teilen. Als Erstes wird der Ansatz von Nuss et
al. [36] zum Unterscheiden zwischen bewegten und nicht bewegten Objekten durch das Einsetzen
von Convolutional Neural Networks (CNNs) erweitert. Dabei besteht die Ausgabe des CNNs aus
einer Klassifikation für jede Rasterzelle. Die Initialisierung der CNNs erfolgt durch publizierte
vortrainierte Netze, welche anschließend durch semi-automatisch erzeugte Daten trainiert werden.
Die publizierten vortrainierten Netze sind mit farbigen Eingabebildern trainiert wurden, was einen
anderen Eingabetyp als in dieser Arbeit darstellt. Aus diesem Grund werden diverse Parameter
des CNNs, wie zum Beispiel die Netzwerkstruktur, die Kombination der Eingabedaten oder die
Gewichtung der Trainingsklassen, optimiert. Als zweiter Teil wird die Orientierungsextraktion der
erkannten Objekte betrachtet, welche bestehende (auf der dynamischen Belegungsrasterkarte basierende) Tracking-Algorithmen verbessern kann. Das Ergebnis der Orientierungsextraktion zeigt
eine leichte Verbesserung basierend auf dem CNN im Vergleich zur Extraktion der Orientierung direkt aus der dynamischen Belegungsrasterkarte. Eine generelle Herausforderung beim Entwickeln
von Maschinellen Lernen Ansätze ist die limitierte Anzahl verfügbarer Trainingsdaten. Aus diesem Grund ist der dritte Teil der Arbeit die Evaluation eines semi-überwachten Lernalgorithmus,
welcher automatisch zusätzliche Daten klassifiziert und damit die Anzahl verfügbarer Trainingsdaten vergrößert. Das Ergebnis zeigt keine Verbesserung des CNNs bei Verwendung der automatisch
erzeugten Trainingsdaten. Zum Schluss werden die besten Ergebnisse der Optimierungen aus dem
ersten Teil vereint und mit dem Ansatz von Nuss et al. [36] verglichen, wobei eine relative Verbesserung von 34, 8% erreicht wird.
vi
Acknowledgment
First, I would like to thank all, who supported and motivated me during my master thesis.
I would also like to acknowledge my supervising professor Prof. Dr.–Ing. J. M. Zöllner, who
supervised me on the part of the Karlsruhe Institute of Technology (KIT) and proposed to write
my thesis at Mercedes-Benz Research and Development North America (MBRDNA). In several
deep discussions, he had passionate participation and motivating input to develop the topic of my
thesis.
Additionally, I would like to thank my on-site supervisor Dipl.–Ing. Timo Rehfeld, who supervised
me on the side of MBRDNA, and Dipl.–Inform. Michael Weber, who supervised me on the side
of FZI Research Center for Information Technology. They supported me every time, whenever I
ran into a problem or had a question about my research or writing. They consistently allowed this
thesis to be my own work, but steered me in the right direction whenever they thought I needed it.
I would also thank Prof. Dr.–Ing. R. Dillmann, who agreed to be my second supervisor at Karlsruhe Institute of Technology (KIT).
In addition I would like to thank MBRDNA and my team "Autonomous Driving - Environment
Perception", who gave me the opportunity to make research in California within an enjoyable
atmosphere and a great infrastructure to train and execute machine learning approaches.
Furthermore, I would like to acknowledge Christian Olms, who supported me to hand in my thesis,
while being located on another continent.
Finally, I must express my very profound gratitude to my parents for providing me with unfailing
support and continuous encouragement throughout my years of study and through the process of
researching. This accomplishment would not have been possible without them.
Thank you.
Sunnyvale,
Florian Pierre Joseph Piewak
June 2016
vii
Contents
1 Introduction
1
2 Background
3
2.1 Trend of Autonomous Driving . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2.2 Representation of the Environment . . . . . . . . . . . . . . . . . . . . . . .
5
2.2.1 Static Occupancy Grid Map . . . . . . . . . . . . . . . . . . . . . . .
5
2.2.2 Dynamic Occupancy Grid Map . . . . . . . . . . . . . . . . . . . . .
6
2.3 Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.3.2 Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2.3.3 Fully Convolutional Network . . . . . . . . . . . . . . . . . . . . . . .
12
2.3.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
2.3.5 Integration of Convolutional Neural Networks . . . . . . . . . . . . .
14
2.4 Semi Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.5 Orientation Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.6 Receiver Operator Characteristic . . . . . . . . . . . . . . . . . . . . . . . .
16
2.7 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
3 Methodology
20
3.1 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
3.1.1 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3.1.2 Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.1.3 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.1.4 Zooming into the Input . . . . . . . . . . . . . . . . . . . . . . . . . .
25
3.1.5 Learning Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
3.1.6 Weight Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
3.2 Orientation Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
3.2.1 Orientation Extraction with Convolutional Neural Networks . . . . .
27
3.2.2 Orientation Extraction over Velocities . . . . . . . . . . . . . . . . . .
29
3.3 Semi-Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4 Realization
31
4.1 Data Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
4.1.1 Recording Route . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
viii
4.1.2 Semi-Automatic Data Labeling . . . . . . . . . . . . . . . . . . . . .
32
4.1.3 Splitting the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
4.1.4 Rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
4.2 Comparison to other Approaches . . . . . . . . . . . . . . . . . . . . . . . .
36
4.3 Framework and Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
5 Evaluation
38
5.1 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
5.1.1 Training Configuration . . . . . . . . . . . . . . . . . . . . . . . . . .
38
5.1.2 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.1.3 Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.1.4 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
5.1.5 Zooming into the input . . . . . . . . . . . . . . . . . . . . . . . . . .
51
5.1.6 Learning Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
5.1.7 Weight Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
5.2 Orientation Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
5.3 Semi-Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
5.4 Comparison to other Approaches . . . . . . . . . . . . . . . . . . . . . . . .
65
6 Conclusion
67
7 Outlook
69
A List of Figures
71
B List of Tables
74
C Bibliography
75
1. Introduction
Every day people are injured by car accidents. Some of these injuries result in death so that 1.25
million road traffic fatalities happened globally in 2013 [55, p. 2], what can be seen in Figure 1.1.
There the increase of traffic deaths has to be observed related to the increase of 16% of registered
vehicles worldwide between 2010 and 2013. As a result the number of relative traffic fatalities
decreased, but is still high and has to be reduced absolutely with new approaches for vehicle safety.
The Swedish Parliament for instance introduced “Vision Zero” in 1997 [48], with the suggestion
that “no one will be killed or seriously injured within the road transport system” [48]. To fulfill this
vision, it is important to conduct research in the field of driver assistance systems and autonomous
driving. Especially because most of the accidents are caused by human errors [53], vehicles should
be able to recognize and react on dangerous situations or drive autonomously.
All the research in the field of driver assistance systems and autonomous driving shows that environment perception is an elementary part for these autonomous or driver assistance systems. Not
only to localize the vehicle in a global map, also to detect and react on dynamic obstacles in the
surrounding of the vehicle such as pedestrians or cars, a stable environment perception should be
achieved. To recognize them, a variety of sensors such as cameras, radars and LiDARs are used.
The information of the different sensors is fused together to obtain a representation of the entire
surrounding. This representation is commonly a grid mapping approach [16, 46]. Based on this
grid map moving objects and non-moving objects, which are called dynamic and static, can be
distinguished with different approaches [36, 50, 51]. This information is used to track the object
over time [56] and calculate reactions.
Road traffic deaths (millions)
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0
2001
2004
2007
2010
2013
Year
Figure 1.1.: Number of worldwide road traffic fatalities [55, p. 2].
1
1. Introduction
Tracking in general works well, but has its limits, especially if the input data can be improved.
For this purpose, Convolutional Neural Networks (CNNs) are used in this thesis to detect moving
obstacles within the grid map. The advantage of CNN approaches is that with sufficient data
the solution of a specified problem would be calculated implicit. The general obstruction is to
generate and label the data. Especially for deep learning approaches like CNNs large amounts of
training samples are required to generate good results. Typically the data is generated by human
annotators, which is a time and cost expensive task. Furthermore, the annotators have to be skilled
to recognize and label the data in a correct way. To reduce this expensive task semi supervised
learning algorithms were introduced in 1965 [40] and became more important since the 1990s [32,
33, 35]. By using this approach, the machine learning algorithms could handle expensive labeled
data as well as cheap unlabeled data. Especially in the current decade, unlabeled data became easy
to generate in a large scale regardless of which area of life. So it is not a problem to get pictures
from Google, videos from YouTube or other unlabeled data for machine learning approaches.
Several goals should be achieved in this thesis. First of all the CNN should improve current
implementations of moving object detections. For this purpose, pre-trained CNNs are used and
optimized. On the one hand the accuracy has to be optimized and on the other hand the processing
time has to be observed. The training data for the fine tuning of the CNN is generated in a semiautomatic manner. A second goal is the evaluation if orientation extraction of moving objects with
the CNN can improve the current orientation extraction. The last goal is to automatically generate
new training data from unlabeled data and to evaluate the improvement of the CNN by restarting
trainings with a larger dataset.
The structure of this thesis correspond to these different objectives. After a background section
(Chapter 2), the methodology of each of these three goals is presented (Chapter 3). Afterwards,
the realization with the semi-automatic data generation is described (Chapter 4). Then the results
of the three goals are presented (Chapter 5). Finally the thesis is summarized (Chapter 6) and an
outlook is given (Chapter 7).
2
2. Background
In this chapter, the background of the thesis and the whole system is described. First, a brief
overview over the development of autonomous driving is given. Then a description of the environment perception with grid maps is provided. Afterwards the CNNs and the semi supervised
learning are introduced and finally an angle representation for the orientation extraction and receiver operator characteristic are discussed.
2.1. Trend of Autonomous Driving
Already in the 1990s the European PROMETHEUS project was one of the first autonomous driving projects with an autonomous drive from Munich in Germany to Odense in Denmark [15, 20,
14]. In this project several vehicles like a Mercedes 500 SEL and a 10-ton bus (see Figure 2.1) of
Daimler-Benz AG were equipped with different sensors. The results were autonomous driving systems for different situations like lane following or lane changing on the freeway entrance without
traffic. Other results were detection and tracking of up to five Obstacles with CCD-cameras.
After that, several projects and challenges were created to promote research in this field. For
example the Defense Advanced Research Projects Agency (DARPA) started a challenge in off-
Figure 2.1.: Testing vehicle within PROMETHEUS
project 1990 [15, p. 1275].
Figure 2.2.: Laser grid map of STANLEY, the winner of DARPA Challenge 2005 [47,
p. 674].
3
2. Background
Figure 2.3.: Winner of the DARPA Urban Challenge 2007 with several sensors on top
of the roof [49, p. 19].
Figure 2.4.: Path of Bertha, an autonomous SClass with close-to-market sensors in
2013 [59, p. 2].
road scenarios in 2004 [47]. The goal of the challenge was to develop an autonomous vehicle
capable of traversing off-road terrain. The autonomous vehicles had to navigate 142 miles through
the Mojave Desert in no more than ten hours. After repetition of the challenge in 2005, five of
23 teams finished the course. The winner used different sensors like radars, lasers, cameras and
inertial measurement units (IMU). These measurements were transferred to a grid map for each
sensor type (see Figure 2.2), which were combined to a fused grid map. Based on this map the
drivable path was generated.
To extend the scenario, the DARPA Urban Challenge was launched in 2007 [49]. There, autonomous vehicles had to drive a 30 miles race through urban traffic in and around the former
George Air Force Base at speeds up to 30 miles per hour (mph). This challenge was especially
created to build autonomous systems that could react to dynamic environments. The tasks are
among other things, obeying traffic rules while driving along roads, merging with moving traffic, parking and independently handling unusual situations. Six of eleven teams reached the end
of the course in 2007 and showed the state-of-the-art of autonomous driving development (see
Figure 2.3).
Also beside such challenges, car manufacturers like Daimler presented their expertise: Bertha, an
autonomous S-Class, was released in 2013 and drove the Bertha Benz Memorial route [59] through
urban streets and rural roads, including traffic lights, roundabouts and pedestrian crossings (see
Figure 2.4). The specialty of this vehicle was the sensor configuration. In contrast to the DARPA
challenges, where each car had specific sensor racks on top of the roof, Bertha was only equipped
with close-to-market sensors like stereo cameras, mono cameras and radars (short and long-range).
Finally, Bertha drove 64 miles without an intervention of a human.
4
2. Background
All these autonomous vehicles were designed for specific situations and scenarios. To integrate
these approaches in mass-production vehicles, the algorithms have to be generalized, which is the
current research of several companies within the automotive or software industry. They are testing
these algorithms for example on roads in California, what is regularly reported by the California
Department of Motor Vehicles (DMV) [44].
2.2. Representation of the Environment
Recognizing the environment is an essential part of autonomous driving. Based on this information object detection, tracking, and path planning algorithms can be applied. A well-established
environment perception approach is grid mapping, which was first introduced in the field of
robotics [16, 46]. Grid maps divide the environment into grid cells and estimate their properties like the occupancy state. The first approaches generated grid maps only for stationary environments and are called static occupancy grid map (SOG). Later approaches detected or tracked
moving grid cells and inserted this information as features into the grid map [11, 37]. These approaches are called dynamic occupancy grid map (DOG). In the next subsections, the SOG and
DOG formulation by Nuss et al. [37] is introduced briefly, as it is the foundation of the work
presented in this thesis.
2.2.1. Static Occupancy Grid Map
The goal of the static occupancy grid map (SOG) is to find occupied and free areas in the environment. This can be provided with a probability P(Ock ) for the occupancy states ock ∈ {O, F} at time
k for each grid cell c in a static environment. The probability can be updated with independent
measurements z by means of a Bayesian filter [46, p. 23]. With an initial estimation probability
P(Ock ) = 0.5 the update equation becomes:
P(Ock+1 ) =
P(Ock )P(Ock+1 |zk+1 )
c
c
c |z
P(Ok )P(Ok+1 |zk+1 ) + P(Fkc )P(Fk+1
k+1 )
.
[2.1]
Here P(Fkc ) represents the probability for the free space, which is the counter hypothesis:
P(Fkc ) = 1 − P(Ock ) .
[2.2]
c |z
P(Ock+1 |zk+1 ) and P(Fk+1
k+1 ) describe the inverse sensor model, which is the likelihood of
the state, given the current measurement zk+1 . For further details, the reader is referred to [46]
and [37].
The problem of the SOG is the movement of objects within the grid map. Especially to detect and
react on these objects in traffic scenarios, they have to be recognized.
5
2. Background
0%
100%
occupancy probability
Figure 2.5.: Example of a dynamic occupancy grid map (DOG): the orientation of the velocity is encoded
in a color (see the circle) and the occupancy probability is encoded with gray scale values.
2.2.2. Dynamic Occupancy Grid Map
In this thesis, the dynamic occupancy grid map approach of Nuss et al. [37] is used. There, a
particle filter predicts dynamic and occupied states for each grid cell, where the cell state Xkc at
time k for a grid cell c is given as
Xkc = (ock , xkc ) .
[2.3]
The term ock is the occupancy state as in the SOG and xkc is the dynamic state, which is defined as
a two-dimensional position px , py and a two dimensional velocity vx , vy , i.e.
x = (px , py , vx , vy )T .
[2.4]
The cell state is represented by particles of the particle filter. Each particle i in a cell c at time
k contain velocities, which corresponds to the dynamic state, and has a weight wc,i
k . The sum of
these weights over a cell correspond to the occupancy probability of the cell. An example of a
DOG can be recognized in Figure 2.5.
The processing of the particle filter is divided into the prediction, update and resampling step,
which are described in the following subsections. The prediction and the update step can also be
recognized in Figure 2.6. For further details, the reader is referred to [37] and [36].
Prediction
Each particle is propagated according to a constant velocity and constant orientation model including a discrete, unbiased, white, Gaussian noise v. This noise produces diversity over the particles,
even if there are some similar particles resampled in further steps. The model transition is defined
as:
6
2. Background
1
0
xk+1 = f (xk , vk ) =
0
0
0 T
1
0
0
1
0
0
0
T
xk + vk .
0
1
[2.5]
T represents the time difference between the time step k and k + 1. With this propagation a set
of particles can move from cell a to cell c, what is denoted as Γc←a
k+1|k . The set of weights of these
c←a
particles is stated as w Γc←a
k+1|k and a probability for the transition event Ek+1|k that the occupancy of
cell a moves to cell c can be defined by
c←a
P(Ek+1|k
)=
∑
w c←a
wa,i
k+1|k ∈ Γk+1|k
wa,i
k+1|k .
[2.6]
Based on this definition the a priori occupancy probability can be specified with the assumption
that all transitions are statistically independent by
P(Ock+1|k ) = 1 −
∏
1≤a≤C
c←a
1 − P(Ek+1|k
) ,
[2.7]
where C defines the total number of grid cells.
Figure 2.6.: Different steps of the particle filter within the DOG: (a) describes the first state which is used
to create the prediction (b). After a measurement (c), the actual state is updated (d). The color
of the grid cell describes the occupancy probability (the darker, the higher the probability is).
The arrows describe the dynamic state of each particle. [36]
7
2. Background
Figure 2.7.: Inverse sensor model for a laser measurement as a measurement grid [37, p. 1078].
Update
To update the prediction of each particle, measurements of several sensors are combined. The
resulting grid map is called measurement grid. To generate a measurement grid from several
sensors, their individual inverse sensor model is used. An example of an inverse sensor model
can be seen in Figure 2.7, where a laser measurement with multiple beams is shown. A high
occupancy probability is created for grid cells in the neighborhood of sensor detections. Low
occupancy probabilities are created for grid cells in between the sensor detection and the vehicle.
All other grid cells are set to an occupancy probability of 0.5 to represent uncertainty about the
occupancy state. Similar to the SOG approach, the measurement grid and the predicted occupancy
states are then combined with a binary Bayes filter to obtain the posteriori occupancy probability:
P(Ock+1|k+1 )
=
P(Ock+1|k )P(Ock+1 |zk+1 )
c
c |z
P(Ock+1|k )P(Ock+1 |zk+1 ) + P(Fk+1|k
)P(Fk+1
k+1 )
.
[2.8]
The dynamic state of a particle is updated by a measurement through a weight update of each
ec,i
particle. The updated unnormalized weight w
k+1|k+1 is
c,i
c,i
c,i
c
ec,i
w
k+1|k+1 = βk+1 wk+1|k+1 = wk+1|k p(zk+1 |xk+1|k ) .
[2.9]
ec,i
Unnormalized means that the sum of each particle weight w
k+1|k+1 within a grid cell does not represent the occupancy probability. To get a normalized updated weight wc,i
k+1|k+1 , the normalization
c
factor βk+1
can be calculated with
c
βk+1
=
1
ec,i
w
c
k+1|k+1
P(Ok+1|k+1 ) c,i
ek+1|k+1 ∈we Γc←
w
k+1|k
∑
,
[2.10]
w c←
where we Γc←
k+1|k describes the set of unnormalized weights of the particles in Γk+1|k .
Resampling
Resampling is a main part of classical particle filters. It is used to eliminate particles and reproduce
others based on their weights. This includes the reproduction of one particle multiple times. As a
8
2. Background
Figure 2.9.: One of the first technical realization of visual recognition of mammal
brains [21, p. 198].
Figure 2.8.: Image processing and reaction in a
monkey brain [45].
result, degeneracy is avoided and convergence to a stable velocity is evoked. At the same time new
particles are inserted. The weight of new spawn particles is dependent of the ratio of measured
and predicted occupancy probability and the weights of survived particles. This can be defined as
follow
c
P(Ok+1|k+1 )
∑ wnew
=γ
.
P(Ock+1|k )
∑ wsurvived
[2.11]
Especially if a high occupancy is measured but no particles were predicted into the corresponding
cell, new particles are produced.
2.3. Convolutional Neural Networks
In this section, the Convolutional Neural Networks (CNNs) are described in context of this thesis.
After a brief motivation different layer types are discussed. Afterwards, the development of Fully
Convolutional Neural Networks (FCNs) followed by basic training algorithms is provided. Then
possible integrations into the current environment perception system are introduced. For further
details, the reader is referred to [23].
2.3.1. Motivation
Since the invention of computers there is an effort to create intelligent machines that are able to
solve problems adaptively without human interaction. The inspiration for the realization of these
ideas mostly came from nature. CNNs are motivated by the visual recognition part of the brain
of mammals. As shown in Figure 2.8, the brain is structured in different sections to recognize
objects and create a body reaction. Each section has a specific task to solve. Especially in the
first levels of recognition (Retina - LGN - V1 - V2 - V4 - PIT - AIT) different levels of object
detection can be found. The first layers are used to detect edges, corners and simple shapes. Based
on these detections more complex shapes can be recognized. Only at the last layers (AIT) real
9
2. Background
C1: feature maps
6@28x28
INPUT
32x32
Convolutions
C3: f. maps 16@10x10
S4: f. maps 16@5x5
S2: f. maps
6@14x14
Subsampling
C5: layer F6: layer
120
84
Convolutions
OUTPUT
10
Gaussian connections
Full connection
Subsampling
Full connection
Figure 2.10.: One of the first Convolutional Neural Networks trained in a supervised manner to recognize
characters [27, p. 7].
object representations are produced. Afterwards, this representation is used to generate decisions
and reactions executed by actors like fingers.
This layered detection and recognition structure is used for several years in the computer vision
field. Two-dimensional filters are used to detect simple structures like corners and edges. Based
on this information a higher level of object detection is calculated [43]. The problem of these
approaches is that the filter types have to be chosen and optimized by hand. In contrast, CNNs
allow to learn and optimize the filters automatically. Based on first technical realization of the
brain visual recognition from Fukushima [21] (see Figure 2.9), LeCun et al. [27] created one of
the first CNNs that was trained in a supervised manner. Their CNN was trained to recognize
handwritten characters, which can be seen in Figure 2.10. Based on this paper several deeper
CNN structures were created to solve classification tasks based on two-dimensional input images
like AlexNet [26], VGG-net [41] or GoogLeNet [42].
2.3.2. Layer
The CNNs consist of three-dimensional feature maps which can be connected with different layer
types. The most important layers (convolutional layer, pooling layer and deconvolutional layer)
are discussed in the following subsections.
Convolutional layer
The convolutional layer includes the filters f (i, j), which are also used in several non-machine
learning computer vision techniques. There a filter is a combination of weights, what represent a
specific feature extractor. The filters of size n1 × n2 , which are usually squared, are used to con-
volve the input feature map g(i, j). The input feature map at the first layer can be a two-dimensional
grayscale image or an RGB-image, where the color channel represents the third dimension. The
convolution equation is given by
h(i, j) = f ∗ g =
Z Z
k l
f (i − k, j − l)g(k, l) .
10
[2.12]
2. Background
Input feature map
m1 × m2 × r
Output feature map
mnew,1 × mnew,2 × k
Filter n1 × n2 × r
1
5
3
8
1
2
5
3
5
8
7
2
4
5
9
7
9
8
7
5
Figure 2.12.: Example of a max pooling layer with
stride two.
Figure 2.11.: Principle of a convolutional layer.
This equation becomes a sum by the discretization of the feature map in pixels:
h(i, j) = f ∗ g = ∑ ∑ f (i − k, j − l)g(k, l) .
k
[2.13]
l
This convolution creates a two-dimensional output, what can be seen in Figure 2.11. By using k
filters per convolutional layer the resulting feature map becomes three-dimensional with the size
mnew × mnew × k. The edge of the resulting feature map mnew can be calculated via
mnew =
m − n + 2p
+1 ,
s
[2.14]
where s describes the stride, i.e. the step size to use for the iteration over the feature map. A
larger stride produces a smaller output map. p describes the padding, which can be used to resize
the feature map to a specific size. The most common padding approach is the zero padding. The
padding is usually used to rebuild the same output size as the input size.
Pooling Layer
With the filters within the convolutional layers several features are extracted. But these features
are related to a specific size of the input feature map. To obtain features of different map sizes,
the feature map can be reduced by using pooling layers. They fuse squares of pixels with different
possible approaches like mean or maximum, what can be seen in Figure 2.12.
Deconvolutional Layer
With the discussed layers the size of the feature maps becomes smaller in deeper layers of the
network. Consequently, this process has to be reversed to obtain a pixelwise prediction as output.
For this purpose Long et al. [28] used the deconvolutional layer, which was already introduced by
Zeiler et al. [57]. This layer uses bilinear upscaling to initialize the deconvolutional weights. They
11
2. Background
Q00
Q10
forward/inference
dy
backward/learning
P
dx
.
ion
g.t
ict
ion
red
at
p
t
n
ise
me
elw
seg
pix
ex
4
6
25
38
96 096 21
4
4
6 40
38 25
96
ey
Q01
21
Q11
Figure 2.14.: One of the first Fully Convolutional
Neural Networks to create a pixelwise
predictions [28, p. 3431].
Figure 2.13.: Bilinear upscaling, the red points are
the original points and the green point
is an example for an interpolated target pixel.
produce a larger output map, what you can see in Figure 2.13. An upscaled pixel value can be
calculated based on the four surrounding pixel values with
ex
dx
Q00 +
Q10
dx + ex
dx + ex
dx
ex
Q01 +
Q11
R1 =
dx + ex
dx + ex
ey
dy
P=
R0 +
R1
dy + ey
dy + ey
R0 =
[2.15]
[2.16]
[2.17]
2.3.3. Fully Convolutional Network
All the networks described in Section 2.3.1 classify the entire image, because of the whole structure with fully connected layers in the end. For semantic segmentation a sliding window approach
could be used to create a pixelwise prediction. This is in general not suitable caused by the computational expense what arises by executing the CNN once for each pixel. For this purpose, the basic
idea of Fully Convolutional Nerual Networks (FCNs) were introduced by Matan et al. [30] for the
one dimensional output case and by Wolf and Platt [54] for the two dimensional output case.
The basic idea of a fully convolutional network structure is the size independence of the input
image. This can only be proceed, if the fully connected layers of conventional CNNs are converted
to convolutional layers with a specific filter size. The fully connected layers can be interpreted as
convolutional layer with a filter size of the last feature map. For this reason, the two types of layer
12
2. Background
can easily be transformed by just copying the weights. The final output of a FCN is a feature map,
which can produce different output feature map sizes in dependence of the input size.
Caused by the convolutions of different layers, a pixelwise prediction can not be established in
the first versions of FCNs. For this reason, Long et al. [28] introduced later a fully convolutional
network structure with deconvolutional layers, which was trained end-to-end for semantic segmentation. This combination of the layers facilitates a two-dimensional output of the neural network
for each class, where the output has the same size as the input (see Figure 2.14). As a result, a
pixelwise classification is produced by executing the CNN only once.
2.3.4. Training
The training of CNNs is responsible for tuning and optimizing the filters of convolutional and
deconvolutional layers, where each filter consists of adaptable weights. This is done by using a
gradient descent algorithm. The applied algorithm is called backpropagation. There, an overall
error is calculated, which is called loss. This loss is used to adapt the weights for example of
the convolutional and deconvolutional layers by propagating back the error through the different
layers. As an example the multinomial logistic loss is described.
The multinomial logistic loss is used to tune and optimize the parameters of the CNN with a multi
class output yi ∈ {1, . . . , K} for K classes. For this purpose a training set { x(1) , y(1) ), . . . , (x(m) , y(m) }
with m samples is given, which contains the input x and the desired output y. The hypothesis hθ
describes the class probability and is defined with
P (y = 1|x; θ )
exp θ (1)> x
P (y = 2|x; θ )
exp
1
hθ (x) =
.
= K
.
.
∑ exp θ ( j)> x
j=1
P (y = K|x; θ )
exp
,
(K)>
θ
x
θ (2)> x
..
.
[2.18]
where θ represents the parameters of the model, what is in this thesis the weights of the CNN. As
a result the cost function J(θ ) is defined as
J(θ ) = −
"
m
K
∑ ∑ 1{y
i=1 k=1
(i)
#
(i)
(i)
= k} log P y = k|x ; θ
,
[2.19]
where 1{.} is the “indicator function”, so that
1{a true statement} = 1 and
[2.20]
1{a f alse statement} = 0 .
[2.21]
13
2. Background
The last part of the equation P y(i) = k|x(i) ; θ represents the probability of a specific class, which
can be extracted from Equation 2.18:
P y(i) = k|x(i) ; θ =
exp θ (k)> x(i)
K
∑ exp
j=1
θ ( j)> x(i)
[2.22]
The cost function has to be minimized to generate an optimal output. For this reason the multinomial loss is propagated back through all layers, to adapt the CNN as a model. For further details,
the reader is referred to [34].
In general a training of CNNs need a large amount of labeled data, especially if the network
is trained from scratch. For this reason pre-trained networks are used to initialize the CNNs.
Pre-trained networks are CNNs, which are trained and optimized on public datasets like Pascal
VOC [18] or Imagenet [13]. These public datasets contains different images for classification or
segmentation approaches. After initializing the weights of the CNN by using pre-trained CNNs,
the fine tuning can be executed, to adapt the CNN to the own dataset.
2.3.5. Integration of Convolutional Neural Networks
One of the most important parts of environment perception is detection of other obstacles in the
surrounding of the vehicle. To achieve that, several sensors like radars, LiDARs and cameras are
installed in autonomous vehicles. The produced sensor data is fused to a general representation
of the surrounding of the vehicle as described in Section 2.2.2. Based on the DOG, clusters of
occupied grid cells are extracted and marked as dynamic or static [39]. These clusters are then
used to instantiate or update object-level tracks which incorporate higher level object knowledge
such as vehicle dynamics [56], that are not taken into account in the DOG.
To distinguish between dynamic and static clusters only the velocities of particles per grid cell
are taken into account [36]. This can cause some problem with noise or incorrect measurements.
To improve the whole toolchain and generate more accurate clusters this thesis proposes a deep
learning approach with convolutional neural networks (CNNs) to create a pixelwise classification
of dynamic and static parts of the grid map. Possible integrations in the toolchain can be seen in
Figure 2.15 and Figure 2.16. In Figure 2.15 the classification of the CNN is used in parallel to the
toolchain to improve the current clustering algorithm. With this approach the clustering algorithm
can generate clusters as before and uses the CNN simply as an improvement. Another approach
is to integrate the CNN completely into the toolchain like in Figure 2.16. As a result, the cluster
algorithm do not need the whole grid map as an input what can end in a reduction of processing
time.
14
2. Background
Dynamic
Occupied
Grid Map
Sensor Data
Sensor Data
Sensor Data
Clustering
Tracking
CNN Labeling
Grid cells
Figure 2.15.: Parallel integration of Convolutional Neural Networks (CNNs).
Sensor Data
Sensor Data
Sensor Data
Dynamic
Occupied
Grid Map
CNN Labeling
Grid Cells
Clustering
Tracking
Figure 2.16.: Full integration of Convolutional Neural Networks (CNNs).
2.4. Semi Supervised Learning
In contrast to static modeling approaches, which solves problems by using manual defined equations as a model, machine learning approaches can potentially determine better results by using
adaptive models, especially when the static model is not trivial to find. For this purpose machine
learning approaches need representative data to extract meaningful representations. In general two
possible learning approaches exist:
• supervised learning
• unsupervised learning
Supervised learning uses labeled data, i.e. input data and the corresponding desired output data.
Based on this, the machine learning approach regresses a representation to map the input to the
desired output. Unsupervised learning uses unlabeled data, i.e input data without any specific
target output. Most of the unsupervised learning approaches recognize clusters or detect structures
within the data. This type of learning is usually used in the field of data mining. These two
presented learning approaches are combined in different ways to more complex approaches like
active learning or semi supervised learning. For further details, the reader is referred to [4].
A major problem with supervised machine learning approaches is the generation of labeled data.
Especially if the problem is slightly different to problems in the literature, public datasets like
Pascal VOC [18], Imagenet [13] or Citycapes [10] cannot be used. An own dataset has to be
created, what ends in a manual annotation. For this annotation the annotators have to be skilled to
recognize and label the data in a correct way, what is a cost and time expensive task. To reduce
15
2. Background
this problem, semi supervised learning algorithms were introduced in 1965 [40] and became more
important since the 1990s [32, 33, 35]. There, a combination of labeled and unlabeled data is used
to train the model. The goal of these semi supervised algorithms is that only a few data is labeled
and together with a large number of unlabeled data the machine learning approach becomes more
accurate. This results in a less cost and time expensive task, especially because unlabeled data is
nowadays available in large amounts.
2.5. Orientation Extraction
Orientation angle extraction from images is usually applied in the field of head pose estimation [2,
5]. There, the orientation is often split in different bins and a classifier is trained for each separate
bin. After that a linear regression over the bins can be calculated. These approaches restrict
the capability of machine learning algorithms, especially if they are able to directly regress a
continuous orientation range. The problem that occurs is the periodicity of the orientation angle
φ . It is important that a smooth transition from 359◦ to 0◦ is encoded. This periodicity causes that
gradient descent approaches (like in the training step of CNNs) cannot be used. For this purpose,
Beyer et al. [3] introduced the Biternion representation, which is based on quaternions. There, the
orientation q is represented as
q = (cos φ , sin φ ) .
[2.23]
With this formulation the periodicity is solved. The difference is that the machine learning approach has to regress two variables (sine and cosine) in range of −1 and 1. These two regressions
could then be recombined to an angle φ with
φ = arctan
sin φ
cos φ
[2.24]
2.6. Receiver Operator Characteristic
For evaluating the parameters of binary classification algorithms and to benchmark different approaches usually the Receiver Operator Characteristic (ROC) curve is used. This curve represents
the relationship between the true positive rate (TPR) and the false positive rate (FPR) of an algorithm (see Figure 2.17). One specific configuration of the classification algorithm produces one
point within the ROC space. To generate a curve, one parameter of the algorithm is iterated over
his whole range. This parameter is usually a threshold of the output probability. The goal of
each classification algorithm is to reach a TPR of 100% and a FPR of 0% what can be seen in
Figure 2.17 as the green line. The red diagonal represents a random process.
16
2. Background
True Positive Rate
1
0
False Positive Rate
1
Figure 2.17.: Receiver Operator Characteristic curve: the red line represents a random process, the green
curve represents a perfect classifier and the blue line represents a classification algorithm,
which can be optimized.
2.7. Related Works
The detection of other obstacles in the surrounding of a vehicle is an elementary part of environment perception. Especially moving objects have to be detected to generate reactions for driver
assistance systems or autonomous driving according to the environment. For this purpose several
publications exist about the detection of moving objects with different sensor types. For example
Vu et al. [50] implemented a dynamic obstacle detection by using a measurement grid of a laser
sensor. They used the movements of the occupied grid cells in dependence of the free space information to detect dynamic obstacles. These dynamic objects were verified with the measurements
of radar sensors. Another approach from Asvadi et al. [1] uses a Velodyne sensor to generate a
2.5 D grid map, where dynamic objects are detected by recognizing movements of grid cells over
time.
The mentioned papers are using specific sensors to distinguish between dynamic and static objects.
Nuss et al. [36] proposed a method to generate a dynamic object detection based on a dynamic
occupancy grid map (DOG), what is described in Section 2.2.2. This allows a sensor independent
recognition of moving objects. Some problems like false positives at clutter occurs by using this
approach in different scenarios. For this purpose this thesis extends this approach by using the
DOG as an input for a Convolutional Neural Network to detect moving objects as one of the goals.
In the last few years deep and Convolutional Neural Networks (CNNs) were in the center of research. Especially with the rising computational power deeper networks were feasible with lower
training and execution time, what Cireşan et al. [7] proved with a speed improvement of factor
50. There a pixelwise labeling of images from an electron microscopy to classify membrane and
17
2. Background
non-membrane pixels was introduced with CNNs. To obtain this pixelwise prediction a sliding
window approach was used, where each sub image created the label for one output pixel.
Other approaches like the Hypercolumns of Hariharan et al. [24] use a CNN for feature generation.
These features are produced over all the CNN layers by upscaling smaller layers and combining
all the layers pixelwise. This produces a feature vector for each pixel, which can be used to
train linear classifiers. A slightly different approach is the Region-based Convolutional Network
from Girshick et al. [22] which creates classified bounding boxes instead of pixelwise predictions.
There, the CNN is used to create feature vectors of generated region proposals which are provided
to support vector machines (SVMs). Based on these SVMs the regions are classified.
One of the problems of these approaches is that for each pixel or each region proposal the classifier has to be executed. That produces a computational overhead, which can be reduced. For this
purpose Long et al. [28] introduced Fully Convolutional Neural Networks (FCNs) in combination
with deconvolutional layer, which were trained end-to-end for segmentation tasks. They transfered the standard AlexNet [26], VGG-net [41], and GoogLeNet [42] to FCNs and compared them
based on the Pascal VOC 2012 dataset [18]. To create a fully convolutional network for pixelwise
prediction, they used deconvolution layers to rebuild the input image size as a 2D output for each
class. The deconvolution layers are implemented as up-sampling layers and the output is a pixelwise class prediction. This reduces the computational overhead so that for a pixelwise prediction
the net has to be executed only once. To refine the segmentation result, several deconvolution layers combined with convolutional layers were stacked. This results in more detailed segmentation
structures at the output layer.
Based on this approach, several extensions and applications were developed. Fischer et al. [19]
applied a similar FCN structure for calculation of optical flow. The input was a combination of
two images to create the optical flow between these pictures. Another extension is the refinement
with Conditional Random Fields (CRFs) [6, 58]. There, a CRF is used to produce more detailed
structure of the segmented objects.
All of these approaches use predefined datasets like Pascal VOC [18], Imagenet [13] or Citycapes [10] to evaluate the neural networks. There are only a few approaches, which uses different
input images. For example Wang et al. [52] used a pre-trained network, which was trained on the
cifar-10 dataset [25], to detect oil tanks on satellite images. Another approach from Maturana et
al. [31] used an three dimensional occupancy grid map generated by an LiDAR sensor. Based on
this grid map different objects like pedestrians or toilets are classified.
In general there was no approach found, which uses two dimensional DOGs (see chapter 2.2.2)
as an input for a CNN. For this reason, the datasets for training and testing has to be created and
public datasets can not be used. Because of the amount of data which is needed for training, a semisupervised learning approach is followed as one goal of the thesis. This is a current topic in the
literature, but most of the approaches use weakly labeled data to generate pixelwise segmentation
labels or use unsupervised pre-training. For example Dai et al. [12] use manual bounding box
18
2. Background
labels to generate a pixelwise ground truth automatically. Another approach is a combination of
weakly- and semi-supervised approaches [38]. There Papandreou et al. combined pixel-level,
image-level and bounding box level annotations to create more data for training. Unsupervised
pre-training approaches use unlabeled data for the initialization of the CNN and train afterwards
on the labeled data, like Masci et al. [29]. However all these approaches need a specific kind of
labels for the main training step (image-level, bounding box, etc.).
In 1965, Scudder [40] introduced a self-learning algorithm, where an “adaptive pattern-recognition
machine” was created. To obtain better results, unlabeled data was labeled by the machine itself.
This new data was then used to train the machine for obtaining more accurate results. After that the
labeling step starts again so that finally the amount of labeled data rises by obtaining better results
with the “adaptive pattern-recognition machine”. This approach will be applied in this thesis to
obtain more accurate results on the one hand and more labeled data on the other hand.
19
3. Methodology
In this chapter, the methodological background of the three goals of this thesis are described.
First the segmentation of the grid map with CNNs and their parameter optimization is introduced.
Afterwards the orientation extraction is presented. Finally the semi-supervised learning approach
is described.
3.1. Segmentation
The first goal of this thesis is to create a pixelwise classification of the DOG, as shown in Figure 3.1, to support clustering of dynamic obstacles. To this end, the target set consists of two
classes, dynamic (moving objects) and static (non-moving objects), where background is part of
the static class. The classification process is divided into three steps. First the DOG is preprocessed
to obtain a specific input for the CNN. Based on this the CNN creates the pixelwise classification.
Finally, the intersection between the segmented image and an image of occupied grid cells is calculated. Doing so produces sharper labeling borders and rejects false positives which appear by
labeling non-occupied grid cells as dynamic.
To achieve the classification task, Fully Convolutional Neural Networks (FCNs) are used to reduce
the computationally expensive task of sliding window approaches (as mentioned in Section 2.3.3).
Based on Long et al. [28], the popular network structures VGG-net [41] and Alexnet [26] are converted to FCNs with deconvolutional layers, which are using bilinear upscaling. This produces a
two-dimensional output of the network and creates the desired pixelwise classification by using
the same input and output size. For initialization of the CNN weights, pre-trained networks are
used for reducing the training time. These pre-trained networks were trained on the public Imagenet [13] dataset, which contains colored images of different categories. Even with a different
type of input data, the pre-trained networks can reduce the training time. This is caused by the pretrained filters of the convolutional layers, which contain specific shape representations and which
should improve the segmentation task.
The transfer of the pre-trained networks to a new scope requires several parameter optimizations,
which are discussed in the next subsections. Each parameter is optimized by its own to finally
merge the optimized parameters together to obtain an optimized FCN. Therefore an independence
of parameters is assumed.
20
Occupancy
Refinement
Convolutional
Neural Network
Preprocessing
3. Methodology
Figure 3.1.: Overview of segmentation steps: the raw DOG is preprocessed, to extract a specific input for
the CNN. Afterwards the output of the network is refined with the occupancy information of
the DOG. The final output contains the static labels (blue) and the dynamic labels (turquoise).
3.1.1. Input
Based on the input data, the CNN calculates the pixelwise classification. This data has to contain
sufficient information to extract an accurate prediction. For this reason different input data type
combinations are provided and evaluated against each other. The following data types can be
provided from the DOG:
• Occupancy probability (with and without freespace information)
• Mean velocity in x direction over all particles in one grid cell
• Mean velocity in y direction over all particles in one grid cell
• Variance of velocity in x direction over all particles in one grid cell
• Variance of velocity in y direction over all particles in one grid cell
• Mahalanobis distance of overall velocity in one grid cell.
As described in Section 2.2.2, the DOG contains several particles in each grid cell. These particles
have weights, which represents the occupancy probability. This probability is combined with the
freespace information, which is provided as well. The result is a value between 0.0 and 1.0, which
represents the occupancy state: 0.0 represents free space and 1.0 represents an obstacle, which is
shown in Figure 2.5. If the freespace information is not provided, the value is only between 0.5
and 1.0.
Each particle describes a dynamic state as well (see Section 2.2.2). This is described with the
velocity in x and y direction. For one grid cell the mean velocity and the variance of the velocities
of the particles can be calculated, which represents an uncertainty for the velocities. Additionally,
21
3. Methodology
a mahalanobis distance m of the mean overall velocity voverall to the velocity vre f erence = 0 can be
calculated, as mentioned in [36]:
m2 = vToverall Σ−1 voverall ,
[3.1]
where Σ represents the covariance matrix of the velocities.
The structure of the used CNN is designed based on public network structures like VGG-net [41]
and Alexnet [26]. This means the input has to be a colored image, otherwise the weights of pretrained networks cannot be used. For this purpose the three color channels (blue, green, and red)
are used to encode the discussed input data. The following combinations of input data are provided
(B - G - R):
1. Occ f ree - vx - vy
2. Occ f ree - vx,norm - vy,norm
3. Occnon f ree - vx,norm - vy,norm
4. Occ f ree - voverall - Varoverall
5. Occ f ree - voverall - m
In every combination the occupancy probability is included. The reason for this is that the CNN
should use this occupancy information to include shape recognition of the obstacles for the prediction. The third combination excludes the freespace information to analyze the influence of
freespace information. All the other parameters are used to represent the movement of grid cells.
Besides the velocity itself, the normalized velocities vx,norm and vy,norm can be provided. The
normalized velocity in x and in y direction can be calculated accordingly, where the normalized
velocity in x direction is defined as
vx
vx,norm = p
.
Var(vx )
[3.2]
The normalization should help to distinguish between clutter and real moving objects. This is
caused by the DOG: grid cells, which belong to clutter, contain particles with different velocities.
This can potentially be detected with the variance over these velocities. Additionally, in combination four and five, the parameters for the uncertainty are provided in a separate input channel. One
of these parameters is the overall variance, which is calculated with
Varoverall = Var(vx ) + 2Cov(vx , vy ) +Var(vy ) .
[3.3]
Another parameter is the mahalanobis distance (see Equation 3.1). The problem of these combinations is the restriction of three input channels. This causes the velocities to be combined to the
overall velocity voverall .
22
3. Methodology
FCN-32s
``tabby cat"
96
FCN-16s
FCN-8s
Ground truth
4 4 6 96 96 00
6 38 38 25 40 40 10
25
convolutionalization
tabby cat heatmap
25
6
96 96 00
4 4 6 40 40 10
38 38 25
Figure 3.3.: Refinement of the segmentation with
the deep jet and a resulting smaller
stride at the last layers. The first images
show the output from the stride 32, 16,
and 8 network [28, p. 3436].
96
Figure 3.2.: Conversion of fully connected layers to
fully convolutional layers [28, p. 3433].
3.1.2. Range
The CNN is initialized with the weight of pre-trained networks, which were trained with colored
images from the public Imagenet [13] dataset. This means that every input channel of the CNN
has a range of [0, 255], i.e. the input values from the previous section have to be discretized to this
range. The occupancy probability with a range of [0, 1] can easily be transferred to the required
range by using a multiplication factor of 255. For the other parameters, a mapping to the required
range has to be applied. For this purpose, the following input limits l are evaluated and transferred
to the range of [0, 255]:
l = [−t,t] ,
t ∈ 5, 10, 15, 20, 25 .
[3.4]
These ranges are applied to the normalized and unnormalized velocities, the variances, and the
mahalanobis distance.
3.1.3. Structure
Based on Long et al. [28], public network stuctures are used and converted to fully convolutional
layers as mentioned in Section 2.3.3. This is executed by transferring weights of fully connected
layers to convolutional layers as shown in Figure 3.2. Two network structures are used in this
thesis: VGG-net [41] and Alexnet [26]. VGG-net is deeper than Alexnet, what means that it
contains more layers. That causes on the one side a longer execution time. On the other side Long
et al. [28] showed a higher accuracy.
The original VGG-net contains 16 weight layers (convolution and fully connected) and five max
pooling layers. Each convolution layer besides the first has a stride of one, a filter size of three,
and a padding of one. This ensures that input and output of a convolutional layer have the same
size. All the pooling layers have a filter size and a stride of two, what produces a reduction of the
input size by factor two. For further details, the reader is referred to [41].
23
3. Methodology
image
conv1
pool1
conv2
pool2
conv3
pool3
conv4
pool4
conv5
pool5
conv6-7
2x conv7
pool4
4x conv7
2x pool4
pool3
32x upsampled
prediction (FCN-32s)
16x upsampled
prediction (FCN-16s)
8x upsampled
prediction (FCN-8s)
Figure 3.4.: Deep jet introduced to the VGG-net: pooling and prediction layers are shown as grids that reveal
relative spatial coarseness, while convolutional layers are shown as vertical lines. First row
(FCN-32s): up-sampling with a stride 32 in a single step. Second row (FCN-16s): Combining
predictions from both the final layer and the pool4 layer, at stride 16, to predict finer details,
while retaining high-level semantic information. Third row (FCN-8s): Additional predictions
from pool3, at stride 8, provide further precision. [28, p. 3435].
image conv1 pool1 conv2 pool2
conv3-5
pool5 conv6-7
2x conv7
pool2
8x conv7
4x pool2
pool1
32x upsampled
prediction (Alex-32s)
16x upsampled
prediction (Alex-16s)
4x upsampled
prediction (Alex-4s)
Figure 3.5.: Deep jet introduced to the Alexnet: pooling and prediction layers are shown as grids that reveal
relative spatial coarseness, while convolutional layers are shown as vertical lines. First row
(Alex-32s): up-sampling with a stride 32 in a single step. Second row (Alex-16s): Combining
predictions from both the final layer and the pool2 layer, at stride 16, to predict finer details,
while retaining high-level semantic information. Third row (Alex-4s): Additional predictions
from pool1, at stride 4, provide further precision. (modeled after [28, p. 3435]).
24
3. Methodology
The original Alexnet is not as stepwise structured as the VGG-net. It contains seven weight layers
(convolution and fully connected) and three pooling layers, but the smallest feature map has the
same size as the VGG-net. This is possible because of the first convolutional layer, which has a
filter size of eleven and a stride of 4. In total Alexnet is much smaller, what results in a shorter
execution time. This should be used as a reference in terms of accuracy.
As last layer, a deconvolutional layer, as described in Section 2.3.2, is introduced. Caused by the
network structure, the single deconvolutional layer has a stride of 32 to upscale the image to the
input size. This limits the scale of details of the segmentation because the inner representation
is 32 times smaller than the final output. To improve this, Long et al. [28] introduced a stepwise
deconvolution, called deep jet. There, each deconvolution step is fused with lower layers to recombine higher semantic information with lower shape information, as shown in Figure 3.4. The
last deconvolution layer ten uses a smaller stride, what produces a finer segmentation output (see
Figure 3.3). For the VGG-net, three variations were introduced:
• VGG-net with a stride of 32 at the last deconvolutional layer (FCN-32s)
• VGG-net with a stride of 16 at the last deconvolutional layer (FCN-16s)
• VGG-net with a stride of 8 at the last deconvolutional layer (FCN-8s)
The training of the structures with smaller strides can be achieved with different approaches, which
are proposed from Long et al. [28]. On the one side the network can be trained non-incremental.
That means the network with the smaller strides is initialized with the pre-trained network and
trained afterwards for certain iterations. On the other side the network can be trained incremental.
That means the network with the largest stride is trained first. Afterwards this network is used to
initialize the network with the next smaller stride and so on.
In this thesis, deep jet is additionally applied to Alexnet to achieve a finer segmentation output.
Because of the different network structure, the possible strides differ from the ones applied to
VGG-net. The main structure of the Alexnet is shown in Figure 3.5. Three possible deep jet
versions are evaluated:
• Alexnet with a stride of 32 at the last deconvolutional layer (ALEX-32s)
• Alexnet with a stride of 16 at the last deconvolutional layer (ALEX-16s)
• Alexnet with a stride of 4 at the last deconvolutional layer (ALEX-4s)
3.1.4. Zooming into the Input
The size of the DOG is predefined. That means the dimension of obstacles like vehicles or pedestrians is predefined as well. Additionally the filter sizes of the CNN are also constant. That causes
the network to recognize only specific sizes of objects, especially with the initialization with the
pre-trained network. To evaluate different sizes of obstacles, the DOG is cropped in the middle
25
3. Methodology
crop
resize
Figure 3.6.: Zooming into the input: first the input is cropped and afterwards it is resized to change the size
of the obstacles by using a constant input size for the CNN.
and scaled afterwards to the origin input dimensions, as shown in Figure 3.6. That produces an
adaptable obstacle size while using a constant image size for the network, so the network structure
has not to be adapted for each input size. The following crop sizes cs are evaluated:
cs ∈ {300 × 300, 400 × 400, 500 × 500, 600 × 600}
[3.5]
where the origin image size is 600 × 600. This implies that the crop size 300 × 300 is a doubling
of the object size and the crop size 600 × 600 is the origin object size.
3.1.5. Learning Rate
Long et al. [28] and Cordts et al. [9] provided learning rates to train the Fully Convolutional
Neural Networks. These learning rates were unnormalized and accordingly dependent on the
input image size. In this thesis, these learning rates are transferred to normalized values, to get
rid of numerical issues and to being able using backpropagation independent of the total number
of pixels or weights of pixels. The transferred learning rates are fixed learning rates, where the
momentum and the weight decay is set according to Long et al. [28].
The type of input images differs from the type used by Long et al. and Cordts et al. For this purpose
the learning rates are evaluated in a smaller and a larger range around the proposed values. The
evaluated learning rates µ are
µ ∈ {2.14 × 10−6 , 7.14 × 10−6 , 2.14 × 10−5 , 3.6 × 10−5 , 7.14 × 10−5 , 2.14 × 10−4 } ,
[3.6]
where the learning rates 2.14 × 10−5 and 3, 6 × 10−5 represent the transferred learning rates of
Cordts et al. and Long et al.
Additionally a step learning rate policy is evaluated, where a higher learning rate can be chosen at
the beginning of the training. Afterwards the learning rate is decreased in steps over the iterations.
It is expected that the pre-trained weights of the first layers can be easier adapted to the different
26
3. Methodology
type of input with a higher learning rate at the beginning of the training. After the reduction of
the learning rate the last layers should be finetuned to reach a convergence. To achieve that, the
start learning rate is chosen ten times higher than the learning rate of the fixed learning rate policy.
Thereby the step size is chosen in relation to the total training iterations, so that the learning rate
decreases to the learning rate of the fixed learning rate policy after half of the training iterations.
3.1.6. Weight Matrix
As it is shown in Figure 2.5 or Figure 3.1, the number of static pixels (including the background)
is much higher than the number of dynamic pixels. That can cause problems with the backpropagation algorithm to a higher affinity to the static class. For this purpose a weight matrix C is
introduced in the cost function J(θ ) of the multinomial logistic loss (see Section 2.3.4):
#
y(i) )
(i)
(i)
(i)
(
,
1{y = k} log P y = k|x ; θ
J(θ ) = − ∑ ∑ C
"
m
K
[3.7]
i=1 k=1
where C contains a weight for each labeled class:
h
i
C = c(1) c(2) . . . c(K) .
[3.8]
3.2. Orientation Extraction
The CNN from Section 3.1 creates a pixelwise prediction for the DOG. This segmentation should
improve a cluster and tracking algorithm as described in Section 2.3.5. To provide more information to the tracking algorithm, the orientation of the object should be extracted. The orientation in
this context is represented by the moving direction. To extract the moving direction of the obstacle
two approaches are compared to each other: orientation extraction through the CNN and over the
mean velocities in x and y direction, what is discussed in the further subsections.
3.2.1. Orientation Extraction with Convolutional Neural Networks
One approach is to extract the orientation through the CNN. This causes some problems with the
periodicity of the angle, as described in Section 2.5. For this purpose the biternion representation
with sine and cosine of Beyer et al. [3] is used. To regress the angle, the network structure is
adapted as shown in Figure 3.7. There, the structure of the pre-trained network is not changed,
besides the transformation to a FCN. The extension with deconvolutional layers in the end of the
network is replicated twice to create the same structure once for the regression of the sine and once
for the regression of the cosine. Only the last layer, which is a softmax layer for the classification,
is skipped for the regression. This produces a pixelwise regression for the sine and cosine, which
can be recombined to an angle. To produce a clusterwise orientation extraction, clusters can be
27
3. Methodology
cosine regression
transferred
layers
sine regression
pre-trained
network structure
classification
deconvolutional
layer
Figure 3.7.: Adaption of CNNs for orientation extraction: two more output maps (deconvolutional layer)
are introduced for the regression of sine and cosine. The pre-trained network structure is not
changed, besides the last layers, which are transfered from fully connected to fully convolutional layers.
28
3. Methodology
Training
CNN
Unlabeled
Data
Labeled
Data
CNN
Execution
False
Positive
Rejection
Clustering
Figure 3.8.: Semi-supervised training loop.
produced over the pixelwise classification output of the network. Then a mean of sine and cosine
can be calculated over the clusters to extract the cluster angles.
3.2.2. Orientation Extraction over Velocities
Another approach is to extract the orientation by directly using the velocity information, which
is provided by the DOG. The output of the pixelwise classification (see Section 3.1) is used to
create clusters of dynamic objects. Based on this information a mean of the velocities v in x and y
direction can be calculated, which can be used to extract the angle φ of the clusters with
φ = arctan
vy
vx
.
[3.9]
3.3. Semi-Supervised Learning
A time and cost expensive task of training deep neural networks is the labeling of data (see Section 2.4). For this purpose a semi-supervised approach based on Scudder et al. [40] is evaluated.
There, the machine learning approach is trained on manually labeled data. Afterwards, unlabeled
data is labeled by the machine learning approach automatically. This produces a larger training
database, which should improve the accuracy. Especially to reduce clutter, the machine learning
approach can generalize over a lager dataset.
Additionally, to reduce false detected objects in the automatic labeling step, the labeled data is
corrected. For this purpose, the pixelwise labeled data is clustered and a parameter is calculated for
each cluster, which is used to reject false positives. As parameter the mean normalized velocity of
29
3. Methodology
a cluster and a combination p of the mahalanobis distance m (see Equation 3.1) and mean velocity
v of a cluster is used.
p = mv
[3.10]
This parameter is chosen, because of the velocity independent measure with the mahalanobis distance on the one side and the velocity itself on the other side. That produces a parameter, where
the value decreases with a smaller velocity and a smaller uncertainty by using the mahalanobis
distance. Finally, obstacles that move either slowly or have a high uncertainty can potentially
be filtered, what increases the quality of the automatic labeled data. The whole semi-supervised
training loop is shown in Figure 3.8.
30
4. Realization
In this chapter the realization of the methods is discussed. First, the data generation is described.
Afterwards, the proceeding for the comparison with other approaches is defined. Finally, the used
framework and the training hardware are presented.
4.1. Data Generation
In this section, the data generation for training and evaluation is defined. First, the route of the
recording vehicle is described. Afterwards, the semi-automatic labeling and the split of data are
introduced. Finally, the rotation of the training data to receive a higher orientation variance is
specified.
4.1.1. Recording Route
For training machine learning approaches, it is important to generate a large database with high
diversity in the data. In order to obtain this data, a route was recorded 4.3 miles through different
areas in Sunnyvale (California), as shown in Figure 4.1. The first area (green ellipse) is a six-lane
street, with a speed limit of 45 miles per hour. There, several passing maneuvers of the recording
vehicle and from other vehicles are captured. Additionally, some traffic lights are located in this
area, where several vehicles crossed the street. After a left turn the recording vehicle entered the
downtown area, which is zoomed in Figure 4.1. There, pedestrians, bicycles, and vehicles with
a lower speed were recorded as well as parked vehicles, which were not moving. The recording
vehicle also entered a limited traffic zone (orange ellipse) and a parking lot of a shopping center
(red ellipse) with several pedestrian crossings to generate enough data. Finally, the recording
vehicle drove back on the six-lane street.
The whole recording contains different scenarios like larger streets, traffic lights, pedestrian crossings, and limited traffic zones, as shown in Figure 4.3. This should result in a certain diversity of
the data to facilitate generalization of urban areas for the machine learning approach.
31
4. Realization
Figure 4.1.: Route for recording the data: the route was defined through different areas in Sunnyvale (California): a six-lane street (green ellipse), a limited traffic zone (orange ellipse), and a parking lot
of a shopping center. The zoomed area is the downtown area of Sunnyvale.
4.1.2. Semi-Automatic Data Labeling
After recording the data, the labels have to be created. For this purpose, a semi-automatic way is
followed. Labels are first generated automatically, clustered, and afterwards corrected manually,
as shown in Figure 4.2.
To distinguish between dynamic and static objects in an automated way, the appraoch of Nuss
et al. [36] is used as baseline method. There, the velocities of the particles within a grid cell of
the DOG and their variances are fused to the mahalanobis distance, which was already defined in
Equation 3.1. The mahalanobis distance describes the distance of a point to a distribution, where
the standard deviation σ is taken into account. Therefore, a mahalanobis distance of 3 corresponds
to a distance of 3σ of the value to the distribution. In case of the DOG, Nuss et al. [36] calculate
the distance of the velocity vre f erence = 0 to the distribution of the particle velocities within a grid
cell. As a result, the mahalanobis distance is used as a measure for the certainty of movement, what
Dynamic
Occupied
Grid Map
Detect Dynamic
Occupied
Grid Cells
Clustering
Manual
Correction
Figure 4.2.: Overview of semi-automatic labeling toolchain.
32
4. Realization
Figure 4.3.: Examples of grid map recordings: the recording vehicle is placed in the middle of each image
and its moving direction is to the right. Moving vehicles are labeled with a red ellipse, moving
pedestrians are labeled with a blue ellipse, and the colored areas encode the velocity information
including the angle, which is shown within the circle. The upper left image shows a six-lane
street, the upper right image shows a left turn on a traffic light, the lower left image shows a
pedetrian crossing, and the lower right image shows a limited traffic zone.
33
velocity in y direction
4. Realization
0
velocity in x direction
Figure 4.4.: Example of the mahalanobis distance: the green point describes a grid cell with a high mean
velocity and a high variance of the particle velocities. The red point describes a grid cell with
a lower mean velocity and a lower variance of the particle velocities. As a result the red point
has a greater mahalanobis distance than the green point.
is shown in Figure 4.4. Based on this information, a threshold to distinguish between dynamic and
static obstacles was calculated by Nuss et al.
The pixelwise decision for dynamic obstacles has to be improved to remove false positives and
false negatives. For this purpose, the output of the pixelwise decision is clustered based on the
DBSCAN algorithm introduced by Ester et al. [17] and transferred to the DOG by Peukert [39].
Afterwards, the convex hull of these clusters is provided to correct the data manually. Additionally,
the orientation for each cluster is extracted by using the mean velocity in x and y direction over
the cluster of grid cells, like in Section 3.2.2. This is used to generate the ground truth for the
evaluation of the orientation extraction.
The manual correction finalizes the ground truth for evaluation of the different goals of this thesis.
During this process, not all of the data is used. Especially if the recording vehicle had to wait on
a traffic light for a left turn or if the recorded images are too similar, the annotation is skipped.
Overall, 3 457 images were labeled with this semi-automatic approach.
4.1.3. Splitting the Data
The goal, while training machine learning approaches, is to generalize and not to overfit over
the training data. Overfitting means in this context, that the machine learning approach adapts
perfectly to the training data, but does not generalize over it, as shown in Figure 4.5.
34
4. Realization
Figure 4.5.: Overfitting on a physical model [8].
dataset
number of images
testing
369
validation
343
training
2745
scenarios
pedestrian crossing, six-lane street,
limited traffic zone
parking of a shopping center, left turn
on traffic light of a six-lane street,
right turn in the downtown area
rest of the recording
Table 4.1.: Split of data with the different scenarios.
To evaluate the generalization, the data has to be split in a training and a test dataset, where the
test dataset is not used for training. Furthermore, the parameters of the CNN are optimized on an
additional dataset. For this purpose a validation dataset is created. This allows to train the CNN on
the training dataset, to optimize the parameters of the CNN by evaluating on the validation dataset,
and finally to evaluate the CNN against other approaches on the test dataset.
The problem of splitting the data is that the number of training samples decreases. Therefore, the
validation and test dataset should not be as large as the training dataset. On the other side, the
validation and test dataset have to be large enough to be representative. In this thesis, a split of
roughly 80% - 10% - 10% for training, validation, and test data is chosen.
The labeled data is generated from one recording, where around 100 milliseconds are between two
images. As a result, the split of the data cannot be established randomly. Otherwise, the datasets
have a high correlation and the split will not be meaningful. For reducing the correlation, the split
is done manually, where two policies are complied. First, the selected data has to be representative
and should contain a large street scenario, pedestrians, and scenarios with a lot of clutter. The
second policy is, that between the different sets of data the appearance should change, so that
correlation decreases. This is ensured by using a minimum time difference between two images
of different datasets. Finally the data is divided into three dataset as mentioned in Table 4.1.
35
4. Realization
Figure 4.6.: Rotation of a training images rotated 140◦ counter clockwise: the corners (red ellipses) are
filled with an uncertainty occupancy P(Ock ) = 0.5 and a velocity, variance and mahalanobis
distance of zero.
4.1.4. Rotation
Within traffic scenarios, vehicles can appear at different positions and with different orientations.
This has also to be encoded in the training dataset. The different positions are captured with the
different scenarios within the recording. For the different orientations, the recordings have to be
much longer to produce different views of orientations. For this purpose the orientation variance
is introduced afterwards in the training dataset by rotating the images in ten degree steps. This
produces on the one side rotation variance in the dataset and on the other side the dataset grows
by the factor of 36. Especially for the orientation extraction (see Section 2.5) it is required to have
high variation of the orientation angle within the dataset.
The rotation of the images is calculated on the input images of the CNN. These images contain the
velocity in certain directions. Thus, the velocity information has to be rotated as well. The corners
of the rotation, where no information is provided, are set to an uncertainty occupancy P(Ock ) = 0.5
and a velocity, variance and mahalanobis distance of zero, as shown in Figure 4.6. As a result, the
datasets size increases according to Table 4.2. The network is trained on images with rotation and
the evaluation is done on original images. That means the rotation of the testing dataset is never
used.
4.2. Comparison to other Approaches
After the optimization of the CNN parameters, the approach of this thesis is compared to Nuss
et al. [36] as a baseline method of grid cell classification of dynamic and static objects, which is
36
4. Realization
number of images
(original)
369
343
2 745
dataset
testing
validation
training
number of images (with
rotation)
not used
12 348
98 820
Table 4.2.: Increse of dataset size by rotation of data: the rotated images are used for training and the
original images are used for evaluation. For the final comparison to other approaches the network
is trained on the rotated training and validation dataset and evaluated on the original testing
dataset. All other optimization and evaluation is done on the original validation dataset, by using
the rotated training dataset for training.
Training
≈ 80%
Validation
≈ 10%
training for
parameter optimization
evaluation of
parameter optimization
training for
final evaluation
Testing
≈ 10%
final evaluation
Figure 4.7.: Utilization of the different datasets.
described in Section 4.1.2. This approach represents the state of the art, where the mahalanobis
distance based on velocities is used as distinguishing criterion. To evaluate the approaches against
each other, the training is based on another dataset than the the optimization of the CNN: the
training dataset and the validation dataset are merged to create the final training dataset. The
evaluation is based on the test dataset, as shown in Figure 4.7. As a result, the final comparison is
decorrelated to the optimization steps of the CNN.
4.3. Framework and Hardware
Timing of the CNN is one part of the thesis. Therefore the used Framework as well as the hardware
is important. The CNN is developed with the branch “future” of the caffe framework of Long et
al. [28]. This branch contains additional layers, like the deconvolutional layer for the FCNs. The
branch is adapted at some point to extend the features, like the weight matrix of Section 3.1.6.
The CNNs are trained on a Nvidia Geforce GTX 980 Ti and the timing is measured on a Nvidia
Geforce GTX Titan X. The used Nvidia driver version is 352.93 with a CUDA version of 7.5 and
the operating system is Linux Ubuntu 14.04.
37
5. Evaluation
In this chapter, the evaluation of the goals of this thesis is discussed. First, the segmentation of the
DOG is evaluated. Afterwards, the results of the angle regression and the semi supervised learning
approach are described. Finally the CNN is compared to the referenced pixelwise classification
approach of Nuss et al [36].
5.1. Segmentation
In this section, the results of the segmentation are presented. First, a general description of the
trainings of the CNNs is given. Afterwards the different optimization parameters of Section 3.1
are evaluated. Each comparison of different configurations is made with ROC curves, which are
already described in Section 2.6. As mentioned in Section 3.1.6, the number of background grid
cells in the DOG is much higher than the number of occupied grid cells, but in this thesis the focus
is the recognition of dynamic objects. For this purpose, the generation of ROC curves uses only
occupied grid cells for the pixelwise evaluation to be independent of background of the DOG.
5.1.1. Training Configuration
All the parameters are optimized in the same way. For every parameter, the pre-trained CNN is
converted to FCN as discussed in Section 2.3.4 and is then trained for 200 000 iterations. Afterwards, the CNN is evaluated with the validation dataset to optimize each parameter. A typical
learning curve is shown in Figure 5.1. There, the convergence of the accuracy, precision and recall
can be recognized. That means the CNN does not overfit to the training data, especially because
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−10, 10]
Network Structure
FCN-32s
Crop Size
500 × 500
fixed policy with 2.14 × 10−5
Learning Rate
Weight Factor
60
Table 5.1.: Network configuration for the learning curve in Figure 5.1.
38
5. Evaluation
1.00
0.40
0.98
0.35
0.96
0.30
0.25
0.92
0.90
0.20
0.88
0.15
0.86
0.10
0.84
Pixel Accuracy
0.82
0.05
Recall Dynamic
200,000
180,000
160,000
140,000
120,000
100,000
80,000
60,000
40,000
Precision Dynamic
20,000
0
0.80
Precision
Accuracy, Recall
0.94
0.00
Iteration
Figure 5.1.: A typical learning curve with the network configuration of Table 5.1.
the accuracy does not decrease in the end of the training. Additionally, the figure shows a quick
convergence, what signifies a fast learning progress.
The learning curve includes also some noise. This noise of the data is caused by the training type:
in each iteration only one training sample is used to adapt the network weights (batchsize = 1).
This results (together with no weight accumulation) in a noisy learning curve.
For every parameter optimization, the CNN is trained once to make the training time feasible. This
can cause problems if the training of the network is not deterministic. For this purpose a network
configuration (see Table 5.1) was trained several times, what is shown in Figure 5.2. There, a
slightly indeterminism can be recognized, what is caused by the random initialization of the newly
inserted layer while converting the pre-trained CNN structure to an FCN, but the differences are
not essential. As a result, a single training of one network configuration will not influence the
outcome of the optimization process.
5.1.2. Input
In this subsection, the results of the input combinations of Section 3.1.1 are discussed. The network
configuration for this evaluation is shown in Table 5.2. Figure 5.3 shows the ROC curves of the
input configurations.
39
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
test run 1
test run 2
0.94
test run 3
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
test run 4
False Positive Rate
Figure 5.2.: ROC curves of multiple trainings of the same network configuration.
1.00
0.98
0.97
0.96
0.95
1. Occ f ree - vx - vy
2. Occ f ree - vx,norm - vy,norm
3. Occnon f ree - vx,norm - vy,norm
0.94
4. Occ f ree - voverall - Varoverall
False Positive Rate
Figure 5.3.: ROC curves of different input configurations.
40
0.07
0.06
0.05
0.04
0.03
0.02
0.01
5. Occ f ree - voverall - m
0.93
0.00
True Positive Rate
0.99
5. Evaluation
Parameter
Value
Input
evaluated in this section
Input Range
[−10, 10]
Network Structure
FCN-8s
Crop Size
500 × 500
fixed policy with 2.14 × 10−5
Learning Rate
Weight Factor
60
Table 5.2.: Network configuration for optimization of the input.
Figure 5.4.: Example of the recognition of a vehicle (upper left L-shape) and a wall (lower right L-shape)
within a DOG by using the freespace information: the green object in the middle is the recording
vehicle.
In general, the combinations with separation of velocities in x and y direction (combination 1 + 2)
and the combination with the mahalanobis distance (combination 5) reach a better result than the
combination with the separate overall variance (combination 4) and the combination without the
freespace information (combination 3). That proves on the one side that the split of the normalized velocity of different direction into the overall velocity and the overall variance creates no
information gain for the CNN. On the other side it proves, that the freespace information is an
elementary information gain for recognizing moving obstacles. This can be explained with an
example showed in Figure 5.4. There it is easy to distinguish between a corner of a wall, where at
the inside of the L-shape is freespace information, and a possible vehicle. Without the freespace
information, the two objects have the same appearance.
41
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
evaluated in this section
Network Structure
FCN-8s
Crop Size
500 × 500
fixed policy with 2.14 × 10−5
Learning Rate
Weight Factor
60
Table 5.3.: Network configuration for optimization of the range.
Some smaller differences can be seen between the combination with the mahalanobis distance
(combination 5) and the normalized and unnormalized approaches (combination 1 + 2). There,
the mahalanobis distance has a slightly lower performance than the other two approaches. A
suggestion for the reason is the fusion of the velocity in x and in y direction to the overall velocity.
There the information of the moving direction is lost. So some specific motion model cannot be
calculated. For example a vehicle can only drive forward and backward. This shows that an object
with a vehicle appearance and with a motion to the side is probably clutter and should be labeled
as static. This cannot be recognized without a moving direction.
The combination with the normalized velocities in x and y direction (combination 2) creates the
best result. It is slightly better than the combination with the unnormalized velocities. That can be
explained with the appearance of clutter. It can be observed, that clutter can reach a high velocity
but has also a high variance. If the velocities are normalized, clutter can easier be detected as
static.
As a result of the input configuration, the combination with the normalized velocity in x and y
direction and with freespace information (combination 2) will be used for the final comparison.
5.1.3. Range
In this subsection the range configurations, defined in Section 3.1.2, are evaluated. The configuration of the network is shown in Table 5.3. The results of the ranges are presented in Figure 5.5
as ROC curves. There, the performance decreases with a larger range. This can be caused by the
discretization of the input values while transferring the input values to discrete pixel values in the
range of [0, 255]. By increasing the range of the input values, the discretization error becomes
greater, what could decrease the accuracy of the pixelwise labeling.
The only range, which does not match to the observed result that the performance decreases by increasing the range size, is the range [−20, 20]. This range is slightly better than the smallest range.
To verify, if there is a correlation between the range and the velocity distribution, Figure 5.6 shows
the normalized velocity distribution in x direction (driving direction of the recording vehicle) of
42
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
range [−5, 5]
range [−10, 10]
range [−15, 15]
0.94
range [−20, 20]
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
range [−25, 25]
False Positive Rate
Figure 5.5.: ROC curves of different range configurations.
dynamic grid cells (green) and static grid cells (red). The distribution of static grid cells can be
described with a Gaussian curve. This is potentially caused by the particle filter of the DOG, where
the process noise (see Equation 2.5) and the measurement noise influence the velocity distribution.
The distribution of the dynamic grid cells has a specific appearance: above zero the distribution
is relatively low and equally whereas below zero a distribution similar to a Gaussian curve with
a negative mean can be recognized. This should be caused by the appearance of the obstacles
in street environments. Vehicles that are following the recording vehicle can be recognized with
different velocities, depending on the velocity of the recording vehicle. That is the reason why
the velocity is relatively equal distributed in the positive x direction. Obstacles that are driving
towards the recording vehicle appear most of the time on the oncoming traffic lane. That causes
the nearly same velocity for these vehicles and produces the Gaussian distribution in the negative
x direction.
The distributions of the normalized velocity in y direction are shown in Figure 5.7. There, the
normalized velocities are all around zero, what should be also caused by the appearance of obstacles in the recorded street environment. Vehicles in crossing direction are only recognized on
intersections, where on the recorded route only lower velocities were recognized.
The normalized velocity histograms do not show a relation to the range [−20, 20]. So the cause of
this unexpected result is not a specific distribution of the normalized velocities. One reason for this
appearance could be the usage of the pre-trained networks. The pre-trained networks are trained
on public datasets with a different type of input data. As a result, the weights of the network are
adapted to a specific appearance of the input, which does not necessarily match the appearance of
43
5. Evaluation
0.40
0.30
0.25
0.20
0.15
0.10
10
0
−10
−20
−30
0.00
30
Dynamic
Static
0.05
20
Relative Number of Grid Cells
0.35
Velocity
Figure 5.6.: Histogram of the normalized velocities in x direction of the occupied grid cells within the
validation dataset.
0.40
0.30
0.25
0.20
0.15
0.10
10
0
−10
−20
−30
0.00
30
Dynamic
Static
0.05
20
Relative Number of Grid Cells
0.35
Velocity
Figure 5.7.: Histogram of the normalized velocities in y direction of the occupied grid cells within the
validation dataset.
44
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−10, 10]
Network Structure
evaluated in this section
Crop Size
500 × 500
Learning Rate
fixed policy with 2.14 × 10−5 (adapted
for incremental trainings)
Weight Factor
60
Table 5.4.: Network configuration for optimization of the structure.
the input data within this thesis. By transferring the velocity to a certain range, the appearance of
the input is changed. This can cause different and unexpected results.
As a result of the range configuration, the range [−20, 20] is slightly better than the range [−5, 5],
what was also recognized in additional evaluations. Thus, the range [−20, 20] will be used for the
final comparison.
5.1.4. Structure
In this subsection, the structure configuration of Section 3.1.3 is evaluated. The configuration of
the network is defined in Table 5.4. The first ROC curves in Figure 5.8 state the results of the
incremental training of the FCN-16s. All the three network configurations are initialized with
the pre-trained network. The FCN-16s non-incremental version means that the FCN-16s is trained
directly from the pre-trained network. For the incremental versions the FCN-32s is trained first and
afterwards the weights are transferred to the FCN-16s to continue the training. The two versions
of the incremental training use different learning rates. Long et al. [28] recommend decreasing the
learning rate in case of the incremental training. This is achieved with a decrease of the learning
rate by factor ten. The other approach is implemented with a constant learning rate while applying
the incremental training.
The incremental training for the FCN-16s shows that the incremental version with the recommendation of Long et al. is slightly better than the non-incremental training. The unchanged learning
rate has the lowest performance in this comparison. The higher learning rate seems to deteriorate
the learned model of the FCN-32s.
In Figure 5.9, the ROC curves of the FCN-8s are shown. There, the opposite can be seen. The
incremental training with an unchanged learning rate gives the best result, while the incremental
training with the decreased learning rate creates a lower performance. The reason for that seems to
be the refinement with the deep jet, which seems to need more adjustment at the FCN-8s compared
to the FCN-16s.
45
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
FCN-16s non-incremental
0.94
FCN-16s incremental LR decreased
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
FCN-16s incremental LR constant
False Positive Rate
Figure 5.8.: ROC curves of different structure configurations for the incremental training of the FCN-16s.
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
FCN-8s non-incremental
0.94
FCN-8s incremental LR decreased
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.01
FCN-8s incremental LR constant
0.93
False Positive Rate
Figure 5.9.: ROC curves of different structure configurations for the incremental training of the FCN-8s.
46
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
FCN-32s
0.94
FCN-16s incremental LR decreased
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
FCN-8s incremental LR constant
False Positive Rate
Figure 5.10.: Best ROC curves of different incremental trainings evaluated with the FCN-32s.
In comparison to the FCN-32s, the best incremental versions are shown in Figure 5.10. There,
the FCN-32s is the best network configuration. This is an interesting result, because Long et
al. [28] proved the opposite. They describe that the FCN-8s version creates finer details within
the segmentation, what creates a better result. In general the finer details can be proved in this
thesis as well, what is shown in Figure 5.11. There, the segmentation output of the FCN-32s,
FCN-16s, FCN-8s, and the input data is provided. The FCN-32s created much coarser results
than the FCN-8s, but it creates also less false positive clusters. Additionally, the filter sizes of the
last deconvolutional layer can be recognized at the marked vehicle (yellow ellipse). There, the
up-sampling artifacts can be recognized, which becomes smaller with a smaller stride as with the
FCN-16s or the FCN-8s. The output of the CNN is not used directly for segmentation in this thesis.
As mentioned in Section 3.1, the output of the CNN is refined with the occupancy information of
the DOG. As a result, the FCN-32s generate a better ROC curve than the FCN-16s or FCN-8s
caused by less false positive. That means the refinement with the occupancy information of the
DOG is better than the refinement with the deep jet.
As mentioned in Section 3.1.3, the resulting quality should also be evaluated against the execution
time. For this purpose, the smaller Alexnet is also evaluated in this context. The results are shown
as ROC curves in Figure 5.12. The ROC curves show that the Alexnet in general has a lower
performance than the FCN-Xs networks. This is as expected caused by the smaller structure of the
network. It can also be recognized in Figure 5.13, where the outputs of the Alexnets are provided.
There, the Alexnet creates a coarser segmentation as the FCN-Xs in Figure 5.11, especially at the
stride 32 version the Alexnet combines two vehicles to one segment. This coarser segmentation is
47
5. Evaluation
Figure 5.11.: Example of the result of deep jet of VGG-net: top left: input image, top right: output of the
FCN-32s, bottom left: output of the FCN-16s, bottom right: output of the FCN-8s. In general,
the smaller stride at the last layers (for example at FCN-8s or FCN-16s) produces finer details
of objects, but increases also the number of false positive clusters.
48
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
FCN-32s
0.95
FCN-16s non-incremental
FCN-8s non-incremental
Alex-32s
0.94
Alex-16s non-incremental
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
Alex-4s non-incremental
False Positive Rate
Figure 5.12.: ROC curves for comparison of the Alexnet and the VGG-net.
Network Structure
FCN-32s
Alex-32s
Preprocessing and Network Execution
Postprocessing
Time
≈ 100ms
≈ 160ms
≈ 20ms
Table 5.5.: Execution time of different network structures.
afterwards also refined with the occupancy information of the DOG. That results in a better output
of the Alex-32s related to the Alex-4s as well as the FCN-32s related to the FCN-8s.
The execution time of the Alexnet is much smaller, what can be seen in Table 5.5. There, the
execution time is divided into two parts. The first part is the pre-processing and post-processing
time, which is needed to create the specific input image for the CNN (see Section 3.1.1) and
to refine the output with the occupancy information of the DOG. This time can be reduced by
parallelizing the execution. The second part of the execution time is the processing of the CNN
itself, which is already highly optimized on the GPU. The shorter execution time has to be seen in
contrast to the quality of the output. As a result the Alexnet can be used for real-time execution,
especially because the output is still good, but for further optimization in this thesis, the FCNs
based on the VGG-net will be used. As a result of the structure configuration, the FCN-32s will
be used for the final comparison.
49
5. Evaluation
Figure 5.13.: Example of the result of deep jet of Alexnet: top left: input image, top right: output of the
Alex-32s, bottom left: output of the Alex-16s, bottom right: output of the Alex-4s. The
smaller stride at the last layers (at Alex-4s) produces finer details of objects, but increases also
the number of false positive clusters.
50
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−10, 10]
Network Structure
FCN-8s
Crop Size
evaluated in this section
Learning Rate
fixed policy with 2.14 × 10−5
Weight Factor
60
Table 5.6.: Network configuration for optimization of the zooming.
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
cropsize 300 × 300
cropsize 400 × 400
0.94
cropsize 500 × 500
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
cropsize 600 × 600
False Positive Rate
Figure 5.14.: ROC curves of different zooming configurations.
5.1.5. Zooming into the input
In this subsection the zooming configuration of the input of Section 3.1.4 is evaluated. The configuration of the network is given in Table 5.6. The result of the zooming is shown in Figure 5.14,
where zooming into the input images creates a lower performance. This seems to be caused by
the size of the object. The goal of the zooming was to enlarge the moving objects for a better
recognition, but at the same time clutter and other objects are zoomed. These objects provoke
false positives, what is shown in Figure 5.15. There, an input image is calculated with the different
crop sizes. In the output image without zooming (crop size 600 × 600), no clutter is labeled as
dynamic. By increasing the zooming, more clutter is labeled, what is marked with red ellipses.
This signifies that the zooming into the images to increase the size of the objects does not improve
the segmentation task.
51
5. Evaluation
Figure 5.15.: Example of different crop sizes for one input image: top left: zoom of the input image and
the detection of a vehicle with a crop size 300 × 300, top right: input image, middle left: crop
size 300 × 300, middle right: crop size 400 × 400, bottom left: crop size 500 × 500, bottom
right: crop size 400 ×400. With a smaller crop size, the number of false positive increases (red
ellipse) and objects are recognized with more details, what can produce a partly recognition
of an object if the velocity information differs over the object (violet ellipse).
52
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−20, 20]
Network Structure
FCN-32s
Crop Size
600 × 600
Learning Rate
evaluated in this section
Weight Factor
60
Table 5.7.: Network configuration for optimization of the learning rate.
Additionally the large objects receive more details by increasing the zooming, what can be seen
at the borders of the recognized vehicle in Figure 5.15. This can also decrease the accuracy of the
result, because the details become so strong, that the vehicle is not recognized as one obstacle. As
a result, moving parts of the vehicles are recognized by itself, what can be seen at the crop size
300 × 300 (violet ellipse). There, the right part of the vehicle has lower velocity information as the
left part of the vehicle (light blue instead of dark blue) within the DOG. This produces a different
appearance of the output labels. In general it is important to choose the right level of details to
recognize the vehicle as one obstacle and not with too many detailed structures. As a result of the
zooming configuration, the crop size 600 × 600 will be used for the final comparison.
5.1.6. Learning Rate
In this subsection, the learning rate configuration from Section 3.1.5 is evaluated. The configuration of the network is defined in Table 5.7. The results of the fixed learning rates are stated in
Figure 5.16. There it can be recognized, that the proposed learning rate of Long et al., transferred
to a normalized learning rate, creates the best results. All other learning rates produce a lower
performance. That is caused by the network structure and the way how to learn. A CNN learns by
adapting its weights of the neurons. For that reason a gradient descent algorithm is used and the
CNN converges to a minimum. The goal is to converge to a global minimum and instead of local
minima. To reach that, the learning rate should be on the one side high enough to create a global
optimization and on the other side low enough to reach the minimum. That can be recognized in
the results, where higher or lower learning rates produce less performances.
The problem while training with a constant learning rate is the conflict between the global optimization and the convergence to a local minimum. For that purpose a step learning rate is evaluated
as described in Section 3.1.5. In Figure 5.17 and Figure 5.18 the two different learning developments are shown. There, the step learning rate converges faster and is more stable in the end of
the training. This is caused by the difference of the learning approaches. On the one side the step
learning rate starts with a ten times higher learning rate, what creates a faster conversion. On the
53
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
2.14 × 10−6
0.95
7.14 × 10−6
2.14 × 10−5
3.6 × 10−5
0.94
7.14 × 10−5
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
2.14 × 10−4
False Positive Rate
Figure 5.16.: ROC curves of different fixed learning rate configurations.
other side the learning rate decreases to a ten times lower learning rate at the end of the training,
what causes the stability of the training in the end.
The step learning rate results are shown in Figure 5.19, where, equally to the fixed learning rate approach, the transferred learning rate of Long et al. generates the best result. The other approaches
receive a lower performance in dependence of the distance to the optimal learning rate. The learning rate µ = 2.14 × 10−3 is not shown in the figure, caused by numerical problems by using such
a high learning rate.
To choose the best learning rate for the final comparison, the best learning rates of the two learning
policies are evaluated against each other. That can be recognized in Figure 5.20. There, the step
learning rate is better than the fixed learning rate approach. As a result the step learning rate of
µ = 3.6 × 10−4 will be used for the final comparison.
5.1.7. Weight Matrix
In this subsection, the weight matrix configuration of Section 3.1.6 is evaluated. The configuration
of the network is defined in Table 5.8. To define the weight matrices for evaluation, the ratio
between moving grid cells and non-moving grid cells of the validation dataset was calculated,
which is nearly 1 : 200. For this reason the following factors c(1) for the moving grid cells were
defined:
c(1) ∈ {1, 20, 40, 60, 80, 100, 120, 140, 160, 180, 200} .
54
[5.1]
5. Evaluation
1.00
0.40
0.98
0.35
0.96
0.30
0.25
0.92
0.90
0.20
0.88
Precision
Accuracy, Recall
0.94
0.15
0.86
0.10
0.84
Pixel Accuracy
0.82
0.05
Recall Dynamic
200,000
180,000
160,000
140,000
120,000
100,000
80,000
60,000
40,000
0
0.80
20,000
Precision Dynamic
0.00
Iteration
Figure 5.17.: Learning curve of the fixed learning rate 3.6 × 10−5 .
1.00
0.40
0.98
0.35
0.96
0.30
0.90
0.20
0.88
0.15
0.86
0.10
0.84
Pixel Accuracy
0.82
Iteration
Figure 5.18.: Learning curve of the step learning rate 3.6 × 10−4 .
55
200,000
180,000
160,000
140,000
120,000
100,000
80,000
60,000
40,000
Precision Dynamic
20,000
0.80
0.05
Recall Dynamic
0.00
Precision
0.25
0.92
0
Accuracy, Recall
0.94
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
2.14 × 10−5
7.14 × 10−5
2.14 × 10−4
0.94
3.6 × 10−4
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
7.14 × 10−4
False Positive Rate
Figure 5.19.: ROC curves of different step learning rate configurations.
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
0.94
fixed learning rate (3.6 × 10−5 )
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
step learning rate (3.6 × 10−4 )
False Positive Rate
Figure 5.20.: Comparison between fixed and step learning rate policies.
56
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−20, 20]
Network Structure
FCN-32s
Crop Size
600 × 600
step policy with 3.6 × 10−4
Learning Rate
Weight Factor
evaluated in this section
Table 5.8.: Network configuration for optimization of the weight matrix.
Note that the factor for the non-moving grid cells is set to c(2) = 1.
The result of the weight matrix adaption is shown in Figure 5.21 and Figure 5.22. There, it is
shown that the factor c(1) = 1 has the lowest performance and that the quality of the result increase
with a growing factor. After a factor of c(1) = 40 the quality decreases until the factor c(1) = 100.
Then it increases slightly before it decreases again. That result demonstrates the improvement
of the quality by using a factor to solve the imbalance of the labeled data. Especially when the
imbalance is as high as in this thesis, it is important to weight the labels. Otherwise the CNN
would learn that the background is most important, as shown in Figure 5.23. There the difference
between a factor of 1, 40 and 180 can be recognized. With an increasing factor the blobs around
the moving objects become larger, but also false positive blobs appear. That is the reason why the
factor has to be adapted to the database.
As a result the factor matrix
C = [1 40]
[5.2]
will be used for the final comparison.
5.2. Orientation Extraction
In this section the orientation extraction of Section 3.2 is evaluated. The configuration of the
network is defined in Table 5.9 and is created by the optimized network of Section 5.1. The
development of the learning process can be recognized in Figure 5.24. There, the two approaches
for the orientation extraction (extraction over the CNN directly and over the velocity information)
can be recognized. The figure shows the result of each approach against the ground truth data with
the mean difference to the ground truth and the deviation of the difference to the ground truth. The
ground truth is generated as described in Section 4.1.2. The goal is to reach a mean difference of
zero and a deviation of zero.
The results show that after large changes (until ≈ 80 000 iterations), the CNN converges to a better
result than the velocity approach. The large changes are caused by the first weight adaption, where
57
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
factor 1
0.95
factor 20
factor 40
factor 60
0.94
factor 80
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
factor 100
False Positive Rate
Figure 5.21.: ROC curves of different weight factor configurations in the range [1, 100].
1.00
True Positive Rate
0.99
0.98
0.97
0.96
factor 100
0.95
factor 120
factor 140
factor 160
0.94
factor 180
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.01
factor 200
0.93
False Positive Rate
Figure 5.22.: ROC curves of different weight factor configurations in the range [100, 200].
58
5. Evaluation
Figure 5.23.: Example of different weight matrices for one input image: top left: input image, top right:
weight factor c(1) = 1, bottom left: weight factor c(1) = 40, bottom right: weight factor c(1) =
180. By using a weight factor c(1) = 1 the CNN interpret the background as most important.
With an increasing weight factor, the blobs around the moving objects become larger, but also
false positives appear.
59
5. Evaluation
Parameter
Value
Input
Occ f ree - vx,norm - vy,norm
Input Range
[−20, 20]
Network Structure
FCN-32s
Crop Size
600 × 600
step policy with 3.6 × 10−4
Learning Rate
Weight Factor
40
Table 5.9.: Network configuration with the optimized parameters of Section 5.1.
the segmentation part (see Section 5.1) and the angle extraction part of the CNN generate weight
updates separately. This produces different weight adaption, which converges over time. The
segmentation part of the CNN is used to create clusters for the orientation extraction, as descibed
in Section 3.2. Caused by the weight adaption of this part, the orientation extraction with the
velocity changes in the first iterations (until ≈ 100 000) and stays relatively constant afterwards.
The training in this situation stops at 200 000 iterations. By observing the development of the
mean difference CNN angle
0.8
deviation difference CNN angle
mean difference velocity angle
deviation difference velocity angle
0.4
0.2
0
Iteration
Figure 5.24.: Development of the angle extraction over the iteration number.
60
200,000
180,000
160,000
140,000
120,000
100,000
80,000
60,000
40,000
20,000
−0.2
0
Angle difference in rad
0.6
5. Evaluation
Figure 5.25.: Examples of the orientation extraction: green line: orientation extraction with the CNN, red
line: ground truth, yellow line: orientation extraction with the velocity information, green line
without comparison to the ground truth: false positive in the segmentation. The CNN seams
to have problems with small slow moving objects (left image), but it uses the occupancy
information to extract a better orientation with uncertainty velocity information (right image).
training, the orientation extraction of the CNN could converge to a better result by training more
iterations, what could be evaluated in future work.
Figure 5.25 shows two examples of the orientation extraction, where the green line displays the
orientation output of the CNN, the red line presents the ground truth, and the yellow line indicates
the velocity approach. Clusters with only a green line are false positives, where no ground truth
is available. The CNN seems to have problems with small slow moving objects like pedestrians.
There, a large difference of the CNN to the ground truth can be recognized. The velocity approach
shows nearly no differences for these kind of objects. On the other side objects with uncertain
velocity information exist. This results in a lower performance of the orientation extraction with
the velocity approach. In this example the CNN can use the occupancy information to extract the
right orientation from the appearance of the object, as shown in Figure 5.25.
All in all the orientation extraction shows a better result with the CNN approach. Consequently,
tracking algorithms can be improved by using the orientation information directly extracted from
the CNN.
5.3. Semi-Supervised Learning
In this section, the proposed semi-supervised learning algorithm of Section 3.3 is evaluated. First,
an algorithm according to Scudder et al. [40] without a false positive suppression is evaluated. For
this purpose, the best network configuration of Section 5.1 (see Table 5.9) is used and a threshold
61
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
non SSL
SSL without FP suppression
0.94
SSL with FP suppression (normalized velocity)
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
SSL with FP suppression (parameter p)
False Positive Rate
Figure 5.26.: ROC curves of the semi supervised learning approach.
for the CNN is extracted from the corresponding ROC curve. The automated labeling is then executed with a recording, which is different to the manually created recording of Section 4.1.1. This
recording produces 9 466 additional images for the training, which are again rotated to receive
340 776 input images. From these automatically labeled images, every fifth image is used in the
first step to be merged with the manual annotated dataset. The result can be seen in Figure 5.26 together with the non semi-supervised learning approach. The number of iterations to train the CNN
is adapted accordingly to the number of training samples. The result shows, that the automated
labeled data without a false positive suppression does not improve the result of the CNN. This can
be caused by too many false positives, which are produced by the automated labeling.
To improve the semi-supervised learning approach a false positive suppression is inserted. For
this purpose, the automated labels are clustered and a threshold for the normalized velocity and
for the combination p of the mahalanobis distance and the mean velocity is extracted to remove
false positives, as described in Section 3.3. Additionally, small clusters are removed. The result is
shown in Figure 5.26 as well, where the false positive suppression with the normalized velocity has
a lower performance than the false positive suppression with the combined parameter p by using
every fifth automated labeled image. The suppression with the combined parameter has nearly
the same result as the non-semi-supervised approach. That means the produced labeled data has
nearly the same quality as the manually created data. By evaluating the changes of the output
labels, it can be observed that the labels of the semi-supervised approach create on the one side a
better output by suppressing small false positive clusters (see Figure 5.27). This can be recognized
at small curb stones. On the other side false positives are inserted (see Figure 5.28). These false
62
5. Evaluation
Figure 5.27.: Suppression of small curb stones with the semi-supervised learning approach: left: input
image, middle: output of the semi-supervised learning approach, right: output of the CNN
without semi-supervised training.
Figure 5.28.: Insertion of false positives with the SSL approach: left: input image, middle: output of
the semi-supervised learning approach, right: output of the CNN without semi-supervised
training.
positives could be the reason why the usage of all the automated labeled data in combination with
the manual created data generates no better results.
For further reducing the false positive, the threshold for the probability of the CNN to generate the
labeled images is variated, what can be recognized in Figure 5.29. There, the best result is still the
optimal threshold of Section 5.1. That means the number of false positives may be reduced with
a variation of the threshold, but the number of false negatives increases instead, what produces a
lower overall performance.
Additionally the training dataset is increased by using every automatic annotated image instead
of every fifth, to see the influence of the labeled data. The result is shown in Figure 5.30, where
the number of iterations is adapted according to the number of training data. As a result, the
automatic labeled data creates a lower performance, by introducing more of them in the training
process. That proves the suggestion, that the false positives of the automatic labeled data have a
negative effect on the training process.
63
5. Evaluation
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
non SSL
SSL threshold = 0.25
0.94
SSL threshold = 0.535
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
SSL threshold = 0.75
False Positive Rate
Figure 5.29.: ROC curves of different threshold of the CNN probability with false positive suppression.
1.00
True Positive Rate
0.99
0.98
0.97
0.96
0.95
non SSL
0.94
SSL with every fifth labeled sample
0.07
0.06
0.05
0.04
0.03
0.02
0.00
0.93
0.01
SSL with all labeled samples
False Positive Rate
Figure 5.30.: ROC curves of different number of generated training data with the semi-supervised learning
approach (threshold = 0.535).
64
5. Evaluation
1.00
0.95
True Positive Rate
0.90
0.85
0.80
0.75
0.70
0.65
CNN
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.00
0.60
0.05
baseline method
False Positive Rate
Figure 5.31.: ROC curves of the final comparison with Nuss et al. [36] as baseline method.
As a result, the semi-supervised learning approach of this thesis cannot be used for improvement
of the CNN. For this reason, the semi-supervised learning approach is not used for the final comparison.
5.4. Comparison to other Approaches
For the final comparison to the baseline method, which is described in Section 4.2, the optimized
network configuration of the segmentation evaluation (see Table 5.9) is used. The CNN is evaluated against the baseline method introduced by Nuss et al. [36] by distinguishing dynamic and
static grid cells. The results are shown in Figure 5.31, where a general improvement can be recognized. The accuracy of 72.1% for the baseline method and 97.2% for the CNN approach, results
in a relative improvement of 34.8%. That means a better performance by using the appearance of
an object while introducing the occupancy information in the decision process of movement.
The baseline method creates a decision for each grid cell independently. That results in a complete
rotation invariance of objects. The CNN internally uses filters at the convolutional layer of a
specific size, which results in a specific field of view. That can affect the rotation invariance of a
CNN. For this reason the input images are rotated for the training as described in Section 4.1.4. To
examine the rotation invariance in this section, the test input images are rotated as well, where one
test sample is shown in Figure 5.32. The figure shows that the precision, recall and the accuracy
do not essentially change over the different rotation angles. Thus, a rotation invariance is proved.
Additionally it can be recognized, that the precision and the recall have in a distance of 90 degrees
65
5. Evaluation
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
Pixel Accuracy
0.1
Precision Dynamic
350
300
250
200
150
100
0
0
50
Recall Dynamic
Angle in ◦
Figure 5.32.: Evaluation of orientation invariance by rotating one input image of the test dataset.
the same curve progression. That is caused by the filters of the convolutional layers, which are
squared. By rotating the input image, the squared filter changes the field of view caused by the
corners of the square. To reduce this and enforce the rotation invariance, a circled filter shape can
be used in future works.
66
6. Conclusion
An elementary part of environment perception in the field of autonomous driving is tracking of
objects. For this purpose a distinction between dynamic and static objects has to be achieved,
what is made a subject in this thesis. The differentiation is based on a dynamic occupancy grid
map (DOG), where for each grid cell a decision has to be made.
In contrast to other approaches, this thesis uses Convolutional Neural Networks (CNNs) to predict
the class affiliation of each grid cell. To achieve that, ground truth datasets had to be created.
For this purpose a semi-automatic labeling toolchain was implemented, where the mahalanobis
distance of each grid cell was used to create clusters of dynamic grid cells. Afterwards the generated data was corrected manually to create the training, validation and test dataset. The primary
goal was to use additional information like the occupancy information of the DOG to generate an
appearance-based decision of the movement and improve other approaches.
The CNN was initialized with the pre-trained VGG-net to shorten the training time and to use
specific filters of the pre-trained networks. Afterwards, the network was transferred to a Fully
Convolutional Network (FCN) to create a pixelwise prediction of the movement. Caused by the
transfer of the pre-trained network from video captured input to the DOG as input, the CNN had
to be optimized at the following parameters:
• type of input data,
• input rage,
• structure of the network,
• zooming into the input,
• learning rate,
• weight matrix.
With this optimization, different problems like the imbalance of the training data or the different appearance of objects were solved. Additionally a smaller network structure (Alexnet) was
evaluated to prove the real-time feasibility.
Another goal of this thesis was the orientation extraction of the detected objects to improve tracking algorithms with this information. This was done with two different approaches: the extraction
of the orientation directly through the CNN and the extraction of the orientation through the velocity information for each detected moving cluster. For the prediction of the orientation angle with
67
6. Conclusion
the CNN, the biternion representation was used, to solve the problem of periodicity. The results
presented a better overall performance of the orientation extraction through the CNN.
A general problem while training machine learning approaches is the need of labeled data. For
this reason a semi-supervised learning algorithm with false positive suppression was implemented.
This algorithm created automatic labeled data based on a trained CNN. The results of the semisupervised learning approach showed no improvement of the CNN when using additional automatic labeled data even after reducing the number of false positives.
Finally a comparison of the trained CNN with the approach of Nuss et al. [36] as a baseline method
of distinction between static and dynamic grid cells was executed. There, a relative improvement
of 34.8% for the accuracy was proven. Additionally a rotation invariance of the CNN was demonstrated, what allows a realization of this CNN in a real environment.
All in all, this thesis showed a feasible approach for grid cell based distinction between dynamic
and static obstacles, which outperforms the current baseline method, which represents the state of
the art. This could improve tracking algorithms in the field of autonomous driving, what helps to
approach the vision zero.
68
7. Outlook
This thesis presented a general improvement of the approach of Nuss et al. [36] to distinguish between dynamic and static objects, but the result can still be improved. First, the network structure
can be adapted or completely developed from scratch to create a network, which has on the one
side a good performance and on the other side a real-time applicable execution time. This could be
implemented by adapting the filter sizes according to the size of obstacles. Especially for recognition of an obstacle as a whole, compared to recognizing different parts, the filter size could be
increased within the first layers. At the same time, the number of layers could be decreased, creating a shorter execution time. Especially the last convolutional layers (before the deconvolutional
layer) could be removed, because these layers were introduced into the public network structures
like VGG-net [41] or Alexnet [26] to create a classification of the whole image. This is not needed
for a pixelwise classification. The rejection of these layers would also cause less upscaling in the
end, what creates a more detailed output.
The adaption of the network structure would cause that pre-trained network cannot be used for
initialization. For this purpose, an own pre-training has to be performed, what needs a larger
number of training samples. To solve this problem, an unsupervised pre-training like an autoencoder could be used. This would create a problem-specific pre-trained network instead of using
a pre-trained network, which was trained on another type of input data. As a result, discretization
or range adaptions would not be needed anymore.
Furthermore, by training from scratch, the types of input can be combined in different ways to
obtain a better result. For example the velocity of the recording vehicle could be provided as a
separate input, what could improve the recognition of clutter caused by a fast moving environment.
Also the restriction of three input channels would be removed, what enables a lot of possibilities
for combinations of the inputs.
Additionally, the structure can be adapted, as discussed in Section 5.4, with circular filter shapes.
That could improve the rotation invariance, which is an elementary requirement in street environment.
A general problem of training neural networks is the limited number of training samples. For this
reason a larger training dataset can be created, by recording more data on different environments,
like a highway with a high speed limit or different cities, where the appearance of the environment
changes. This should further improve the results of the CNN. Also the labeling process itself can
be improved by using further semi-automatic labeling approaches, which use the CNN for prelabeling new recordings. This would create more labeled data by a supervision approach, what
69
7. Outlook
could improve the performance of the CNN. For example a recording with non-moving objects
can be recorded, what should reduce the false positives. For this kind of recording the labeled data
can easily be generated.
Finally, the semi-supervised learning (SSL) approach can be improved to generate automatic labeled data for increasing the performance of the CNN. Especially the false positive rejection can
be improved to create labeled data with a higher quality.
70
A. List of Figures
1.1
Number of worldwide road traffic fatalities . . . . . . . . . . . . . . . . . . . . .
1
2.1
Testing vehicle within PROMETHEUS project 1990 . . . . . . . . . . . . . . .
3
2.2
Laser grid map of STANLEY, the winner of DARPA Challenge 2005 . . . . . . .
3
2.3
Winner of the DARPA Urban Challenge 2007 with several sensors on top of the roof
4
2.4
Path of Bertha, an autonomous S-Class with close-to-market sensors in 2013 . . .
4
2.5
Example of a dynamic occupancy grid map (DOG) . . . . . . . . . . . . . . . .
6
2.6
Different steps of the particle filter within the DOG . . . . . . . . . . . . . . . .
7
2.7
Inverse sensor model for a laser measurement as a measurement grid . . . . . . .
8
2.8
Image processing and reaction in a monkey brain . . . . . . . . . . . . . . . . .
9
2.9
One of the first technical realization of visual recognition of mammal brains . . .
9
2.10 One of the first Convolutional Neural Networks trained in a supervised manner to
recognize characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2.11 Principle of a convolutional layer. . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.12 Example of a max pooling layer with stride two. . . . . . . . . . . . . . . . . . .
11
2.13 Bilinear upscaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.14 One of the first Fully Convolutional Neural Networks to create a pixelwise predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.15 Parallel integration of Convolutional Neural Networks (CNNs). . . . . . . . . . .
15
2.16 Full integration of Convolutional Neural Networks (CNNs). . . . . . . . . . . . .
15
2.17 Receiver Operator Characteristic curve . . . . . . . . . . . . . . . . . . . . . . .
17
3.1
Overview of segmentation steps . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3.2
Conversion of fully connected layers to fully convolutional layers . . . . . . . . .
23
3.3
Refinement of the segmentation with the deep jet . . . . . . . . . . . . . . . . .
23
3.4
Deep jet applied to the VGG-net . . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.5
Deep jet applied to the Alexnet . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.6
Zooming into the input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
3.7
Adaption of CNNs for orientation extraction . . . . . . . . . . . . . . . . . . . .
28
3.8
Semi-supervised training loop. . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4.1
Route for recording the data . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
4.2
Overview of semi-automatic labeling toolchain. . . . . . . . . . . . . . . . . . .
32
4.3
Examples of grid map recordings . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.4
Example of the mahalanobis distance . . . . . . . . . . . . . . . . . . . . . . . .
34
71
4.5
Overfitting on a physical model . . . . . . . . . . . . . . . . . . . . . . . . . . .
140◦
35
4.6
Rotation of a training image rotated
counter clockwise . . . . . . . . . . . .
36
4.7
Utilization of the different datasets. . . . . . . . . . . . . . . . . . . . . . . . . .
37
5.1
A typical learning curve with the network configuration of Table 5.1. . . . . . . .
39
5.2
ROC curves of multiple trainings of the same network configuration. . . . . . . .
40
5.3
ROC curves of different input configurations. . . . . . . . . . . . . . . . . . . .
40
5.4
Example of the recognition of a vehicle and a wall within a DOG by using the
freespace information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
5.5
ROC curves of different range configurations. . . . . . . . . . . . . . . . . . . .
43
5.6
Histogram of the normalized velocities in x direction of the occupied grid cells
within the validation dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7
Histogram of the normalized velocities in y direction of the occupied grid cells
within the validation dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8
44
ROC curves of different structure configurations for the incremental training of the
FCN-16s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.9
44
46
ROC curves of different structure configurations for the incremental training of the
FCN-8s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.10 Best ROC curves of different incremental trainings evaluated with the FCN-32s. .
47
5.11 Example of the result of deep jet of VGG-net . . . . . . . . . . . . . . . . . . .
48
5.12 ROC curves for comparison of the Alexnet and the VGG-net. . . . . . . . . . . .
49
5.13 Example of the result of deep jet of Alexnet . . . . . . . . . . . . . . . . . . . .
50
5.14 ROC curves of different zooming configurations. . . . . . . . . . . . . . . . . .
51
5.15 Example of different crop sizes for one input image . . . . . . . . . . . . . . . .
52
5.16 ROC curves of different fixed learning rate configurations. . . . . . . . . . . . .
54
5.17 Learning curve of the fixed learning rate
3.6 × 10−5 .
. . . . . . . . . . . . . . . .
55
5.18 Learning curve of the step learning rate 3.6 × 10−4 . . . . . . . . . . . . . . . . .
55
5.19 ROC curves of different step learning rate configurations. . . . . . . . . . . . . .
56
5.20 Comparison between fixed and step learning rate policies. . . . . . . . . . . . . .
56
5.21 ROC curves of different weight factor configurations in the range [1, 100]. . . . .
58
5.22 ROC curves of different weight factor configurations in the range [100, 200]. . . .
58
5.23 Example of different weight matrices for one input image . . . . . . . . . . . . .
59
5.24 Development of the angle extraction over the iteration number. . . . . . . . . . .
60
5.25 Examples of the orientation extraction . . . . . . . . . . . . . . . . . . . . . . .
61
5.26 ROC curves of the semi supervised learning approach. . . . . . . . . . . . . . .
62
5.27 Suppression of small curb stones with the semi-supervised learning approach . .
63
5.28 Insertion of false positives with the semi-supervised learning approach . . . . . .
63
5.29 ROC curves of different threshold of the CNN probability with false positive suppression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
5.30 ROC curves of different number of generated training data with the semi-supervised
learning approach (threshold = 0.535). . . . . . . . . . . . . . . . . . . . . . .
72
64
5.31 ROC curves of the final comparison with Nuss et al. [36] as baseline method. . .
65
5.32 Evaluation of orientation invariance . . . . . . . . . . . . . . . . . . . . . . . .
66
73
B. List of Tables
4.1
Split of data with the different scenarios. . . . . . . . . . . . . . . . . . . . . . .
35
4.2
Increse of dataset size by rotation of data . . . . . . . . . . . . . . . . . . . . . .
37
5.1
Network configuration for the learning curve in Figure 5.1. . . . . . . . . . . . .
38
5.2
Network configuration for optimization of the input. . . . . . . . . . . . . . . . .
41
5.3
Network configuration for optimization of the range. . . . . . . . . . . . . . . .
42
5.4
Network configuration for optimization of the structure. . . . . . . . . . . . . . .
45
5.5
Execution time of different network structures. . . . . . . . . . . . . . . . . . .
49
5.6
Network configuration for optimization of the zooming. . . . . . . . . . . . . . .
51
5.7
Network configuration for optimization of the learning rate. . . . . . . . . . . . .
53
5.8
Network configuration for optimization of the weight matrix. . . . . . . . . . . .
57
5.9
Network configuration with the optimized parameters of Section 5.1. . . . . . . .
60
74
C. Bibliography
[1] A. Asvadi, P. Peixoto, and U. Nunes. Detection and Tracking of Moving Objects Using
2.5D Motion Grids. In IEEE International Conference on Intelligent Transportation Systems
(ITSC), 2015.
[2] R. H. Baxter, M. J. V. Leach, S. S. Mukherjee, and N. M. Robertson. An Adaptive Motion
Model for Person Tracking with Instantaneous Head-Pose Features. IEEE Signal Processing
Letters, 22(5):578–582, 2015.
[3] L. Beyer, A. Hermans, and B. Leibe. Biternion Nets: Continuous Head Pose Regression from
Discrete Training Labels. In German Conference on Pattern Recognition (GCPR), 2015.
[4] O. Chapelle, B. Schoelkopf, and A. Zien. Semi-Supervised Learning. MIT Press, Massachusetts, 2006.
[5] C. Chen and J. M. Odobez. We are not Contortionists: Coupled Adaptive Learning for Head
and Body Orientation Estimation in Surveillance Video. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2012.
[6] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille. Semantic Image
Segmentation with Deep Convolutional Nets and Fully Connected CRFs. In International
Conference on Learning Representations (ICLR), 2015.
[7] D. C. Ciresan, A. Giusti, L. M. Gambardella, and J. Schmidhuber. Deep Neural Networks
Segment Neuronal Membranes in Electron Microscopy Images. In Conference on Neural
Information Processing Systems (NIPS), 2012.
[8] P. Cloutier, C. Tibirna, B. P. Grandjean, and J. Thibault. NNFit version 2.0 User’s Manual,
1997.
[9] 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. In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
[10] M. Cordts, M. Omran, S. Ramos, T. Scharwächter, M. Enzweiler, R. Benenson, U. Franke,
S. Roth, and B. Schiele. The Cityscapes Dataset. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2015.
75
[11] C. Coue, C. Pradalier, C. Laugier, T. Fraichard, and P. Bessiere. Bayesian Occupancy Filtering for Multitarget Tracking: An Automotive Application. The International Journal of
Robotics Research, 25(1):19–30, 2006.
[12] J. Dai, K. He, and J. Sun. BoxSup: Exploiting Bounding Boxes to Supervise Convolutional
Networks for Semantic Segmentation. In IEEE International Conference on Computer Vision
(ICCV), 2015.
[13] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2009.
[14] E. Dickmanns, R. Behringer, D. Dickmanns, T. Hildebrandt, M. Maurer, F. Thomanek, and
J. Schiehlen. The Seeing Passenger Car ’VaMoRs-P’. In IEEE Intelligent Vehicles Symposium (IV), 1994.
[15] E. D. Dickmanns, B. Mysliwetz, and T. Christians. An Integrated Spatio-Temporal Approach
to Automatic Visual Guidance of Autonomous Vehicles. In IEEE Transactions on Systems,
Man and Cybernetics (SMC), 1990.
[16] A. Elfes. Using Occupancy Grids for Mobile Robot Perception and Navigation. IEEE Computer Magazine, 22(6):46–57, 1989.
[17] M. Ester, H. P. Kriegel, J. Sander, and X. Xu. A Density-Based Algorithm for Discovering
Clusters in Large Spatial Databases with Noise. In International Conference on Knowledge
Discovery and Data Mining (KDD), 1996.
[18] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL
Visual Object Classes Challenge, 2012.
[19] P. Fischer, A. Dosovitskiy, E. Ilg, P. Haeusser, C. Hazirbas, V. Golkov, P. van der Smagt,
D. Cremers, and T. Brox. FlowNet: Learning Optical Flow with Convolutional Networks. In
IEEE International Conference on Computer Vision (ICCV), 2015.
[20] U. Franke, S. Mehring, A. Suissa, and S. Hahn. The Daimler-Benz Steering Assistant: a
Spin-off from Autonomous Driving. In IEEE Intelligent Vehicles Symposium (IV), 1994.
[21] K. Fukushima. Neocognitron: A Self-organizing Neural Network Model for a Mechanism of
Pattern Recognition Unaffected by Shift in Position. Biological Cybernetics, 36(4):193–202,
1980.
[22] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Region-based Convolutional Networks for
Accurate Object Detection and Segmentation. IEEE Transactions on Pattern Analysis and
Machine Intelligence (TPAMI), 38(1):142–158, 2016.
[23] I. J. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. Book in preparation for MIT
Press available at http://www.deeplearningbook.org, 2016.
76
[24] B. Hariharan, P. Arbel, and R. Girshick. Hypercolumns for Object Segmentation and Finegrained Localization. In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2015.
[25] A. Krizhevsky. Learning Multiple Layers of Features from Tiny Images. Masterthesis, University of Toronto, 2009.
[26] A. Krizhevsky, I. Sulskever, and G. E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Conference on Neural Information Processing Systems (NIPS),
2012.
[27] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE (Journal), 86(11):2278–2324, 1998.
[28] J. Long, E. Shelhamer, and T. Darrell. Fully Convolutional Networks for Semantic Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
[29] J. Masci, U. Meier, D. Ciresan, and J. Schmidhuber. Stacked Convolutional Auto-Encoders
for Hierarchical Feature Extraction. In International Conference on Artificial Neural Networks (ICANN), 2011.
[30] O. Matan, C. J. C. Burges, Y. LeCun, and J. S. Denker. Multi-Digit Recognition Using
a Space Displacement Neural Network. In Conference on Neural Information Processing
Systems (NIPS), 1991.
[31] D. Maturana and S. Scherer. VoxNet: A 3D Convolutional Neural Network for Real-Time
Object Recognition. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2015.
[32] T. Mitchell and A. Blum. Combining Labeled and Unlabeled Data with Co-Training. In
Conference on Computational Learning Theory (COLT), 1998.
[33] T. M. Mitchell. The Role of Unlabeled Data in Supervised Learning. In International Colloquium on Cognitive Science (ICCS), 1999.
[34] A. Y. Ng, J. Ngiam, C. Y. Foo, Y. Mai, and C. Suen. Unsupervised Feature Learning and
Deep Learning Tutorial - Softmax Regression, 2013.
[35] K. Nigam, A. Mccallum, S. Thrun, and T. Mitchell. Learning to Classify Text from Labeled
and Unlabeled Documents. In Conference on Innovative Applications of Artificial Intelligence (IAAI), 1998.
[36] D. Nuss, S. Reuter, M. Thom, T. Yuan, G. Krehl, M. Maile, A. Gern, and K. Dietmayer. A
Random Finite Set Approach for Dynamic Occupancy Grid Maps with Real-Time Application. ArXiv e-prints, 2016.
77
[37] D. Nuss, T. Yuan, G. Krehl, M. Stuebler, S. Reuter, and K. Dietmayer. Fusion of Laser
and Radar Sensor Data with a Sequential Monte Carlo Bayesian Occupancy Filter. In IEEE
Intelligent Vehicles Symposium (IV), 2015.
[38] G. Papandreou, L.-C. Chen, K. Murphy, and A. L. Yuille. Weakly- and Semi-Supervised
Learning of a DCNN for Semantic Image Segmentation. IEEE International Conference on
Computer Vision (ICCV), 2015.
[39] J. Peukert. Object Generation from a Dynamic Occupancy Gridmap Considering Autocorrelation. Masterthesis, Karlsruhe Institute of Technology, 2016.
[40] H. Scudder. Probability of Error of Some Adaptive Pattern-Recognition Machines. IEEE
Transactions on Information Theory, 11(3):363–371, 1965.
[41] K. Simonyan and A. Zisserman. Very Deep Convolutional Networks for Large-Scale Image
Recognition. In International Conference on Learning Representations (ICLR), 2015.
[42] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and
A. Rabinovich. Going Deeper with Convolutions. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2015.
[43] R. Szeliski. Computer Vision: Algorithms and Applications. Springer Science & Business
Media, London, 2010.
[44] The California Department of Motor Vehicles (DMV). Autonomous Vehicle Disengagement
Reports, 2016.
[45] S. J. Thorpe and M. Fabre-Thorpe. Seeking Categories in the Brain. Science, 291(5502):260–
263, 2001.
[46] S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics (Intelligent Robotics and Autonomous Agents). MIT Press, Cambridge, 2005.
[47] S. Thrun, M. Montemerlo, H. Dahlkamp, D. Stavens, A. Aron, J. Diebel, P. Fong, J. Gale,
M. Halpenny, and G. Hoffmann. Stanley, the Robot that Won the DARPA Grand Challenge.
Journal of Field Robotics, 23(9):661–692, 2006.
[48] C. Tingvall and N. Haworth. Vision Zero - An ethical approach to safety and mobility. In
Road Safety & Traffic Enforcement, 1999.
[49] C. Urmson, C. Baker, J. Dolan, P. Rybski, B. Salesky, W. Whittaker, D. Ferguson, and
M. Darms. Autonomous Driving in Traffic: Boss and the Urban Challenge. AI Magazine,
30(2):17–28, 2009.
[50] T.-D. Vu, O. Aycard, and N. Appenrodt. Online Localization and Mapping with Moving
Object Tracking in Dynamic Outdoor Environments. In IEEE Intelligent Vehicles Symposium
(IV), 2007.
78
[51] C. Wang, C. Thorpe, and S. Thrun. Online Simultaneous Localization and Mapping with
Detection and Tracking of Moving Objects: Theory and Results from a Ground Vehicle
in Crowded Urban Areas. In IEEE International Conference on Robotics and Automation
(ICRA), 2003.
[52] Q. Wang, J. Zhang, X. Hu, and Y. Wang. Automatic Detection and Classification of Oil
Tanks in Optical Satellite Images Based on Convolutional Neural Network. In International
Conference on Image and Signal Processing (ICISP), 2016.
[53] J. Wieczner. Why self-driving cars are crashing: humans. Time Inc., 2015.
[54] R. Wolf and J. C. Platt. Postal Address Block Location Using A Convolutional Locator. In
Conference on Neural Information Processing Systems (NIPS), 1994.
[55] World Health Organization (WHO). Global Status Report On Road Safety. World Health
Organization (WHO), Villars-sous-Yens, 2015.
[56] T. Yuan, B. Duraisamy, T. Schwarz, and M. Fritzsche. Track Fusion with Incomplete Information for Automotive Smart Sensor Systems. In IEEE Radar Conference (RadarConf),
2016.
[57] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus. Deconvolutional Networks. In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2010.
[58] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, Z. Su, D. Du, C. Huang, and P. Torr.
Conditional Random Fields as Recurrent Neural Networks. In IEEE International Conference on Computer Vision (ICCV), 2015.
[59] J. Ziegler, P. Bender, M. Schreiber, H. Lategahn, T. Strauss, C. Stiller, T. Dang, U. Franke,
N. Appenrodt, C. G. Keller, E. Kaus, R. G. Herrtwich, C. Rabe, D. Pfeiffer, F. Lindner,
F. Stein, F. Erbs, M. Enzweiler, C. Knoppel, J. Hipp, M. Haueis, M. Trepte, C. Brenk,
A. Tamke, M. Ghanaat, M. Braun, A. Joos, H. Fritz, H. Mock, M. Hein, and E. Zeeb. Making
Bertha Drive - An Autonomous Journey on a Historic Route. IEEE Intelligent Transportation
Systems Magazine, 6(2):8–20, 2014.
79
| 1cs.CV
|
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
AN IMPROVED GENETIC ALGORITHM WITH A
LOCAL OPTIMIZATION STRATEGY AND AN EXTRA
MUTATION LEVEL FOR SOLVING TRAVELING
SALESMAN PROBLEM
Keivan Borna1 and Vahid Haji Hashemi2
1
Faculty of Mathematics and Computer Science, Kharazmi University, Tehran, Iran
2
Faculty of Engineering, Kharazmi University, Tehran, Iran
ABSTRACT
The Traveling salesman problem (TSP) is proved to be NP-complete in most cases. The genetic algorithm
(GA) is one of the most useful algorithms for solving this problem. In this paper a conventional GA is
compared with an improved hybrid GA in solving TSP. The improved or hybrid GA consist of
conventional GA and two local optimization strategies. The first strategy is extracting all sequential
groups including four cities of samples and changing the two central cities with each other. The second
local optimization strategy is similar to an extra mutation process. In this step with a low probability a
sample is selected. In this sample two random cities are defined and the path between these cities is
reversed. The computation results show that the proposed method also finds better paths than the
conventional GA within an acceptable computation time.
KEYWORDS
Travelling salesman problem, Genetic algorithm, mutation, complexity, NP-complete.
1. INTRODUCTION
The GA maps a set of individual objects or elements, each with a specified value, into a new set
of the population [1]. This algorithm attempts to find an approximately good solution to the
system by genetically breeding the set of individuals over a series of iterations [2, 3]. The GA
algorithm starts by choosing a random set defined as initial population of individuals (a set of
solutions) and precedes them in a generational way [4]. During each generation, individuals
with high fitness value in the current population are selected to be part of the population formed
in the next generation [5, 6]. Generally, the algorithm stops after a fixed number of generations
or when an acceptable fitness level has been reached for the last population [7]. The aim of
traveling salesman problem (TSP) is to find the shortest tour that passed each city once and
exactly once in a known map with different distances between cities. TSP has been widely
studied in the fields of artificial intelligence, graph theory, mathematics and computer science
due to its applications in real world [14, 15]. However, there are no polynomial algorithms for
the NP Complete problems [16]. Exact, approximate and very intelligent methods are
extensively designed for TSP. The exact methods waste time and memory and usually are
unreachable so local search rules are used to find an approximately good answer. This local
search rules are efficient and able to find the shortest or semi shortest path in a polynomial
computation time. The local search rules, such as the neighbourhood information [17], may be
DOI : 10.5121/ijcseit.2014.4405
47
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
trap into the local minima and do not find a good answer. Therefore, the quality of the solutions
cannot be evaluated due to its random nature and the lack of answer. The intelligent algorithms
such as GA are other methods for solving TSP. They find the best or approximate solutions
based on the evolutionary rules that differ from the local search rules. In the meantime
analytical and some of other intelligent methods can be combined with GA for enhancing its
performance. The genetic algorithm is improved by [18] with the reinforcement mutation which
relies on the reinforcement learning. The genetic algorithm, firefly method, simulated annealing,
ant colony, bee and particle swarm optimization are some of intelligent methods that can be
combined with GA for solving TSP [19]. The disadvantage of GA based methods is trap into the
local minima. Simple or traditional mutation cannot correct this problem, for this reason GA
shortest path is usually has so bigger than the best path. In this paper two local optimization
strategies try to improve GA accuracy.
The article is organized as follows. In section 2 GA algorithm phases including crossover and
mutation operator studied in TSP. Two local optimization strategies are described with details in
Section 3. Section 4 provides an overview of results on a standard TSP dataset. Section 5
highlights the main results of proposed method and indicates further research.
2. The genetic algorithm
For solving the TSP with a genetic algorithm, we need a coding, a crossover method, and a
mutation method. First of all, algorithm should generate a permutation of integer numbers that
each number refers to the ith city in the tour. In this permutation every number may only occur
exactly once and belong to interval [1 k], otherwise we do not have a complete tour [8, 9]. The
conventional GA one-point crossover method is not inappropriate to do this [10] and some other
crossover methods compatible with TSP suggested in [11, 12, 13].
2.1. Crossover
2.1.1. Partially Mapped Crossover
Partially mapped crossover (PMX) tries to keep Childs as similar as parents. To achieve this
goal, a substring is swapped look like two-point crossover and the values in all other nonconflicting situations are kept. The conflicting positions are changed with the values which
swapped to other positions.
An example:
p1 = (1 2 3 4 5 6 7 8 9)
p2 = (4 5 2 1 8 7 6 9 3)
Assume that positions 4–7 are selected for swapping. Then the two offspring’s are given as
follows if we omit the conflicting positions:
o1 = (* 2 3 | 1 8 7 6 | * 9) o2 = (* * 2 | 4 5 6 7 | 9 3)
Now we take the conflicting positions and fill in what was swapped to the other offspring. For
instance, 1 and 4 were swapped. Therefore, we have to replace the 1 in the first position of o1 by
4, and so on:
o1 = (4 2 3 1 8 7 6 5 9)
o2 = (1 8 2 4 5 6 7 9 3)
48
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
2.1.2. Order Crossover
Order crossover (OX) is based on this principle that the order of cities is important in compare
with its positions in the tour. Similar to PMX, OX swaps two aligned substrings. The
computation of the remaining substrings is done with the following way that differs from PMX
way. In order to illustrate the OX method, consider the above example (p1, p2) as for PMX.
Simply swapping two substrings and omitting all other positions, the result is:
o1 = (* * * | 1 8 7 6 | * *) o2 = (* * * | 4 5 6 7 | * *)
For computing the open positions of o2, let us write down the positions in p1, but starting from
the position after the second crossover site:
934521876
If we discard all those values which are already remain in the offspring after swapping (4, 5, 6,
and 7), the shortened result is:
93218
Now OX insert this list into o2 starting after the second crossover position and the updated o2
will be:
o2 = (2 1 8 4 5 6 7 9 3).
Repeating the above process to o1 the following result obtained:
o1 = (3 4 5 1 8 7 6 9 2)
2.1.3. Cycle Crossover
PMX and OX usually introduce cities outside the crossover sites which have not been present in
either parent. As an example, for instance, the 3 in the first position of o1 in the OX example
appears neither in p1 nor in p2. Cycle crossover (CX) tries to overcome this problem and
guarantee that every string position in any tour belongs to one of the two parents. Let us
continue with the following example:
p1 = (1 2 3 4 5 6 7 8 9) p2 = (4 1 2 8 7 6 9 3 5)
CX starts from the first position of o1:
o1 = (1 * * * * * * * *) o2 = (* * * * * * * * *)
o2 may only have a 4 in the first position, because method do not want new values to be
introduced there:
o1 = (1 * * * * * * * *) o2 = (4 * * * * * * * *)
Since the 4 is already fixed for o2 first position, CX keep it in the same position for o1 in order
to guarantee that no new positions for the 4 are introduced. We have to keep the 8 in the fourth
position of o2 for the same reason:
o1 = (1 * * 4 * * * * *) o2 = (4 * * 8 * * * * *)
49
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
This process must be repeated for all cities until end up in a value which has previously been
considered to complete a cycle:
o1 = (1 2 3 4 * * * 8 *)
o2 = (4 1 2 8 * * * 3 *)
For the second cycle, CX can start with a value from p2 and insert it into o1:
o1 = (1 2 3 4 7 * * 8 *) o2 = (4 1 2 8 5 * * 3 *)
After the same computations, the result is obtained as following:
o1 = (1 2 3 4 7 * 9 8 5)
o2 = (4 1 2 8 5 * 7 3 9)
The last cycle is a simple replication and the final offspring’s are given as follows:
o1 = (1 2 3 4 7 6 9 8 5)
o2 = (4 1 2 8 5 6 7 3 9)
2.2 Mutation
Unlike the selection and crossover, in all GA variants similar to real life, the mutation
probability is set to a small value [8]. If the mutation probability is set to a large value, the GA
is rarely converged and if it is set to a small value, the GA will easily trap into local optima [8].
In this paper, we assign pm equal to 0.05. In the proposed mutation process a random number
assign to each city of a child. If this number is lower than pm, the corresponding city is changed
with the second defined city that has a random number lower than pm. If the number of cities
lower than pm in a child was odd, the last of them is discarded. In this process the mutated cities
is swapped with each other in a sequential order. This method ensure that no duplicate occur in
samples.
3. The two local optimization strategy
3.1. First local optimization strategy
The first local optimization strategy is extracting all sequential groups including four cities of
samples and changing the two central cities with each other. The name of this strategy is four
vertices and three lines inequality, which is applied to all samples and the shortest path in each
sample, is selected. Based on only two changes in each sample in comparison with main
sample, it is not necessary to compute all distances in each process. As a heuristic only the three
distances of the selected group should be computed. If this number is lower than initial group,
the main sample is reconstruct with this new arrangement and if this number is higher than
initial group, algorithm check the next group. For a tour including N cities we have N-3 group
with four cities and only three computations is needed to compare sub samples with each other.
This means the computational burden of this step is acceptable and can be discarded in total
computation time.
3.2. Second local optimization strategy
In the ideal situation algorithm should extract and compare all sequential groups including 4 to
N-3 cities in each sample and checking all combinations of them with each other. In a large
grid this means a terrible run time. To reduce this complexity, we proposed a new mutation
scheme as second local optimization strategy. In the proposed method a random number is
assigned to each sample. If this number is lower than pm2, that is selected 0.02 in this paper, the
50
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
sample was selected for this level of optimization. After selecting a sample, two integer random
number between 2 and N-1 are generated (N is the number of cities) and all cities between these
numbers will be reversed.
If this new sample tour is lower than initial group, the main sample is reconstructed with this
new arrangement and if this number is higher than initial group, algorithm discards it. In
simulation process, we can manage run time and accuracy with the pm2 value.
The pseudo code of our proposed method
Choose initial paths
Evaluate each path's length
Determine path's average length
Repeat
Select best-ranking paths to reproduce
Mute pairs at random
Apply crossover operator
Apply level 1 mutation operator
Apply level 2 mutation operator
Evaluate each path's length
Determine path’s average lengths
Until terminating condition
(E.g. until at least one path has the desired length or enough generations have passed)
4. Simulation result
The main parameters of the GA set as follows:
Initial samples are created randomly in search space. The number of initial samples is set to
256. Only N/2 samples with lower tour distance are chosen and other samples are discarded
(Ideal selection). We used partially mapped crossover that described in 2.1.1.
The mutation process is similar to 2.2 and pm is set to 0.05. In the second local strategy the
value of pm2 is equal to 0.02. Ftv170 as a standard complicated database is selected from
TSPLIB for testing the proposed method and we will compare its efficiency with traditional GA
methods. The main parameters for comparing two methods are runtime and answer accuracy.
Based on random nature of the total process, we run the two algorithms 30 times and the
average results are shows as final results.
Figure 1 shows this result for two algorithms. The tour length in modified method is about 30%
lower than its value in conventional GA and obviously shows the improvement in algorithm.
The average runtime of proposed method in 30 times run 1000 iterations, in a Core i5 CPU with
4GB ram is about 140.5 seconds that in compare with 70.4 seconds of conventional GA is about
two times higher. Notice that the total run time is negligible and thus the time is not a critical
parameter in this comparison.
5. Conclusion
In this paper an improved hybrid GA method is used for solving TSP. The proposed method
consists of conventional GA and two local optimization strategies. The first local optimization
strategy is extracting all sequential groups including four cities of samples and changing the two
51
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
central cities with each other and is applied to all samples and the shortest path in each sample
is selected. The second local optimization strategy is similar to an extra mutation process. In this
step with a low probability a sample is selected. In this sample two random cities are defined
and the path between these cities is reversed. The computation results show that the proposed
method also find the better paths than the conventional GA within an acceptable computation
time. In the future we plan to use other meta-heuristic algorithms instead of GA and apply our
methods.
Fig. 1: The tour length vs. iteration number in proposed method and conventional GA
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
ESHELMAN, L., AND SCHAFFER, J. Real-coded genetic algorithms and interval-schemata. In Foundations
of Genetic Algorithms 2, L. D. Whitley, Ed. Morgan Kaufmann, 1993, pp. 187–202.
FOGEL, D. B. Evolving Artificial Intelligence. PhD thesis, University of California, San Diego, 1992.
FOGEL, D. B. Evolutionary Computation. IEEE Press, New York, 1995.
GOLDBERG, D. E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
HERRERA, F., LOZANO, M., AND VERDEGAY, J. L. Tackling real coded genetic algorithms: Operators
and tools for behavioral analysis. Artificial Intelligence Review 12 (1998), 265–319.
KOZA, J. R. Genetic Programming: On the Programming of Computers by Means of Natural Selection. The
MIT Press, Cambridge, MA, 1992.
MICHALEWICZ, Z. Genetic Algorithms + Data Structures = Evolution Programs, third extended ed. Springer,
Heidelberg, 1996.
J. Majumdar, A.K. Bhunia, "Genetic algorithm for asymmetric traveling salesman problem with imprecise
travel
times", Journal of Computational and Applied Mathematics, Vol.235, No.9, 2011, pp. 3063-3078.
T.D.G wiazda, Genetic algorithms reference: Crossover for single-objective numerical optimization problems,
Berlin: Springer Press, 2006.
C. Reeves, "Genetic algorithms and neighborhood search", Evolutionary Computing, AISB Workshop, Lecture
Notes in Computer Science, vol.865, 1994, pp. 115-130.
M. Srinivas and L. M. Patnaik, “Genetic Algorithms: A Survey”, IEEE journal on Computer, vol. 27, No.6,
1994, pp.11-26.
M. Črepinšek, S.H Liu and M. Mernik, “Exploration and exploitation in evolutionary algorithms: A survey”,
Computing Surveys (CSUR), ACM, Vol.45 No.3, 2013.
A. Sharma and A. Mehta, “Review Paper of Various Selection Methods in Genetic Algorithm”, International
journal of Advanced Research in Computer Science and Software Engineering, vol. 3, No.7, 2013, pp.14761479.
A. Rodríguez, R. Ruiz, "The effect of the asymmetry of road transportation networks on the traveling salesman
problem", Computers & Operations Research, Vol.39, No.7, 2012, pp. 1566-1576.
Y. Wang and J. H. Liu, "Chaotic Particle Swarm Optimization for Assembly Sequence Planning, Robotics and
Computer-Integrated Manufacturing, Vol.26, No.2, 2010, pp. 212-222.
52
International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol. 4, No.4, August 2014
[16] P. Berman,M, Karpinski, "8/7-Approximation Algorithm for (1,2)-TSP", In: SODA’06, Miami, FL, 2006, pp.
641-648.
[17] Y.H. Liu, "Diversified local search strategy under scatter search framework for the probabilistic traveling
salesman problem", European Journal of Operational Research, Vol.191, No.2, 2008, pp. 332-346.
[18] F. Liu, G. Z. Zeng, "Study of genetic algorithm with reinforcement learning to solve the TSP", Expert System
with Applications, Vol.36, No.3, 2009, pp. 6995-7001.
[19] S. M. Chen, C. Y. Chien, "Solving the traveling salesman problem based on the genetic simulated annealing
ant colony system with particle swarm optimization techniques", Expert System with Applications, Vol.38,
No.12, 2011, pp. 14439-14450.
Authors
Dr. Keivan Borna joined the Department of Computer Science at the Faculty of
Mathematics and Computer Science of Kharazmi University as an Assistant Professor in
2008. He earned his Ph.D. in Computational Commutative Algebra from the Department of
Mathematics, Statistics and Computer Science of the University of Tehran; where he
previously received an M.Sc. in the same field. He also was a visiting scholar in the
Dipartemento di Matematicha, Universita' di Genova- Italia and the Department of
Mathematik and Informatik at Essen University, Germany, from Sep. 2007 to Apr. 2008. His research
interests include Computer Algebra, Cryptography, Approximation Algorithms, and Computational
Geometry. He is the author of the "Advanced Programming in JAVA" (in Persian) and is a life member of
"Elite National Foundation of Iran".
Vahid Hajihashemi is currently a master student of Computer Engineering at Faculty of
Engineering at Kharazmi University of Tehran. His research interests include artificial
intelligence and evolutionary computations.
53
| 9cs.NE
|
Optimum Search Schemes
for Approximate String Matching Using Bidirectional FM-Index
Kiavash Kianfar1,∗ , Christopher Pockrandt2,3 , Bahman Torkamandi1 ,
Haochen Luo1 , and Knut Reinert2,3,∗∗
1
arXiv:1711.02035v2 [cs.DS] 5 Mar 2018
2
Department of Industrial and Systems Engineering, Texas A&M University
*[email protected]
Department of Computer Science and Mathematics, Freie Universität Berlin, Germany
3
Max Planck Institute for Molecular Genetics, Berlin, Germany
**[email protected]
Abstract. Finding approximate occurrences of a pattern in a text using a full-text index is a central
problem in bioinformatics and has been extensively researched. Bidirectional indices have opened new
possibilities in this regard allowing the search to start from anywhere within the pattern and extend in
both directions. In particular, use of search schemes (partitioning the pattern and searching the pieces
in certain orders with given bounds on errors) can yield significant speed-ups. However, finding optimal
search schemes is a difficult combinatorial optimization problem.
Here for the first time, we propose a mixed integer program (MIP) capable to solve this optimization
problem for Hamming distance with given number of pieces. Our experiments show that the optimal search
schemes found by our MIP significantly improve the performance of search in bidirectional FM-index upon
previous ad-hoc solutions. For example, approximate matching of 101-bp Illumina reads (with two errors)
becomes 35 times faster than standard backtracking. Moreover, despite being performed purely in the index,
the running time of search using our optimal schemes (for up to two errors) is comparable to the best stateof-the-art aligners, which benefit from combining search in index with in-text verification using dynamic
programming. As a result, we anticipate a full-fledged aligner that employs an intelligent combination
of search in the bidirectional FM-index using our optimal search schemes and in-text verification using
dynamic programming that will outperform today’s best aligners. The development of such an aligner,
called FAMOUS (Fast Approximate string Matching using OptimUm search Schemes), is ongoing as our
future work.
Keywords: FM index, bidirectional, read mapping, approximate matching, mixed integer programming, optimization
1
Introduction
Finding approximate occurrences of a string in a large text is a fundamental problem in computer
science with numerous applications. The approximate string matching (ASM) problem for Hamming
distance considered in this paper is defined as follows: Given a number of mismatches K, a string (here
referred to as a read ) of length R, and a text of length T , composed of characters from an alphabet
of size σ, find a substring of the text whose Hamming distance to the read is at most K. A similar
definition can be provided for ASM for edit distance where in addition to mismatches, insertions and
deletions, are also considered.
Solving the ASM problem has become especially important in bioinformatics due to the advances in
sequencing technology during the last years. The mainstream second generation sequencing techniques
like Illumina produce reads of length 150-250 with an error rate of about 1%, mostly substitutions
caused by the sequencing technology. Other sequencing technologies, e.g., Pacific Bioscience or Oxford
Nanopore, produce much longer reads but with a higher error rate (in the range of 15%) containing
both substitutions and insertions/deletions. A standard problem is to map the reads back to a reference
genome while taking into account the errors introduced by the sequencing technology as well as those
caused by biological variation, such as SNPs or small structural variations. Such a problem is almost
always modeled as the ASM problem for Hamming or edit distance.
There are two main algorithmic strategies to address the ASM problem for large input sizes (in
number of reads and size of the text): filtering and indexing. In this work, we focus on the indexing
approach. Here, the main idea is to preprocess the reference sequence, the set of reads, or both, in a
more intricate way. Such preprocessing into full-text string indices has the benefit that we usually do
not have to scan the whole reference, but can conduct queries much faster at the expense of larger
memory consumption. String indices that are currently used are the suffix array [12] and enhanced
suffix array [1], and affix arrays [11,18], as well as the FM-index [3], a data structure based on the
Burrows-Wheeler Transform (BWT) [2] and some auxiliary tables. For an in-depth discussion see
[16]. Such indices are used to compute exact matches between a query and a text as a subroutine in
backtracking approaches. For the ASM problem for Hamming or edit distance, the existing algorithms
all have exponential complexity in K (e.g. [5,19]), and are hence only suited for small K.
Lam et al. [7] introduced bidirectional FM indices to speed up ASM for Hamming distance. For
the cases K = 1 and 2, they partitioned the read into K + 1 equal pieces, and argued that performing
approximate matching on a certain combination of these pieces in a bidirectional index amounts to
faster approximate matching of the whole read. This combination is such that all possible mismatch
patterns, i.e., all possible distributions of K mismatches among the pieces, are covered. The main idea
behind improved speed is that a bidirectional index not only can start the search from the beginning
(or end) of the read, but also from the beginning (or end) of any of the pieces. Therefore, we can start
the search from a middle piece and then expand it to the left or right into adjacent pieces in any order
we like. By choosing multiple appropriate orderings of pieces for this purpose, we can perform a much
faster ASM compared to a unidirectional search because we can enforce exact or near-exact searches
on the first pieces in the partition, significantly reducing the number of backtrackings, while using
different orderings of pieces to ensure all possible mismatch patterns are still covered.
Kucherov et al. [6] formalized and generalized this idea by defining the concept of search schemes.
Assume a read can be partitioned into a given number of pieces, denoted by P (not necessarily equal to
K + 1). The pieces are indexed from left to right. A search scheme S = {(πs , Ls , Us ), s = 1, . . . , S} is a
collection of S searches, where each search s is designated by a triplet (πs , Ls , Us ). πs is a permutation
of 1, . . . , P and denotes the order in which the pieces of the partition are searched in search s. If
πs,i = j, then piece j is searched at position i in the order (shortly referred to as iteration i in this
paper). Due to the way a bidirectional index works, the permutation πs must satisfy the so-called
connectivity condition, i.e, a piece j can appear at iteration i > 1 in the permutation only if at least
one of pieces j − 1 or j + 1 have appeared at an iteration before i. Ls and Us each are strings of P
numbers. Ls,i , is the lower bound on the cumulative number of mismatches allowed at iteration i of
search s, and Us,i is the upper bound on this value.
Having this formal framework, the answer to the Optimal Search Scheme problem, defined as
follows, can potentially have a great impact on improving the running time of ASM using a Bidirectional
index (ASM-B).
Optimal Search Scheme Problem: What is the search scheme that minimizes the number of steps
in ASM-B while ensuring all possible mismatch patterns are covered?
It turns out that this is a very difficult combinatorial optimization problem due to several reasons:
There are a large number of attributes that define a solution (including S, P , size of each piece, and
(πs , Ls , Us ) for each search) with a large number of possibilities for each attribute; the solution must
satisfy complex combinatorial constraints; and, calculating the objective function, i.e., number of steps
in the ASM-B algorithm, for a given solution is complicated.
Kucherov et al. [6] presented some interesting results contributing initial insight into this key
problem. More specifically, they assumed the number of steps in the ASM-B algorithm with a given
search scheme, is a constant factor of the (weighted) total number of substrings enumerated by the
algorithm in all searches. Assuming that a randomly generated read is to be matched to a randomly
generated text, they presented a method to calculate this objective function for a given search scheme.
They then showed that unequal pieces in the partition can potentially improve the objective function
compared to equal pieces, and presented a dynamic programming (DP) algorithm that for a single
1
prespecified search, with given P and (π, L, U ), finds the optimal sizes of pieces assuming that we only
calculate the objective function as the total number of substrings up to a limited length (justified by
total randomness of the read and the text); see [6] for more details. In fact, the superiority of this DP
over explicit enumeration is only due to this assumption. Nevertheless, this DP is very inefficient, and
most importantly, it only finds the optimal piece sizes for a prespecified search. In other words, it does
not address the problem of finding an optimal search scheme which calls for determining S and all
attributes of each search in the search scheme, and ensuring that they cover all mismatch patterns.
Kucherov et al. [6] also presented solutions for another limited problem, i.e., lexicographically
minimizing the lexicographically maximal U string (critical U string) in a search scheme, only for
P = K + 1 or K + 2 and assuming that the L strings for all searches contain only zeros. The usefulness
of these solutions is justified by the high probability that the search with the critical U string has the
largest share in the objective function; see [6] for details. Again from the perspective of finding an
optimal search scheme, this result has similar limitations. Only one of the attributes (U ) of one of the
searches for two specific values of P are optimized by fixing all L strings, which is far from designing
a globally optimal search scheme as defined above. Consequently, in their computational experiments,
Kucherov et al. [6] use a greedy algorithm based on this limited result to construct search schemes
with unknown quality and only optimize the piece sizes for these schemes using their DP.
In this paper, for the first time, we propose a method to solve the optimal search scheme problem
for ASM-B with Hamming distance, for any given P and equal-size pieces. Our method is based on a
novel and powerful mixed integer linear program (MIP) that gets K, R, P , and an upper bound on S,
denoted by S, as input, and provides, as its solution, all the attributes of the exact optimal search
scheme (MIP methodology for optimization has been addressed in many references such as [13,20]).
We present our MIP and provide the results of our computational study on the characteristics of its
optimal solution and its running time, for different values of its input parameters.
Next we conduct an experiment using a bidirectional search (for Hamming and edit distance)
that performs the search based on the optimal search schemes we have obtained from our MIP. This
bidirectional index search is implemented in SeqAn [15] and uses a recent fast implementation of
bidirectional indices [14] based on EPR dictionaries. We show that, for practical ranges of various
input parameters, the number of substrings for the optimal search schemes found by our MIP can
reduce to as small as half the number of substrings in the unidirectional complete backtracking. In
another experiment, we search for all occurrences of Illumina reads in the human genome completely
in the bidirectional index using our optimal search schemes and show that our time is is up to 3, 14,
and 35 times faster than the standard backtracking search in the index for K = 1, 2, and 3 Hamming
distance errors, respectively. Although our MIP finds the optimal search schemes for Hamming distance,
when we used its optimal schemes with the edit distance, we got similar improvements.
The drastic improvement over standard backtracking gained by using our optimal search schemes
for bidirectional search in index suggests that the performance of read mappers that utilize an
index can be significantly improved. To gauge this potential, we even challenged our optimal search
schemes by performing a pure index-based search using them and comparing the performance with
the full-fledged state-of-the-art aligners that benefit from using a combination of search in index and
verification in text using dynamic programming. We noticed that our optimal schemes are so effective
that although they are employed in the pure index-based search, the resulting times for K = 1, 2,
and 3, are competitive with the state-of-the-art aligners which use a combination of index search
and verification in text, in strata mode. In the all-mapping mode, the results are better for K = 1,
and competitive for K = 2. These observations suggest that a full-fledged aligner that employs an
intelligent combination of search in the bidirectional FM-index using our optimal search schemes
and verification in text using dynamic programming can outperform today’s best approximate read
2
mappers. The development of such an aligner, called FAMOUS (Fast Approximate string Matching
using OptimUm search Schemes) is ongoing in our group as future work.
We will introduce our MIP for solving the optimal search scheme problem in Section 2 and discuss
our computational studies on solving this MIP. Then in Section 3, we present the computational
impact of using the optimal search schemes obtained from the MIP on solving ASM-B using various
realistic scenarios. In Section 4, we will motivate the potential impact of our optimal search schemes in
developing a full-fledged read mapper and mention FAMOUS as our future work. We will conclude in
Section 5 by summarizing our contributions and raising several open problems and possible extensions.
2
Solving Optimal Search Scheme Problem using MIP
In this section, we provide our MIP-based methodology for finding optimal search schemes after
presenting some preliminaries. We will then follow with a brief computational report on solving our
MIP in order to find the optimal search schemes, including its optimal objective value as a function of
its input parameters, its solution running time, and its convergence rate to optimal solution).
Our MIP is for Hamming distance, but as mentioned before, based on our computational experiments (Section 4), its optimal schemes for Hamming distance are very good (but not necessarily
optimal) search schemes for the edit distance as well.
2.1
Preliminaries
Our MIP presented in Section 2.2 will solve the optimal search scheme problem assuming P is given
as an input (is not a decision variable in optimization) and all P pieces of the partition are equal
in length, i.e., R = mP , where m denotes the length of any piece. Note that these assumptions
pose no practical restrictions. Solving problems which include P as a decision variable and allow
unequal pieces is part of our future research plan. Given the upper bound on Hamming distance K
(maximum number of mismatches) as an input, a mismatch pattern is a particular distribution of h
mismatches among the P pieces, for any h ≤ K. Specifically, the mismatch pattern
PP q is a string of
P integers aq,1 . . . aq,P such that aq,j ∈ {0, . . . , min{m, K}} for j = 1, . . . , P , and j=1 aq,j = h. For
given KPand P , we denote
the set of all possible mismatch patterns by M. Note that if K ≤ m then
h+P −1
|M| = K
.
Given
a search s = (πs , Ls , Us ), a mismatch pattern q is said to be covered
h=0
h
P
by s if at every iteration i = 1, . . . , P of s, Ls,i ≤ it=1 aq,πs,t ≤ Us,i , i.e., the cumulative number of
mismatches up to iteration i is between the allowed lower and upper bounds of search s. A search
scheme S is feasible if and only if every mismatch pattern in M is covered by at least one search in S.
A search scheme can be visualized by representing each of its searches as a trie that captures all
substrings enumerated by the search. Each edge at a level of the trie corresponds to a character of the
alphabet at that level of search. A vertical edge represents a match, and a diagonal edge represents
a mismatch. Fig. 1(a) shows the tries associated with the search scheme presented by Lam et al.
[7] for K = 2 and P = 3, SLam , applied on the six-character read “abbaaa” from alphabet {a, b}
(note that the tries are slightly different from the ones given in [6], which contained a small error).
Fig. 1(b) shows a search scheme with a single unidirectional search (complete backtracking), SUni ,
for the same problem, and Fig. 1(c) shows the optimal search scheme, SOpt , found by our MIP, for
the same problem. Each one of the three schemes in Fig. 1 covers all 10 mismatch patterns, namely
{000, 001, 010, 100, 011, 101, 110, 002, 020, 200}. Interestingly, the three searches sf , sb , sbi in SOpt cover
the mismatch patterns {002, 011}, {000, 010, 100, 110, 020, 200}, and {001, 101}, respectively, which is
indeed a partition of all mismatch patterns (see open problems in Section 5), whereas in SLam , the
searches sf and sb both cover 000 and 010 redundantly.
Following the method of Kucherov et al. [6], we define the performance of a search scheme as the
number of forward and backward steps taken by the ASM-B algorithm, which is equal to the total
3
a
P1
b
P2
P1
P1
P3
P2
P2
P3
P2
P2
P2
P3
P1
P1
P3
P3
P1
P1
a
b
P2
a
P3
b
a
b
a
b
a
a
b a
b a
a
b
a
b a a
b a
a
b
P3
sf=(123,000,022)
sb=(321,000,012)
sf =(123,000,222)
sbi=(231,001,012)
(a)
sf=(123,002,012)
(b)
sb=(321,000,022)
sbi=(231,011,012)
(c)
Fig. 1. (a) The search of Lam et al. [7] as described by Kucherov [6] for K = 2 and P = 3, i.e., SLam = {sf =
(123, 000, 022), sb = (321, 000, 012), sbi = (231, 001, 012)}, shown for the read “abbaaa” from the alphabet {a,b}, i.e.,
R = 6 and σ = 2. The read is partitioned into P1 =ab, P2 =ba, and P3 =aa. Partition borders are shown by horizontal
lines. A vertical and a diagonal edge represent a match and a mismatch, respectively. Edge labels are only shown
for sf for a cleaner picture. The search corresponding to each trie is designated underneath it by its (π, L, U ). The
number of edges in SLam tries is 71. (b) The unidirectional search scheme SUni = {sf = (123, 000, 222)} for the same
problem. The number of edges in SUni is 62, i.e., for this particular problem, in which R is very small, SLam enumerates
even more substrings than SUni (if all possible substrings are present in the text). Of course, if R gets larger, the
situation is reversed, making SLam more efficient than SUni as reported by Lam et al. [7]. (c) The optimal search scheme
SOpt = {sf = (123, 002, 012), sb = (321, 000, 022), sbi = (231, 011, 012)} for the same problem, found by our MIP. The
total number of edges in SOpt (optimal number of edges) is 59, which is less than that of SUni , and significantly less than
that of SLam . As shown in Section 4, for bigger problems, the reduction in the total number of edges of the optimal
search scheme found by our MIP compared to the unidirectional search is much more significant (up to 50%).
number of substrings enumerated by all searches in the scheme. We assume a single step of forward or
backward search in the bidirectional index takes the same amount of time. The tries of any search
scheme in Fig. 1 contain all possible substrings of length R. The number of substrings in each trie is
equal to the number of edges (or total number of non-root nodes). If the text contains all substrings
of length R, the search enumerates all substrings in the tries; hence, the performance of the search
scheme can be measured by the total number of edges in the search scheme. Otherwise, only a subset
of the substrings in the tries will be enumerated depending on whether they occur in the text or not.
To address the performance measure in this latter case, Kucherov et al. [6] assumed the read and the
text are randomly and independently drawn from the alphabet according to a uniform distribution,
and hence, calculated the expected number of substrings enumerated by the scheme as the sum, over
all non-root nodes of the tries, of the probability that the corresponding substring appears in the text.
As a result, they presented a weighted sum of number of edges as the measure of performance. Due to
the assumption of complete randomness and independence of the read and the text, they show that
the weights of the edges at levels lower than dlogσ T e + cσ of the tries, where cσ is that ((σ − 1)/σ)cσ
is sufficiently small, are almost zero meaning that they can be dropped from the weighted summation.
For the main application of our interest, i.e. ASM of DNA sequence reads to reference genomes, the
assumption of randomness and independence of the read and the text is far from reality. Calculating the
expected number of substrings enumerated by a scheme calls for significant more study on determining
probabilities that DNA sequence reads of particular length from a sample occur in the reference
genomes. As currently there is no trivial answer to this problem, in this paper, we use the same
performance measure of total number of edges in the tries of the search scheme even for the case
where not all substrings occur in the text. Of course, our MIP can be easily modified to incorporate
any other weighting scenario which might be proposed in the future.
Adapting the method from [6], the total number of edges in the search scheme is calculated by
XS
s=1
XR XK
l=1
d=0
ns,l,d ,
(1)
where ns,l,d is defined as the number of edges at level l of the trie of search s that end at nodes
corresponding to substrings with d cumulative mismatches up to that level. The value of ns,l,d can be
4
calculated using the following recursive equation, which is an adaptation of the formula in [6]:
for l ≥ 1 and Ls,l ≤ d ≤ Us,l ,
ns,l,d = ns,l−1,d + (σ − 1)ns,l−1,d−1
(2)
where, by definition, ns,0,0 = 1, ns,0,−1 = 0 and ns,0,d = 0, for d ≥ 1, s = 1, . . . , S, and Ls,l and
Us,l denote the smallest and largest cumulative number of mismatches that can occur at level l of
the trie of search s, respectively, calculated as Ls,l = max{Ls,dl/me−1 , Ls,dl/me − mdl/me + l} and
Us,l = min{Us,dl/me , Us,l−1 + 1}. Here dl/me, the smallest integer greater than or equal to l/m, would
be the index of the iteration in which level l falls, and by definition, Ls,0 = Us,0 = 0, for s = 1, . . . , S.
For example, for search sbi of SOpt , we have Lsbi = (0, 0, 0, 1, 1, 1) and Usbi = (0, 0, 1, 1, 2, 2).
2.2
MIP Formulation of Optimal Search Scheme Problem
Our MIP formulation, presented below, solves the optimal search scheme problem assuming P is given
as an input and pieces are all equal in length. More specifically, for given K, R, P , and S, this MIP
finds the search scheme with minimum total number of edges among all feasible search schemes that
have at most S searches. The optimal solution to the MIP provides the (π, L, U ) of all searches in
the optimal search scheme. The objective value of this optimal solution provides the minimum total
number of edges (substrings) achievable among all feasible search schemes.
min
XS
XR XK
s=1
l=1
d=0
ns,l,d
(3)
subject to
XP
xs,i,j = 1
for all s and j
(4a)
xs,i,j = 1
for all s and i
(4b)
for all s, i = 2, . . . , P − 1,
j = 1, . . . , P + 1
(5a)
for all s, i = 2, . . . , P − 1
(5b)
d − (Ls,dl/me − mdl/me + l) + 1 ≤ (R + 1)z s,l,d
for all s, l, and d
(6a)
Us,dl/me + 1 − d ≤ (K + 1)z s,l,d
l
(σ − 1)d (z s,l,d + z s,l,d − 2) ≤ ns,l,d − ns,l−1,d − (σ − 1)ns,l−1,d−1
d
for all s, l, and d
(6b)
for all s, l, and d
(6c)
i=1
XP
j=1
Xi
xs,h,j −
h=1
XP +1
j=1
Xi
h=1
−
xs,h,j−1 = t+
s,i,j − ts,i,j
−
(t+
s,i,j + ts,i,j ) = 2
Ls,i ≤ Ls,i+1
for all s, and i = 1, . . . , P − 1
(7a)
Us,i ≤ Us,i+1
for all s, and i = 1, . . . , P − 1
(7b)
for all q, s, and i
(8a)
for all q
(8b)
for all q, s, i, j, l, and d
(9a)
for all s and i
(9b)
for all q, s, i, j, l, and d
(9c)
Ls,i + K(λq,s − 1) ≤
XS
s=1
Xi
h=1
XP
j=1
aq,j xs,h,j ≤ Us,i + K(1 − λq,s )
λq,s ≥ 1
ns,l,d ≥ 0
Ls,i , Us,i ≥ 0
Integer
−
xs,i,j , λq,s , z s,l,d , z s,l,d , t+
s,i,j , ts,i,j
∈ {0, 1}
The objective function (3) minimizes the total number of edges as calculated by (1) with ns,l,d as
defined before. The binary variables xs,i,j capture the assignment of pieces to iterations, i.e., xs,i,j = 1
if piece j is searched at iteration i of search s, and xs,i,j = 0 otherwise. We define xs,i,0 = xs,i,P +1 = 0
to simplify presentation of constraints. At optimality, these variables determine the πs values for the
optimal search scheme. Constraints (4a) and (4b) make sure that for any search s, only one piece is
assigned to an iteration and only one iteration is assigned to a piece.
5
Constraints (5a)-(5b) ensure the connectivity of the pieces and are in fact linearization of the
−
following constraint using auxiliary binary variables t+
s,i,j and ts,i,j :
XP
Xi
j=1
h=1
xs,h,j −
Xi
h=1
xs,h,j−1 = 2
for all s and i = 2, . . . , P − 1,
(10)
Pi
which is one way to enforce connectivity of pieces. The term h=1 xs,h,j will have a binary value which
denotes
whether or not piece j has been searched at any of iterations 1 to i of search s. The term
Pi
x
piece j − 1. If at any iteration all searched pieces form a
h=1 s,h,j−1 captures the same notion for
Pi
P
connected block on the read, the value of h=1 xs,h,j − ih=1 xs,h,j−1 will be equal to 1 only for one
j, −1 for another j, and 0 for all other j’s, which is ensured by (10), and hence its linearization.
Constraints (6a)-(6c) enforce calculation of ns,l,d based on the recursive equation (2) with the
help of binary variables z s,l,d and z s,l,d . Due to (6a), if d ≥ Ls,dl/me − mdl/me + l, then z s,l,d = 1,
and due to (6b), if d ≤ Us,dl/me , then z s,l,d = 1. Calculation of equation (2) is then enforced by
(6c). When z s,l,d = z s,l,d = 1, (6c) reduces to nsld − ns,l−1,d − (σ − 1)ns,l−1,d−1 ≥ 0, which implies
nsld − ns,l−1,d − (σ − 1)ns,l−1,d−1 = 0 since the objective function is to be minimized. If any of z s,l,d or
z s,l,d is equal to 0, (6c) does not enforce anything as − dl (σ − 1)d is a lower bound on the right-hand
side of (6c). Constraints (7a)-(7b) ensure Ls,i and Us,i are non-decreasing as they are cumulative values.
Constraints (8a)-(8b) ensure feasibility of the search scheme. λq,s is a binary variable designating
whether or not mismatch pattern q is covered by search s. Constraint (8a) forces λq,s = 0 if search s
does not cover mismatch pattern q and constraint (8b) ensures every mismatch pattern q is covered
by at least one search, for q = 1, . . . , |M|.
Constraints (4a)-(9c) are enough to formulate the MIP; however, we have noticed that imposing
the additional constraints
x1P P = 1
XS Xj−1
XPt=s
−i+1
j=1
k=1
xsij
(11a)
xt,1,k ≤ (S − s + 1)(1 − xs,1,j )
XP
+
xsij = 1
j=i
for all s and j = 2, . . . , P
(11b)
for all s and i ≥ dP/2e + 1
(12)
strengthens the formulation while preserving at least one optimal solution, resulting in faster solution
time for the MIP. Constraints (11a) and (11b) eliminate some symmetry in the solution space. For
every search scheme, there is an equivalent search scheme obtained by reversing all πs , s = 1, . . . , S.
Constraint (11a) eliminates one of these two equivalent solutions in each pair by forcing piece P to be
assigned to iteration P in the first search, eliminating the solutions in which piece 1 is assigned to
iteration P . For any search scheme, another equivalent search scheme can be obtained by permuting
the indices of searches within the scheme. Existence of only one of the search schemes obtained by
this index permutation in the feasible solution set is enough. This can be achieved by sorting (in
ascending order) the searches based on the piece assigned to their first iteration. This is done by
constraint (11b), which does not allow pieces 1, . . . , j − 1 to be assigned to the first iteration of searches
s, . . . , S if piece j is assigned to the first iteration of search s. In addition to symmetry elimination,
notice that the connectivity condition of pieces implies that the piece assigned to iteration P is either
piece 1 or piece P , and in general, the piece assigned to iteration i ≥ dP/2e + 1 is one of pieces
1, . . . , P − i + 1, i, . . . , P . Constraint (12) enforces this property, which strengthens the formulation,
and according to our computational tests, reduces the running time of the MIP.
Remark 1. A powerful feature of our MIP is that S is an upper bound on the number of searches, i.e.,
our MIP finds the optimal search scheme among all schemes with at most S searches. In our MIP,
variables are defined for S searches, and if the optimal search scheme has S ∗ < S searches, our MIP
generates S − S ∗ empty searches, i.e. searches in which Ls,i > Us,i for some i.
t
u
2.3 Solving MIP
We used CPLEX 12.7.1 solver [4] to solve our MIP by implementing the code in C++ using CPLEX
Callable Library. All instances were run over four 28-core nodes (2.4 GHz Intel Broadwell) with 64GB
6
of memory per node. We ran our MIP solver for instances generated for a broad range of parameters
K, R, S, and P and gave each instance a 3-hour time limit. Fig. 2(a) in the Appendix is a small
representative of our results. It shows the optimal objective value (total number of edges) for R = 100,
K = 1, . . . , 4, P = 5, 6, and S = 1, . . . , 5. If the problem is not solved to optimality in 3 hours, the
best solution found within this time limit is shown. The optimal objective value does not show a
consistent change pattern in terms of change in P ; however, as expected, it increases as K increases,
as R increases (not shown), and as S decreases. In all instances, the optimal objective value shows a
sharp drop from S = 1 to S = 2, then a modest drop to S = 3, and negligible change beyond S = 3,
generating empty searches in many cases. Therefore, as long as S = 5, it is advisable to use S = 3
instead if we would like to reduce the MIP running time and still find an optimal or near-optimal
solution for S = 5. We also noticed that the optimal search scheme obtained by our MIP is not
sensitive to the value of R (see open problems in Section 5). Therefore, when R is large, it is advisable
to solve the MIP for a much smaller reasonable value of R, e.g., R = KP , in order to get a solution
that is most probably optimal for the large R in a much shorter amount of time.
Using the MIP formulation, we were able to solve considerable size problems to optimality. For
instance, we were able to solve a problem with K = 4, R = 100, P = 3, and S = 3 to optimality in 5802
seconds. However, more complicated cases reached the time limit of 3 hours without proving solution
optimality. Consequently, it is important to investigate the rate of convergence of the solutions found
during execution of MIP to the optimal solution. Fig. 2(b) illustrates the ratio of the best solutions
found by MIP during its execution to the final optimal objective value plotted against running time
for some instances which reached optimality. We observe that in all cases, within 0.1% to 1% of the
total running time, the MIP finds a solution which is finally proved to be optimal or very close to
the optimal after MIP execution is complete. In other words, the MIP solver finds optimal or near
optimal solutions very early on and spends the rest of its time ensuring that no better solution exists.
This can be partly due to the remaining symmetry in the solution space. Nevertheless, from practical
perspective, this is an attractive property because, when the input parameters are much larger, we
can run the MIP for a short time and find solutions which are most probably optimal or near-optimal.
3
Search-in-Index Computational Gains Achieved by Optimum Schemes
In this section, we present the computational advantages achieved by using our optimal search schemes
in ASM-B (ASM performed completely in bidirectional FM-index).
While our optimal search schemes can be found for any alphabet size and read length, we chose to
concentrate on parameter values relevant to standard sequencing reads, e.g., Illumina reads. In Table
1, for a number of relevant parameter values, we have shown how the total number of edges using the
optimal search schemes found by our MIP is reduced compared to the unidirectional backtracking
scheme. It can be seen that the reduction is between 42% and 49%. Also, for K = 2 and K = 3, the
optimal search scheme with P = K + 2 has fewer edges than the one with P = K + 1.
Although the reduction factors in total number of edges obtained by our optimal search schemes
in Table 1 are very significant in themselves, due to the stochastic nature of occurrence of errors
in sequencing reads and occurrence of approximate matches in the reference genome, the real-case
ASM speed-up factors achieved by these optimal search schemes compared to backtracking can be
yet much more significant. To gain insight into this speed-up, we performed an experiment searching
for all approximate matches (for K = 1, 2, and 3) of 100, 000 real Illumina reads of length R = 101
(SRA accession number ERX1959065) in the human genome hg38 and compared the running time of
ASM-B performed with optimal search schemes obtained by our MIP for P = K + 1 and P = K + 2
to that of unidirectional backtracking for Hamming and edit distance. All of our tests were conducted
on Debian GNU/Linux 7.1 with Intel Xeon E5-2667V2 CPUs at fixed frequency of 3.3 GHz to prevent
dynamic overclocking effects. All data was stored on tmpfs, a virtual file system in main memory to
7
Table 1. Total number of edges in the optimal search schemes found by our MIP for K = 1, 2, 3 and P = K + 1,
P = K + 2 and P = K + 3 compared to full backtracking. The factor column shows the ratio of total number of edges in
each scheme to that in backtracking. The optimal search schemes are listed in the Appendix in Table 3.
Distance
Hamming
Edit
Search Scheme
Backtracking
Optimal (P = K + 1)
Optimal (P = K + 2)
Optimal (P = K + 3)
Backtracking
Optimal (P = K + 1)
Optimal (P = K + 2)
Optimal (P = K + 3)
K=1
Edges
Factor
15,554
1.00
8,004
0.51
8,922
0.57
8,004
0.51
41,208
1.00
20,908
0.51
23,356
0.57
20,908
0.51
K=2
Edges
Factor
1,560,854
1.00
892,769
0.57
854,303
0.55
835,213
0.54
11,154,036
1.00
6,315,779
0.57
6,025,907
0.54
5,892,667
0.53
K=3
Edges
Factor
116,299,379
1.00
67,888,328
0.58
65,116,676
0.56
64,060,718
0.55
2,264,515,748
1.00
1,299,709,022
0.57
1,246,126,103
0.55
1,226,903,544
0.54
K=4
Edges
6,862,924,649
4,064,852,156
3,916,700,994
3,887,857,820
367,846,294,116
213,296,122,595
205,509,484,572
203,270,363,390
Factor
1.00
0.59
0.57
0.57
1.00
0.58
0.56
0.55
prevent loading data just on demand during the search and thus affecting the speed of the search by
I/O operations. All tools were run with a single thread to make the results comparable. The results
are shown in Table 2. We can see that for both Hamming and edit distance, employing our optimal
search schemes, is much faster than backtracking, verifying our expectation. The respective speed-ups
for K = 1, 2, and 3 are 3.1, 14.3, and 35.2 for Hamming distance, and 4.1, 11.1, and 21.3 for edit
distance, much more significant than reduction in the total number of edges reported in Table 1.
Edit
Hamm. Dist.
Table 2. Running time comparison of searching all approximate matches of 100, 000 Illumina reads (R = 101) using
optimal bidirectional scheme with P = K + 1 and P = K + 2 versus backtracking for Hamming and edit distance. The
factor column is the speed-up ratio versus backtracking in each category.
Search Tool
Backtracking
Optimal-scheme
Optimal-scheme
Backtracking
Optimal-scheme
Optimal-scheme
bidirect. (P = K + 1)
bidirect. (P = K + 2)
bidirect. (P = K + 1)
bidirect. (P = K + 2)
K=1
Time
Factor
22.80s
1.00
7.73s
2.95
7.39s
3.09
43.59s
1.00
11.21s
3.89
10.66s
4.09
K=2
Time
Factor
269.24s
1.00
19.78s
13.61
18.81s
14.31
1245.70s
1.00
120.70s
10.32
112.23s
11.10
K=3
Time
Factor
2417.06s
1.00
74.62s
32.39
68.69s
35.19
27889.40s
1.00
1338.61s
20.83
1307.23s
21.33
We also compared the optimal search schemes with Hamming distance against Bowtie1 [9] by
searching for all alignments with at most K mismatches (-v <K> -a). It turns out that our search
schemes are significantly faster (Bowtie1 takes 23, 68 and 180 seconds for 1, 2 and 3 errors).
4
Towards a full-fledged aligner
Due to the exponential complexity of ASM using FM-index in terms of K, the state-of-the-art
aligners do not perform ASM completely in index but rather use a combination of search in the index
and verification in text using dynamic programming (DP). Pure index-based search using standard
backtracking is very slow for larger values of K. However, the drastic improvement, over standard
backtracking, gained by using our optimal search schemes for search in an bidirectional index, as
observed in Section 3, suggests a potential for significant improvement in the performance of read
mappers that utilize index-based search. To acquire a sense of this potential, we decided to even
challenge our optimal search schemes by using them in a pure index-based search and compare the
results against the full-fledged state-of-the-art aligners that have the advantage of using a combination
of index-based search and in-text verification using DP.
We performed our first set of comparisons with BWA [10], Yara [17], as well as an available
implementation of the 01∗ 0-filter scheme combined with dynamic programming (named Bwolo) [19] in
the all-mapping mode. We did these comparisons for the edit distance only as all these tools work
for edit distance. As before, 100, 000 Illumina reads of length R = 101 were aligned. BWA was run
8
with the options -N -n <K>, and Yara was run with the options -e <K> -s <K> -y full -t 1.
We note that Bowtie2 [8] is not designed with all-mapping in mind (for our data set, it did not
terminate in 3 hours with default configuration and -a option). Moreover, imposing an all-mapping
with maximum K errors in Bowtie2 in a way that its results are comparable to other tools is difficult.
Bowtie2 settings used in [19] do not enforce this, and nonetheless, led to a very long running time for
our data set (did not terminate in 13 hours). Consequently, we did not use Bowtie2 in this study.
Remark 2. We would like to note that while BWA and Yara are standard read mapping tools for
Illumina reads, Bwolo is interesting in the context of our approach. Vroland et al. [19] presented this
fast method for searching in an index by partitioning the read into K + 2 pieces for K errors and
then exploiting the fact that the K + 2 pieces must contain the mismatch pattern 01∗ 0. Their method
searches for an occurrence of the 01∗ 0 mismatch pattern in the read and then verifies the remaining
pieces of the read partially in the index (they use a unidirectional index) and partially in the text via
dynamic programming. Interestingly, their 01∗ 0 method can be formulated as a search scheme. For
example, for K = 2, and hence P = 4, their method can be expressed as the following search scheme:
S01∗ 0 = {(4321, 0000, 0122), (3214, 0000, 0122), (2134, 0000, 0022)}. The optimal search scheme found
by our MIP for these parameters is SOpt = {(4321, 0002, 0122), (3214, 0000, 0112), (2134, 0011, 0022)},
which is quite similar, but has yet fewer edges. Indeed, we verified the practical superiority of SOpt to
S01∗ 0 by searching 100, 000 real Illumina reads in the human genome with K = 2 Hamming and edit
distance and noticed that SOpt took less time than S01∗ 0 (18.8s vs. 22.0s for Hamming, and 112s vs.
160s for edit distance, respectively).
t
u
Back to our first set of comparisons, for K = 1, BWA, Bwolo, and Yara, took 12.67s, 20.79s, and
11.53s, respectively. This is while, according to Table 2, the pure index-based ASM using our optimal
schemes for P = K + 1 = 2 and P = K + 2 = 3, took 11.21s, and 10.66s, respectively, i.e., our optimal
schemes are so effective that although the search is performed completely in index, the resulting times
are better than these full-fledged state-of-the-art aligners, which use a combination of index search
and in-text verification. Note that pure index-based search using standard backtracking takes 43.59s
which is much worse than all aforementioned times. For K = 2, BWA, Bwolo, and Yara times took
143.32s, 64.09s, and 70.56s, while pure index-based ASM using our optimal schemes for P = K + 1 = 3
and P = K + 2 = 4, took 120.70s, and 112.23s, respectively, i.e., they performed better than BWA but
worse than Bwolo and Yara. For K = 3, the benefit of using in-text verification in full-fledged aligners
catches up and all of them work faster than the pure index-based search using our optimal schemes.
We find these results for our optimal search schemes very impressive. To our knowledge, this is
the first time that, for K = 1 and K = 2, ASM of reads of this size (R = 101) performed completely
in index has been reported to compete in running time with the best full-fledged aligners, which
use combination of index search and in-text DP verification. This implies the power of our optimal
search schemes. We note that the results are even better in strata mode. We performed a second set
of comparisons with Yara, this time in strata mode. The 0-strata search means we first search the
reads with 0 errors, then search all the reads with no exact match, with 1 error, and so on, until K is
reached. This strategy can be generalized to s-strata, where s ≤ K. This means that, for b = 0 to
K − s, for all reads with a b-error best match, we compute all occurrences with up to b + s errors.
We ran Yara in 1-strata mode using -e <K> -s 0 -y full -t 1 and compared the time with the
pure index-based search using our optimal schemes in the same mode. For K = 1, 2, and 3, Yara took
3.57s, 8.00s, and 38.48s, respectively, and the pure index-based search took 4.36s, 8.57s, and 53.74s,
respectively. That is for K = 1, 2, and 3, the pure index-based search using our optimal search scheme
is about as fast as the full-fledged aligner Yara, which uses index search and in-text DP verification.
Of course, we did not expect to be able to outperform full-fledged aligners for larger values of K,
because for larger K and this read length, verification in index is too costly, especially if the number of
successful verifications is low (in our case almost all reads occur only once in the genome). Although
9
Vroland et al. [19] raised the option of using a bidirectional index for verification, they only used
in-text DP verification for the last pieces as they had only a unidirectional index at hand. Nevertheless,
interestingly, for their data set (40bp, exactly 3 errors), pure index-based search using our optimal
search schemes outperforms Bwolo by a factor of almost 1.5 (data not shown), so the read length
matters. Although the optimal scheme found by our MIP is superior to the 01∗ 0 scheme of Vroland et
al. [19] for search in the index, for a larger K, one has to take into account the number of remaining
verifications versus the number of edges in the trie for the remaining pieces. If that ratio is low, it
does not pay off to verify in the index instead of verification in the text as our comparisons showed.
Our observations in this section about the power of our optimal search schemes suggest that a
full-fledged aligner that employs an intelligent combination of search in the bidirectional FM-index
using our optimal search schemes and in-text verification using DP can outperform today’s best
approximate read mappers. The development of such an aligner, called FAMOUS (Fast Approximate
string Matching using OptimUm search Schemes) is ongoing in our group as future work.
5
Conclusions
In this paper, we contributed to the approximate string matching research as follows:
1. We proposed, for the first time, a method to solve the optimal search scheme problem for ASM-B
for Hamming distance (using a MIP formulation).
2. We demonstrated that our MIP approach can solve the optimum search scheme problem to
optimality in a reasonable amount of time for input parameters of considerable size, and enjoys
very quick convergence to optimal or near-optimal solutions for input parameters of larger size.
3. We showed that approximate search in a bidirectional FM-index can be performed significantly
faster if the optimal schemes obtained from our MIP are used. This was demonstrated based
on number of edges in the search tries as well as actual running time of in-index search on real
Illumina reads (up to 35 times faster than standard backtracking for 3 errors). We also showed that
although our MIP solutions are for Hamming distance they perform equally well for edit distance.
4. We showcased the power of our optimal search schemes by demonstrating that for K = 1 and
2 errors, approximate string matching of reads of size R = 101 performed completely in index
compete in running time with the best full-fledged aligners, which benefit from combining search
in index with in-text dynamic programming verification. This suggests that a full-fledged aligner
that intelligently combines search in bidirectional index using our optimal search schemes with
in-text verification using DP can outperform today’s best approximate aligners. The development
of such an aligner, called FAMOUS is ongoing as our future work.
Moreover, our approach in this research has raised some interesting open problems:
1. Our computational experiments in Section 2.3 showed that our current MIP has two attractive
properties: the early solutions it finds are optimal or near-optimal, and its optimal search scheme
is insensitive to the value of R (we ask: “is this insensitivity to R a theoretically provable fact?”).
This makes our current MIP quite powerful in practice because, even if all input parameters K, R,
P , S are quite large, we can run the MIP for a short time with a much smaller R to get a solution
that is most probably optimal or near-optimal for the original problem. Nevertheless, solving the
MIP completely to ascertain optimality is of great interest and currently consumes considerable
computational resources for large instances, especially when S > 5, K > 4, P > 6, R > 100. We
ask “can the solution time be improved by introducing other MIP formulations, or strengthening
the current formulation using strong cutting planes or further elimination of symmetric solutions?”
2. Our current MIP is restricted to optimizing over equal-size pieces, with P and S given as part
of the input. We ask “are there (MIP) approaches to optimize over the number of pieces in the
partition, and/or unequal piece lengths, and/or with no upper bound on the number of searches?”
10
3. Our optimal search scheme, SOpt , for the example of Lam et al. [7] covers every possible mismatch
pattern only once, ensuring that no duplicate computational effort is spent on the same mismatch
pattern. We, however, believe that this is not always the case for every optimal search scheme. We
ask “how would enforcing this requirement on the solution of MIP affect its solution time and the
performance of optimal search schemes it obtains?”
4. We demonstrated that solutions found by our MIP, although for Hamming distance, perform very
well for the edit distance as well. We ask “are there (MIP) approaches to find the actual optimal
search scheme for the edit distance?”
5. We demonstrated that the verification of few occurrences with high errors in the index is worse
than in-text DP verification. We ask “what is the best point to stop verification in the index and
start verifying in the text instead?.” This can be individually decided for each pattern. We are
addressing this problem in the ongoing development of our upcoming full-fledged aligner FAMOUS.
Acknowledgments
We acknowledge Texas A&M University High Performance Research Computing (HPRC) for providing resources to
perform parts of computational experiments. The second author also acknowledges the support of the International
Max-Planck Research School for Computational Biology and Scientific Computing (IMPRS-CBSC).
References
1. Abouelhoda, M.I., Kurtz, S., Ohlebusch, E.: Replacing suffix trees with enhanced suffix arrays. Journal of Discrete
Algorithms 2(1) (2004) 53–86
2. Burrows, M., Wheeler, D.J.: A block-sorting lossless data compression algorithm. Technical Report 124, Digital
SRC Research Report (1994)
3. Ferragina, P., Manzini, G.: Opportunistic data structures with applications. In: FOCS ’00. (2000) 390–398
4. IBM-ILOG: Cplex 12.7.1, https://www.ibm.com/support/knowledgecenter/en/sssa5p 12.7.1/ilog.odms.studio.help/
optimization studio/topics/cos home.html (Accessed on Nov. 2, 2017).
5. Karkkainen, J., Na, J.C.: Faster filters for approximate string matching. In: ALENEX ’07. (2007) 84–90
6. Kucherov, G., Salikhov, K., Tsur, D.: Approximate string matching using a bidirectional index. Theoretical
Computer Science 638 (2016) 145–158
7. Lam, T.W., Li, R., Tam, A., Wong, S., Wu, E., Yiu, S.M.: High throughput short read alignment via bi-directional
bwt. In: IEEE BIBM ’09. 31–36
8. Langmead, B., Salzberg, S.: Fast gapped-read alignment with Bowtie 2. Nature Methods 9(4) (2012) 357–359
9. Langmead, B., Trapnell, C., Pop, M., Salzberg, S.L.: Ultrafast and memory-efficient alignment of short dna sequences
to the human genome. Genome biology 10(3) (2009) R25
10. Li, H., Durbin, R.: Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25(14)
(2009) 1754–1760
11. Maaß, M.G.: Linear bidirectional on-line construction of affix trees. Algorithmica 37(1) (2003) 43–74
12. Manber, U., Myers, E.W.: Suffix arrays: a new method for on-line string searches. In: SODA ’90. (1990) 319–327
13. Nemhauser, G.L., Wolsey, L.A.: Integer and combinatorial optimization. Wiley, New York (1988)
14. Pockrandt, C., Ehrhardt, M., Reinert, K.: EPR-Dictionaries: A Practical and Fast Data Structure for Constant
Time Searches in Unidirectional and Bidirectional FM Indices. In: RECOMB ’17. (2017) 190–206
15. Reinert, K., Dadi, T.H., Ehrhardt, M., Hauswedell, H., Mehringer, S., Rahn, R., Kim, J., Pockrandt, C., Winkler,
J., Siragusa, E., Urgese, G., Weese, D.: The SeqAn C++ template library for efficient sequence analysis: A resource
for programmers. Journal of Biotechnology 261 (2017) 157–168
16. Reinert, K., Langmead, B., Weese, D., Evers, D.J.: Alignment of Next-Generation Sequencing Reads. Annual review
of genomics and human genetics 16 (2015) 133–151
17. Siragusa, E.: Approximate string matching for high-throughput sequencing. PhD thesis, Freie Universität Berlin
(2015)
18. Strothmann, D.: The affix array data structure and its applications to rna secondary structure analysis. Theoretical
Computer Science 389(1-2) (2007) 278–294
19. Vroland, C., Salson, M., Bini, S., Touzet, H.: Approximate search of short patterns with high error rates using the
01*0 lossless seeds. Journal of Discrete Algorithms (2016) 3–16
20. Wolsey, L.A.: Integer programming. Wiley, New York (1998)
11
Appendix
K=1
obj
1.4
1.2
1.0
0.8
R = 100
P
5
6
●
1
●
●
●
●
●
3_
S
4
5
K=3
x108
1.1
1.0
0.9
0.8
0.7
0.6
●
●
2
1.6
1.4
1.2
1.0
0.8
●
●
1
●
●
●
●
●
●
2
3_
S
4
5
7
●
8
K
8
●
●
6
●
●
●
●
4
●
●
4
●
●
●
●●
●
● ●
●● ●
● ● ●
● ●
●●
0.1 1
1
2
3
4
●
●
6
2
● ●●
● ●
●
●
●●
●
●
●
10 102 103 104
● ● ●
●
●
●
●
●●
●
●
●
10 102 103 104
time (seconds)
0.1
1
time (seconds)
R = 75
R = 100
●
7.5
●
R = 50
●
2
●
K=4
x109
●
R = 35
K=2
x106
●
obj /(opt obj)
x104
1.6
●
●
7.5
●
●
●
●
6
5.0
●
5.0
●
●
●
5
2.5
●
●
●
1
2
●
●
3_
S
●
●
4
●
●
4
5
●
1
2
●
●
3_
S
●
●
4
●
5
2.5
●
●
●
●
● ●●
●
0.1
●● ● ●
●
●●
●
●●
●
●
●
●
●
●
10 102 103
1
time (seconds)
(a)
●
●
● ●●
● ●
0.1
●
●
●
●
●
● ●●
●
● ●
●●
1 10 102 103 104
time (seconds)
(b)
Fig. 2. (a). Sensitivity of optimal objective value to parameters R, K, S, and P . For some cases due to memory overflow,
there is no data point. (b) Rapid convergence of feasible solutions to the optimal solution.
Table 3. Search schemes found by our MIP for K = 1, 2, 3 and P = K + 1, P = K + 2 and P = K + 3 used for
experiments in Tables 1 and 2. The schemes for K = 1 and 2 are optimal schemes with S = 5. To control the running
time of MIP, the schemes for K = 3 and 4 are best solutions found by running the MIP for 2 hours with S = 3. These
schemes are most probably optimal for S = 3.
K=1
Optimal (P = K + 1)
(12, 00, 01)
(21, 01, 01)
Optimal (P = K + 2)
(123, 001, 001)
(321, 000, 011)
Optimal (P = K + 3)
(1234, 0000, 0011)
(4321, 0001, 0011)
K=2
(123, 002, 012)
(321, 000, 022)
(231, 011, 012)
(2134, 0011, 0022)
(3214, 0000, 0112)
(4321, 0002, 0122)
(21345, 00011, 00222)
(43215, 00000, 00112)
(54321, 00002, 01122)
12
K=3
(1234, 0003, 0233)
(2341, 0000, 1223)
(3421, 0022, 0033)
(12345, 00022, 00333)
(43215, 00000, 11223)
(54321, 00003, 02233)
(123456, 000003, 022233)
(234561, 000000, 111223)
(654321, 000022, 003333)
K=4
(12345, 00004, 03344)
(23451, 00000, 22334)
(54321, 00033, 00444)
(123456, 000004, 033344)
(234561, 000000, 222334)
(654321, 000033, 004444)
(1234567, 0111111, 3333334)
(1234567, 0000000, 0044444)
(7654321, 0000004, 0333344)
| 8cs.DS
|
DiSAN: Directional Self-Attention Network for
RNN/CNN-Free Language Understanding
Tao Shen†
Jing Jiang†
Tianyi Zhou‡
Shirui Pan†
Guodong Long†
Chengqi Zhang†
arXiv:1709.04696v3 [cs.CL] 20 Nov 2017
†Centre of Artificial Intelligence, FEIT, University of Technology Sydney
‡Paul G. Allen School of Computer Science & Engineering, University of Washington
[email protected], [email protected]
{guodong.long, jing.jiang, shirui.pan, chengqi.zhang}@uts.edu.au
Abstract
Recurrent neural nets (RNN) and convolutional neural nets
(CNN) are widely used on NLP tasks to capture the long-term
and local dependencies, respectively. Attention mechanisms
have recently attracted enormous interest due to their highly
parallelizable computation, significantly less training time,
and flexibility in modeling dependencies. We propose a novel
attention mechanism in which the attention between elements
from input sequence(s) is directional and multi-dimensional
(i.e., feature-wise). A light-weight neural net, “Directional
Self-Attention Network (DiSAN)”, is then proposed to learn
sentence embedding, based solely on the proposed attention
without any RNN/CNN structure. DiSAN is only composed
of a directional self-attention with temporal order encoded,
followed by a multi-dimensional attention that compresses
the sequence into a vector representation. Despite its simple
form, DiSAN outperforms complicated RNN models on both
prediction quality and time efficiency. It achieves the best
test accuracy among all sentence encoding methods and improves the most recent best result by 1.02% on the Stanford
Natural Language Inference (SNLI) dataset, and shows stateof-the-art test accuracy on the Stanford Sentiment Treebank
(SST), Multi-Genre natural language inference (MultiNLI),
Sentences Involving Compositional Knowledge (SICK), Customer Review, MPQA, TREC question-type classification
and Subjectivity (SUBJ) datasets.
1
Introduction
Context dependency plays a significant role in language
understanding and provides critical information to natural
language processing (NLP) tasks. For different tasks and
data, researchers often switch between two types of deep
neural network (DNN): recurrent neural network (RNN)
with sequential architecture capturing long-range dependencies (e.g., long short-term memory (LSTM) (Hochreiter and
Schmidhuber 1997) and gated recurrent unit (GRU) (Chung
et al. 2014)), and convolutional neural network (CNN) (Kim
2014) whose hierarchical structure is good at extracting local or position-invariant features. However, which network
to choose in practice is an open question, and the choice relies largely on the empirical knowledge.
Recent works have found that equipping RNN or CNN
with an attention mechanism can achieve state-of-the-art
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
performance on a large number of NLP tasks, including neural machine translation (Bahdanau, Cho, and Bengio 2015;
Luong, Pham, and Manning 2015), natural language inference (Liu et al. 2016), conversation generation (Shang, Lu,
and Li 2015), question answering (Hermann et al. 2015;
Sukhbaatar et al. 2015), machine reading comprehension
(Seo et al. 2017), and sentiment analysis (Kokkinos and
Potamianos 2017). The attention uses a hidden layer to compute a categorical distribution over elements from the input sequence to reflect their importance weights. It allows
RNN/CNN to maintain a variable-length memory, so that
elements from the input sequence can be selected by their
importance/relevance and merged into the output. In contrast to RNN and CNN, the attention mechanism is trained
to capture the dependencies that make significant contributions to the task, regardless of the distance between the elements in the sequence. It can thus provide complementary
information to the distance-aware dependencies modeled by
RNN/CNN. In addition, computing attention only requires
matrix multiplication, which is highly parallelizable compared to the sequential computation of RNN.
In a very recent work (Vaswani et al. 2017), an attention mechanism is solely used to construct a sequence to
sequence (seq2seq) model that achieves a state-of-the-art
quality score on the neural machine translation (NMT) task.
The seq2seq model, “Transformer”, has an encoder-decoder
structure that is only composed of stacked attention networks, without using either recurrence or convolution. The
proposed attention, “multi-head attention”, projects the input sequence to multiple subspaces, then applies scaled dotproduct attention to its representation in each subspace,
and lastly concatenates their output. By doing this, it can
combine different attentions from multiple subspaces. This
mechanism is used in Transformer to compute both the
context-aware features inside the encoder/decoder and the
bottleneck features between them.
The attention mechanism has more flexibility in sequence
length than RNN/CNN, and is more task/data-driven when
modeling dependencies. Unlike sequential models, its computation can be easily and significantly accelerated by existing distributed/parallel computing schemes. However, to the
best of our knowledge, a neural net entirely based on attention has not been designed for other NLP tasks except NMT,
especially those that cannot be cast into a seq2seq problem.
Compared to RNN, a disadvantage of most attention mechanisms is that the temporal order information is lost, which
however might be important to the task. This explains why
positional encoding is applied to the sequence before being
processed by the attention in Transformer. How to model order information within an attention is still an open problem.
The goal of this paper is to develop a unified and
RNN/CNN-free attention network that can be generally utilized to learn the sentence encoding model for different NLP
tasks, such as natural language inference, sentiment analysis, sentence classification and semantic relatedness. We focus on the sentence encoding model because it is a basic
module of most DNNs used in the NLP literature.
We propose a novel attention mechanism that differs from
previous ones in that it is 1) multi-dimensional: the attention w.r.t. each pair of elements from the source(s) is a vector, where each entry is the attention computed on each feature; and 2) directional: it uses one or multiple positional
masks to model the asymmetric attention between two elements. We compute feature-wise attention since each element in a sequence is usually represented by a vector, e.g.,
word/character embedding (Kim et al. 2016), and attention
on different features can contain different information about
dependency, thus to handle the variation of contexts around
the same word. We apply positional masks to attention distribution since they can easily encode prior structure knowledge such as temporal order and dependency parsing. This
design mitigates the weakness of attention in modeling order
information, and takes full advantage of parallel computing.
We then build a light-weight and RNN/CNN-free neural
network, “Directional Self-Attention Network (DiSAN)”,
for sentence encoding. This network relies entirely on the
proposed attentions and does not use any RNN/CNN structure. In DiSAN, the input sequence is processed by directional (forward and backward) self-attentions to model context dependency and produce context-aware representations
for all tokens. Then, a multi-dimensional attention computes
a vector representation of the entire sequence, which can
be passed into a classification/regression module to compute the final prediction for a particular task. Unlike Transformer, neither stacking of attention blocks nor an encoderdecoder structure is required. The simple architecture of
DiSAN leads to fewer parameters, less computation and easier parallelization.
In experiments1 , we compare DiSAN with the currently
popular methods on various NLP tasks, e.g., natural language inference, sentiment analysis, sentence classification,
etc. DiSAN achieves the highest test accuracy on the Stanford Natural Language Inference (SNLI) dataset among
sentence-encoding models and improves the currently best
result by 1.02%. It also shows the state-of-the-art performance on the Stanford Sentiment Treebank (SST), MultiGenre natural language inference (MultiNLI), SICK, Customer Review, MPQA, SUBJ and TREC question-type classification datasets. Meanwhile, it has fewer parameters and
exhibits much higher computation efficiency than the mod1
Codes and pre-trained models for experiments can be found at
https://github.com/taoshen58/DiSAN
els it outperforms, e.g., LSTM and tree-based models.
Annotation: 1) Lowercase denotes a vector; 2) bold lowercase denotes a sequence of vectors (stored as a matrix);
and 3) uppercase denotes a matrix or a tensor.
2
2.1
Background
Sentence Encoding
In the pipeline of NLP tasks, a sentence is denoted by a sequence of discrete tokens (e.g., words or characters) v =
[v1 , v2 , . . . , vn ], where vi could be a one-hot vector whose
dimension length equals the number of distinct tokens N .
A pre-trained token embedding (e.g., word2vec (Mikolov
et al. 2013b) or GloVe (Pennington, Socher, and Manning
2014)) is applied to v and transforms all discrete tokens to
a sequence of low-dimensional dense vector representations
x = [x1 , x2 , . . . , xn ] with xi ∈ Rde . This pre-process can
be written as x = W (e) v, where word embedding weight
matrix W (e) ∈ Rde ×N and x ∈ Rde ×n .
Most DNN sentence-encoding models for NLP tasks take
x as the input and further generate a vector representation ui
for each xi by context fusion. Then a sentence encoding is
obtained by mapping the sequence u = [u1 , u2 , . . . , un ] to
a single vector s ∈ Rd , which is used as a compact encoding
of the entire sentence in NLP problems.
2.2
Attention
The attention is proposed to compute an alignment score between elements from two sources. In particular, given the token embeddings of a source sequence x = [x1 , x2 , . . . , xn ]
and the vector representation of a query q, attention computes the alignment score between xi and q by a compatibility function f (xi , q), which measures the dependency between xi and q, or the attention of q to xi . A softmax function then transforms the scores [f (xi , q)]ni=1 to a probability
distribution p(z|x, q) by normalizing over all the n tokens
of x. Here z is an indicator of which token in x is important
to q on a specific task. That is, large p(z = i|x, q) means xi
contributes important information to q. The above process
can be summarized by the following equations.
n
a = [f (xi , q)]i=1 ,
p(z|x, q) = softmax(a).
(1)
(2)
exp(f (xi , q))
p(z = i|x, q) = Pn
.
i=1 exp(f (xi , q))
(3)
Specifically,
The output of this attention mechanism is a weighted sum
of the embeddings for all tokens in x, where the weights
are given by p(z|x, q). It places large weights on the tokens
important to q, and can be written as the expectation of a
token sampled according to its importance, i.e.,
s=
n
X
p(z = i|x, q)xi = Ei∼p(z|x,q) (xi ),
(4)
i=1
where s ∈ Rde can be used as the sentence encoding of x.
Additive attention (or multi-layer perceptron attention)
(Bahdanau, Cho, and Bengio 2015; Shang, Lu, and Li
3.1
Multi-dimensional Attention
Multi-dimensional attention is a natural extension of additive attention (or MLP attention) at the feature level. Instead
of computing a single scalar score f (xi , q) for each token xi
as shown in Eq.(5), multi-dimensional attention computes a
feature-wise score vector for xi by replacing weight vector
w in Eq.(5) with a matrix W , i.e.,
f (xi , q) = W T σ W (1) xi + W (2) q ,
(7)
(a)
(b)
Figure 1: (a) Traditional (additive/multiplicative) attention
and (b) multi-dimensional attention. zi denotes alignment
score f (xi , q), which is a scalar in (a) but a vector in (b).
2015) and multiplicative attention (or dot-product attention)
(Vaswani et al. 2017; Sukhbaatar et al. 2015; Rush, Chopra,
and Weston 2015) are the two most commonly used attention mechanisms. They share the same and unified form of
attention introduced above, but are different in the compatibility function f (xi , q). Additive attention is associated with
f (xi , q) = wT σ(W (1) xi + W (2) q),
(5)
de
where σ(·) is an activation function and w ∈ R is a weight
vector. Multiplicative attention uses inner product or cosine
similarity for f (xi , q), i.e.,
D
E
f (xi , q) = W (1) xi , W (2) q .
(6)
In practice, additive attention often outperforms multiplicative one in prediction quality, but the latter is faster and more
memory-efficient due to optimized matrix multiplication.
2.3
Self-Attention
Self-Attention is a special case of the attention mechanism
introduced above. It replaces q with a token embedding xj
from the source input itself. It relates elements at different
positions from a single sequence by computing the attention
between each pair of tokens, xi and xj . It is very expressive and flexible for both long-range and local dependencies, which used to be respectively modeled by RNN and
CNN. Moreover, it has much faster computation speed and
fewer parameters than RNN. In recent works, we have already witnessed its success across a variety of NLP tasks,
such as reading comprehension (Hu, Peng, and Qiu 2017)
and neural machine translation (Vaswani et al. 2017).
3
Two Proposed Attention Mechanisms
In this section, we introduce two novel attention mechanisms, multi-dimensional attention in Section 3.1 (with two
extensions to self-attention in Section 3.2) and directional
self-attention in Section 3.3. They are the main components
of DiSAN and may be of independent interest to other neural
nets for other NLP problems in which an attention is needed.
where f (xi , q) ∈ Rde is a vector with the same length as xi ,
and all the weight matrices W, W (1) , W (2) ∈ Rde ×de . We
further add two bias terms to the parts in and out activation
σ(·), i.e.,
f (xi , q) = W T σ W (1) xi + W (2) q + b(1) + b. (8)
We then compute a categorical distribution p(zk |x, q) over
all the n tokens for each feature k ∈ [de ]. A large p(zk =
i|x, q) means that feature k of token i is important to q.
We apply the same procedure Eq.(1)-(3) in traditional attention to the k th dimension of f (xi , q). In particular, for
each feature k ∈ [de ], we replace f (xi , q) with [f (xi , q)]k ,
and change z to zk in Eq.(1)-(3). Now each feature k in each
token i has an importance weight Pki , p(zk = i|x, q). The
output s can be written as
hXn
ide
de
s=
Pki xki
= Ei∼p(zk |x,q) (xki ) k=1 . (9)
i=1
k=1
We give an illustration of traditional attention and multidimensional attention in Figure 1. In the rest of this paper,
we will ignore the subscript k which indexes feature dimension for simplification if no confusion is possible. Hence,
the output
Pn s can be written as an element-wise product
s = i=1 P·i xi
Remark: The word embedding usually suffers from the
polysemy in natural language. Since traditional attention
computes a single importance score for each word based on
the word embedding, it cannot distinguish the meanings of
the same word in different contexts. Multi-dimensional attention, however, computes a score for each feature of each
word, so it can select the features that can best describe the
word’s specific meaning in any given context, and include
this information in the sentence encoding output s.
3.2
Two types of Multi-dimensional Self-attention
When extending multi-dimension to self-attentions, we have
two variants of multi-dimensional attention. The first one,
called multi-dimensional “token2token” self-attention, explores the dependency between xi and xj from the same
source x, and generates context-aware coding for each element. It replaces q with xj in Eq.(8), i.e.,
f (xi , xj ) = W T σ W (1) xi + W (2) xj + b(1) + b. (10)
Similar to P in vanilla multi-dimensional attention, we compute a probability matrix P j ∈ Rde ×n for each xj such that
j
Pki
, p(zk = i|x, xj ). The output for xj is
n
X
sj =
P·ij xi
(11)
i=1
In DiSA, we first transform the input sequence x =
[x1 , x2 , . . . , xn ] to a sequence of hidden state h =
[h1 , h2 , . . . , hn ] by a fully connected layer, i.e.,
(14)
h = σh W (h) x + b(h) ,
where x ∈ Rde ×n , h ∈ Rdh ×n , W (h) and b(h) are the learnable parameters, and σh (·) is an activation function.
We then apply multi-dimensional token2token selfattention to h, and generate context-aware vector representations s for all elements from the input sequence. We make
two modifications to Eq.(10) to reduce the number of parameters and make the attention directional.
First, we set W in Eq.(10) to a scalar c and divide the part
in σ(·) by c, and we use tanh(·) for σ(·), which reduces the
number of parameters. In experiments, we always set c = 5,
and obtain stable output.
Second, we apply a positional mask to Eq.(10), so the attention between two elements can be asymmetric. Given a
mask M ∈ {0, −∞}n×n , we set bias b to a constant vector Mij 1 in Eq.(10), where 1 is an all-one vector. Hence,
Eq.(10) is modified to
f (hi , hj ) =
c · tanh [W (1) hi + W (2) hj + b(1) ]/c + Mij 1.
Figure 2: Directional self-attention (DiSA) mechanism.
Here, we use li,j to denote f (hi , hj ) in Eq. (15).
The output of token2token self-attention for all elements
from x is s = [s1 , s2 , . . . , sn ] ∈ Rde ×n .
The second one, multi-dimensional “source2token” selfattention, explores the dependency between xi and the entire
sequence x, and compresses the sequence x into a vector. It
removes q from Eq.(8), i.e.,
f (xi ) = W T σ W (1) xi + b(1) + b.
(12)
The probability matrix is defined as Pki , p(zk = i|x)
and is computed in the same way as P in vanilla multidimensional attention. The output s is also same, i.e.,
s=
n
X
P·i
xi
(13)
i=1
We will use these two types (i.e., token2token and
source2token) of multi-dimensional self-attention in different parts of our sentence encoding model, DiSAN.
3.3
Directional Self-Attention
Directional self-attention (DiSA) is composed of a fully
connected layer whose input is the token embeddings x,
a “masked” multi-dimensional token2token self-attention
block to explore the dependency and temporal order, and a
fusion gate to combine the output and input of the attention
block. Its structure is shown in Figure 2. It can be used as
either a neural net or a module to compose a large network.
(15)
To see why a mask can encode directional information, let
us consider a case in which Mij = −∞ and Mji = 0, which
results in [f (hi , hj )]k = −∞ and unchanged [f (hj , hi )]k .
Since the probability p(zk = i|x, xj ) is computed by
softmax, [f (hi , hj )]k = −∞ leads to p(zk = i|x, xj ) = 0.
This means that there is no attention of xj to xi on feature k.
On the contrary, we have p(zk = j|x, xi ) > 0, which means
that attention of xi to xj exists on feature k. Therefore, prior
structure knowledge such as temporal order and dependency
parsing can be easily encoded by the mask, and explored in
generating sentence encoding. This is an important feature
of DiSA that previous attention mechanisms do not have.
For self-attention, we usually need to disable the attention
of each token to itself (Hu, Peng, and Qiu 2017). This is the
same as applying a diagonal-disabled (i.e., diag-disabled)
mask such that
0,
i 6= j
diag
Mij =
(16)
−∞, i = j
Moreover, we can use masks to encode temporal order
information into attention output. In this paper, we use two
masks, i.e., forward mask M f w and backward mask M bw ,
0,
i<j
fw
Mij
=
(17)
−∞, otherwise
0,
i>j
bw
Mij
=
(18)
−∞, otherwise
In forward mask M f w , there is the only attention of later
token j to early token i, and vice versa in backward mask.
We show these three positional masks in Figure 3.
Given input sequence x and a mask M , we compute
f (xi , xj ) according to Eq.(15), and follow the standard procedure of multi-dimensional token2token self-attention to
(a) Diag-disabled mask
Figure 4: Directional self-attention network (DiSAN)
(b) Forward mask
(c) Backward mask
Figure 3: Three positional masks: (a) is the diag-disabled
mask M diag ; (b) and (c) are forward mask M f w and backward mask M bw , respectively.
compute the probability matrix P j for each j ∈ [n]. Each
output sj in s is computed as in Eq.(11).
The final output u ∈ Rdh ×n of DiSA is obtained by combining the output s and the input h of the masked multidimensional token2token self-attention block. This yields a
temporal order encoded and context-aware vector representation for each element/token. The combination is accomplished by a dimension-wise fusion gate, i.e.,
F = sigmoid W (f 1) s + W (f 2) h + b(f )
(19)
u=F
h + (1 − F )
s
(20)
where W (f 1) , W (f 2) ∈ Rdh ×dh and b(f ) ∈ Rdh are the
learnable parameters of the fusion gate.
4
Directional Self-Attention Network
We propose a light-weight network, “Directional SelfAttention Network (DiSAN)”, for sentence encoding. Its architecture is shown in Figure 4.
Given an input sequence of token embedding x, DiSAN
firstly applies two parameter-untied DiSA blocks with forward mask M f w Eq.(17) and M bw Eq.(18), respectively.
The feed-forward procedure is given in Eq.(14)-(15) and
Eq.(19)-(20). Their outputs are denoted by uf w , ubw ∈
Rdh ×n . We concatenate them vertically as [uf w ; ubw ] ∈
R2dh ×n , and use this concatenated output as input to a multidimensional source2token self-attention block, whose output sdisan ∈ R2dh computed by Eq.(12)-(13) is the final
sentence encoding result of DiSAN.
Remark: In DiSAN, forward/backward DiSA blocks
work as context fusion layers. And the multi-dimensional
source2token self-attention compresses the sequence into a
single vector. The idea of using both forward and backward
attentions is inspired by Bi-directional LSTM (Bi-LSTM)
(Graves, Jaitly, and Mohamed 2013), in which forward and
backward LSTMs are used to encode long-range dependency from different directions. In Bi-LSTM, LSTM combines the context-aware output with the input by multi-gate.
The fusion gate used in DiSA shares the similar motivation.
However, DiSAN has fewer parameters, simpler structure
and better efficiency.
5
Experiments
In this section, we first apply DiSAN to natural language
inference and sentiment analysis tasks. DiSAN achieves
the state-of-the-art performance and significantly better efficiency than other baseline methods on benchmark datasets
for both tasks. We also conduct experiments on other NLP
tasks and DiSAN also achieves state-of-the-art performance.
Training Setup: We use cross-entropy loss plus L2 regularization penalty as optimization objective. We minimize
it by Adadelta (Zeiler 2012) (an optimizer of mini-batch
SGD) with batch size of 64. We use Adadelta rather than
Adam (Kingma and Ba 2015) because in our experiments,
DiSAN optimized by Adadelta can achieve more stable performance than Adam optimized one. Initial learning rate is
set to 0.5. All weight matrices are initialized by Glorot Initialization (Glorot and Bengio 2010), and the biases are initialized with 0. We initialize the word embedding in x by
300D GloVe 6B pre-trained vectors (Pennington, Socher,
and Manning 2014). The Out-of-Vocabulary words in training set are randomly initialized by uniform distribution between (−0.05, 0.05). The word embeddings are fine-tuned
during the training phrase. We use Dropout (Srivastava et
al. 2014) with keep probability 0.75 for language inference
and 0.8 for sentiment analysis. The L2 regularization decay factors γ are 5 × 10−5 and 10−4 for language inference
and sentiment analysis, respectively. Note that the dropout
keep probability and γ varies with the scale of corresponding dataset. Hidden units number dh is set to 300. Activation functions σ(·) are ELU (exponential linear unit) (Clevert, Unterthiner, and Hochreiter 2016) if not specified. All
models are implemented with TensorFlow2 and run on sin2
https://www.tensorflow.org
Model Name
|θ|
T(s)/epoch
Unlexicalized features (Bowman et al. 2015)
+ Unigram and bigram features (Bowman et al. 2015)
Train Accu(%)
Test Accu(%)
49.4
99.7
50.4
78.2
84.8
83.9
98.8
83.3
89.2
86.4
82.5
84.5
86.2
77.6
80.6
81.4
82.1
83.2
83.3
83.4
84.2
84.6
100D LSTM encoders (Bowman et al. 2015)
300D LSTM encoders (Bowman et al. 2016)
1024D GRU encoders (Vendrov et al. 2016)
300D Tree-based CNN encoders (Mou et al. 2016)
300D SPINN-PI encoders (Bowman et al. 2016)
600D Bi-LSTM encoders (Liu et al. 2016)
300D NTI-SLSTM-LSTM encoders (Munkhdalai and Yu 2017b)
600D Bi-LSTM encoders+intra-attention (Liu et al. 2016)
300D NSE encoders (Munkhdalai and Yu 2017a)
0.2m
3.0m
15m
3.5m
3.7m
2.0m
4.0m
2.8m
3.0m
Word Embedding with additive attention
Word Embedding with s2t self-attention
Multi-head with s2t self-attention
Bi-LSTM with s2t self-attention
DiSAN without directions
0.45m
0.54m
1.98m
2.88m
2.35m
216
261
345
2080
592
82.39
86.22
89.58
90.39
90.18
79.81
83.12
84.17
84.98
84.66
Directional self-attention network (DiSAN)
2.35m
587
91.08
85.62
Table 1: Experimental results for different methods on SNLI. |θ|: the number of parameters (excluding word embedding part).
T(s)/epoch: average time (second) per epoch. Train Accu(%) and Test Accu(%): the accuracy on training and test set.
gle Nvidia GTX 1080Ti graphic card.
5.1
Natural Language Inference
The goal of Natural Language Inference (NLI) is to reason the semantic relationship between a premise sentence
and a corresponding hypothesis sentence. The possible relationship could be entailment, neutral or contradiction. We
compare different models on a widely used benchmark,
Stanford Natural Language Inference (SNLI)3 (Bowman et
al. 2015) dataset, which consists of 549,367/9,842/9,824
(train/dev/test) premise-hypothesis pairs with labels.
Following the standard procedure in Bowman et al.
(2016), we launch two sentence encoding models (e.g.,
DiSAN) with tied parameters for the premise sentence and
hypothesis sentence, respectively. Given the output encoding
sp for the premise and sh for the hypothesis, the representation of relationship is the concatenation of sp , sh , sp − sh
and sp sh , which is fed into a 300D fully connected layer
and then a 3-unit output layer with softmax to compute a
probability distribution over the three types of relationship.
For thorough comparison, besides the neural nets proposed in previous works of NLI, we implement five extra
neural net baselines to compare with DiSAN. They help
us to analyze the improvement contributed by each part of
DiSAN and to verify that the two attention mechanisms proposed in Section 3 can improve other networks.
• Word Embedding with additive attention.
• Word Embedding with s2t self-attention: DiSAN with
DiSA blocks removed.
• Multi-head with s2t self-attention: Multi-head attention
(Vaswani et al. 2017) (8 heads, each has 75 hidden units)
with source2token self-attention. The positional encoding
3
https://nlp.stanford.edu/projects/snli/
method used in Vaswani et al. (2017) is applied to the input sequence to encode temporal information. We find our
experiments show that multi-head attention is sensitive to
hyperparameters, so we adjust keep probability of dropout
from 0.7 to 0.9 with step 0.05 and report the best result.
• Bi-LSTM with s2t self-attention: a multi-dimensional
source2token self-attention block is applied to the output
of Bi-LSTM (300D forward + 300D backward LSTMs).
• DiSAN without directions: DiSAN with the forward/backward masks M f w and M bw replaced with two
diag-disabled masks M diag , i.e., DiSAN without forward/backward order information.
Compared to the results from the official leaderboard
of SNLI in Table 1, DiSAN outperforms previous works
and improves the best latest test accuracy (achieved by a
memory-based NSE encoder network) by a remarkable margin of 1.02%. DiSAN surpasses the RNN/CNN based models with more complicated architecture and more parameters
by large margins, e.g., +2.32% to Bi-LSTM, +1.42% to BiLSTM with additive attention. It even outperforms models
with the assistance of a semantic parsing tree, e.g., +3.52%
to Tree-based CNN, +2.42% to SPINN-PI.
In the results of the five baseline methods and DiSAN at
the bottom of Table 1, we demonstrate that making attention
multi-dimensional (feature-wise) or directional brings substantial improvement to different neural nets. First, a comparison between the first two models shows that changing
token-wise attention to multi-dimensional/feature-wise attention leads to 3.31% improvement on a word embedding
based model. Also, a comparison between the third baseline
and DiSAN shows that DiSAN can substantially outperform
multi-head attention by 1.45%. Moreover, a comparison between the forth baseline and DiSAN shows that the DiSA
block can even outperform Bi-LSTM layer in context encoding, improving test accuracy by 0.64%. A comparison
between the fifth baseline and DiSAN shows that directional
self-attention with forward and backward masks (with temporal order encoded) can bring 0.96% improvement.
Additional advantages of DiSAN shown in Table 1 are
its fewer parameters and compelling time efficiency. It is
×3 faster than widely used Bi-LSTM model. Compared to
other models with competitive performance, e.g., 600D BiLSTM encoders with intra-attention (2.8M), 300D NSE encoders (3.0M) and 600D Bi-LSTM encoders with multidimensional attention (2.88M), DiSAN only has 2.35M parameters.
5.2
Nonetheless, DiSAN still outperforms these fancy models,
such as NCSL (+0.62%) and LR-Bi-LSTM (+1.12%).
Sentiment Analysis
Model
Test Accu
MV-RNN (Socher et al. 2013)
RNTN (Socher et al. 2013)
Bi-LSTM (Li et al. 2015)
Tree-LSTM (Tai, Socher, and Manning 2015)
CNN-non-static (Kim 2014)
CNN-Tensor (Lei, Barzilay, and Jaakkola 2015)
NCSL (Teng, Vo, and Zhang 2016)
LR-Bi-LSTM (Qian, Huang, and Zhu 2017)
44.4
45.7
49.8
51.0
48.0
51.2
51.1
50.6
Word Embedding with additive attention
Word Embedding with s2t self-attention
Multi-head with s2t self-attention
Bi-LSTM with s2t self-attention
DiSAN without directions
47.47
48.87
49.14
49.95
49.41
DiSAN
51.72
Table 2: Test accuracy of fine-grained sentiment analysis on
Stanford Sentiment Treebank (SST) dataset.
Figure 5: Fine-grained sentiment analysis accuracy vs. sentence length. The results of LSTM, Bi-LSTM and TreeLSTM are from Tai, Socher, and Manning (2015) and the
result of DiSAN is the average over five random trials.
It is also interesting to see the performance of different
models on the sentences with different lengths. In Figure 5,
we compare LSTM, Bi-LSTM, Tree-LSTM and DiSAN on
different sentence lengths. In the range of (5, 12), the length
range for most movie review sentences, DiSAN significantly
outperforms others. Meanwhile, DiSAN also shows impressive performance for slightly longer sentences or paragraphs
in the range of (25, 38). DiSAN performs poorly when the
sentence length ≥ 38, in which however only 3.21% of total
movie review sentences lie.
5.3
Sentiment analysis aims to analyze the sentiment of a
sentence or a paragraph, e.g., a movie or a product review. We use Stanford Sentiment Treebank (SST)4 (Socher
et al. 2013) for the experiments, and only focus on the
fine-grained movie review sentiment classification over five
classes, i.e., very negative, negative, neutral, positive and
very positive. We use the standard train/dev/test sets split
with 8,544/1,101/2,210 samples. Similar to Section 5.1, we
employ a single sentence encoding model to obtain a sentence representation s of a movie review, then pass it into
a 300D fully connected layer. Finally, a 5-unit output layer
with softmax is used to calculate a probability distribution
over the five classes.
In Table 2, we compare previous works with DiSAN on
test accuracy. To the best of our knowledge, DiSAN improves the last best accuracy (given by CNN-Tensor) by
0.52%. Compared to tree-based models with heavy use of
the prior structure, e.g., MV-RNN, RNTN and Tree-LSTM,
DiSAN outperforms them by 7.32%, 6.02% and 0.72%,
respectively. Additionally, DiSAN achieves better performance than CNN-based models. More recent works tend
to focus on lexicon-based sentiment analysis, by exploring sentiment lexicons, negation words and intensity words.
4
https://nlp.stanford.edu/sentiment/
Experiments on Other NLP Tasks
Multi-Genre Natural Language Inference Multi-Genre
Natural Language Inference (MultiNLI)5 (Williams, Nangia, and Bowman 2017) dataset consists of 433k sentence
pairs annotated with textual entailment information. This
dataset is similar to SNLI, but it covers more genres of spoken and written text, and supports a distinctive cross-genre
generalization evaluation. However, MultiNLI is a quite new
dataset, and its leaderboard does not include a session for
the sentence-encoding only model. Hence, we only compare
DiSAN with the baselines provided at the official website.
The results of DiSAN and two sentence-encoding models
on the leaderboard are shown in Table 3. Note that the prediction accuracies of Matched and Mismatched test datasets
are obtained by submitting our test results to Kaggle open
evaluation platforms6 : MultiNLI Matched Open Evaluation
and MultiNLI Mismatched Open Evaluation.
Semantic Relatedness The task of semantic relatedness
aims to predict a similarity degree of a given pair of sentences. We show an experimental comparison of different
5
https://www.nyu.edu/projects/bowman/multinli/
https://inclass.kaggle.com/c/multinli-matched-openevaluation and https://inclass.kaggle.com/c/multinli-mismatchedopen-evaluation
6
Method
Matched
Model
Mismatched
cBoW
Bi-LSTM
0.65200
0.67507
0.64759
0.67248
DiSAN
0.70977
0.71402
Table 3: Experimental results of prediction accuracy for different methods on MultiNLI.
methods on Sentences Involving Compositional Knowledge
(SICK)7 dataset (Marelli et al. 2014). SICK is composed
of 9,927 sentence pairs with 4,500/500/4,927 instances for
train/dev/test. The regression module on the top of DiSAN is
introduced by Tai, Socher, and Manning (2015). The results
in Table 4 show that DiSAN outperforms the models from
previous works in terms of Pearson’s r and Spearman’s ρ
indexes.
Pearson’s r Spearman’s ρ
Model
a
Meaning Factory
.8268
ECNUb
.8414
DT-RNNc
.7923 (.0070)
SDT-RNNc
.7900 (.0042)
Cons. Tree-LSTMd .8582 (.0038)
Dep. Tree-LSTMd .8676 (.0030)
DiSAN
.7721
/
.7319 (.0071)
.7304 (.0042)
.7966 (.0053)
.8083 (.0042)
MSE
.3224
/
.3822 (.0137)
.3848 (.0042)
.2734 (.0108)
.2532 (.0052)
.8695 (.0012) .8139 (.0012) .2879 (.0036)
Table 4: Experimental results for different methods on SICK
sentence relatedness dataset. The reported accuracies are the
mean of five runs (standard deviations in parentheses). Cons.
and Dep. represent Constituency and Dependency, respectively. a (Bjerva et al. 2014), b (Zhao, Zhu, and Lan 2014),
c
(Socher et al. 2014), d (Tai, Socher, and Manning 2015).
Sentence Classifications The goal of sentence classification is to correctly predict the class label of a given sentence
in various scenarios. We evaluate the models on four sentence classification benchmarks of various NLP tasks, such
as sentiment analysis and question-type classification. They
are listed as follows. 1) CR: Customer review (Hu and Liu
2004) of various products (cameras, etc.), which is to predict whether the review is positive or negative; 2) MPQA:
Opinion polarity detection subtask of the MPQA dataset
(Wiebe, Wilson, and Cardie 2005); 3) SUBJ: Subjectivity
dataset (Pang and Lee 2004) whose labels indicate whether
each sentence is subjective or objective; 4) TREC: TREC
question-type classification dataset (Li and Roth 2002). The
experimental results of DiSAN and existing methods are
shown in Table 5.
5.4
MPQA
SUBJ
TREC
DiSAN
84.8 (2.0) 90.1 (0.4) 94.2 (0.6) 94.2 (0.1)
Table 5: Experimental results for different methods on various sentence classification benchmarks. The reported accuracies on CR, MPQA and SUBJ are the mean of 10-fold
cross validation, the accuracies on TREC are the mean of
dev accuracies of five runs. All standard deviations are in
parentheses. a (Mikolov et al. 2013a), b (Kiros et al. 2015),
c
(Kalchbrenner, Grefenstette, and Blunsom 2014), d (Zhao,
Lu, and Poupart 2015), e (Lei and Zhang 2017).
particular, we will focus primarily on the probability in forward/backward DiSA blocks (Figure 6), forward/backward
fusion gates F in Eq.(19) (Figure 7), and the probability
in multi-dimensional source2token self-attention block (Figure 8). For the first two, we desire to demonstrate the dependency at token level, but attention probability in DiSAN
is defined on each feature, so we average the probabilities
along the feature dimension.
We select two sentences from SNLI test set as examples
for this case study. Sentence 1 is Families have some dogs
in front of a carousel and sentence 2 is volleyball match is in
progress between ladies.
(a) Sentence 1, forward
(b) Sentence 1, backward
(c) Sentence 2, forward
(d) Sentence 2, backward
Case Study
To gain a closer view of what dependencies in a sentence
can be captured by DiSAN, we visualize the attention probability p(z = i|x, xj ) or alignment score by heatmaps. In
7
CR
cBoWa
79.9
86.4
91.3
87.3
Skip-thoughtb
81.3
87.5
93.6
92.2
DCNNc
/
/
/
93.0
AdaSentd
83.6 (1.6) 90.4 (0.7) 92.2 (1.2) 91.1 (1.0)
SRUe
84.8 (1.3) 89.7 (1.1) 93.4 (0.8) 93.9 (0.6)
Wide CNNse 82.2 (2.2) 88.8 (1.2) 92.9 (0.7) 93.2 (0.5)
http://clic.cimec.unitn.it/composes/sick.html
Figure 6: Attention probability in forward/backward DiSA
blocks for the two example sentences.
Figure 6 shows that1) semantically important words such
as nouns and verbs usually get large attention, but stop
words (am, is, are, etc.) do not; 2) globally important words,
e.g., volleyball, match, ladies in sentence 1 and dog, front,
carousel in sentence 2, get large attention from all other
words; 3) if a word is important to only some of the other
words (e.g. to constitute a phrase or sense-group), it gets
large attention only from these words, e.g., attention between progress, between in sentence1, and attention between
families, have in sentence 2.
This also shows that directional information can help to
generate context-aware word representation with temporal
order encoded. For instance, for word match in sentence 1,
its forward DiSA focuses more on word volleyball, while
its backward attention focuses more on progress and ladies,
so the representation of word match contains the essential
information of the entire sentence, and simultaneously includes the positional order information.
In addition, forward and backward DiSAs can focus on
different parts of a sentence. For example, the forward one
in sentence 2 pays attention to the word families, whereas the
backward one focuses on the word carousel. Since forward
and backward attentions are computed separately, it avoids
normalization over multiple significant words to weaken
their weights. Note that this is a weakness of traditional attention compared to RNN, especially for long sentences.
(a) Sentence 1, forward
(b) Sentence 1, backward
(c) Sentence 2, forward
(d) Sentence 2, backward
Figure 7: Fusion Gate F in forward/backward DiSA blocks.
In Figure 7, we show that the gate value F in Eq.(19). The
gate combines the input and output of masked self-attention.
It tends to selects the input representation h instead of the
output s if the corresponding weight in F is large. This
shows that the gate values for meaningless words, especially
stop words is small. The stop words themselves cannot contribute important information, so only their semantic relations to other words might help to understand the sentence.
Hence, the gate tends to use their context features given by
masked self-attention.
(a) glass in pair 1
(b) close in pair 2
Figure 8: Two pairs of attention probability comparison of
same word in difference sentence contexts.
In Figure 8, we show the two multi-dimensional
source2token self-attention score vectors of the same word
in the two sentences, by their heatmaps. The first pair has
two sentences: one is The glass bottle is big, and another is
A man is pouring a glass of tea. They share the same word is
glass with different meanings. The second pair has two sentences: one is The restaurant is about to close and another
is A biker is close to the fountain. It can be seen that the
two attention vectors for the same words are very different
due to their different meanings in different contexts. This
indicates that the multi-dimensional attention vector is not
redundant because it can encode more information than one
single score used in traditional attention and it is able to capture subtle difference of the same word in different contexts
or sentences. Additionally, it can also alleviate the weakness
of the attention over long sequence, which can avoid normalization over entire sequence in traditional attention only
once.
6
Conclusion
In this paper, we propose two novel attention mechanisms,
multi-dimensional attention and directional self-attention.
The multi-dimensional attention performs a feature-wise selection over the input sequence for a specific task, and the
directional self-attention uses the positional masks to produce the context-aware representations with temporal information encoded. Based on these attentions, Directional
Self-Attention Network (DiSAN) is proposed for sentenceencoding without any recurrent or convolutional structure.
The experiment results show that DiSAN can achieve stateof-the-art inference quality and outperform existing works
(LSTM, etc.) on a wide range of NLP tasks with fewer parameters and higher time efficiency.
In future work, we will explore the approaches to using
the proposed attention mechanisms on more sophisticated
tasks, e.g. question answering and reading comprehension,
to achieve better performance on various benchmarks.
7
Acknowledgments
This research was funded by the Australian Government
through the Australian Research Council (ARC) under grant
1) LP160100630 partnership with Australia Government
Department of Health, and 2) LP150100671 partnership
with Australia Research Alliance for Children and Youth
(ARACY) and Global Business College Australia (GBCA).
References
Bahdanau, D.; Cho, K.; and Bengio, Y. 2015. Neural machine
translation by jointly learning to align and translate. In ICLR.
Bjerva, J.; Bos, J.; Van der Goot, R.; and Nissim, M. 2014. The
meaning factory: Formal semantics for recognizing textual entailment and determining semantic similarity. In SemEval@ COLING,
642–646.
Bowman, S. R.; Angeli, G.; Potts, C.; and Manning, C. D. 2015. A
large annotated corpus for learning natural language inference. In
EMNLP.
Bowman, S. R.; Gauthier, J.; Rastogi, A.; Gupta, R.; Manning,
C. D.; and Potts, C. 2016. A fast unified model for parsing and
sentence understanding. In ACL.
Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y. 2014. Empirical
evaluation of gated recurrent neural networks on sequence modeling. In NIPS.
Clevert, D.-A.; Unterthiner, T.; and Hochreiter, S. 2016. Fast and
accurate deep network learning by exponential linear units (elus).
In ICLR.
Glorot, X., and Bengio, Y. 2010. Understanding the difficulty of
training deep feedforward neural networks. In Proceedings of the
Thirteenth International Conference on Artificial Intelligence and
Statistics, 249–256.
Graves, A.; Jaitly, N.; and Mohamed, A.-r. 2013. Hybrid speech
recognition with deep bidirectional lstm. In Automatic Speech
Recognition and Understanding (ASRU), 2013 IEEE Workshop on,
273–278. IEEE.
Hermann, K. M.; Kocisky, T.; Grefenstette, E.; Espeholt, L.; Kay,
W.; Suleyman, M.; and Blunsom, P. 2015. Teaching machines to
read and comprehend. In NIPS.
Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation 9(8):1735–1780.
Hu, M., and Liu, B. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international
conference on Knowledge discovery and data mining, 168–177.
ACM.
Hu, M.; Peng, Y.; and Qiu, X. 2017. Reinforced mnemonic reader
for machine comprehension. arXiv preprint arXiv:1705.02798.
Kalchbrenner, N.; Grefenstette, E.; and Blunsom, P. 2014. A convolutional neural network for modelling sentences. arXiv preprint
arXiv:1404.2188.
Kim, Y.; Jernite, Y.; Sontag, D.; and Rush, A. M. 2016. Characteraware neural language models. In AAAI.
Kim, Y. 2014. Convolutional neural networks for sentence classification. In EMNLP.
Kingma, D., and Ba, J. 2015. Adam: A method for stochastic
optimization. In ICLR.
Kiros, R.; Zhu, Y.; Salakhutdinov, R. R.; Zemel, R.; Urtasun, R.;
Torralba, A.; and Fidler, S. 2015. Skip-thought vectors. In NIPS.
Kokkinos, F., and Potamianos, A. 2017. Structural attention
neural networks for improved sentiment analysis. arXiv preprint
arXiv:1701.01811.
Lei, T., and Zhang, Y. 2017. Training rnns as fast as cnns. arXiv
preprint arXiv:1709.02755.
Lei, T.; Barzilay, R.; and Jaakkola, T. 2015. Molding cnns for text:
non-linear, non-consecutive convolutions. In EMNLP.
Li, X., and Roth, D. 2002. Learning question classifiers. In ACL.
Li, J.; Luong, M.-T.; Jurafsky, D.; and Hovy, E. 2015. When
are tree structures necessary for deep learning of representations?
arXiv preprint arXiv:1503.00185.
Liu, Y.; Sun, C.; Lin, L.; and Wang, X. 2016. Learning natural language inference using bidirectional lstm model and inner-attention.
arXiv preprint arXiv:1605.09090.
Luong, M.-T.; Pham, H.; and Manning, C. D. 2015. Effective approaches to attention-based neural machine translation. In EMNLP.
Marelli, M.; Menini, S.; Baroni, M.; Bentivogli, L.; Bernardi, R.;
and Zamparelli, R. 2014. A sick cure for the evaluation of compositional distributional semantic models. In LREC.
Mikolov, T.; Chen, K.; Corrado, G.; and Dean, J. 2013a. Efficient
estimation of word representations in vector space. arXiv preprint
arXiv:1301.3781.
Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J.
2013b. Distributed representations of words and phrases and their
compositionality. In NIPS.
Mou, L.; Men, R.; Li, G.; Xu, Y.; Zhang, L.; Yan, R.; and Jin, Z.
2016. Natural language inference by tree-based convolution and
heuristic matching. In ACL.
Munkhdalai, T., and Yu, H. 2017a. Neural semantic encoders. In
EACL.
Munkhdalai, T., and Yu, H. 2017b. Neural tree indexers for text
understanding. In EACL.
Pang, B., and Lee, L. 2004. A sentimental education: Sentiment
analysis using subjectivity summarization based on minimum cuts.
In ACL.
Pennington, J.; Socher, R.; and Manning, C. D. 2014. Glove:
Global vectors for word representation. In EMNLP.
Qian, Q.; Huang, M.; and Zhu, X. 2017. Linguistically regularized
lstms for sentiment classification. In ACL.
Rush, A. M.; Chopra, S.; and Weston, J. 2015. A neural attention
model for abstractive sentence summarization. In EMNLP.
Seo, M.; Kembhavi, A.; Farhadi, A.; and Hajishirzi, H. 2017. Bidirectional attention flow for machine comprehension. In ICLR.
Shang, L.; Lu, Z.; and Li, H. 2015. Neural responding machine for
short-text conversation. In ACL.
Socher, R.; Perelygin, A.; Wu, J. Y.; Chuang, J.; Manning, C. D.;
Ng, A. Y.; Potts, C.; et al. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP.
Socher, R.; Karpathy, A.; Le, Q. V.; Manning, C. D.; and Ng, A. Y.
2014. Grounded compositional semantics for finding and describing images with sentences. Transactions of the Association for
Computational Linguistics 2:207–218.
Srivastava, N.; Hinton, G. E.; Krizhevsky, A.; Sutskever, I.; and
Salakhutdinov, R. 2014. Dropout: a simple way to prevent neural
networks from overfitting. Journal of Machine Learning Research
15(1):1929–1958.
Sukhbaatar, S.; Weston, J.; Fergus, R.; et al. 2015. End-to-end
memory networks. In NIPS.
Tai, K. S.; Socher, R.; and Manning, C. D. 2015. Improved semantic representations from tree-structured long short-term memory networks. In ACL.
Teng, Z.; Vo, D.-T.; and Zhang, Y. 2016. Context-sensitive lexicon
features for neural sentiment analysis. In EMNLP.
Vaswani, A.; Shazeer; Noam; Parmar, N.; Uszkoreit, J.; Jones, L.;
Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention is all
you need. In NIPS.
Vendrov, I.; Kiros, R.; Fidler, S.; and Urtasun, R. 2016. Orderembeddings of images and language. In ICLR.
Wiebe, J.; Wilson, T.; and Cardie, C. 2005. Annotating expressions
of opinions and emotions in language. Language resources and
evaluation 39(2):165–210.
Williams, A.; Nangia, N.; and Bowman, S. R. 2017. A broadcoverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426.
Zeiler, M. D. 2012. Adadelta: an adaptive learning rate method.
arXiv preprint arXiv:1212.5701.
Zhao, H.; Lu, Z.; and Poupart, P. 2015. Self-adaptive hierarchical
sentence model. In IJCAI.
Zhao, J.; Zhu, T.; and Lan, M. 2014. Ecnu: One stone two birds:
Ensemble of heterogenous measures for semantic relatedness and
textual entailment. In SemEval@ COLING, 271–277.
| 2cs.AI
|
1
A modified sequence domain impedance definition
and its equivalence to the dq-domain impedance
definition for the stability analysis of AC power
electronic systems
arXiv:1605.00526v1 [cs.SY] 28 Apr 2016
Atle Rygg, Marta Molinas, Zhang Chen and Xu Cai
Abstract—Representations of AC power systems by frequency
dependent impedance equivalents is an emerging technique in the
dynamic analysis of power systems including power electronic
converters. The technique has been applied for decades in
DC-power systems, and it was recently adopted to map the
impedances in AC systems. Most of the work on AC systems
can be categorized in two approaches. One is the analysis
of the system in the dq-domain, whereas the other applies
harmonic linearization in the phase domain through symmetric
components. Impedance models based on analytical calculations,
numerical simulation and experimental studies have been previously developed and verified in both domains independently.
The authors of previous studies discuss the advantages and
disadvantages of each domain separately, but neither a rigorous
comparison nor an attempt to bridge them has been conducted.
The present paper attempts to close this gap by deriving the
mathematical formulation that shows the equivalence between
the dq-domain and the sequence domain impedances. A modified
form of the sequence domain impedance matrix is proposed,
and with this definition the stability estimates obtained with the
Generalized Nyquist Criterion (GNC) become equivalent in both
domains. The second contribution of the paper is the definition
of a Mirror Frequency Decoupled (MFD) system. The analysis
of MFD systems is less complex than that of non-MFD systems
because the positive and negative sequences are decoupled. This
paper shows that if a system is incorrectly assumed to be MFD,
this will lead to an erroneous or ambiguous estimation of the
equivalent impedance.
Index Terms—dq-domain, Impedance, Power Electronic Systems, Sequence Domain, Stability Analysis.
I. I NTRODUCTION
The stability of AC power systems with a high penetration
of power electronics is very difficult to analyze. The combination of multiple non-linearities and fast dynamics stemming
from controllers adds significant complexity to the analysis.
Impedance-based analysis of AC power systems is a relevant
and practical tool in this respect because it reduces the system
into a source and load subsystem, and analyses the dynamic
interactions between the two subsystem equivalents [1] [2].
The method is based on existing techniques for DC-systems,
first applied in [3].
A. Rygg and M. Molinas are with the Department of Engineering Cybernetics, Norwegian University of Science and Technology, O. S. Bragstads
plass 2D, 7032 Trondheim, Norway, e-mail: [email protected], tel: +47
95977295
Z. Chen and X. Cai are with the State Energy Smart Grid R&D Center, Shanghai Jiao Tong University, Shanghai, China, e-mail: [email protected]
This method has some highly appealing properties. First,
it considers the subsystems as “black-boxes”, i.e. detailed
knowledge of the parameters and properties of the system
is not required as long as measurements can be obtained
at its terminals. Furthermore, the impedance equivalents can
be extracted based on measured signals in a real system.
The most accurate method for this purpose is based on
frequency scanning [4]-[6]. However, this method requires
advanced and dedicated equipment, and has limited realtime applicability. There are several alternative methods which
can estimate impedance closer to real-time, and with low or
zero additional hardware requirements. Examples are binary
sequence injection [7], impulse response [8], Kalman filtering
[9] and recurrent neural networks [10]. However, the accuracy
of these methods has not been extensively investigated in any
comparative or validation studies.
When an impedance equivalent is established, it can be
used for several purposes. Analytical impedance models for
relatively simple systems were derived in [11]-[14]. System
stability can be assessed based on these models through the
Generalized Nyquist Criterion (GNC) [15]. Other stability
criteria based on impedance models are described in [2],[16][19]. Impedance equivalents have also been verified through
experimental studies [11],[12],[20]-[22].
Previous work in this field can be grouped into two approaches. The first analyzes the system in the sequence domain
using symmetric components (e.g. [1],[7],[8],[23]), whereas
the other applies the synchronous (dq) reference frame (e.g.
[2],[4],[11],[24]). Both domains have certain advantages and
disadvantages, but neither a rigorous comparison nor an attempt to bridge them have yet been conducted. The present
paper shows mathematically how the two impedance domains
are related to each other, and that they can be viewed as
equivalent in terms of stability.
This paper makes two contributions. The first involves the
proposed modified definition of the 2x2 sequence domain
impedance matrix. In this matrix, the positive and negative
sequences are shifted with twice the fundamental frequency.
The coupling between these two frequencies is important
in power electronic systems, and is defined as the mirror
frequency effect. The equivalence between the proposed matrix
and the well-established 2x2 dq domain impedance matrix is
derived, and it is proven that the Generalized Nyquist Criterion
(GNC) estimates is equivalent for both matrices.
2
The second contribution involves the definition of the Mirror
Frequency Decoupled (MFD) system, which is a sufficient
condition to avoid the mirror frequency effect. It is shown
how, in such systems, the impedance matrices become reduced. Furthermore, it is shown that the original definition
of sequence domain impedances [1] is ambiguous unless the
system is MFD.
II. R ELATIONSHIP BETWEEN dq AND SEQUENCE DOMAINS
A. Assumption: Sequence domain balanced systems
This work is based on systems that are sequence domain
balanced. In other words, if a positive sequence current at
an arbitrary frequency is injected anywhere into the system,
there will be no induced negative sequence components at
the same frequency. This assumption has been applied in
almost all stability analyses in previous studies (e.g. [8], [11],
[13]). Although sequence domain unbalanced systems are not
treated in the present paper, the theory explained here can be
extended to cover them. However, the extension of this theory
may create a more complex and abstract representation of the
system.
Even in systems that are sequence-domain balanced, it will
be shown later that a current injection at a given frequency can
induce a voltage shifted by twice the fundamental frequency.
This phenomenon is denoted the mirror frequency effect in this
paper, and will occur in most power systems. Subsystems that
do not contribute to the mirror frequency effect are defined
in this work as Mirror Frequency Decoupled (MFD). This
definition and its applications are presented in section IV.
B. Sequence domain impedance extraction in previous work
In previous work, the sequence domain impedances have
been obtained in two ways. The most complete method takes
into account sequence domain unbalances as defined above
[25]. The resulting impedance relation is similar to that
proposed in this paper (14), but does not consider the mirror
frequency effect defined in section IV. Similarly, although
(14) takes into account mirror frequency effects, it neglects
sequence domain unbalances.
The other, more simple method for obtaining sequence
domain impedance assumes that the positive and negative
sequence are decoupled [1],[7],[12]. This is equivalent to the
following definition, hereafter denoted the original sequence
domain impedance definition:
Vp
Ip
Vn
Zn =
In
Zp =
(1)
where Zp is denoted the positive sequence impedance
and Zn is denoted the negative sequence impedance. This
definition takes into account neither the sequence domain
unbalances nor the mirror frequency effect. Nonetheless, it can
be accurate in many cases, in particular at high frequencies as
will be shown by simulations in section V.
C. Harmonic current and voltage equations
The first step in deriving the impedance relationship between dq and the sequence domain is to relate current and
voltage components to each other in the two domains. This
has been done in previous studies, e.g. [26], but the purpose of
this work was not related with impedance equivalents. Another
study presents an impedance derivation procedure similar to
the one in the present paper [23]. The derivation was here
applied to a case with diagonal dq-impedance matrix, and is
hence a special case of the relations derived in the following
sections.
First, the dq coordinate system is defined according to the
Parks transform:
va
r
2π
2π
2
cos(θ + 3 )
cos(θ)
cos(θ − 3 )
vd
vb
=
2π
vq
3 − sin(θ) − sin(θ − 2π
3 ) − sin(θ + 3 )
vc
(2)
For a given three-phase set of signals, va , vb , vc , the
corresponding dq-domain signals are given by (2). θ is the
transformation angle typically obtained from a Phase Lock
Loop (PLL) or from an oscillator. In steady-state, θ = ω1 t,
where ω1 is the fundamental frequency. (2) is here expressed in
the time-domain, but is valid in the frequency-domain as well.
From now on, all equations are expressed in the frequency
domain. The relationship between a time domain waveform
and its frequency domain components is
v(t) =
∞
X
vk cos(ωk t + φk )
(3)
k=0
where vk is the amplitude of the frequency component at
frequency ωk , and φk is the corresponding phase angle. The
complex number
Vk = vk ejφk
(4)
is defined as the harmonic phasor at frequency ωk . From this
point, the subscript k is omitted, and the relevant frequency
will be indicated.
The sequence domain, also denoted the symmetric component domain, is widely applied in power system analysis
because it can decompose unbalanced three-phase systems into
three balanced and decoupled subsystems:
• A positive sequence subsystem (p)
• A negative sequence subsystem (n)
• A zero sequence subsystem (0)
The effect of zero sequence components is disregarded
in this paper. This is the equivalent of assuming a system
without a neutral wire, as well as neglecting transmission line
capacitance. Note that in most applications of the sequence
domain, only the fundamental frequency components are considered, whereas in this work the phasors can be related to
any frequency. abc-phasors can be related to sequence domain
phasors by the symmetric component transform:
V
Vp
1 a a2 a
Vb
=
(5)
Vn
1 a2 a
Vc
3
where Vp is the positive sequence voltage phasor, and Vn is
the negative sequence voltage phasor at an arbitrary frequency.
2π
a = ej 3 is the complex number corresponding to a 120o
phase shift. It can be shown that sequence domain phasors are
related to dq-domain phasors as follows:
1 Vd + jVq
Vp
, ωp = ωdq + ω1
=√
0
Vn
6
1
Vp
0
=√
, ωn = ωdq − ω1
(6)
Vn
V
−
jVq
d
6
In other words, a general dq voltage phasor at frequency
ωdq is equivalent to two sequence domain voltage phasors at
different frequencies, shifted by the fundamental:
• Positive sequence voltage at ωp = ωdq + ω1
• Negative sequence voltage at ωn = ωdq − ω1
Similarly, it can be shown that for a given positive or negative sequence voltage phasor, the corresponding dq-domain
phasors are:
(7)
By definition, (2)-(7) also applies to currents.
D. Illustration of harmonic phasor relations
Equation (6) is illustrated in Figure 1. A perturbation in a
dq-domain current waveform at a randomly selected frequency
of 80 Hz is modelled as follows:
)
Id = 0.06∠80o
, ωdq = 2π · 80
(8)
Iq = 0.09∠30o
The waveform is expressed in the abc-domain using the
inverse of (2). FFT is used to calculate the frequency domain
harmonic phasors in both domains. These are drawn in the
complex plane in Figure 1. It is seen that the single dqtone at 80 Hz is transformed into a 30 Hz and a 130 Hz
component in the abc-domain. Moreover, it is seen that the
30 Hz waveforms are pure negative sequence, whereas the 130
Hz waveforms are pure positive sequence, which is consistent
with (6). The sequence domain phasors are calculated from
the dq-domain phasors by this equation. Note also that the
sequence domain phasors satisfies the following equation for
balanced abc-components: Iap = Ip and Ian = In .
6
1
In
Zdq
6
Ip
Id
r
3
1
Vd
Vp
, ωdq = ωp − ω1
=
−j
Vq
2
r
3
Vd
1
Vn
, ωdq = ωn + ω1
=
j
Vq
2
1
Ip
Iq
I d jI q
I d jI q
In
2S 80
Zp
2S 130
Zn
2S 30
Ibn
I an
Ian
Icn
Zn
In
2S 30
I ap
Ip
Iap
Ibp
Icp
Zp
2S 130
Fig. 1: Harmonic phasors corresponding to (8). The 80 Hzwaveform at ωdq equals the sum of a positive sequence
waveform at ωp = ωdq +ω1 and a negative sequence waveform
at ωn = ωdq − ω1
TABLE I: Overview of currents, voltages, and impedances and
the corresponding frequencies at which they are defined
Parameter
Vd / I d
Vq / Iq
Zdq
Vp / I p
Vn / I n
Zpp
Znn
Zpn
Znp
Zp
Zn
Frequency
ωdq
ωdq
ωdq
ωp = ωdq + ω1
ωn = ωdq − ω1
ωp
ωn
ωn → ωp
ωp → ωn
ωp
ωn
To provide a better overview, all current, voltages and
impedances are summarized in Table I. These definitions
highlight a key point in the paper: positive and negative
sequences are defined at two different frequencies, shifted by
twice the fundamental frequency. The two frequencies are also
denoted mirror frequencies later in the paper.
A. Definition and fundamental equations
DEFINITION
The basis for the following derivations is the generalized
Ohms law in the dq-domain:
Vd
Zdd Zdq Id
I
=
= Zdq d
(9)
Vq
Zqd Zqq Iq
Iq
In the previous section, current and voltage phasors were
considered separately. This section will relate them to each
other through impedance. The main contribution of the paper
is then outlined, which is a 2x2 impedance matrix composed
of positive and negative sequence impedances at two different
frequencies. The proposed 2x2 matrix has many appealing
properties, as will be highlighted in further sections.
where Zdq is a 2x2 matrix of complex numbers as a
function of frequency. This equation will now be transformed
into the sequence domain.
As shown by (7), any set of dq-domain phasors can be
written as the sum of positive sequence phasors at ωdq + ω1
and negative sequence phasors at ωdq − ω1 . Substituting from
(7) gives a modified version of (9):
III. M ODIFIED SEQUENCE DOMAIN IMPEDANCE
4
mirror frequencies
dq
n dq 1
p dq 1
frequency
S S
Z dq
Iq
L
I dS Vd I d
S
Z dd
L L
Z dq
Iq
S S
Z qd
Id
L
Z dd
I inj , d
L
I qS Vq I q
S
Z qq
L L
Z qd
Id
I inj ,q
L
Z qq
dq domain
sequence domain
S S
Z np
Ip
L
I nS Vn I n
L L
Z np
Ip
I nS I nL
S
Z nn
S S
Z pn
In
NB: positive sequence shunt
current injection assumed
L
Z nn
S
Z pp
L
I pS V p I p
L L
Z pn
In
I inj , p
L
Z pp
Fig. 2: Circuit equivalents of impedance equations in the dq and sequence domain. Mirror frequency coupling illustrated by
voltage sources in the sequence domain.
Zpn : Measures the positive sequence voltage at ωp induced by negative sequence current at ωn
• Znp : Measures the negative sequence voltage at ωn
induced by positive sequence current at ωp
• Znn : Measures the negative sequence voltage at ωn
induced by a negative sequence current at ωn
(13) can be rewritten in a more compact form as a linear
transformation by the unitary matrix AZ :
•
Vp
1
1
1
1
+ Vn
= Zdq Ip
+ In
−j
j
−j
j
(10)
The definitions shown in Table I are based on this convenient representation.
The next step is to multiply (10)
with 12 1 j and 12 1 −j , leading to the following two
equations:
1
1
1
1 j Zdq
1
Vp = Ip
+ In
−j
2
2
1
1
1
1 −j Zdq
1
Vn = Ip
+ In
−j
2
2
1
j Zdq
j
1
−j Zdq
j
(11)
(11) can conveniently be rewritten with matrix notation by
defining the modified sequence domain impedance matrix:
Vp
Zpp Zpn Ip
I
=
= Zpn p
(12)
Vn
Znp Znn In
In
with the four impedances substituted from (11):
Zpn
1
1 j Zdq
1
−j
=
1
2
1 −j Zdq
−j
1
1 j Zdq
j (13)
1
1 −j Zdq
j
The impedances have the following physical interpretation:
• Zpp : Measures the positive sequence voltage at ωp induced by a positive sequence current at ωp
Zpn = AZ · Zdq · A−1
Z
Zdq = A−1
Z · Zpn · AZ
1
j
1 1
−1
∗
, AZ = AZ = √
(14)
−j
2 −j j
where * denotes complex conjugate transpose.
The corresponding admittance equations can be obtained in
the same way by interchanging voltages with currents in the
derivation process:
1 1
AZ = √
2 1
Ypn = AZ · Ydq · A−1
Z
Ydq = A−1
Z · Ypn · AZ
(15)
The generalized Ohms law and the mirror frequency effect
are illustrated as circuit equivalents in Figure 2. The figure
assumes positive sequence shunt current injection at frequency
ωp , but corresponding equivalents can be made for other
injection choices. In the dq-domain, all signals are expressed
using the same frequency ωdq . The cross coupling between
the d- and q-axis is represented by current dependent voltage
sources.
5
TABLE II: Example of dq- and sequence domain frequencies
∗
ωdq
ωp
ωn
500
550
450
120
170
70
65
115
15
40
90
-10→10∗
15
65
-35→35∗
negative sequence frequencies where the impedances are transformed to
positive sequence as explained in section III-B
B. Positive sequence impedances below the fundamental frequency
In this paper, the modified sequence domain impedance
definition has not thus far been defined for positive sequence
at frequencies below the fundamental frequency. Given that
ωp = ωdq + ω1 , the dq-domain frequency ωdq will be negative
for 0 ≤ ωp < ω1 . To extend the impedance definition to
positive sequence below the fundamental frequency, it is first
important to note that a balanced three-phase signal with a
negative frequency is equivalent to a balanced three-phase
signal with positive frequency at the same absolute value. Only
the phase order needs to be changed, i.e. a positive sequence
signal becomes a negative sequence, and vice versa.
Secondly, note that when 0 ≤ ωdq < ω1 , the negative
sequence frequency ωn = ωdq − ω1 is negative. Consequently,
the negative sequence impedance Znn is associated with a
negative frequency. Based on the discussion above, the negative sequence impedance at a negative frequency is equal to
the positive sequence impedance at a positive frequency with
the same absolute value. Thus, a positive sequence impedance
below the fundamental frequency can be defined.
To summarize, an example of dq-domain and corresponding
sequence domain frequencies are given in Table II.
C. Nyquist stability criterion equivalence
The Generalized Nyquist stability Criterion (GNC) has been
widely applied in previous studies to analyze the stability of
power electronics systems (e.g. [13],[18],[21]). The criterion
is mathematically formulated in [15], and was first applied to
AC power electronic systems in [2]. It will now be shown that
when the GNC is applied to the dq and the modified sequence
domains, the results are identical.
Assuming the dq-domain, the basis for the stability criterion
is the system transfer function between source and load:
S
L
h = inv(I + Zdq
Ydq
)
(16)
For convenience, the minor-loop gain Ldq is defined as:
S
L
Ldq = Zdq
Ydq
−1
S
L
= AZ · Zdq
Ydq
· A−1
Z = AZ · Ldq · AZ
(19)
The following equations prove that λdq , the eigenvalues of
Ldq , are equal to λpn , the eigenvalues of Lpn .
0 = det (Lpn − λpn I) = det AZ · Ldq · A−1
Z − λpn I
−1
= det AZ · Ldq · AZ − AZ · (λpn I) · A−1
Z
= det(AZ ) · det (Ldq − λpn I) · det A−1
Z
= det (Ldq − λpn I) = 0
⇒ λpn = λdq
(20)
Consequently, the stability analysis by GNC gives identical
results in the dq- and sequence domains when the modified
definition (14) is applied. By contrast, if the original definition
(1) is used, and one or both subsystems are not mirror
frequency decoupled (see section IV), the calculated stability
is different from the dq-domain stability calculations.
IV. T HE MIRROR FREQUENCY EFFECT
The term mirror frequency effect is defined in this paper
to provide further insight into the properties of impedance
models. When subjected to a harmonic disturbance at a
given frequency, power electronic converters will respond
with induced current/voltages at the same frequency, but also
at the so-called mirror frequency. The mirror frequency is
shifted with twice the fundamental frequency, and the direction
depends on whether the disturbance is a positive or negative
sequence. This effect complicates the analysis of the system
because it violates the assumption of linearity. However, when
a system is assumed to be sequence domain balanced as
discussed in section II, linearity can be regained by either
•
•
applying the dq-domain
applying the modified sequence domain defined by (14)
Of note, sequence domain balanced systems may still
contain the mirror frequency effect as these two properties
are independent. The remainder of this section will focus
on sequence domain balanced systems lacking the mirror
frequency effect. Such systems are hereafter denoted Mirror
Frequency Decoupled (MFD).
(17)
The eigenvalues of Ldq can be found by solving:
det (Ldq + λdq I) = 0
The minor-loop gain in the sequence domain can be expressed as:
S
L
S
L
Lpn = Zpn
Ypn
= AZ · Zdq
· A−1
AZ · Ydq
· A−1
Z
Z
A. Mirror Frequency Decoupled (MFD) systems
(18)
Assume in the following that the source is stable when
connected to an ideal load, and that the load is stable when
connected to an ideal source. The GNC then states that the
system is stable if and only if the characteristic loci of Ldq
do not encircle the point (−1, 0) when drawn in the complex
plane.
A subsystem is said to be Mirror Frequency Decoupled
(MFD) if, when subjected to a harmonic disturbance at an
arbitrary frequency, it only responds with current/voltages
at the same frequency. With reference to Figure 2, this is
equivalent to removing the current-dependent voltage sources
in the sequence domain. In other words, Zpn = Znp = 0. MFD
systems have several interesting properties presented below.
6
B. Impedance matrices of MFD systems
MFD subsystems have sequence- and dq-domain impedance
matrices of the following form:
if the reluctance in d- and q-axes differ. Although mirror
frequency coupling is independent of power electronics, it
is clear from the bullet points above that power electronics
systems introduces many instances for this to occur.
Zpn
Zdq
MF D
MF D
Zpp
0
=
0
Znn
Zx Zy
=
−Zy Zx
V. VALIDATION BY NUMERIC SIMULATION
A. Obtaining impedances through simulation
(21)
where Zpn = Znp = 0 by the definition in the previous
section. The dq-domain matrix is skew symmetric with Zdd =
Zqq = Zx and Zdq = −Zqd = Zy . The proof of these relations
is presented in Appendix C. In a previous study a similar result
was found under the assumption of Zdq = Zqd = 0 [23].
Since Zpn is diagonal, the original impedance definition (1)
is equivalent to the modified (14) for MFD-systems.
C. dq impedance extraction in MFD systems
It has been argued that sequence domain impedances are
easier to obtain than dq-domain impedances due to the decoupling between positive and negative sequence [1],[12].
Furthermore, the sequence domain impedance can be obtained
from a single measurement with no need for matrix inversions
because of this decoupling. However, it has been shown
in the previous section that sequence domain impedances
can been assumed to be decoupled only if the subsystem
is MFD. In contrast with the statements of previous work,
the following equations show that dq-domain impedances can
also be obtained from a single measurement in this case.
Combining (9) with (21) gives:
Vd Id + Vq Iq
Id2 + Iq2
Vd Iq − Vq Id
=
Id2 + Iq2
Zx = Zdd = Zqq =
Zy = Zdq = −Zqd
(22)
Consequently, only a single measurement is needed to
obtain the dq impedance matrix in a MFD subsystem.
D. Sources to mirror frequency coupling
Mirror frequency coupling is introduced in all parts of the
power system where (21) is not satisfied. For example:
• Phase-Lock-Loops (PLL)
• Converter current controllers with unequal structure
and/or parameter values in the d- and q-axis
• DC-link voltage control systems
• Active and reactive power controllers
• Salient-pole synchronous machines
The analytical impedance calculation in the dq-domain is
described in e.g. [11], where the coupling related to the first
four bullet-points can be identified. Note that all transfer
functions must be identical in d- and q-axis in order for the
subsystem to be MFD. Furthermore, all cross-coupling between d- and q-axis must have opposite sign. The synchronous
machine is also mentioned because it is a vital part of many
power systems, and it possesses mirror frequency coupling
The method for obtaining impedances through simulation
is best explained by the flowchart in Figure 3. This method
is able to calculate both the dq-domain and sequence domain
impedances in an integrated process. The first step is to select
a vector of frequencies fdq,tab , i.e. the frequencies at which
the impedances shall be calculated. Note that these frequencies
are expressed in the dq-domain.
The system can be simulated under either shunt current
or series voltage injection. The difference between these two
methods is illustrated in Figure 4. If shunt current is used, the
following three-phase perturbation signals will be injected:
sin ([ωinj + ω1 ] t)
iinj1 = Iinj sin [ωinj + ω1 ] t − 2π
3
sin [ωinj + ω1 ] t + 2π
3
sin ([ωinj − ω1 ] t)
(23)
iinj2 = Iinj sin [ωinj − ω1 ] t + 2π
3
2π
sin [ωinj − ω1 ] t − 3
If series voltage injection is applied, i can be replaced with
v. The two sets of signals need to have different frequencies
because linear independent injections are required when solving for the impedance matrices (24)-(25). The selection of
injection signals is discussed in [4].
The needed output from simulations are the current and
voltage signals shown in Figure 4. Note that the injection
signal itself is not needed in impedance calculations. After
converting time-domain signals to the frequency domain as
described in the flowchart, the following equations can be used
to find the impedances in the two domains:
−1
Zdd Zdq
V
Vd2 Id1 Id2
= d1
Zqd Zqq
Vq1 Vq2 Iq1 Iq2
−1
Zpp Zpn
V
Vp2 Ip1 Ip2
= p1
Znp Znn
Vn1 Vn2 In1 In2
(24)
(25)
After these two matrices are established, all other impedance
expressions in the paper can be derived based on them.
B. Case study description
In this section, the validity of the previously derived expressions is checked through numeric simulations. Simulation
cases A and B are developed in MATLAB/Simulink, see
Figure 5 and Figure 6. Both cases consist of a source converter
and a load converter. The control systems operates in the dqdomain.
In Case A the source converter controls the voltage v
according to set-points and the virtual inductances Lvd and
Lvq . The converter is synchronized to the fixed clock signal
θS = 2πfn · t. The load converter operates with DC-voltage
7
control and reactive current control, and a current source Idc
consumes power at the DC-side. The converter is synchronized
to the grid by a Phase Lock Loop (PLL).
Select f dq ,tab
For f dq in f dq ,tab
Sim with
Sim with
iinj iinj1 iinj iinj 2
or
or
vinj vinj1 vinj vinj 2
Store iS1 , iL1 , vS1 , v L1
iS 2 , iL 2 , vS 2 , v L 2
Transform to dqdomain
Perform FFT (dq)
Perform FFT (abc)
Extract components
at f f dq
Extract components at
at f f dq f1
Calculate dqimpedance at f dq
Calculate pnimpedance at f dq
More frequencies?
Y
N
The following sources to mirror frequency coupling is
present in Case A source subsystem:
• Lvd 6= Lvq
• Kpvd 6= Kpvq
• Tivd 6= Tivq
In Case A load subsystem, the following sources to mirror
frequency coupling is present:
• DC-link voltage controller
• PLL is connected to a non-stiff point in the grid
• Kpid 6= Kpiq
• Tpid 6= Tpiq
To better illustrate the findings in the paper, Case A has been
divided into two subcases, Case A1 and Case A2. In Case A1
both subsystems are mirror frequency coupled according to the
bullet-lists above. However, in Case A2 all mirror frequency
couplings in the source subsystem are removed by setting
Lvd = Lvq , Kpvd = Kpvq and Tivd = Tivq . Hence, in Case
A2 the source subsystem is MFD, while the load subsystem is
not. The differences between these two cases will be discussed
in the result section.
In Case B all mirror frequency coupling is removed in both
subsystems, leading to a complete MFD system. The source
subsystem is identical to the one in Case A2. In the load
subsystem the converter is connected to a constant voltage
at the DC-side, which eliminates the need for DC voltage
control. The control system consists of current controllers
with set-points i∗Ld and i∗Ld . The PI-controller parameters are
identical in d- and q-axis. Furthermore, the converter does
not contain a PLL, but is instead synchronized to the fixed
ramp θL = 2πfn · t in the same way as the source converter.
Consequently, all mirror frequency coupling sources from the
above bullet-list have been removed.
Parameter values applied in the simulation cases are given
in Appendix A.
end
C. Simulation results - Zdq and Zpn
Fig. 3: Illustration of simulation method to obtain both dq and
sequence domain impedances as a function of frequency. The
flowchart is valid for both shunt current and series voltage
injection.
=
VSdRth
LIthSd
V Z S,dq I
Sq
Sq
iS v
iinj
Source subsystem
VSdRth
LIthSd
Z
V
S,dq
Sq
I Sq
Source subsystem
i vS
vinj
vL
I Ld
iL VLd
V Z L,dq I
Lq
Lq
Load subsystem
VLd
I Ld
V Z L,dq I
Lq
Lq
Load subsystem
Fig. 4: Illustration of the two injection methods (shunt and
series) for a general source and load subsystem.
The resulting impedance curves for the three cases are
shown in Figures 7-9 for the dq-domain, and in Figures
10-12 for the modified sequence domain. Only magnitude
is presented in these figures for simplicity, but is has been
verified that the angles are consistent with the conclusions.
Impedances for the load and source subsystem are plotted in
the same graph. In the sequence domain plots, impedances
have been obtained in two ways, denoted by subscript a and
b:
• Za : Direct simulation of Zpn using (25)
• Zb : Based on simulated Zdq from (24) and the transform
given by by Zpn = AZ · Zdq · A−1
Z (14)
The following observations support the claims in the previous sections:
• The two ways of obtaining modified sequence domain
impedances (Za and Zb ) produce identical results in all
cases. This confirms the transformation relationship from
dq- to sequence domain (14).
• In Case A1, there are no symmetries in the impedance
curves, neither in dq nor sequence domain. This is
8
Source subsystem
ZS
=
VSdc
Load subsystem
ZL
v
iS
=
iL
+
vd*
+
-
vd isq Lvq
vq*
+
K pvd
-
++
-
iSd
vq
K pvq
Cdc
d L , abc
d S , abc
Power circuit
S 2 f n t
1 Tivd s
Tivd s
d S , abc
dq
1 Tivq s
abc
Tivq s
Lvd
Source converter control
vabc
*
VLdc
+
K pvdc
-
VLdc
I dc
VLdc
1 Tivdc s
+
Tivdc s i Ld
*
iLq
K piq
L
1 Tiid s
Tiid s
K pid
+-
PLL
d L , abc
dq
1 Tiiq s
abc
Tiiq s
Load converter
control
iLq
Fig. 5: Detailed schematic of Case A
Source subsystem Load subsystem
ZS
=
VSdc
ZL
v
iS
=
iL
VLdc
d L , abc
d S , abc
Power circuit
vd*
+
+
-
vd isq Lvq
vq*
+
-
vq
iSd
K pvd
++
K pvq
Lvd
*
iLd
S 2 f n t
d S , abc
dq
1 Tivq s
abc
Tivq s
Source converter control
+-
K pid
iLd
*
iLq
1 Tivd s
Tivd s
-
+-
iLq
K piq
1 Tiid s
Tiid s
1 Tiiq s
Tiiq s
L 2 f n t
dq
d L , abc
abc
Load converter
control
Fig. 6: Detailed schematic of Case B
•
•
expected since both subsystems have mirror frequency
coupling.
S
S
In Case A2, Zpn
≈ Znp
≈ 0, as expected since the
source subsystem is MFD (see also (21)).
In Case B, Zpn ≈ Znp ≈ 0 for both subsystems. This
is expected since both subsystems are MFD. It can also
be observed that Zdq = −Zqd and Zdd = Zqq for both
subsystems, again according to (21).
9
10
10 0
ZL
-2
10
|Zpp|
Magnitude [+]
10 0
Z
10 1
Magnitude [+]
|Zdq|
S
10 2
10
-2
10 3
|Zqd|
0
10 1
10
10 -2
10 2
|Zqq|
0
10 -2
10 1
10 2
10 3
10 1
Frequency [Hz]
10 2
Magnitude [+]
S
10 2
10
-2
10 3
|Zqd|
10 1
10 0
10 0
10 -2
10 -2
10 2
|Zqq|
10 -2
10 -2
1
10 2
10 3
|Znn|
10 0
10
2
10
3
10
1
10
2
10
3
Frequency [Hz]
Fig. 10: Case A1 sequence domain impedances for both
subsystems. a and b indicate two calculation methods.
|Zpn|
10 0
10
10 0
ZL
ZS
a
L
Zb
-2
10 1
10 3
1
10
2
10
3
10
Frequency [Hz]
1
10
2
a
S
10
Zb
10 2
-2
ZS
:0
pn
10 3
10 1
|Znp|
10
10 2
10 3
|Znn|
0
10
10 -2
0
10 -2
Z S :0
10
Magnitude [+]
1
ZL
Z
10 1
10 2
|Zqd|
0.01
S
10 3
1
10 1
10
10 2
|Zqq|
0
10 -2
0.01
10 1
10 2
Frequency [Hz]
10 3
10 1
10 2
10
3
10
1
10 3
Frequency [Hz]
Fig. 9: Case B dq-domain impedances for both subsystems
10
2
10
3
Frequency [Hz]
|Zpn|
10 0
10 0
ZL
10 -2
a
ZLb
10 1
10 3
10 2
ZS
a
ZS
b
10 -2
10 3
ZS
:Z Lpn:0
pn
10 1
|Znp|
Magnitude [+]
-2
2
|Zpp|
dq
0
10
Fig. 11: Case A2 sequence domain impedances for both
subsystems. a and b indicate two calculation methods.
|Z |
dd
10
1
Frequency [Hz]
Frequency [Hz]
|Z |
10
10
3
Fig. 8: Case A2 dq-domain impedances for both subsystems
Magnitude [+]
10 1
np
10
Magnitude [+]
10 3
Frequency [Hz]
Magnitude [+]
Z
10 1
10 2
10 -2
10 0
10
Magnitude [+]
Magnitude [+]
ZL
ZS
b
|Zpp|
10 0
-2
ZS
a
ZLb
0
|Znp|
|Zdq|
10 0
ZLa
10 1
Frequency [Hz]
|Zdd|
10
10 -2
10 3
Fig. 7: Case A1 dq-domain impedances for both subsystems
10
10
|Zpn|
0
10 3
Magnitude [+]
Magnitude [+]
|Zdd|
10 2
10 3
10 2
10 3
|Znn|
10 0
10 0
10 -2
10 -2
ZS
: ZLnp: 0
np
10 1
10 2
Frequency [Hz]
10 3
10 1
Frequency [Hz]
Fig. 12: Case B sequence domain impedances for both subsystems. a and b indicate two calculation methods.
10
Zp,shunt,a
10 1
Zp,shunt,b
Zp,series,a
10 0
Zp,series,b
Z
10 -1
10 2
10 3
Zn,shunt,a
10 1
Z
10
10
pp,a
Zpp,b
10 1
Magnitude [+]
The purpose of this section is to investigate the validity of
the equations derived in Appendix D, and to determine how
the original and modified sequence domain impedances relate
to each other. Figure 13-15 presents the same comparison for
each of the three simulation cases. Only the load subsystem
impedance is included in the comparison. The original sequence domain impedances, Zp and Zn , are compared with the
diagonal elements in the modified sequence domain impedance
matrix, Zpp and Znn . Furthermore, the original sequence
domain impedances are simulated both by shunt current and
series voltage injection. Additionally, they are estimated in two
ways to validate the equations derived in Appendix D. Table
III summarizes the notation and estimation methods.
Magnitude [+]
D. Simulation results - Modified vs. original
n,shunt,b
Zn,series,a
0
Zn,series,b
Znn,a
-1
10
Znn,b
1
10
2
10
3
Frequency [Hz]
ZpL
Zp,shunt,b
ZpL
Zp,series,a
ZpL
Direct simulation using (1)
Calculation based on simulated Zpn
from (25) and the formula (30)
Direct simulation using (1)
Calculation based on simulated Zpn
from (25) and the formula (33)
Direct simulation using (25)
Calculation based simulated Zdq
from (24) and the formula (14)
Zp,series,b
ZpL
Zpp,a
L
Zpp
Zpp,b
L
Zpp
shunt
shunt
series
series
First, it can be noted that the methods based on direct
simulations always overlap with the calculated one for all three
cases, which validates (30), (33) and (14).
In Case A1, the original impedances obtained by shunt
and series injection are not equal. This is expected based
on the difference between (30) and (33). The difference is
noticeable at frequencies below 100 Hz. It is also observed
that the modified sequence impedances Zpp and Znn deviate
substantially from the original ones at frequencies up to
≈ 500Hz; however, after this point they are close to equal.
Consequently, the load subsystem can be assumed to be MFD
for frequencies above 500 Hz.
In Case A2, it is clear that the load subsystem impedances
obtained from shunt and series injection are equal. This is
consistent with (35), because the source subsystem is MFD in
this case. The same equation give Zpp 6= Zp and Znn 6= Zn ,
which is verified by the figure. However, the difference is close
to zero at frequencies above ≈ 500Hz, similar to Case A1.
In Case B, all impedance estimates coincide. From (36) we
have that Zp = Zpp and Zn = Znn whenever Zpn = Znp = 0.
Both systems are now MFD, and hence the assumption of
decoupled sequence domain is valid.
Magnitude [+]
Zp,shunt,a
Fig. 13: Comparison of load subsystem original and modified
sequence domain impedances in Case A1
Zp,shunt,a
10 1
Zp,shunt,b
10
Zp,series,a
0
Zp,series,b
Zpp,a
10 -1
10
Magnitude [+]
Estimation mehtod
Zpp,b
1
10
2
10
3
Zn,shunt,a
10 1
Zn,shunt,b
10
Zn,series,a
0
Zn,series,b
Znn,a
10 -1
Znn,b
10 1
10 2
10 3
Frequency [Hz]
Fig. 14: Comparison of load subsystem original and modified
sequence domain impedances in Case A2
Magnitude [+]
Eq. notation
Legend notation
Zp,shunt,a
10
Zp,shunt,b
Zp,series,a
1
Zp,series,b
Zpp,a
0.1
10
Magnitude [+]
TABLE III: Explanation of the legends in Figures 13-15
Zpp,b
1
10
2
10
3
Zn,shunt,a
10
Zn,shunt,b
Zn,series,a
1
Zn,series,b
Znn,a
0.1
10
Znn,b
1
10
2
10
3
Frequency [Hz]
Fig. 15: Comparison of load subsystem original and modified
sequence domain impedances in Case B
E. Simulation results - Generalized Nyquist Criterion
It has been shown in section III-C that the Generalized
Nyquist Criterion will give same result in the dq-domain and
11
the modified sequence domain. Furthermore, it has been shown
that the stability analysis based on original sequence domain
impedances will give different results unless both subsystems
are MFD. This has been investigated by applying the GNC
to the impedance curves found in Figure 7 to Figure 15. The
resulting Nyquist plots are presented in Figure 16. Note that
only the most critical eigenvalue is plotted, corresponding to
the operating point of isd = 1.1pu.
In all cases, the dq-domain gives exactly the same result as
the modified sequence domain. On the other hand, the original
sequence domain impedances do not give the same Nyquist
plot in Case A1 and A2. In Case B all methods give the same
Nyquist curves.
F. Time-domain analysis
To complement the stability analysis from the previous
section, a time-domain simulation has been conducted and
presented in Figure 17. Cases A1 and A2 are used for the
simulation, see Figure 5 for the system block diagram. The
power consumed by the load converter is stepwise increased
by applying step changes to Idc . Figure 17 shows the resulting
dq-currents at the source converter. The transient oscillations
in both isd and isq are gradually increasing until instability
occurs at the reference value of 1.2 p.u.. Note that a reference
value of 1.1 p.u. is the basis for the Nyquist plot in Figure
16 as indicated by arrows. This operation point has poorly
damped oscillations, supporting the fact that the Nyquist plot
is close to encircling the point (−1, 0). It can be observed that
Case A2 has stronger oscillations, which is also indicated by
its Nyquist plot being closer to (−1, 0) compared to Case A1.
Grid converter current [p.u.]
Case A1
Grid converter current [p.u.]
0
0
Case A2
1.2
isd
1
isq
0.8
Case A1
0.6
0.4
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
1.2
1.4
1.6
1.8
2
Time [s]
1.2
isd
1
isq
0.8
Case A2
0.6
0.4
0
0.2
0.4
0.6
0.8
1
Time [s]
0
Fig. 17: Time-domain analysis of Case A1 and Case A2. Grid
converter dq-currents during stepwise increase in Idc . Arrows
indicate the operation point used for impedance calculations.
0
Case B
Im
0.2
0
6
6
-0.2
dq
pn,modified
6pn,original,shunt
6pn,original,series
Unity Circle
-0.2
0
0.2
Re
Fig. 16: Comparison of Nyquist plots for cases A1, A2 and B
VI. D ISCUSSION
After the derivations and proofs presented in the previous
sections it is useful to discuss the advantages and disadvantages of the two impedance domains. In this discussion, the
modified sequence domain is assumed because the original
definition has been shown to be ambiguous and can be
inaccurate for systems that are not MFD.
The general statement is that the two domains are equivalent. The impedance matrix in one domain can be obtained
from the other through a linear transformation where the eigenvalues are preserved. Hence, both domains should provide
the same results. However, from a more practical viewpoint,
there are a few advantages with the sequence domain over dq,
especially when impedances are obtained from simulation or
measurements:
1) There is no need to perform dq-transformations to any
measured signal, and hence no need for a reference
transformation angle.
12
2) The off-diagonal terms in the sequence domain
impedance matrix will often have low values, and are
equal to zero for a MFD system. In other words, the
sequence domain impedance matrix is “close to decoupled”.
3) It can be argued that the sequence domain can be
intuitively associated to a physical meaning and is less
abstract than the dq-domain.
The main advantage with the dq-domain is that most previous studies on the control and stability of power electronic
converters have been performed in dq-coordinates. Regarding
analytical impedance models, it is also possible that the
derivations are less complex in the dq-domain, but this has
not been yet extensively investigated.
In most of the previous work on stability analysis based
on the Nyquist criterion, it has been argued that off-diagonal
impedance matrix elements can be neglected. Based on the
impedance matrix structure for MFD systems (21), it is clear
that such simplifications are correct in the sequence domain
but not in the dq-domain.
The paper has shown that original sequence domain
impedances depends on injection type (shunt vs. series) when
both subsystems are non-MFD. In general, the load subsystem
prefer series injection, while the source subsystem prefer shunt
injection. This is due to the fact that the load subsystem
has higher impedance at most frequencies. When choosing
injection type for a given system, the system with more mirror
frequency coupling should be prioritized. It is expected that
this is often the load subsystem, in this case series injection
is more accurate.
VII. C ONCLUSION
Stability analysis of AC power electronics systems through
frequency dependent impedance equivalents is a relatively
new field of research. Both dq-domain and sequence domain
analysis have been reported in previous works. However,
limited effort has been dedicated to the understanding of their
equivalence with respect to the stability estimates they provide.
This paper attempts to contribute in this direction by reporting
the following findings:
1) A modified definition for the sequence domain
impedance matrix, which extends the original sequence
domain impedance definition. The extension is related
with the ability to account for induced frequency components shifted by twice the fundamental frequency
2) The relationship between the well-established dqdomain impedance matrix and the modified sequence
domain impedance matrix. This can be viewed as a
linear transformation where many essential properties
are preserved.
3) The choice of impedance domain does not affect the
analysis of stability using the Generalized Nyquist Criterion (GNC).
4) Definition of the terms mirror frequency effect and
Mirror Frequency Decoupled (MFD) systems.
5) The modified sequence domain impedance matrix is
diagonal for MFD systems. The dq-domain impedance
matrix is skew symmetric.
6) The dq-domain impedance matrix can be obtained from
single measurements without the need for matrix inversions when both subsystems are MFD.
7) The original sequence domain impedance is shown to be
ambiguous in the general case; it depends on injection
type, and the source state variables appear in the load
impedance expression, and vice versa.
8) The relationship between the original and the modified
sequence domain impedance was derived under the
assumption of both ideal series and ideal shunt injection.
9) The original sequence domain impedance no longer
depends on injection type when one of the subsystems
is MFD. However, the source state variables still appear
in the load impedance expression (assuming the source
is MFD).
All equations were derived mathematically on a general
basis, and were verified by numeric simulations. Points 1-5
were examined in simulation section V-C, whereas points 6-9
were examined in simulation section V-D.
13
R EFERENCES
[1] J. Sun, “Small-signal methods for ac distributed power systems;a review,” Power Electronics, IEEE Transactions on,
vol. 24, no. 11, pp. 2545–2554, Nov 2009.
[2] M. Belkhayat, Stability criteria for AC power systems with
regulated loads. Purdue University, 1997.
[3] R. D. Middlebrook, “Input filter considerations in design and
application of switching regulators,” in IEEE Industry Applications Society Annual Meeting, 1976.
[4] G. Francis, R. Burgos, D. Boroyevich, F. Wang, and K. Karimi,
“An algorithm and implementation system for measuring
impedance in the d-q domain,” in Energy Conversion Congress
and Exposition (ECCE), 2011 IEEE, Sept 2011, pp. 3221–3228.
[5] Y. Familiant, J. Huang, K. Corzine, and M. Belkhayat, “New
techniques for measuring impedance characteristics of threephase ac power systems,” Power Electronics, IEEE Transactions
on, vol. 24, no. 7, pp. 1802–1810, July 2009.
[6] J. Huang, K. Corzine, and M. Belkhayat, “Small-signal
impedance measurement of power-electronics-based ac power
systems using line-to-line current injection,” Power Electronics,
IEEE Transactions on, vol. 24, no. 2, pp. 445–455, Feb 2009.
[7] T. Roinila, M. Vilkko, and J. Sun, “Online grid impedance
measurement using discrete-interval binary sequence injection,”
Emerging and Selected Topics in Power Electronics, IEEE
Journal of, vol. 2, no. 4, pp. 985–993, Dec 2014.
[8] M. Cespedes and J. Sun, “Adaptive control of grid-connected
inverters based on online grid impedance measurements,” Sustainable Energy, IEEE Transactions on, vol. 5, no. 2, pp. 516–
523, April 2014.
[9] N. Hoffmann and F. Fuchs, “Minimal invasive equivalent grid
impedance estimation in inductive-resistive power networks
using extended kalman filter,” Power Electronics, IEEE Transactions on, vol. 29, no. 2, pp. 631–641, Feb 2014.
[10] P. Xiao, G. Venayagamoorthy, K. Corzine, and J. Huang, “Recurrent neural networks based impedance measurement technique for power electronic systems,” Power Electronics, IEEE
Transactions on, vol. 25, no. 2, pp. 382–390, Feb 2010.
[11] B. Wen, D. Dong, D. Boroyevich, R. Burgos, P. Mattavelli, and
Z. Shen, “Impedance-based analysis of grid-synchronization stability for three-phase paralleled converters,” Power Electronics,
IEEE Transactions on, vol. PP, no. 99, pp. 1–1, 2015.
[12] M. Cespedes and J. Sun, “Impedance modeling and analysis of
grid-connected voltage-source converters,” Power Electronics,
IEEE Transactions on, vol. 29, no. 3, pp. 1254–1261, March
2014.
[13] X. Wang, F. Blaabjerg, and W. Wu, “Modeling and analysis
of harmonic stability in an ac power-electronics-based power
system,” 2014.
[14] R. Turner, S. Walton, and R. Duke, “A case study on the
application of the nyquist stability criterion as applied to interconnected loads and sources on grids,” Industrial Electronics,
IEEE Transactions on, vol. 60, no. 7, pp. 2740–2749, July 2013.
[15] C. Desoer and Y.-T. Wang, “On the generalized nyquist stability
criterion,” Automatic Control, IEEE Transactions on, vol. 25,
no. 2, pp. 187–196, Apr 1980.
[16] Z. Liu, J. Liu, W. Bao, and Y. Zhao, “Infinity-norm of
impedance-based stability criterion for three-phase ac distributed power systems with constant power loads,” Power
Electronics, IEEE Transactions on, vol. 30, no. 6, pp. 3030–
3043, June 2015.
[17] C. Wildrick, F. Lee, B. Cho, and B. Choi, “A method of defining
the load impedance specification for a stable distributed power
system,” Power Electronics, IEEE Transactions on, vol. 10,
no. 3, pp. 280–285, May 1995.
[18] J. Sun, “Impedance-based stability criterion for grid-connected
inverters,” Power Electronics, IEEE Transactions on, vol. 26,
no. 11, pp. 3075–3078, Nov 2011.
[19] R. Burgos, D. Boroyevich, F. Wang, K. Karimi, and G. Francis,
“On the ac stability of high power factor three-phase rectifiers,”
[20]
[21]
[22]
[23]
[24]
[25]
[26]
in Energy Conversion Congress and Exposition (ECCE), 2010
IEEE. IEEE, 2010, pp. 2047–2054.
D. Dong, B. Wen, D. Boroyevich, P. Mattavelli, and Y. Xue,
“Analysis of phase-locked loop low-frequency stability in threephase grid-connected power converters considering impedance
interactions,” Industrial Electronics, IEEE Transactions on,
vol. 62, no. 1, pp. 310–321, Jan 2015.
B. Wen, D. Boroyevich, R. Burgos, P. Mattavelli, and Z. Shen,
“Small-signal stability analysis of three-phase ac systems in the
presence of constant power loads based on measured d-q frame
impedances,” Power Electronics, IEEE Transactions on, vol. 30,
no. 10, pp. 5952–5963, Oct 2015.
——, “Analysis of d-q small-signal impedance of grid-tied
inverters,” Power Electronics, IEEE Transactions on, vol. 31,
no. 1, pp. 675–687, Jan 2016.
Z. Bing, “Three-phase ac-dc converters for more-electric aircraft,” Ph.D. dissertation, Rensselaer Polytechnic Institute,
2010.
V. Valdivia, A. Lazaro, A. Barrado, P. Zumel, C. Fernandez, and
M. Sanz, “Black-box modeling of three-phase voltage source
inverters for system-level analysis,” Industrial Electronics, IEEE
Transactions on, vol. 59, no. 9, pp. 3648–3662, Sept 2012.
M. Cespedes and J. Sun, “Three-phase impedance measurement
for system stability analysis,” in Control and Modeling for
Power Electronics (COMPEL), 2013 IEEE 14th Workshop on.
IEEE, 2013, pp. 1–6.
D. Zmood, D. Holmes, and G. Bode, “Frequency-domain analysis of three-phase linear current regulators,” Industry Applications, IEEE Transactions on, vol. 37, no. 2, pp. 601–610, Mar
2001.
A PPENDIX
A. Parameter values used in simulations
Vbase = 690 V
VSdc = 1 p.u.
Cdc = 11.5 mF
Lvd = 0.0 p.u.
Lvq = 0.2 p.u.
Kpid = 1.59 p.u.
Kpiq = 2.07 p.u.
Tivdc = 0.0036 s
i∗Lq = 0.4 p.u.
Sbase = 1 M W
∗
VLdc
= 1 p.u.
Idc = 1.1 p.u.
Kpvd = 1 p.u.
Kpvq = 1.3 p.u.
Tiid = 0.047 s
Tiiq = 0.033 s
vd∗ = 1.0 p.u.
Vdc,base = 1400 V
ZS = 0.007 + j0.15 p.u.
ZL = 0.02 + j0.25 p.u.
Tivd = 0.1 s
Tivq = 0.2 s
fn = 50 Hz
Kpvdc = 8.33 p.u.
vq∗ = 0.0 p.u.
TABLE IV: Parameter values applied in Case A1
Vbase = 690 V
VSdc = 1 p.u.
Cdc = 11.5 mF
Lvd = 0.1 p.u.
Lvq = 0.1 p.u.
Kpid = 1.59 p.u.
Kpiq = 2.07 p.u.
Tivdc = 0.0036 s
i∗Lq = 0.4 p.u.
Sbase = 1 M W
∗
VLdc
= 1 p.u.
Idc = 1.1 p.u.
Kpvd = 1 p.u.
Kpvq = 1 p.u.
Tiid = 0.047 s
Tiiq = 0.033 s
vd∗ = 1.0 p.u.
Vdc,base = 1400 V
ZS = 0.007 + j0.15 p.u.
ZL = 0.02 + j0.25 p.u.
Tivd = 0.1 s
Tivq = 0.1 s
fn = 50 Hz
Kpvdc = 8.33 p.u.
vq∗ = 0.0 p.u.
TABLE V: Parameter values applied in Case A2. Parameters
in RED are different from Case A1.
B. Proof of equal determinants
The following relation prove that the determinant of Zdq is always
equal to the determinant of Zpn :
det(Zpn ) = det AZ · Zdq · A−1
Z
= det(AZ ) · det (Zdq ) · det(A−1
Z )
= −j · det (Zdq ) · j = det (Zdq )
(26)
14
Vbase = 690 V
VSdc = 1 p.u.
Cdc = 11.5 mF
Lvd = 0.1 p.u.
Lvq = 0.1 p.u.
Kpid = 1.59 p.u.
Kpiq = 1.59 p.u.
vd∗ = 1.0 p.u.
Sbase = 1 M W
VLdc = 1 p.u.
i∗Ld = 1.1 p.u.
Kpvd = 1 p.u.
Kpvq = 1 p.u.
Tiid = 0.047 s
Tiiq = 0.047 s
vq∗ = 0.0 p.u.
Vdc,base = 1400 V
ZS = 0.007 + j0.15 p.u.
ZL = 0.02 + j0.25 p.u.
Tivd = 0.1 s
Tivq = 0.1 s
fn = 50 Hz
i∗Lq = 0.4 p.u.
TABLE VI: Parameter values applied in Case B. Parameters
in RED are different from Case A2.
ZpL
ZpS
C. Proof of MFD impedance matrices relations (21)
When a subsystem is assumed MFD, the off-diagonal elements
in Zpn are equal to zero by definition. Hence Zpn = Znp = 0.
Substituting from (13) then gives:
1
1
1
−j
Zpn =
Zdq
=0
−j
2
1
1
1 j Zdq
=0
(27)
Znp =
j
2
=
L
S
Zpp
DS + Zpp
DL
Vp
= S
L
L
S
L ZS
Ip
D + Znn Zpp − Zpn
np
=
S
L
Zpp
DL + Zpp
DS
Vp
= L
S
S
L
S ZL
Ip
D + Znn Zpp − Zpn
np
shunt
shunt
(30)
where DS and DL are the determinants of the source and load
S
L
dq-domain impedance matrices Zdq
and Zdq
, respectively:
S
S
S
S
DS = Zdd
Zqq
− Zdq
Zqd
L
L
L
L
DL = Zdd
Zqq
− Zdq
Zqd
(31)
It is shown in Appendix B that the determinant of Zpn equals the
determinant of Zdq .
A corresponding expression can be derived for series voltage
positive sequence injection. The last five equations in (29) are then
modified to:
Expanding these expressions by substituting from (9) gives:
Zdd = Zqq = Zx
Zdq = −Zqd = Zy
Zx − jZy
Zpp =
2
Zx + jZy
Znn =
2
Zx and Zy are defined in (21).
depend on the choice of injection type, as well as the choice of
positive or negative sequence injection. The voltages in the two
subsystems are equal if the injection is shunt type. Furthermore, the
sum of negative sequence current must be zero because the injected
perturbation is assumed to be pure positive sequence. Solving (29)
gives the following original impedance:
IpL = −IpS = Ip
(28)
InL = −InS = In
VnS = VnL
D. Relationship between modified and original sequence domain impedance definitions
1) General case: The relationship between the modified sequence domain impedance definition (14) and the original (1) can
be derived by solving (11) for the source and load subsystem
simultaneously. When specifying the set of equations, one must
choose between:
• Shunt current or series voltage injection
.. and between
• Positive or negative sequence injection
One should choose positive sequence injection in order to find the
positive sequence impedance Zp , and negative sequence injection to
find Zn . The following set of equations should be solved to obtain
the impedance Zp for shunt current injection. This is equivalent to
solving the circuit presented in Figure 2.
L
L
VpL = IpL Zpp
+ InL Zpn
L
L
VnL = IpL Znp
+ InL Znn
ZpL
ZpS
ZpL
ZpS
ZpL
ZpS
shunt
shunt
=
=
VnS =
−InL
ZnL
Vp
= L
Ip
Vp
= S
Ip
ZnL
Vn
ZnS
(29)
The superscript L denotes load subsystem, whereas S denotes source
subsystem. The first four equations are the Generalized Ohms Law
with the modified sequence domain definition. The last five equations
series
(32)
L
S
L
S
VpL
Zpp
Znn
− Zpn
Znp
+ DL
=
S
L
Ip
Znn + Znn
S
L
S
L
VpS
Zpp
Znn
− Zpn
Znp
+ DS
=
=
L
S
Ip
Znn + Znn
=
series
series
(33)
Given that (33) clearly differs from (30), it can be concluded that
the original sequence domain impedance is not well defined in the
general case because it depends on the injection type. This has been
illustrated by simulations in Figure 13.
The additional equations for negative sequence are given in (34).
VpL = VpS = Vp
VnL
InS
series
Solving the set of equations gives:
S
S
VpS = IpS Zpp
+ InS Zpn
S
S
VnS = IpS Znp
+ InS Znn
VpL
Ip
VpS
=
Ip
=
ZnS
=
Vn
Z L DS + Z S DL
= S nn L S nn L S
L
In
D + Zpp Znn − Znp Zpn
=
L
S
L
S
Znn
Zpp
− Znp
Zpn
+ DL
VnL
=
S + ZL
In
Zpp
pp
=
Vn
Z S DL + Z L DS
= L nn S L nn S L
S
In
D + Zpp Znn − Znp Zpn
=
S
L
S
L
Znn
Zpp
− Znp
Zpn
+ DS
VnS
=
L + ZS
In
Zpp
pp
shunt
series
shunt
series
(34)
In section V these analytic expressions are validated through a
comparison where the original sequence impedances are obtained
directly by simulation.
15
2) Special case with one MFD subsystem: In the special case
where one subsystem is MFD, the expressions from the previous
sections can be simplified. If the source subsystem is MFD, i.e.
S
S
Zpn
= Znp
= 0, then, (30), (33) and (34) are reduced to:
ZpL
ZpS
ZnL
ZnS
shunt
shunt
shunt
= ZpL
= ZpS
= ZnL
=
shunt
ZnS
series
series
series
series
L
= Zpp
−
L
L
Zpn
Znp
S + ZL
Znn
nn
S
= Zpp
L
= Znn
−
L
L
Znp
Zpn
S
L
Zpp + Zpp
S
= Znn
(35)
Three important observations are obtained from (35). As expected,
in the source subsystem the original and modified impedances are
S
equal, i.e. ZpS = Zpp
because this subsystem is MFD. Second,
in the load subsystem, the original sequence domain impedance no
longer depends on injection type, i.e. ZpL
= ZpL
. The
shunt
series
L
third observation is that ZpL 6= Zpp
. The difference between them is
L
L
proportional to Zpn
Znp
, and also depends on the source impedance
S
Znn
. These observations can be seen in the simulation result shown
Figure 14.
E. Special case in which both subsystems are MFD
In this case, (30), (33) and (34) are reduced to:
ZpL
ZpS
ZnL
ZnS
shunt
shunt
shunt
shunt
= ZpL
= ZpS
= ZnL
= ZnS
series
series
series
series
L
= Zpp
S
= Zpp
L
= Znn
S
= Znn
(36)
In other words, the original and modified sequence domain
impedances are equal. This was also shown by (21), and demonstrates
the fact that MFD is a sufficient assumption for the original sequence
domain impedances to be uniquely defined. The corresponding simulation result is shown in Figure 15.
| 3cs.SY
|
Active Information Acquisition for Linear Optimization
Shuran Zheng∗1 , Bo Waggoner†2 , Yang Liu‡1 , and Yiling Chen§1
1
arXiv:1709.10061v1 [cs.DS] 28 Sep 2017
2
Harvard University
University of Pennsylvania
September 29, 2017
Abstract
We consider partially-specified optimization problems where the goal is to actively, but efficiently,
acquire missing information about the problem in order to solve it. An algorithm designer wishes to
solve a linear program (LP), max cT x s.t. Ax ≤ b, x ≥ 0, but does not initially know some of the
parameters. The algorithm can iteratively choose an unknown parameter and gather information in the
form of a noisy sample centered at the parameter’s (unknown) value. The goal is to find an approximately
feasible and optimal solution to the underlying LP with high probability while drawing a small number
of samples.
We focus on two cases. (1) When the parameters c of the objective are initially unknown, we take an
information-theoretic approach and give roughly matching upper and lower sample complexity bounds,
with an (inefficient) successive-elimination algorithm. (2) When the parameters b of the constraints are
initially unknown, we propose an efficient algorithm combining techniques from the ellipsoid method
for LP and confidence-bound approaches from bandit algorithms. The algorithm adaptively gathers
information about constraints only as needed in order to make progress. We give sample complexity
bounds for the algorithm and demonstrate its improvement over a naive approach via simulation.
1 Introduction
In many real-world settings, the goal is to solve an optimization problem, but some parameters of the problem are initially unknown. For example, consider a delivery company that wishes to plan driver routes, but
does not initially know average congestion or travel time of various links of the network.
To capture such settings, this paper proposes a model of optimization wherein the algorithm can iteratively choose a parameter and draw a “sample” that gives information about that parameter; specifically, the
sample is an independent draw from a subgaussian random variable centered at the true value of the parameter. This models, for instance, observing the congestion on a particular segment of road on a particular day.
Drawing each sample is presumed to be costly, so the goal of the algorithm is to draw the fewest samples
necessary in order to find a solution that is approximately feasible and approximately optimal.
Thus, the challenge falls under an umbrella we term active information acquisition for optimization
(AIAO). The key feature of the AIAO setting is the structure of the optimization problem itself, i.e. the
objective and constraints. The challenge is to understand how the difficulty of information acquisition relates
∗
[email protected]
[email protected]
‡
[email protected]
§
[email protected]
†
1
to this underlying structure. For example, are there information-theoretic quantities relating the structure to
the sample complexity? Meanwhile, the opportunity of AIAO is to exploit algorithms for the underlying
optimization problem. For example, can one interface with the algorithm to reduce sample complexity by
only acquiring the information needed, as it is needed?
These are the questions investigated in this paper, which focuses on active information acquisition for
linear optimization problems. Specifically, we consider linear programs in the form
max cT x, s.t. Ax ≤ b, x ≥ 0
x
(1)
with A ∈ Rm×n , c ∈ Rn , and b ∈ Rm . For most of the paper, we will consider either the case that the b in
the constraints is unknown or else the case that the c in the objective is unknown, with all other parameters
initially known to the algorithm. The algorithm can iteratively choose an unknown parameter, e.g. bi , and
draw a “sample” from it, e.g. observing bi + η for an independent, zero-mean, subgaussian η. The algorithm
must eventually output a solution x such that, with probability 1 − δ, Ax ≤ b + ε1 1 and cT x ≥ cT x∗ − ε2 ,
where x∗ is the optimal solution. The goal is for the algorithm to achieve this while using as few total
samples as possible. (In Appendix A, we give some motivating examples of practical problems that fit this
model.)
There is a natural “naive” or “static” approach: draw samples for all unknown parameters until they are
known to high accuracy with high probability, then solve the “empirical” linear program. However, we can
hope to improve by leveraging known algorithms and properties of linear programs. If a linear program
has an optimal solution, it has an optimal solution that is an extreme point (a corner point of the feasible
region); and at this extremal optimal solution, several constraints are binding. These suggest that it is more
important to gather information on the differing objective values of extreme points and to focus on binding
constraints. Algorithms developed in this paper leverage these properties of linear programs to decide how
much information to acquire for each unknown parameter.
The setting considered in this paper, active information acquisition for optimization, is related at a high
level to a large number of lines of work (in theoretical computer science (Balkanski et al., 2016), machine
learning (Balcan et al., 2006, 2007; Castro and Nowak, 2008), and artificial intelligence (Braziunas, 2006;
Blum et al., 2004)) that deal with optimization and uncertainty. However, most of them solve a different
conceptual problem, failing to model either the optimization aspect or the active aspect. At a technical level,
some works (Chen et al., 2014; Gabillon et al., 2016; Chen et al., 2016a,b, 2017) are relatively close to our
model, but do not provide useful solutions to the model of linear AIAO studied here. We discuss the related
work in details in Appendix B.
1.1 Approaches and Results
Two settings and our approaches. The paper investigates two settings: unknown c but known b, and
unknown b but known c. We always suppose A is known (see the discussion in Section 6 on how our
algorithms may extend to unknown-A) and assume that the linear program has an optimal solution.
It might initially appear that these cases are “equivalent” via duality theory, so we briefly digress to
explain why this is not the case (indeed, we argue that the two cases are quite different). Given a primal
linear program of the form (1), the dual program is given by miny bT y s.t. AT y ≥ c, y ≥ 0, which is easily
transformed into the maximization format of (1). In particular, the parameters c in the objective function
of a primal LP becomes the parameters in the constraints of the dual LP. By duality theory, the (exact)
optimal solutions to the primal and dual are connected by complementary slackness conditions, which allow
one to use the optimal dual solution y∗ to determine which of the primal constraints are binding at the
optimal primal solution x∗ . However, this approach breaks down in the approximate setting for two reasons.
First, approximately optimal solutions do not satisfy complementary slackness; and second, even knowing
2
which constraints bind does not suffice to determine the optimal solution x∗ when some of the constraint or
objective parameters are unknown.1 We hence take two different approaches toward our two settings.
Unknown-c case. In the unknown-objective setting, we take an information-theoretic perspective. Because the feasible region is known exactly, our (inefficient) algorithm focuses only on the set of extreme
points of the feasible region as one of them must be an optimal solution to the LP. For each of the extreme
point, there are a set of possible values for c such that if c takes any value in the set, this extreme point is
the optimal solution to the LP. The algorithm hence draws just enough samples to determine with high probability which is actually the case for the true c. In fact, the algorithm therefore produces the exact optimal
solution x∗ to the LP with high probability.
For this setting, we define an information-theoretic measure, Low(I) for an instance I. We show that
this quantity essentially characterizes the sample complexity of a problem instance and we give an algorithm,
not necessarily efficient, for achieving it up to low-order factors.
Theorem 1.1 (Informal). Given an instance I, let K(I) be the number of extreme points of the LP’s feasibility region. For the uknown-c problem with failure probability δ:
(i) On any instance I, at least Ω(Low(I) ln δ−1 ) samples
are requiredin expectation.
samples with high probability.
(ii) On any instance I, Algorithm 1 draws at most O Low(I) ln K(I)
δ
K(It )
(iii) There are families of instances {It }∞
on
which
any
algorithm
must
draw
Ω
Low(I
)
ln
t
t=1
δ
samples on instance It .
Unknown-b case. In contrast, in the unknown b setting, the uncertainty is over the feasible region. An
algorithm cannot initially identify extreme points but, with known c, it does know the direction that improves the objective function value. We hence take an “interior point” approach. Our algorithm, which we
call UCB-Ellipsoid, starts with a large relaxed feasible region and an interior point in that region and draws
samples to confidently “cut” the region in iterations. This leverages the key idea of the ellipsoid algorithm
for solving linear programs: as we advance in the direction of the optimal solution, we can ignore irrelevant
constraints and focus on either those that help cut the region or else (approximately) bind at an (approximately) optimal solution. Meanwhile, it also leverages the idea of “confidence bounds” from multi-armed
bandits algorithms such as UCB (Upper Confidence Bound). For each unknown parameter, we maintain
upper and lower confidence bounds which shrink as we gather more samples from the parameter. Each iteration of the ellipsoid algorithm, we will adaptively sample from various constraints until we are confident
that some constraint is violated or that no constraint is. To do so, we sample the constraint that is most
violated in an “optimistic” fashion, using the extreme value of the confidence bound.
We show the following bound on the sample complexity of UCB-Ellipsoid, along with a natural lower
bound:
Theorem 1.2 (Informal). For the unknown-b problem on n variables and m constraints (for simplicity take
m ≥ n), with approximation tolerance ε and failure probability δ:
1
−2 .
·
ε
(i) The sample complexity of any algorithm in the worst
case
is
Ω
n
ln
δ
m
log
(ii) Algorithm 2 draws at most O ∆(I)
a problem-specific measure of difficulty.
m
δ
· ε−2 samples with high probability, where ∆(I) ≥ 1 is
1
Nor does knowing which constraints bind even necessarily help, as approximately satisfying them may still lead to large
violations of other constraints. Thus, while we do not rule out some future approach that connects approximate solutions of the
primal and dual, the evidence suggests to us that the two settings are quite different and we approach each differently in this paper.
3
To get an idea of the sample complexity of this method in practice, we implemented it on a variety
of synthetic linear programming problems. In our simulations, UCB-Ellipsoid far outperforms the naive
approach of sampling all parameters to high accuracy, and approaches the performance of an oracle that
knows the binding constraints in advance and needs only to sample these.
2 Model and Preliminaries
2.1 The AIALO problem
We now formally define an instance I of the active information acquisition for linear optimization (AIALO)
problem. We then describe the format of algorithms for solving this problem. Note that one can easily
extend this into a more general formal definition of AIAO, for more general optimization problems, but we
leave this for future work.
An instance I consists of three components. The first component consists of the parameters of the
underlying linear program on n variables and m constraints: a vector c ∈ Rn , a vector b ∈ Rm , and a
matrix A ∈ Rn×m . Naturally, these specify the program2
max cT x s.t. Ax ≤ b, x ≥ 0.
x
(2)
We assume for simplicity in this paper that all linear programs are feasible and are known a priori to have
a solution of norm at most R. The second component specifies which parameters are initially known and
which are initially unknown. The third and final component specifies, for each unknown parameter (say ci ),
of a σ 2 -subgaussian distribution with mean equal to the value of the parameter.3
Given I, we define the following sets of approximately-feasible, approximately-optimal solutions.
Definition 2.1. Given an instance I, let x∗ be an optimal solution to the LP. Define OP T (I; ε1 , ε2 ) to be
the set of solutions x satisfying cT x ≥ cT x∗ − ε1 and Ax ≤ b + ε2 1. We use OP T (I) as shorthand for
OP T (I; 0, 0).
2.2 Algorithm specification
An algorithm for the AIALO problem, run on an instance I, functions as follows. The algorithm is given
as input n (number of variables), m (number of constraints), and σ 2 (subgaussian parameter). It is also
given the second component of the instance I, i.e. a specification of which parameters are known and
which are unknown. For each parameter that is specified as “known”, the algorithm is given the value of
that parameter, e.g. it is given “A11 = 42”. Finally, the algorithm is given an optimality parameter ε1 , a
feasibility parameter ε2 , and a failure probability parameter δ.
The algorithm may iteratively choose an unknown parameter and sample that parameter: observe an
independent and identically-distributed draw from the distribution corresponding to that parameter (as specified in the third component of the instance I). At some point, the algorithm stops and outputs a solution
x ∈ Rn .
Definition 2.2 ((δ, ε1 , ε2 )-correct algorithm). An algorithm A is (δ, ε1 , ε2 )-correct if for any instance I and
inputs (δ, ε1 , ε2 ), with probability at least 1 − δ, A outputs a solution x ∈ OP T (I; ε1 , ε2 ). In the case
ε1 = ε2 = 0, we say A is δ-correct.
2
Note that any linear program can be transformed into the given format with at most a factor 2 increase in n and m.
2 2
Distribution D with mean µ is σ 2 -subgaussian if, for X ∼ D, we have E[et(X−µ) ] ≤ eσ t /2 for all t. The family of
sub-Gaussian distributions with parameter σ encompasses all distributions that are supported on [0, σ] as well as many unbounded
distributions such as Gaussian distributions with variance σ 2 .
3
4
For the purposes of this paper, the sample complexity of an algorithm on I is the maximum, over all
possible random outcomes, of the number of samples the algorithm draws when run on instance I. The
sample complexity with respect to n, m, and σ 2 refers to the maximum sample complexity of the algorithm
over all instances with these parameters (recalling that we only consider instances for which an optimal
solution exists and has norm at most R).
2.3 Cases in this paper
Unknown objective. We will first consider the unknown-c case. Here, every parameter of the objective c
is initially unknown, and all other parameters are initially known. Geometrically, the algorithm is initially
given an exact description of the feasible polytope, in the form of Ax ≤ b and x ≥ 0, but no information
about the “direction” of the objective.
Because the constraints are known exactly, we focus on exact feasibility in this setting, i.e. ε2 = 0. We
also focus on an information-theoretic understanding of the problem, and produce an essentially optimal but
computationally inefficient algorithm.
Information-theoretically, any algorithm for this case begins with the set of feasible points (a polytope),
at least one of which is exactly optimal and hence in OP T (I; ε1 , 0). Information obtained over time implicitly eliminates far-from-optimal points until those that remain are ε1 -optimal with high confidence. For
a more combinatorial slant, note that the same argument is true just for the corners of the polytope, the
extreme points, of which there is a finite (though exponential) number, and at least one of which is exactly
optimal. In fact, if there is a unique optimal solution, then by iteratively eliminating extreme points it is
possible to achieve (δ, 0, 0)-correctness for the unknown-c case. We will focus on this goal. (In the case
where there is not a unique optimal solution, a tweak to our algorithm can achieve (δ, ε1 , 0)-correctness.)
Unknown constraint. Second, we will consider the unknown-b case where every parameter of the constraint vector b is initially unknown, and all other parameters are initially known. Geometrically, the algorithm is given an objective “direction” (c) and a set of constraint “orientations” (A), but does not initially
know the “offset” or displacement bi of each constraint i.
In this setting, we do not expect to attain either exact feasibility or exact optimality, as the exact constraints can never be known, and in general an arbitrarily small change in constraints of an LP leads to
a nonzero change in the value of the optimal solution. This represents a qualitative difference from the
unknown-c case.4
Other cases. We briefly discuss in Section 6 the case of both unknown constraint and (partially) unknown
A. Variants on the observational model are also interesting in this case. For instance, perhaps instead of
being able to sample a particular parameter Aij or bi , the algorithm can only propose an x and observe a
random sample centered on the quantity Ai x − bi , While we briefly discuss how Ellipsoid-UCB can extend
to solve such cases, they go beyond the scope of this paper and we leave sample complexity bounds and/or
empirical investigation to future work.
3 The Unknown Objective Function Case
In this section, consider instances of AIALO where A and b are initially known, but c is initially unknown.
We focus on the case where there is a unique optimal solution x∗ , and consider the problem of finding an
4
In particular, it may be possible to hope for some sort of duality-based reduction from the unknown-b case to the unknown-c
case, because the latter can be solved to exact feasibility and optimality; but the reverse seems less likely. This is an exciting
direction for future work, especially in terms of duality relationships between approximate solutions.
5
exact optimal solution with confidence δ (i.e., a δ-correct algorithm). For this section, which focuses more
on information-theoretic structure than practical algorithms, we also make the simplifying assumption that
each parameter’s distribution is a Gaussian of variance 1 (in particular is 1-subgaussian). Our results are
relatively straightforward to extend to the more general setting.
We first introduce a function Low(I) that characterizes the sample complexity required for an LP instance I. The function Low(I) is defined by the solution of a convex program. We then give an instancewise lower bound in terms of the Low(I) function and the failure probability parameter δ. We also formulate
a worst-case lower bound of the problem, which is polynomially related to the instance-wise lower bound.
Finally, we give an algorithm based on successive elimination that matches the worst-case lower bound
within a factor of ln(1/∆), where ∆ is the gap between the objective function value of the optimal extreme
point (x∗ ) and the second-best.
3.1 Lower Bounds
The function Low(I) is defined as follows.
Definition 3.1 (Low(I)). For any instance I of AIALO (or more generally, for any linear program), we
define Low(I) ∈ R to be the optimal solution to the following convex program.
min
τ
s.t.
n
X
i=1
n
X
τi
(3)
(k)
(si
i=1
τi ≥ 0, ∀i
2
− x∗i )2 T ∗
≤ c (x − s(k) ) , ∀k
τi
Here x∗ is the optimal solution to the LP in I and s(1) , . . . , s(k) are the extreme points of the feasible region
{x : Ax ≤ b, x ≥ 0}.
For intuition about Low(I), consider a thought experiment where we are given an extreme point x∗ ,
and we want to check whether or not x∗ is the optimal solution using as few samples as possible. Given our
empirical estimate b
c we would like to have enough samples so that with high probability, for each s(k) 6= x∗ ,
we have
b
cT (x∗ − s(k) ) > 0 ⇐⇒ cT (x∗ − s(k) ) > 0.
This will hold by a standard concentration bound (Lemma C.4) if enough samples of each parameter are
drawn; in particular, “enough” is given by the k-th constraint in (3).
Theorem 3.1 (Instance lower bound). Let I be an instance of AIALO in the unknown-c case. For 0 < δ <
0.1, any δ-correct algorithm A must draw
Ω(Low(I) ln δ−1 )
samples in expectation on I.
Proof Sketch: Given I with optimal solution x∗ , we construct an instance I ′ which differs only in the
unknown parameters c, modified just enough so that the extreme point s(k) becomes the optimal solution
to I. We then appeal to Lemma C.1, which relies on a multi-armed-bandits-style result of Kaufmann et al.
(2016) showing that, if two algorithms are given samples from distributions that are similar enough, then
their outputs are also similar unless enough samples are drawn from the various distributions. This allows us
6
to conclude that, since the δ-correct algorithm must be able to distinguish I from I ′ , its vector of expected
number of samples per parameter must satisfy the k-th constraint in the definition of Low(I). Repeating this
argument for each s(k) shows that the expected vector of samples must be feasible for the convex program
in the definition of Low(I), so the expected number of samples must exceed Ω(Low(I) ln δ−1 ).
We believe that it is unlikely for an algorithm to match the instance-wise lower bound without knowing
the value of c and x∗ in the definition of Low(I). To formally prove this claim, for any δ-correct algorithm
A, we construct a group of LP instances that share the same feasible region Ax ≤ b, x ≥ 0 but have different objective functions and different optimal solutions. We prove that A will have unmatched performance
on at least one of these LP instances.
Our worst-case lower bound can be stated as follows.
Theorem 3.2 (Worst-case lower bound for unknown c). Let n be a positive integer and δ ∈ (0, 0.1). For
any δ-correct algorithm A, there exists an infinite sequence of LP instances with n variables, I1 , I2 , . . . ,
such that A takes
(1)
Ω Low(Ik )(ln |Sk | + ln δ−1 )
(1)
samples in expectation on Ik , where Sk
Low(Ik ) goes to infinity.
is the set of all extreme points of the feasible region of Ik , and
We postpone the proof of the worst-case lower bound to Appendix C.3.
3.2 A Successive Elimination Algorithm
Before the description of the algorithm, we first define a function LowAll(S, ε, δ) that indicates the number
of samples we should take for each ci , such that the difference in objective value between any two points
in S can be estimated to an accuracy ε with probability 1 − δ. Define LowAll(S, ε, δ) to be the optimal
solution of the following convex program,
min
τ
s.t.
n
X
τi
(4)
i=1
n
X
(xi − yi )2
i=1
τi
≤
ε2
, ∀x, y ∈ S
2 ln(2/δ)
τi ≥ 0, ∀i.
Our algorithm starts with a set S (1) that contains all extreme points of the feasible region {x : Ax ≤ b, x ≥
0}, which is the set of all possible optimal solutions. We first draw samples so that the difference between
each pairs in S (1) is estimated to accuracy ε(1) . Then we delete all points that are not optimal with high
probability. In the next iteration, we halve the accuracy ε(2) = ε(1) /2 and repeat the process. The algorithm
terminates when the set contains only one point.
7
Algorithm 1 A successive elimination algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
S (1) ← set of all extreme points of feasible region {x : Ax ≤ b, x ≥ 0}
r←1
λ ← 10
while |S (r) | > 1 do
ε(r) ← 2−r , δ(r) ← δ/(10r 2 |S (1) |2 )
(r)
(r)
(t1 , . . . , tn ) ← LowAll(S (r) , ε(r) /λ, δ(r) )
(r)
(r)
ci be the empirical mean
Sample ci for ti times, let b
Let x(r) be the optimal solution in S (r) with respect to b
c(r)
(r)
(r)
(r)
Eliminate the points in S that are ε /2 + 2ε /λ worse than x(r) when the objective function is
b
c(r) ,
r ←r+1
11: end while
12: Output x ∈ S (r)
10:
S (r+1) ← {x ∈ S (r) : hx, b
c(r) i ≥ hx(r) , b
c(r) i − ε(r) /2 − 2ε(r) /λ}
(5)
The algorithm has the following sample complexity bound.
Theorem 3.3 (Sample complexity of Algorithm 1). For the AIALO with unknown-c problem, Algorithm 1
is δ-correct and, on instance I, with probability 1 − δ draws at most the following number of samples:
O Low(I) ln ∆−1 (ln |S (1) | + ln δ−1 + ln ln ∆−1 ) ,
where S (1) is the set of all extreme points of the feasible region and ∆ is the gap in objective value between
the optimal extreme point and the second-best,
∆ = max cT x −
x∈S (1)
max
x∈S (1) \x∗
cT x.
Proof Sketch: We define a good event E to be the event that for each iteration r and each pair of extreme
points in S (r) , the difference between them is estimated within error ε(r) /λ. We first show that E holds
with probability at least 1 − δ, giving correctness. Conditioning on E, we can prove that the algorithm
will not delete the optimal solution and will terminate before ⌊log(∆−1 )⌋ + 1 iterations. Then, we bound
the number of samples used in each iteration separately. We show that in each iteration r, the optimal
solution of Low(I) times α(r) = 32λ2 ln(2/δ(r) ) is a feasible solution of the convex program that defines LowAll(S (r) , ε(r) /λ, δ(r) ). Therefore the number of samples used in iteration r is no more than
α(r) Low(I). Together with the upper bound of iteration number, this completes the proof.
This matches the worst-case lower bound within a problem-dependent factor ln(1/∆). Notice, however,
that the size of |S (1) | can be exponentially large, and so is the size of the convex program (4). So Algorithm 1
is computationally inefficient if implemented straightforwardly, and it remains open whether the algorithm
can be implemented in polynomial time or an alternative algorithm with similar sample complexity and
better performance can be found.
8
4 The Unknown Constraint Case
In this section, we study the AIALO problem with unknown-b. Since we can never know the exact feasible
region in this case, we choose non-zero ε1 and ε2 and try to find a solution of I in OP T (I; ε1 , ε2 ) using as
few samples as possible.
The section begins with a lower bound in terms of the binding constraints. Then we introduce UCBEllipsoid algorithm and give an upper bound of its sample complexity. Finally we compare our algorithm
with the lower bound, as well as a naive approach using simulations.
Theorem 4.1 (Lower bound for unknown b). Suppose we have a (δ, ε1 , ε2 )-correct algorithm A where
δ ∈ (0, 0.1), ε1 > 0, ε2 > 0. Then for any n > 0, there exists infinitely many instances of the AIALO
problem with unknown-b with n variables with objective function kck∞ = 1 such that A must draw at least
Ω n ln(1/δ) · max{ε1 , ε2 }−2
samples in expectation on each of them.
The idea of the lower bound (proof in Appendix D.1) is that in the worst case, an algorithm must
accurately estimate at least all n binding constraints (in general with n variables, up to n constraints bind
at the optimal solution). It remains open whether we can get a tighter lower bound which also captures the
difficulty of ruling out non-binding constraints.
4.1 Ellipsoid-UCB Algorithm
Background. Before giving our algorithm, we describe the standard ellipsoid algorithm for linear programming. The high-level idea is to maintain a boundary, using an “ellipsoid” shape, around the space
where the optimal solution must lie; then repeatedly cut this space using either the constraints or the ob(0)
jective. The algorithm begins
withT an−1ellipsoid E known to contain the optimal solution. Formally, an
ellipsoid is the set of points x : x P x ≤ 1 specified by a positive definite matrix P.
The algorithm sets x(0) to be the center of this ellipsoid (which can be computed efficiently). Then, it
checks two cases: (1) x(0) is feasible, or (2) it is not feasible. If (2), say it violates constraint i, then the
algorithm considers the halfspace defined by the constraint Ai x(0) ≤ bi . If (1), the algorithm considers the
halfspace defined by the “constraint” cT x ≥ cT x(0) , as the optimal solution must satisfy this constraint. In
either case, it updates to a new ellipsoid E (1) defined as the minimal ellipsoid containing the intersection of
E (0) with the halfspace under consideration.
The obstacle is that, now, b is initially unknown. When the ellipsoid algorithm needs to find a violated
constraint (or certify that none is violated), how should we sample the constraints bi ?
A first observation is that we only need to find a single violated constraint, so there may be no need to
sample most parameters at a given round. A second observation is that it suffices to find the most violated
constraint. This can be beneficial as it may require only a few samples to find the most violated constraint;
and in the event that no constraint is violated, we still need to find an upper bound on “closest to violated”
constraint in order to certify that no constraint is violated.
To do so, we draw inspiration from algorithms for bandits problems (whose details are not important
2 , and
to this paper). Suppose we have m distributions with means µ1 , . . . , µm and variances σ12 , . . . , σm
we wish to find the largest µi . After drawing a few samples from each distribution, we obtain estimates
µ
bi along with confidence intervals given by tail bounds. Roughly, an “upper confidence bound” (UCB)
algorithm (see e.g. Jamieson and Nowak (2014)) for finding maxi µi proceeds by always sampling the i
whose upper confidence bound is the highest.
In our case, at each stage of the ellipsoid algorithm, we wish to find the largest Ai x − bi or ensure that
all are nonpositive. We therefore will propose a UCB-style approach to doing so, but with the advantage
that we can re-use any samples from earlier stages of the ellipsoid algorithm.
9
Algorithm and results. Ellipsoid-UCB is given in Algorithm 2. At each round k = 1, 2, . . . , we choose
the center point x(k) of the current ellipsoid E (k) and call the subroutine Algorithm 3 to draw samples and
check for violated constraints. We use the result of the oracle to cut the current space exactly as in the
standard ellipsoid method, and continue.
Some notation: t is used to track the total number of samples drawn (from all parameters) and Ti (t)
denotes the number of samples of bi drawn up to “time” t. The average of these samples is:
Definition 4.1. Let Xi,s denote the s-th sample of bi and let Ti (t) denote the number of times bi is sampled
in the first t samples. Define
Ti (t)
X
bbi,T (t) =
Xi,s /Ti (t)
(6)
i
s=1
to be the empirical mean of bi up to “time” t.
Algorithm 2 Modified ellipsoid algorithm
Let E (0) be the initial ellipsoid containing the feasible region.
Draw one sample for each bi , i ∈ [m].
Let k = 0 and t = m.
Let Ti (t) = 1 for all i.
while stopping criterion is not met5 do
Let x(k) be the center of E (k)
Call UCB method to get constraint i or “feasible”
if x(k) is feasible then
Let x ← x(k) if x is not initialized or cT x(k) > cT x.
y ← −c
else
y ← ATi
end if
Let E (k+1) be the minimal ellipsoid that contains E (k) ∩ {p : yT p ≤ yT x(k) }
Let k ← k + 1
end while
Output x or “failure” if it was never set.
The performance of the algorithm is measured by how many samples (observations) it needs to draw.
To state our theoretical results, define Vi (k) = Ai x(k) − bi to be the amount by which the i-th constraint is
violated by x(k) , and V ∗ (k) = maxi Vi (k). Define gapi,ε (k) = max{|Vi (k)|, V ∗ (k) − Vi (k), ε} and ∆i,ε
= mink gapi,ε (k).
Theorem 4.2 (Ellipsoid-UCB algorithm). The Ellipsoid-UCB algorithm is (δ, ǫ1 , ǫ2 )-correct and with probability 1 − δ, draws at most the following number of samples:
!!
m
m
X
m X σi2
σi2
σi2
+
.
log
log log
O
2
2
2
δ
∆
∆
∆
i,ε
/2
i,ε
/2
i,ε
/2
2
2
2
i=1
i=1
Specifically, the number of samples used for bi is at most
5
σi2
∆2i,ε /2
2
log(m/δ) + log log(σi2 /∆2i,ε2 /2 ) .
Our stopping criterion is exactly the√same as in the standard ellipsoid algorithm, for which there are a variety of possible
criteria that work. In particular, one is cT P−1 c ≤ min{ε1 , ε2 }, where P is the matrix corresponding to ellipsoid E (k) as
discussed above.
10
Algorithm 3 UCB-method
Input x(k)
Output either index j of a violated constraint, or “feasible”.
2/3
δ
Set δ′ = 20m
loop
1. Let j be the constraint with the largest index,
j = arg max Ai x(k) − bbj,Tj (t) + Ui (Ti (t)),
i
q
2σi2 log(log(3s/2)/δ′ )
and bbj,Tj (t) as in Definition 4.1.
where Ui (s) = 3
s
(k)
2. If Aj x − bbj,Tj (t) − Uj (Tj (t)) > 0 return j.
3. If Aj x(k) − bbj,T (t) + Uj (Tj (t)) < 0 return “feasible”.
j
4. If Uj (Ti (t)) < ε2 /2 return “feasible”.
5. Let t ← t + 1
6. Draw a sample of bj .
7. Let Tj (t) = Tj (t − 1) + 1.
8. Let Ti (t) = Ti (t − 1) for all i 6= j.
end loop
Proof Sketch: Our analysis is inspired by the techniques used in Jamieson et al. (2014). Define event A to
be the event that bbi,s − bi ≤ Ui (s) for all s ≥ 0 and i ∈ [m]. According to Lemma 3 in Jamieson et al.
(2014), A holds with probability at least 1 − δ. We prove the correctness and the sample complexity of the
algorithm conditioning on that A holds.
Correctness: If UCB-method always gives a correct answer, the ellipsoid algorithm will be able to find an
ε1 -suboptimal solution. So we only need to prove the correctness of the UCB-method. Conditioning
on event A holds, our UCB method will only return a constraint that is violated (line 2) and when it
returns “feasible”, no constraint is violated more than ε2 (line 3 and 4).
Number of samples: We bound the number of samples used on each constraint separately. Consider a
fixed ellipsoid iteration k in which UCB method is given input x(k) , the key idea is to prove that if
bi is sampled in this iteration at “time” t, Ui (Ti (t)) should be larger than gapi,ε2 /2 (k). Since Ui is a
decreasing function, this immediately gives us an upper bound of Ti (t), the number of samples used
on constraint i before sample number t. Thus for each ellipsoid iteration k, we have an upper bound
of Ti (t). Taking the maximum of them, we get the final result.
Discussion. To understand the bound, suppose for simplicity that each σi = 1. We observe that the first
term will dominate in all reasonable parameter settings, so we can ignore the second summation in this
discussion.
Next, note that each term in the sum reflects a bound on how many times constraint i must be sampled
over the course of the algorithm. This depends inversely on ∆i,ε2 /2 , which is a minimum over all stages k
of the “precision” we need of constraint i at stage k. We only need a very precise estimate if both of the
following conditions are satisfied:
• |Vi (k)| is small, meaning that the ellipsoid center x(k) is very close to binding constraint i.
11
• There is no other constraint that is significantly violated, meaning that i is very close to the mostviolated constraint for x(k) if any.
Because this is unlikely to happen for most constraints, we expect ∆i,ε2 /2 to generally be large (leading to
a good bound), although we do not have more precise theoretical bounds. The only constraints where we
might expect ∆i,ε2 /2 to be small are the binding constraints, which we expect to come close to satisfying
the above two conditions at some point. Indeed, this seems inevitable for any algorithm, as we explore in
our experiments.
Again suppose each σi = 1 for simplicity. Note that each ∆i,ε2/2 ≥
, ignoring the dominated second
ε2 /2. This implies that our bound is always better than O m log(m/δ)
2
ε2
term.
The static approach is to measure each bi with enough samples to obtain a good precision so that relaxed feasibility can be satisfied with high probability, then solve the linear program using the estimated
samples. (This number comes from using tail bounds to ensure good
constraints. This uses 4m log(m/δ)
ε22
precision is achieved on every bi .)
Therefore, the UCB-Ellipsoid algorithm dominates the static approach up to some constant factors and
can show dramatic instance-wise improvements. Indeed, in some simple cases, such as the number of
variables equal to the number of constraints, we do not expect any algorithm to be able to improve over the
static approach. However, a nice direction for future work is to show that, if m is very large compared to n,
then the UCB-Ellipsoid algorithm (or some other algorithm) is guaranteed to asymptotically improve on the
static approach.
Comparison to static approach.
5 Experiments
In this section, we investigate the empirical number of samples used by Ellipsoid-UCB algorithm for the
unknown-b case of AIALO. We fix δ = 0.1 and focus on the impact of the other parameters6 , which are
more interesting.
We compare three algorithms on randomly generated LP problems. The first is Ellipsoid-UCB. The second is the naive “static approach”, namely, draw 4σ 2 log(m/δ)/ε22 samples of each constraint, then solve the
LP using estimated means of the parameters. (This is the same approach mentioned in the previous section,
except that previously we discussed the case σ = 1 for simplicity.) The third is designed to intuitively match
the lower bound of Theorem 4.1: Draw 4σ 2 log(d/δ)/ε22 samples of each of only the binding constraints,
where there are d of them, then solve the LP using estimated means of the bi . (For a more fair comparison,
we use the same tail bound to derive the number of samples needed for high confidence, so that the constants
match more appropriately.)
We generate instances as follows. c is sampled from [−10, 10]n uniformly at random. b is uniformly
drawn from [0, 10]n . Each Ai is sampled from unit ball uniformly at random. Notice that the choice of bi ≥
0 guarantees feasibility because the origin is always a feasible solution. We also add additional constraints
xi ≤ 500 to make sure that the LP generated is bounded. When the algorithm makes an observation, a
sample is drawn from Gaussian distribution with variance σ 2 .
In Figure 1, each algorithm’s number of samples (average of 50 instances) is plotted as function of
different parameters. The number of samples used by Ellipsoid-UCB is proportional to n, σ 2 and ε−2 .
However, it does not change much as m increases.7 This will not be surprising if ellipsoid uses most of its
6
99.5 percent of the outputs turn out to satisfy relaxed feasibility and relaxed optimality.
Indeed, the standard ellipsoid algorithm for linear programming requires a number of iterations that is bounded in terms of the
number of variables regardless of the number of constraints.
7
12
samples on binding constraints, just as the lower bound does. This is shown in Table 1, where it can be seen
that Ellipsoid-UCB requires much fewer samples of non-binding constraints than binding constraints.
× 105
2.5
Static approach
Ellipsoid-UCB
Lower bound
2
Number of measurements
Number of measurements
2.5
1.5
1
0.5
0
30
× 105
Static approach
Ellipsoid-UCB
Lower bound
2
1.5
1
0.5
0
40
50
60
70
4
80
6
8
(a)
14
(b)
× 106
14
Static approach
Ellipsoid-UCB
Lower bound
12
10
Number of measurements
Number of measurements
12
Number of variables n
Number of constraints m
14
10
8
6
4
2
× 106
Static approach
Ellipsoid-UCB
Lower bound
12
10
8
6
4
2
0
0
0
2
4
6
0
8
20
40
60
80
1/ ǫ
σ
(c)
(d)
Figure 1: Number of samples as we vary m, n, σ and 1/ε. Every data point is the mean of 50 randomly drawn
problem instances. The baseline parameters are m = 80, n = 6, σ = 1, ε1 = ε2 = 0.1. In figure (d), ε1 = ε2 = ε.
Figure 2 addresses the variance in the number of samples drawn by Ellipsoid-UCB by plotting its empirical CDF over 500 random trials. The horizontal axis is the ratio of samples required by Ellipsoid-UCB
to those of the lower bound. For comparison, we also mention R, the ratio between the performances of the
static approach and the lower bound. These results suggest that the variance is quite moderate, particularly
when the total number of samples needed grows.
13
Static approach
Ellipsoid-UCB
Lower bound
Binding
2674
3325
1476
Non-binding
2674
11.7
0
Table 1: Average number of samples used per binding constraint and per non-binding constraint. Numbers are average
from 100 trials. Here, m = 80, n = 4, σ = 1 ε1 = ε2 = 0.1.
1.2
1
0.8
0.6
0.4
m=50 n=4
m=50 n=8
m=50 n=8
m=50 n=8
0.2
0
1.5
2
2.5
3
3.5
σ=1
σ=1
σ=5
σ=1
4
ǫ=0.1 R=35.27
ǫ=0.1 R=11.75
ǫ=0.1 R=11.75
ǫ=0.05 R=11.75
4.5
5
Figure 2: Empirical cumulative distribution function of Ellipsoid-UCB’s number of samples, in units of the “lower
bound”, over 500 trials. Note that the lower bound varies when parameters change. R = md log(m)
log(d) is the ratio between
the number of samples used by static approach and lower bound.
6 Discussion and Future Work
One question is whether our approach can extend when the constraint matrix A is unknown as well as b.
The goal is again to solve the problem with few total observations. One model would be that the algorithm
can request samples from any Aij or bi at any time. In this case, our algorithm can extend (although we
have not yet proven theoretical bounds), by maintaining confidence bounds around each Aij and bi . Once
enough certainty is obtained that a given constraint is violated, many samples are drawn for bi and Aij (∀j),
allowing an accurate update to the ellipsoid algorithm.
Another model would only allow samples from the violation of a constraint, i.e. the algorithm provides
an x and chooses a constraint i and learns a noisy estimate of bi − Ai x. Here we can still determine which
constraint is violated (if any), but more creativity is required to determine the entries of Ai so as to continue
the ellipsoid algorithm. We have a proposal involving requestion bi − Ai x for a large set of different x,
giving noisy linear equations which can be solved to learn Ai accurately.
A second extension to the model would allow algorithms access to varying qualities of samples for varying costs. For instance, perhaps some crowd workers can give very low-variance estimates for high costs,
while some workers can give cheaper estimates, but have larger variance. In this case, some preliminary
theoretical investigations suggest picking the worker that minimizes the product (price)(variance). A direction for future work is for the algorithm to select samples dynamically depending on the payment-variance
tradeoffs currently available. A final interested direction is a more mechanism-design approach where the
designer collects bids from the agents and selects a winner whose data is used to update the algorithm.
14
References
Maria-Florina Balcan, Alina Beygelzimer, and John Langford. Agnostic active learning. In ICML, 2006.
Maria-Florina Balcan, Andrei Broder, and Tong Zhang. Margin based active learning. In COLT, 2007.
Rui M Castro and Robert D Nowak. Minimax bounds for active learning. Information Theory, IEEE
Transactions on, 54(5):2339–2353, 2008.
Eric Balkanski, Aviad Rubinstein, and Yaron Singer. The power of optimization from samples. In Advances
in Neural Information Processing Systems, pages 4017–4025, 2016.
Shouyuan Chen, Tian Lin, Irwin King, Michael R Lyu, and Wei Chen. Combinatorial pure exploration of
multi-armed bandits. In Advances in Neural Information Processing Systems, pages 379–387, 2014.
Victor Gabillon, Alessandro Lazaric, Mohammad Ghavamzadeh, Ronald Ortner, and Peter Bartlett. Improved learning complexity in combinatorial pure exploration bandits. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, pages 1004–1012, 2016.
Wei Chen, Wei Hu, Fu Li, Jian Li, Yu Liu, and Pinyan Lu. Combinatorial multi-armed bandit with general
reward functions. In Advances in Neural Information Processing Systems, pages 1659–1667, 2016a.
Lijie Chen, Anupam Gupta, and Jian Li. Pure exploration of multi-armed bandit under matroid constraints.
In Conference on Learning Theory, pages 647–669, 2016b.
Lijie Chen, Anupam Gupta, Jian Li, Mingda Qiao, and Ruosong Wang.
Nearly optimal sampling algorithms for combinatorial pure exploration.
CoRR, abs/1706.01081, 2017.
URL
http://arxiv.org/abs/1706.01081.
Darius Braziunas. Computational approaches to preference elicitation. Technical report, 2006.
Avrim Blum, Jeffrey Jackson, Tuomas Sandholm, and Martin Zinkevich. Preference elicitation and query
learning. Journal of Machine Learning Research, 5(Jun):649–667, 2004.
Sarah OConnell, Barry OSullivan, and Eugene C Freuder. Strategies for interactive constraint acquisition.
In Proceedings of the CP-2002 Workshop on User-Interaction in Constraint Satisfaction, pages 62–76,
2002.
Christian Bessiere, Frédéric Koriche, Nadjib Lazaar, and Barry O’Sullivan. Constraint acquisition. Artificial
Intelligence, 2015.
Daniel P Heyman and Matthew J Sobel. Stochastic models in operations research: stochastic optimization,
volume 2. Courier Corporation, 2003.
Michael J Neely. Stochastic network optimization with application to communication and queueing systems.
Synthesis Lectures on Communication Networks, 3(1):1–211, 2010.
Aharon Ben-Tal, Laurent El Ghaoui, and Arkadi Nemirovski. Robust optimization. Princeton University
Press, 2009.
Dimitris Bertsimas, Dessislava Pachamanova, and Melvyn Sim. Robust linear optimization under general
norms. Operations Research Letters, 32(6):510–516, 2004.
15
Emilie Kaufmann, Olivier Cappé, and Aurélien Garivier. On the complexity of best arm identification in
multi-armed bandit models. Journal of Machine Learning Research, 17(1):1–42, 2016.
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.
Kevin G Jamieson, Matthew Malloy, Robert D Nowak, and Sébastien Bubeck. lil’ucb: An optimal exploration algorithm for multi-armed bandits. In COLT, volume 35, pages 423–439, 2014.
John Duchi. Derivations for linear algebra and optimization.
16
A
Motivating Examples
Our settings can be used to model many real-world optimization problems. In this section, we expand on
some example real-world problems that fall into our framework:
(i) A company who wants to decide a production plan to maximize profit faces a linear program. But
when entering a new market, the company may not initially know the average unit price ci for their
different products in this market. Sampling ci corresponds to surveying a consumer on his willingness
to buy product i.
(ii) A delivery company who wants to plan driver routes may not know the average traffic/congestion of
road segments. Each segment e (an edge in the graph) has a true average travel time be , but any given
day it is be + noise. One can formulate shortest paths as an LP where b is the vector of edge lengths.
Sampling be corresponds to sending an observer to the road segment e to observe how long it takes to
traverse on a given day.
(iii) A ride sharing company (e.g. Uber) wants to decide a set of prices for rides request but it may not
know customers’ likelihood of accepting the prices ci . Sampling ci in this setting corresponds to
posting different prices to collect information.
(iv) For the purpose of recommending the best route in real time, a navigation App, e.g., Waze8 , may want
to collect traffic information or route information from distributed driver via their App.
B Related Work
Related work: Theoretical computer science and machine learning. Much work in active learning
considers acquiring data points iteratively with a goal of low sample complexity Balcan et al. (2006, 2007);
Castro and Nowak (2008).The key difference to AIAO is between data and parameters. In learning, the
goal is to minimize the average or expectation of some loss function over a distribution of data points. Other
than its likelihood, each data point plays the same role in the problem. Here, the focus is on how much
information about each of various parameters is necessary to solve a structured optimization problem to
a desired level of accuracy. In other words, the key question here, which is how much an optimization
algorithm needs to know about the parameters of the problem it is solving, does not apply in active learning.
Broadly related is recent work on optimization from samples Balkanski et al. (2016), which considers
the sample complexity of a two-stage process: (1) draw some number of i.i.d. data points; (2) optimize
some loss function or submodular function on the data. In that setting, the algorithm sees a number of inputoutput pairs of the function, randomly distributed, and must eventually choose a particular input to optimize
the function. Therefore, it is quite different from our setting in both important ways: (1) the information
collected are data points (and evaluations), as in ML above, rather than parameters as in our problem; (2)
(so far) it is not active, but acquires information in a batch.
A line of work that is closely related to our problem is the study of combinatorial pure exploration (CPE)
problem, where a learner collects samples of unknown parameters of an objective function to identify the
optimal member in a solution set. The problem was first proposed in Chen et al. (2014), and subsequently
studied by Gabillon et al. (2016); Chen et al. (2016a,b, 2017). CPE only considers combinatorial optimization problems whose solution set contains only binary vectors of length n. A recent work by Chen et al.
(2017) extended CPE to a General-Sampling problem by allowing general solution sets. Our unknown-c
problem can be fitted into the setting of General-Sampling. Nevertheless, our algorithm for unknown-c
8
www.waze.com
17
leverages the structure of LP and hence has better sample complexity performance than directly treating it
as a General-Sampling problem. The General-Sampling problem does not encompass all AIAO settings,
e.g., our unknown-b case.
Related work: artificial intelligence. Several existing lines of work in the artificial intelligence literature,
deal with actively acquiring information about parameters of an optimization problem in order to solve it.
Preference elicitation (Braziunas, 2006; Blum et al., 2004) typically focuses on acquiring information about
parameters of the objective by querying a user about his preferences, this is similar to our goal for the
unknown-c setting. But often the existing literatures do not analyze an active information acquisition setting;
neither do they analyze the sample complexity. Relevant to our unknown-b case, for more combinatorial
problems, the constraint acquisition literature (OConnell et al., 2002; Bessiere et al., 2015) is closer to our
problem in some respects, as it posits an optimization problem with unknown constraints that must be
learned via interactive querying.
We emphasize that a central feature of the model in this paper is noisy observations: the observations of the algorithm are only noisy samples of a true underlying constraint. The key challenge is to
choose how many repeated samples of each parameter to draw so as to gain confidence in using or discarding its constraint. This aspect of the problem is not to our knowledge present in preference elicitation or
model/constraint acquisition. This also differs from the literatures on stochastic programming (Heyman and Sobel,
2003; Neely, 2010), chance-constrained programming (Ben-Tal et al., 2009), and robust optimization (Ben-Tal et al.,
2009; Bertsimas et al., 2004).
C
Proofs for the Unknown Objective Function Case
C.1 Change of Distribution Lemma
A key element to derive the lower bounds is the Change of Distribution lemma, which was first formulated
in Kaufmann et al. (2016). The lemma provides a general relation between the expected number of draws
and Kullback-Leibler divergences of the arms distributions. The core elements of the model that the lemma
can be applied to are almost the same as the classical bandit model. We will state it here and explain the
applicability of our setting. In the bandit model, there are n arms, with each of them being characterized
by an unknown distribution νi , i = 1, 2, ..., n. The bandit model consists of a sequential strategy that selects
a subset of arms of size K. Upon selection, each arm reveals a reward generated from its corresponding
distribution. The rewards for each arm form an i.i.d. sequence. The selection strategy/algorithm invokes a
stopping time T when the algorithm will terminate and output a solution.
If we consider a LP instance I with unknown parameters d as a bandit model, an unknown parameter
di will correspond to an arm a in the Lemma C.2, and thus νa is just the Gaussian distribution with mean
di and variance 1 (both being unknown). Each step, we select K = 1 coefficient to sample with. Let T be
a sufficiently large number such that algorithm A terminates before time T almost surely. Then we will be
able to safely apply the Change of Distribution lemma to our setting. The lemma can be stated as follows in
our setting of the problem.
Lemma C.1. Let A be a (δ, ǫ1 , ǫ2 )-correct algorithm with δ ∈ (0, 0.1). Let I, I ′ be two LP instances
that are equal on all known parameters, and let d, d′ be their respective vectors of unknown parameters.
Suppose each instance has samples distributed Gaussian with variance 1. Suppose OP T (I; ǫ1 , ǫ2 ) and
OP T (I ′ ; ǫ1 , ǫ2 ) are disjoint. Then letting τi be the number of samples A draws for parameter di on input
I, we have
X
1
E
τi (di − d′i )2 ≥ 0.8 ln .
δ
i
18
Proof. We use a result on bandit algorithms by Kaufmann et al. (2016), which is restated as follows.
Lemma C.2 (Kaufmann et al. (2016)). Let ν and ν ′ be two bandit models with n arms such that for all arm
a, the distribution νa and νa′ are mutually absolutely continuous. For any almost-surely finite stopping time
T with respect to (Ft ),
n
X
i=1
Eν [Na (T )]KL(νa , νa′ ) ≥ sup d(Pr(E), Pr′ (E)),
E∈FT
ν
ν
where d(x, y) = x ln(x/y) + (1 − x) ln ((1 − x)/(1 − y)) is the binary relative entropy function, Na (T )
is the number of samples drawn on arm a before time T and KL(νa , νa′ ) is the KL-divergence between
distribution νa and νa′ .
Let I and I ′ be the two bandit models in Lemma C.2. Applying above lemma we have
n
X
i=1
EA,I [τi ]KL(N (di , 1), N (d′i , 1)) ≥ d( Pr (E), Pr′ (E)), for all E ∈ FT ,
A,I
A,I
where N (µ, σ) is the Gaussian distribution with mean µ and variance σ, PrA,I [E] is the probability of
event E when algorithm A is given input I, and EA,I [X] is the expected value of random variable X
when algorithm A is given input I. According to the result in Duchi, the KL-divergence for two Gaussian
distribution with mean µ1 , µ2 and variance σ1 , σ2 is equal to
log
σ2 σ12 + (µ1 − µ2 )2
.
+
σ1
2σ22
Thus we have KL(N (di , 1), N (d′i , 1)) = 21 (di − d′i )2 . We further define event E to be the event that algorithm A finally outputs a solution in set OP T (I; ε1 , ε2 ), then since A is (δ, ε1 , ε2 )-correct and OP T (I; ε1 , ε2 )
is disjoint from OP T (I ′ ; ε1 , ε2 ), we have PrA,I (E) ≥ 1 − δ and PrA,I ′ (E) ≤ δ. Therefore
n
X
i=1
1
EA,I [τi ] (di − d′i )2 ≥ d(1 − δ, δ) ≥ 0.4 ln δ−1 .
2
The last step uses the fact that for all 0 < δ < 0.1,
d(1 − δ, δ) = (1 − 2δ) ln
1−δ
1
≥ 0.8 ln √ = 0.4 ln δ−1 .
δ
δ
C.2 Proof for Theorem 3.1
We restate the instance-wise lower bound for unknown objective function LP problems.
Theorem 3.1 (Instance lower bound). Let I be an instance of AIALO in the unknown-c case. For 0 < δ <
0.1, any δ-correct algorithm A must draw
Ω(Low(I) ln δ−1 )
samples in expectation on I.
19
Let I be a LP instance max{x:Ax≤b} cT x, and A be a δ-correct algorithm, where 0 < δ < 0.1. Define
ti to be the expected number of samples that algorithm will draw for ci when the input is I.
We only need to show that 5t/ ln(1/δ) is a feasible solution of the convex program (3) that computes
Low(I).
Consider a constraint in (3)
2
≤ cT (x∗ − s(k) ) ,
n
(k)
X
(s − x∗ )2
i
i
τi
i=1
where x∗ is the optimal solution of I and s(k) is a corner point of the feasible region of I. To prove that
5t/ ln(1/δ) satisfies this constraint, we will construct a new LP instance I∆ by adding ∆ to the objective
function c, such that s(k) becomes a better solution than x∗ . We construct vector ∆ as follows,
(k)
∆i =
D(x∗i − si )
,
ti
and
−2cT (x∗ − s(k) )
D= P
.
(k)
(si −x∗i )2
n
i=1
ti
It is not difficult to verify that x∗ is no longer the optimal solution of I∆ :
hc + ∆, x∗ − s(k) i = hc, x∗ − s(k) i + h∆, x∗ − s(k) i
∗
(k)
= hc, x − s
∗
n
(k)
X
2cT (x∗ − s(k) ) x∗i − si
(k)
· (x∗i − si )
·
i−
∗ )2
Pn (s(k)
t
−x
i
i
i
i=1
(k)
= −hc, x − s
< 0.
i=1
ti
i
Then by Lemma C.1,
0.8 ln(1/δ) ≤
=
n
X
i=1
n
X
i=1
n
X
ti · ∆2i
(k)
D(x∗i − si )
ti
ti ·
!2
(x∗i − s(k) )2
· D2
ti
i=1
2
n
∗
(k)
2
T
∗
(k)
X
(xi − s ) −2c (x − s )
=
· P
(k)
(si −x∗i )2
ti
n
=
i=1
i=1
(cT (x∗
=4· P
n
−
s(k) ))2
(k)
(si −x∗i )2
i=1
ti
which is equivalent to
n
(k)
X
(s − x∗ )2
i
i=1
i
5ti / ln(1/δ)
ti
,
≤ (cT (x∗ − s(k) ))2 .
Therefore 5t/ ln(1/δ) is a feasible solution of the convex program (3), which completes our proof.
20
C.3 Proof for Theorem 3.2
We prove the worst case lower bound for unknown c case.
Theorem 3.2 (Worst-case lower bound for unknown c). Let n be a positive integer and δ ∈ (0, 0.1). For
any δ-correct algorithm A, there exists an infinite sequence of LP instances with n variables, I1 , I2 , . . . ,
such that A takes
(1)
Ω Low(Ik )(ln |Sk | + ln δ−1 )
(1)
samples in expectation on Ik , where Sk
Low(Ik ) goes to infinity.
is the set of all extreme points of the feasible region of Ik , and
The following lemma will be used in the construction of desired LP instances.
Lemma C.3. Let n be a positive integer. There exists a constant c, a positive integer l = Ω(n) and z = 2cn
sets W1 , . . . , Wz ⊆ [n] such that
• For all i ∈ [z], we have |Wi | = l = Ω(n).
• For all i 6= j, |Wi ∩ Wj | ≤ l/2.
Proof. Define l = n/10. Let each Wi be a uniformly random subset of [n] with size l. Then it is satisfied
that
Pr[|Wi ∩ Wj | > l/2] ≤ 2−Ω(n)
for all 1 ≤ i, j ≤ n, i 6= j. So we can choose sufficiently small c such that
Pr[∃i 6= j, |Wi ∩ Wj | > l/2] ≤ z 2 2−Ω(n) < 1,
which implies the existence of a desired sequence of subsets.
Now for any δ-correct algorithm A, we prove the existence of LP instances I1 , I2 , . . . , which all have
n variables.
For simplicity, all the linear program instances we construct in this proof share the same feasible region,
which we define as follows. Let W1 , . . . , Wz ⊆ [n] be the sequence of subsets in Lemma C.3. For a subset
W ⊆ [n], we define a point pW
1, if i ∈ W ;
W
pi =
0, otherwise.
The feasible region we are going to use throughout this proof is the convex hull of pW1 , . . . , pWz .
To find a desired LP instance Ik , we first choose an arbitrary constant ∆k . We construct z different
LP instances I∆k ,W1 , . . . , I∆k ,Wz and show that at least one of them satisfies the condition in the theorem.
Define the objective function cWj of I∆k ,Wj to be
W
ci j
=
∆k ,
−∆k ,
if i ∈ Wj ;
otherwise.
Then clearly the optimal solution of I∆k ,Wj is point pWj . We define Pr[A(I∆,Wi ) = pWj ] to be the
probability that algorihtm A outputs pWj when the input is I∆,Wi . Then we have
Pr[A(I∆,Wi ) = pWi ] ≥ 1 − δ,
21
and
X
j:j6=i
Pr[A(I∆,Wi ) = pWj ] ≤ δ.
Thus there must exists Wk such that
Pr[A(I∆,Wi ) = pWk ] ≤ 2δ/z.
Let T be the number of samples used by algorithm A when the input is I∆k ,Wk . Since A is a δ-correct
algorithm, Pr[A(I∆,Wk ) = pWk ] ≥ 1 − δ > 0.9. So if we define event E to be the event that A outputs
pWk and apply Lemma C.2,
E[T ] · (2∆2 ) ≥ d Pr[A(I∆,Wk ) = pWk ], Pr[A(I∆,Wi ) = pWk ]
≥ Ω(ln(z/δ)
= Ω(ln z + ln(1/δ)).
Here we use the following property of d(1 − δ, δ) function: for 0 < δ < 0.1, d(1 − δ, δ) ≥ 0.4 ln(1/δ). So
we get a lower bound for E[T ],
E[T ] ≥ Ω ∆−2 (ln z + ln(1/δ)) .
Meanwhile if we look at the Instance Lower Bound, Low(I∆,Wk ),
n
X
min
τ
s.t.
i=1
n
X
τi
Wj
(pi
i=1
k 2
− pW
i )
≤ hcWk , (pWk − pWj )i2 , ∀j
τi
τi ≥ 0,
8n
It is easy to verify that τi = l∆8 2 for all i is a feasible solution. So we have Low(I∆,Wk ) = Θ( l∆
2) =
−2
−1
Θ(∆ ). Therefore the number of samples that A will use on I∆,Wk is Ω Low(I∆,Wk )(ln z + ln(δ ))
in expectation.
By simply setting ∆k = k1 , we will get an infinite sequence of LP instances as stated in the theorem.
C.4 Proof for Theorem 3.3
In this section, we prove the sample complexity of our successive elimination algorithm for unknown c case.
Theorem 3.3 (Sample complexity of Algorithm 1). For the AIALO with unknown-c problem, Algorithm 1
is δ-correct and, on instance I, with probability 1 − δ draws at most the following number of samples:
O Low(I) ln ∆−1 (ln |S (1) | + ln δ−1 + ln ln ∆−1 ) ,
where S (1) is the set of all extreme points of the feasible region and ∆ is the gap in objective value between
the optimal extreme point and the second-best,
∆ = max cT x −
x∈S (1)
The following lemma will be used in our proof.
22
max
x∈S (1) \x∗
cT x.
Lemma C.4. Given a set of Gaussian arms with unit variance and mean c1 , . . . , cn . Suppose we take τi
samples for arm i. Let Xi be the empirical mean. Then for an arbitrary vector p,
T
ε2
T
Pr |p X − p c| ≥ ε ≤ 2 exp − P 2
2 pi /τi
Proof. By definition, pT X − pT c follows Gaussian distribution with mean 0 and variance
P
2
i pi /τi .
We define a good event E to be the event that |(x − y)T (b
c(r) − c))| ≤ ε(r) /λ for all r and x, y ∈ S (r) .
According to Lemma C.4,
Pr[E] ≥ 1 −
X X
r x∈S (r) y∈S
Since τ satisfies the constraints in (4),
(ε/λ)2
.
2 exp − P
2 (xi − yi )2 /τi
(r)
X X X
(ε/λ)2
(r)
2 exp − P
≤
2
exp
−
ln(2/δ
)
2 (xi − yi )2 /τi
r x∈S (r) y∈S (r)
x∈S (r) y∈S (r)
X X X
=
δ(r)
X X
r
X
X
r x∈S (r) y∈S (r)
≤δ
Therefore Pr[E] ≥ 1 − δ.
We first prove the correctness of the algorithm conditioning on E.
Lemma C.5. When the good event E holds, the optimal LP solution x∗ = maxAx≤b cT x will not be deleted.
Proof. Suppose to the contrary x∗ is deleted in iteration r, i.e., x∗ ∈ S (r) but x∗ ∈
/ S (r+1) . Then according
(r)
∗
(r)
(r)
c , x is at least ε /2 − 2ε /λ worse than x(r) ,
to (5), when the objective function is b
hx(r) − x∗ , b
c(r) i > ε(r) /2 + 2ε(r) /λ.
By the definition of the optimal solution x∗ ,
hc, x∗ − x(r) i > 0.
Combining the two inequalities will give
hc − b
c(r) , x∗ − x(r) i > ε(r) /2 + 2ε(r) /λ > ε(r) /λ,
contradictory to that event E holds.
We then bound the number of samples conditioning on E. We first prove the following lemma.
Lemma C.6. When event E holds, all points s in set S (r+1) satisfies
hc, x∗ − si < ε(r) .
after the r th iteration.
23
Proof. Suppose when entering the r th iteration, there exists s ∈ S (r) such that hc, x∗ − si > ε(r) . Then
since E holds and λ = 10,
hc, x∗ − si > hb
c(r) , x∗ − si − ε(r) /λ
> (1 − 1/λ)ε(r)
> ε(r) /2 + 2ε(r) /λ.
By Lemma C.5, we have x∗ ∈ S (r) . Therefore s will be deleted in this iteration.
Now consider a fixed iteration r. Let τ ∗ be the optimal solution of the convex program (3) that computes
low(I). Define α = 32λ2 ln(2/δ(r) ). We show that t = ατ ∗ is a feasible solution in the convex program (4)
that computes LowAll(S (r) , ε(r) , δ(r) ). For any x, y ∈ S (r) ,
X (xi − yi )2
1 X (xi − yi )2
=
ti
α
τi∗
1 X (xi − x∗i + x∗i − yi )2
=
α
τi∗
1 X 2(xi − x∗i )2 + 2(x∗i − yi )2
≤
α
τi∗
due to the fact that (a + b)2 ≤ 2a2 + 2b2 for all a, b ∈ R.
Since τ ∗ satisfies the constraints in Low(I) function (3),
1 X 2(xi − x∗i )2 + 2(x∗i − yi )2
2
≤
(cT (x∗ − x))2 + (cT (x∗ − y))2
∗
α
τi
α
And because of Lemma C.6,
2
(ε(r) )2
4
.
(cT (x∗ − x))2 + (cT (x∗ − y))2 ≤ (ε(r−1) )2 = 2
α
α
2λ ln(2/δ(r) )
So we have proved that t = ατ ∗ is a feasible solution of the convex program that computes LowAll(S (r) , ε(r) , δ(r) ).
P
(r)
Thus the number of samples used in iteration r, ni=1 ti , is no more than
n
X
i=1
(r)
ti
≤
n
X
ti = α
i=1
X
i
τi∗ = O(Low(I)(ln |S (r) | + ln δ−1 + ln r)
Conditioning on E, the algorithm will terminate before ⌊log(∆−1 )⌋+1 iterations according to Lemma C.6.
Therefore the total number of samples is
O Low(I) ln ∆−1 (ln |S (1) | + ln δ−1 + ln ln ∆−1 ) .
D
The Unknown Constraints Case
D.1 Proof for Theorem 4.1
Theorem 4.1 (Lower bound for unknown b). Suppose we have a (δ, ε1 , ε2 )-correct algorithm A where
δ ∈ (0, 0.1), ε1 > 0, ε2 > 0. Then for any n > 0, there exists infinitely many instances of the AIALO
problem with unknown-b with n variables with objective function kck∞ = 1 such that A must draw at least
Ω n ln(1/δ) · max{ε1 , ε2 }−2
samples in expectation on each of them.
24
Let A be a (δ, ε1 , ε2 )-correct algorithm. For a positive integer n, consider the following LP instance I
with n variables and n constraints,
max
s.t.
x1
x1 ≤ C,
x1 + xi ≤ C,
x ≥ 0.
∀2 ≤ i ≤ n,
Clearly the optimal solution is x∗1 = C and x∗i = 0 for i > 1. Every constraint is a binding constraint.
Now we prove that for any k ∈ [n], algorithm A should take at least Ω ln(1/δ) · max{ε1 , ε2 }−2 for the
kth constraint. We construct a new LP I ′ by subtracting the right-hand side of kth constraint by 2(ε1 + ε2 ).
Then OP T (I; ε1 , ε2 ) and OP T (I ′ ; ε1 , ε2 ) must be disjoint, since for any x ∈ OP T (I ′ ; ε1 , ε2 ), x will not
violate the kth constraint of I ′ by more than ε2 ,
x1 ≤ C − 2(ε1 + ε2 ) + ε2 < C − 2ε1 ,
which means that x ∈
/ OP T (I; ε1 , ε2 ). According to Lemma C.1,
E[τk ] · 4(ε1 + ε2 )2 ≥ 0.8 ln(1/δ)
And since 2 max{ε1 , ε2 } ≥ ε1 + ε2 ,
E[τk ] = Ω(max{ε1 , ε2 }−2 · ln(1/δ)).
D.2 Proof for Theorem 4.2
Recall that our algorithm and the sample complexity theorem works as follows:
Algorithm 4 Modified ellipsoid algorithm
Let E (0) be the initial ellipsoid containing the feasible region.
Draw one sample for each bi , i ∈ [m].
Let k = 0 and t = m.
Let Ti (t) = 1 for all i.
while stopping criterion is not met9 do
Let x(k) be the center of E (k)
Call UCB method to get constraint i or “feasible”
if x(k) is feasible then
Let x ← x(k) if x is not initialized or cT x(k) > cT x.
y ← −c
else
y ← ATi
end if
Let E (k+1) be the minimal ellipsoid that contains E (k) ∩ {t : yT t ≤ yT x(k) }
Let k ← k + 1
end while
Output x or “failure” if it was never set.
9
Our stopping criterion is exactly the√same as in the standard ellipsoid algorithm, for which there are a variety of possible
criteria that work. In particular, one is cT P−1 c ≤ min{ε1 , ε2 }, where P is the matrix corresponding to ellipsoid E (k) as
discussed above.
25
Algorithm 5 UCB-method
Input x(k)
2/3
δ
Set δ′ = 20m
loop
1. Let j be the constraint with the largest index,
j = arg max Ai x(k) − bbj,Tj (t) + Ui (Ti (t)),
i
q
2σi2 log(log(3s/2)/δ′ )
and bbj,Tj (t) as in Definition 4.1.
where Ui (s) = 3
s
2. If Aj x(k) − bbj,Tj (t) − Uj (Tj (t)) > 0 return j.
3. If Aj x(k) − bbj,T (t) + Uj (Tj (t)) < 0 return “feasible”.
j
4. If Uj (Ti (t)) < ε2 /2 return “feasible”.
5. Let t ← t + 1
6. Draw a sample of bj .
7. Let Tj (t) = Tj (t − 1) + 1.
8. Let Ti (t) = Ti (t − 1) for all i 6= j.
end loop
Theorem 4.2 (Ellipsoid-UCB algorithm). The Ellipsoid-UCB algorithm is (δ, ǫ1 , ǫ2 )-correct and with probability 1 − δ, draws at most the following number of samples:
!!
m
m
X
m X σi2
σi2
σi2
+
log
log log
.
O
δ
∆2i,ε2 /2
∆2i,ε2 /2
∆2i,ε2/2
i=1
i=1
Specifically, the number of samples used for bi is at most
σi2
∆2i,ε
2 /2
log(m/δ) + log log(σi2 /∆2i,ε2 /2 ) .
Our analysis is inspired by the techniques used in Jamieson et al. (2014). The following lemma is the
same as Lemma 3 in Jamieson et al. (2014), and is simplified by setting ε = 1/2. We choose 1/2 only for
simplicity. It will not change our result asymptotically. The constant in this lemma can be optimized by
selecting parameters carefully.
Lemma D.1. Let X1 , X2 , . . . be i.i.d. sub-Gaussian random variables
with scale parameter σ and mean
P
µ. We has probability at least 1 − 20 · δ3/2 for all t ≥ 1, 1t ts=1 Xs − µ ≤ L(t, δ), where L(t, δ) =
q
2
3 2σ log(log(3t/2)/δ)
.
t
Define event A to be the event that b
bi,t − bi ≤ Ui (t) for all t ≥ 0 and i ∈ [m]. Since our definition of
Ui (t) is the same as L(t, (δ/20m)2/3 ) in Lemma D.1 with scale parameter σi , the probability that event A
holds is at least 1 − δ according to union bound.
We prove the correctness and the sample number of the algorithm conditioning on that A holds.
Correctness: We first prove that the output of our algorithm satisfies relaxed feasibility and relaxed optimality when A holds. If our UCB-method always gives correct answer, the ellipsoid algorithm will be able
to find an ε1 -suboptimal solution. So we only need to prove the correctness of the UCB-method.
26
• When UCB method returns a violated constraint j in line 2, it is indeed a violated one: since |bbj,Tj (t) −
bj | ≤ Uj (Tj (t)),
Aj xk − bj
≥Aj xk − bbj,T
j (t)
>0.
− Uj (Tj (t))
• When it returns “feasible” in line 3, no constraint is violated:
Ai xk − bi
≤Ai xk − bbi,T (t) + Ui (Ti (t))
i
≤Aj xk − bbj,Tj (t) + Uj (Tj (t))
<0,
∀i ∈ [m].
• When it returns ”feasible” in line 4, no constraint is violated more than ε2 :
Ai xk − bi
≤Ai xk − bbi,T (t) + Ui (Ti (t))
i
≤Aj xk − bbj,Tj (t) + Uj (Tj (t))
≤Aj xk − bbj,Tj (t) − Uj (Tj (t)) + 2Uj (Tj (t))
≤0 + ε2 ,
∀i ∈ [m].
Therefore the relaxed feasibility should be satisfied and the relaxed optimality is guaranteed by ellipsoid
algorithm.
Number of samples: We bound the number of samples used on each constraint separately. The number
of samples used on constraint i can be stated as the maximum Ti (t) where t is a mini-stage in which a
sample of bi is drawn. We bound Ti (t) by showing that Ui (Ti (t)) should be larger than a certain value if
bi is sampled at mini-stage t. This immediately give us an upper bound of Ti (t) since Ui (t) is a decreasing
function of t. Suppose bi is sampled at mini-stage t in ellipsoid iteration k. Let i∗ be the constraint with
largest violation. Conditioning on A holds, the fact that constraint i have a larger index than i∗ gives
Vi (k) + 2Ui (Ti (t))
≥Ai xk − bbi,T (t) + Ui (Ti (t))
i
≥Ai∗ xk − bbi∗ ,Ti∗ (t) + Ui∗ (Ti∗ (t))
≥Vi∗ (k).
(7)
which implies 2Ui (Ti (t)) ≥ V ∗ (k) − Vi (k). Now look at line 2 in UCB-method. If a sample of bi is drawn,
we should not quit in this step. So if Vi (k) > 0, we must have
Vi (k) − 2Ui (Ti (t))
≤Ai xk − bbi,T (t) − Ui (Ti (t))
i
≤0.
27
(8)
Similarly, because of line 3 in UCB-method, if Vi (k) ≤ 0, it should be satisfied that
Vi (k) + 2Ui (Ti (t))
≥Ai xk − bbi,T (t) + Ui (Ti (t))
i
≥0.
(9)
Putting inequality (7), (8) and (9) and Ui (Ti (t)) ≥ ε2 /2 together, we get the conclusion that 2Ui (Ti (t)) ≥
max{V ∗ (k) − Vi (k), |Vi (k)|, ε2 /2} = gapi,ε2 /2 (k) should be satisfied if we draw a sample of bi at ministage t in ellipsoid iteration k.
Then we do some calculation,
⇒
2Ui (Ti (t)) ≥ gapi,ε2 /2 (k)
s
2σi2 log(log(3Ti (t)/2)/δ′
6
≥ gapi,ε2 /2 (k)
Ti (t)
2
⇒
⇒
gapi,ε2 /2 (k)
log(log(3Ti (t)/2)/δ′ )
≥
Ti (t)
72σi2
20m
108σi2
log
Ti (t) ≤
δ
gap2i,ε2 /2 (k)
72σi2
log log
+
gap2i,ε2 /2 (k)
108σi2
gap2i,ε2 /2 (k)δ′
!
.
(10)
In the last step, we use the fact that for 0 < δ ≤ 1, c > 0,
1
log(3t/2)
· log
≥c
t
δ
2 log(3/(2cδ))
1
.
⇒ t ≤ log
c
δ
Take maximum of (10) over all k and according to the definition of ∆i,ε2/2 ,
108σ 2
Ti (t) ≤ 2 i log
∆i,ε2 /2
72σi2
log log
∆2i,ε2/2
20m
δ
108σi2
∆2i,ε2 /2 δ′
+
!
Therefore the overall number of samples is at most
O
X
i
σi2
∆2i,ε2 /2
m X σi2
+
log
log log
δ
∆2i,ε2 /2
i
28
σi2
∆2i,ε2/2
!!
.
| 8cs.DS
|
A simple solution to the word problem for virtual
braid groups
Paolo Bellingeri, Bruno A. Cisneros de la Cruz, Luis Paris
arXiv:1506.05283v2 [math.GR] 4 Mar 2016
March 7, 2016
Abstract
We show a simple and easily implementable solution to the word problem for virtual braid
groups.
AMS Subject Classification. Primary: 20F36. Secondary: 20F10, 57M25.
1
Introduction
Virtual braid groups were introduced by L. Kauffman in his seminal paper on virtual knots and
links [12]. They can be defined in several ways, such as in terms of Gauss diagrams [2, 9], in
terms of braids in thickened surfaces [9], and in terms of virtual braid diagrams. The latter will
be our starting point of view.
A virtual braid diagram on n strands is a n-tuple β = (b1 , . . . , bn ) of smooth paths in the plane
R2 satisfying the following conditions.
(a) bi (0) = (i, 0) for all i ∈ {1, . . . , n}.
(b) There exists a permutation g ∈ Sn such that bi (1) = (g(i), 1) for all i ∈ {1, . . . , n}.
(c) (p2 ◦ bi )(t) = t for all i ∈ {1, . . . , n} and all t ∈ [0, 1], where p2 : R2 → R denotes the
projection on the second coordinate.
(d) The bi ’s intersect transversely in a finite number of double points, called the crossings of
the diagram.
Each crossing is endowed with one of the following attributes: positive, negative, virtual. In
the figures they are generally indicated as in Figure 1.1. Let V BDn be the set of virtual
braid diagrams on n strands, and let ∼ be the equivalence relation on V BDn generated by
ambient isotopy and the virtual Reidemeister moves depicted in Figure 1.2. The concatenation
of diagrams induces a group structure on V BDn / ∼. The latter is called virtual braid group on
n strands, and is denoted by V Bn .
It was observed in [11, 18] that V Bn has a presentation with generators σ1 , . . . , σn−1 , τ1 , . . . , τn−1 ,
and relations
τi2 = 1
σi σj = σj σi , σi τj = τj σi , and τi τj = τj τi
σi σj σi = σj σi σj , σi τj τi = τj τi σj , and τi τj τi = τj τi τj
1
for 1 ≤ i ≤ n − 1
for |i − j| ≥ 2
for |i − j| = 1
positive
negative
virtual
Figure 1.1. Crossings in a virtual braid diagram.
Figure 1.2. Virtual Reidemeister moves.
A solution to the word problem for virtual braid groups was shown in [10]. However, this
solution is quite theoretical and its understanding requires some heavy technical knowledge on
Artin groups. Therefore, it is incomprehensible and useless for most of the potential users,
including low dimensional topologists. Moreover, its implementation would be difficult. Our
aim here is to show a new solution, which is simpler and easily implementable, and whose
understanding does not require any special technical knowledge. This new solution is in the
spirit of the one shown in [10], in the sense that one of the main ingredients in its proof is the
study of parabolic subgroups in Artin groups.
We have not calculated the complexity of this algorithm, as this is probably at least exponential
because of the inductive step 3 (see next section). Nevertheless, it is quite efficient for a limited
number of strands (see the example at the end of Section 2), and, above all, it should be useful
to study theoretical questions on V Bn such as the faithfulness of representations of this group
in automorphism groups of free groups and/or in linear groups. Note that the faithfulness of
such a representation will immediately provide another, probably faster, solution to the word
problem for V Bn .
The Burau representation easily extends to V Bn [18], but the question whether V Bn is linear
or not is still open. A representation of V Bn in Aut(Fn+1 ) was independently constructed in
[3] and [14], but such a representation has recently been proven to be not faithful for n ≥ 4
[8, Proposition 5.3] (see the example at the end of Step 1). So, we do not know yet any
representation on which we can test our algorithm.
In [8], Chterental shows a faithful action of V Bn on a set of objects that he calls “virtual curve
2
diagrams”. We have some hope to use this action to describe another explicit solution to the
word problem for V Bn . But, for now, we do not know any formal definition of this action, and
how it could be encoded in an algorithm.
Acknowledgments. The research of the first author was partially supported by French grant
ANR-11-JS01-002-01.
2
The algorithm
Our solution to the word problem for V Bn is divided into four steps. In Step 1 we define
a subgroup KBn of V Bn and a generating set S for KBn , and we show an algorithm (called
Algorithm A) which decides whether an element of V Bn belongs to KBn and, if yes, determines a
word over S ±1 which represents this element. For X ⊂ S, we denote by KBn (X ) the subgroup
of KBn generated by X . The other three steps provide a solution to the word problem for
KBn (X ) which depends recursively on the cardinality of X . Step 2 is the beginning of the
induction. More precisely, the algorithm proposed in Step 2 (called Algorithm B) is a solution
to the word problem for KBn (X ) when X is a full subset of S (the notion of ”full subset” will
be also defined in Step 2; for now, the reader just need to know that singletons are full subsets).
In Step 3 we suppose given a solution to the word problem for KBn (X ), and, for a given subset
Y ⊂ X , we show an algorithm which solves the membership problem for KBn (Y) in KBn (X )
(Algorithm C). In Step 4 we show an algorithm which solves the word problem for KBn (X )
when X is not a full subset, under the assumption that the group KBn (Y) has a solvable word
problem for any proper subset Y of X (Algorithm D).
2.1
Step 1
Recall that Sn denotes the group of permutations of {1, . . . , n}. We denote by θ : V Bn → Sn the
epimorphism which sends σi to 1 and τi to (i, i + 1) for all 1 ≤ i ≤ n − 1, and by KBn the kernel
of θ. Note that θ has a section ι : Sn → V Bn which sends (i, i + 1) to τi for all 1 ≤ i ≤ n − 1, and
therefore V Bn is a semi-direct product V Bn = KBn o Sn . The following proposition is proved
in Rabenda’s master thesis [15] which, unfortunately, is not available anywhere. However, its
proof can also be found in [4].
Proposition 2.1 (Rabenda [15]). For 1 ≤ i < j ≤ n we set
δi,j
= τi τi+1 · · · τj−2 σj−1 τj−2 · · · τi+1 τi ,
δj,i = τi τi+1 · · · τj−2 τj−1 σj−1 τj−1 τj−2 · · · τi+1 τi .
Then KBn has a presentation with generating set
S = {δi,j | 1 ≤ i 6= j ≤ n} ,
and relations
δi,j δk,` = δk,` δi,j
for i, j, k, ` distinct
δi,j δj,k δi,j = δj,k δi,j δj,k
for i, j, k distinct
The virtual braids δi,j and δj,i are depicted in Figure 2.1.
3
i
j
i
δi,j
j
δj,i
Figure 2.1. Generators for KBn .
The following is an important tool in the forthcoming Algorithm A.
Lemma 2.2 (Bardakov, Bellingeri [4]). Let u be a word over {τ1 , . . . , τn−1 }, let ū be the element
of V Bn represented by u, and let i, j ∈ {1, . . . , n}, i 6= j. Then ūδi,j ū−1 = δi0 ,j 0 , where i0 =
θ(ū)(i) and j 0 = θ(ū)(j).
−1
are
Note that τi−1 = τi , since τi2 = 1, for all i ∈ {1, . . . , n − 1}. Hence, the letters τ1−1 , . . . , τn−1
not needed in the above lemma and below.
±1
, τ1 , . . . , τn−1 }, decides
Now, we give an algorithm which, given a word u over {σ1±1 , . . . , σn−1
whether the element ū of V Bn represented by u belongs to KBn . If yes, it also determines a
±
word u0 over S ±1 = {δi,j
| 1 ≤ i 6= j ≤ n} which represents ū. The fact that this algorithm is
correct follows from Lemma 2.2.
±1
, τ1 , . . . , τn−1 }. We write u in the form
Algorithm A. Let u be a word over {σ1±1 , . . . , σn−1
u = v0 σiε11 v1 · · · v`−1 σiε`` v` ,
where v0 , v1 , . . . , v` are words over {τ1 , . . . , τn−1 }, and ε1 , . . . , ε` ∈ {±1}. On the other hand, for
a word v = τj1 · · · τjk over {τ1 , . . . , τn−1 }, we set θ(v) = (j1 , j1 + 1) · · · (jk , jk + 1) ∈ Sn . Note
that θ(ū) = θ(v0 ) θ(v1 ) · · · θ(v` ). If θ(ū) 6= 1, then ū 6∈ KBn . If θ(ū) = 1, then ū ∈ KBn , and ū
is represented by
u0 = δaε11 ,b1 δaε22 ,b2 · · · δaε`` ,b` ,
where
ak = θ(v0 · · · vk−1 )(ik ) and bk = θ(v0 · · · vk−1 )(ik + 1)
for all k ∈ {1, . . . , `}.
Example. In [8] it was proven that the Bardakov-Manturov representation of V Bn in Aut(Fn+1 )
(see for instance [3] for the definition) is not faithful, showing that the element ω = (τ3 σ2 τ1 σ2−1 )3
is non-trivial in V B4 while the corresponding automorphism of F5 is trivial. In [8] the nontriviality of ω is shown by means of an action on some curve diagrams, but this fact can be
easily checked with Algorithm A. Indeed, θ(ω) = ((3, 4)(1, 2))3 = (3, 4)(1, 2) 6= 1, hence ω 6= 1.
4
2.2
Step 2
Let S be a finite set. A Coxeter matrix over S is a square matrix M = (ms,t )s,t∈S , indexed by
the elements of S, such that ms,s = 1 for all s ∈ S, and ms,t = mt,s ∈ {2, 3, 4, . . . } ∪ {∞} for all
s, t ∈ S, s 6= t. We represent this Coxeter matrix with a labelled graph Γ = ΓM , called Coxeter
diagram. The set of vertices of Γ is S. Two vertices s, t ∈ S are connected by an edge labelled
by ms,t if ms,t 6= ∞.
m
If a, b are two letters and m is an integer ≥ 2, we set ha, bim = (ab) 2 if m is even, and
m−1
ha, bim = (ab) 2 a if m id odd. In other words, ha, bim denotes the word aba · · · of length m.
The Artin group of Γ is the group A = A(Γ) defined by the following presentation.
A = hS | hs, tims,t = ht, sims,t for all s, t ∈ S, s 6= t and ms,t 6= ∞i .
The Coxeter group of Γ, denoted by W = W (Γ), is the quotient of A by the relations s2 = 1,
s ∈ S.
Example. Let VΓn be the Coxeter diagram defined as follows. The set of vertices of VΓn is S.
If i, j, k, ` ∈ {1, . . . , n} are distinct, then δi,j and δk,` are connected by an edge labelled by 2. If
i, j, k ∈ {1, . . . , n} are distinct, then δi,j and δj,k are connected by an edge labelled by 3. There
is no other edge in VΓn . Then, by Proposition 2.1, KBn is isomorphic to A(VΓn ).
Let Γ be a Coxeter diagram. For X ⊂ S, we denote by ΓX the subdiagram of Γ spanned by
X, by AX the subgroup of A = A(Γ) generated by X, and by WX the subgroup of W = W (Γ)
generated by X. By [13], AX is the Artin group of ΓX , and, by [6], WX is the Coxeter group of
ΓX .
For X ⊂ S, we denote by KBn (X ) the subgroup of KBn generated by X . By the above,
KBn (X ) has a presentation with generating set X and relations
• st = ts if s and t are connected in VΓn by an edge labelled by 2,
• sts = tst if s and t are connected in VΓn by an edge labelled by 3.
Definition. We say that a subset X of S is full if any two distinct elements s, t of X are
connected by an edge of VΓn . Recall that the aim of Step 2 is to give a solution to the word
problem for KBn (X ) when X is full.
We denote by Fn = F (x1 , . . . , xn ) the free group of rank n freely generated by x1 , . . . , xn . For
i, j ∈ {1, . . . , n}, i 6= j, we define ϕi,j ∈ Aut(Fn ) by
ϕi,j (xi ) = xi xj x−1
i , ϕi,j (xj ) = xi , and ϕi,j (xk ) = xk for k 6∈ {i, j} .
It is easily checked from the presentation in Proposition 2.1 that the map S → Aut(Fn ), δi,j 7→
ϕi,j , induces a representation ϕ : KBn → Aut(Fn ). For X ⊂ S, we denote by ϕX : KBn (X ) →
Aut(Fn ) the restriction of ϕ to KBn (X ). The following will be proved in Section 3.
Proposition 2.3. If X is a full subset of S, then ϕX : KBn (X ) → Aut(Fn ) is faithful.
Notation. From now on, if u is a word over S ±1 , then ū will denote the element of KBn
represented by u.
Algorithm B. Let X be a full subset of S, and let u = sε11 · · · sε` ` be a word over X ±1 . We have
ϕX (ū) = ϕX (s1 )ε1 · · · ϕX (s` )ε` . If ϕ(ū) = Id, then ū = 1. Otherwise, ū 6= 1.
5
2.3
Step 3
Let G be a group, and let H be a subgroup of G. A solution to the membership problem for H in
G is an algorithm which, given g ∈ G, decides whether g belongs to H or not. In the present step
we will assume that KBn (X ) has a solution to the word problem, and, from this solution, we will
give a solution to the membership problem for KBn (Y) in KBn (X ), for Y ⊂ X . Furthermore,
if the tested element belongs to KBn (Y), then this algorithm will determine a word over Y ±1
which represents this element.
Let u be a word over S. (Remark: here the alphabet is S, and not S ±1 .)
• Suppose that u is written in the form u1 ssu2 , where u1 , u2 are words over S and s is an
element of S. Then we say that u0 = u1 u2 is obtained from u by an M -operation of type I.
• Suppose that u is written in the form u1 stu2 , where u1 , u2 are words over S and s, t are
two elements of S connected by an edge labelled by 2. Then we say that u0 = u1 tsu2 is
obtained from u by an M -operation of type II(2) .
• Suppose that u is written in the form u1 stsu2 , where u1 , u2 are words over S and s, t are
two elements of S connected by an edge labelled by 3. Then we say that u0 = u1 tstu2 is
obtained from u by an M -operation of type II(3) .
Let Y be a subset of S.
• Suppose that u is written in the form tu0 , where u0 is a word over S and t is an element of
Y. Then we say that u0 is obtained from u by an M -operation of type IIIY .
We say that u is M -reduced (resp. MY -reduced ) if its length cannot be shortened by M operations of type I, II(2) , II(3) (resp. of type I, II(2) , II(3) , IIIY ). An M -reduction (resp. MY reduction) of u is an M -reduced word (resp. MY -reduced word) obtained from u by M -operations
(resp. MY -operations). We can easily enumerate all the words obtained from u by M -operations
(resp. MY -operations), hence we can effectively determine an M -reduction and/or an MY reduction of u.
Let Y be a subset of S. From a word u = sε11 · · · sε` ` over S ±1 , we construct a word πY (u) over
Y ±1 as follows.
+
• For i ∈ {0, 1, . . . , `} we set u+
i = s1 · · · si (as ever, u0 is the identity).
• For i ∈ {0, 1, . . . , `} we calculate an MY -reduction vi+ of u+
i .
• For a word v = t1 · · · tk over S, we denote by op(v) = tk · · · t1 the anacycle of v. Let i ∈
+
+
{1, . . . , `}. If εi = 1, we set wi+ = vi−1
·si ·op(vi−1
). If εi = −1, we set wi+ = vi+ ·si ·op(vi+ ).
• For all i ∈ {1, . . . , `} we calculate an M -reduction ri of wi+ .
• If ri is of length 1 and ri ∈ Y, we set Ti = riεi . Otherwise we set Ti = 1.
• We set πY (u) = T1 T2 · · · T` .
The proof of the following is given in Section 4.
Proposition 2.4. Let Y be a subset of S. Let u, v be two words over S ±1 . If ū = v̄, then
πY (u) = πY (v). Moreover, we have ū ∈ KBn (Y) if and only if ū = πY (u).
6
Algorithm C. Take two subsets X and Y of S such that Y ⊂ X , and assume given a solution
to the word problem for KBn (X ). Let u be a word over X ±1 . We calculate v = πY (u). If
uv −1 6= 1, then ū 6∈ KBn (Y). If uv −1 = 1, then ū ∈ KBn (Y) and v is a word over Y ±1 which
represents the same element as u.
We can use Algorithm C to show that the representation ϕ : KBn → Aut(Fn ) of Step 2 is
not faithful. Indeed, let α = δ1,3 δ3,2 δ3,1 and β = δ2,3 δ1,3 δ3,2 . A direct calculation shows that
ϕ(α) = ϕ(β). Now, set X = S and Y = {δ1,3 , δ3,2 , δ3,1 }. We have πY (δ1,3 δ3,2 δ3,1 ) = δ1,3 δ3,2 δ3,1 ,
hence α ∈ KBn (Y), and we have πY (δ2,3 δ1,3 δ3,2 ) = 1 and β 6= 1, hence β 6∈ KBn (Y). So, α 6= β.
2.4
Step 4
Now, we assume that X is a non-full subset of S, and that we have a solution to the word
problem for KBn (Y) for any proper subset Y of X (induction hypothesis). We can and do
choose two proper subsets X1 , X2 ⊂ X satisfying the following properties.
(a) X = X1 ∪ X2 .
(b) Let X0 = X1 ∩ X2 . There is no edge in VΓn connecting an element of X1 \ X0 to an element
of X2 \ X0 .
It is easily seen from the presentations of the KBn (Xi )’s given in Step 2 that we have the
amalgamated product
KBn (X ) = KBn (X1 ) ∗KBn (X0 ) KBn (X2 ) .
Our last algorithm is based on the following result. This is well-known and can be found for
instance in [16, Chap. 5.2].
Proposition 2.5. Let A1 ∗B A2 be an amalgamated product of groups. Let g1 , . . . , g` be a
sequence of elements of A1 t A2 different from 1 and satisfying the following condition:
if gi ∈ A1 (resp. gi ∈ A2 ), then gi+1 ∈ A2 \B (resp. gi+1 ∈ A1 \B), for all i ∈ {1, . . . , `−1}.
Then g1 g2 · · · g` is different from 1 in A1 ∗B A2 .
Algorithm D. Let u be a word over X ±1 . We write u in the form u1 u2 · · · u` , where
• ui is either a word over X1±1 , or a word over X2±1 ,
• if ui is a word over X1±1 (resp. over X2±1 ), then ui+1 is a word over X2±1 (resp. over X1±1 ).
We decide whether ū is trivial by induction on `. Suppose that ` = 1 and u = u1 ∈ KBn (Xj )
(j ∈ {1, 2}). Then we apply the solution to the word problem for KBn (Xj ) to decide whether ū
is trivial or not. Suppose that ` ≥ 2. For all i we set vi = πX0 (ui ). If ui vi−1 6= 1 for all i, then
ū 6= 1. Suppose that there exists i ∈ {1, . . . , `} such that ui vi−1 = 1. Let u0i = v1 u2 if i = 1,
u0i = u`−1 v` if i = `, and u0i = ui−1 vi ui+1 if 2 ≤ i ≤ ` − 1. Set v = u1 · · · ui−2 u0i ui+2 · · · u` . Then
ū = v̄ and, by induction, we can decide whether v represents 1 or not.
7
2.5
Example
In order to illustrate our solution to the word problem for KBn , we turn now to give a more
detailed and efficient version of the algorithm for the group KB4 . We start with the following
observation.
Remark. For X ⊂ S, we denote by VΓn (X ) the full subgraph of VΓn spanned by X . Let X , Y
be two subsets of S. Note that an injective morphism of Coxeter graphs VΓn (Y) ,→ VΓn (X )
induces an injective homomorphism KBn (Y) ,→ KBn (X ). So, if we have a solution to the word
problem for KBn (X ), then such a morphism would determine a solution to the word problem
for KBn (Y).
The Coxeter graph VΓ4 is depicted in Figure 2.2. Our convention in this figure is that a full
edge is labelled by 3 and a dotted edge is labelled by 2. Note that there are two edges that go
through “infinity”, one connecting δ2,1 to δ4,3 , and one connecting δ1,4 to δ3,2 .
Figure 2.2. Coxeter graph VΓ4 .
Consider the following subsets of S.
X (1) = {δ1,2 , δ2,3 , δ3,4 , δ4,1 , δ3,1 }, X1 (1) = {δ1,2 , δ2,3 , δ3,4 , δ4,1 }, X2 (1) = {δ1,2 , δ2,3 , δ3,1 }.
X (2) = X (1) ∪ {δ4,2 }, X1 (2) = X (1), X2 (2) = {δ4,2 , δ3,4 , δ2,3 , δ3,1 }.
X (3) = X (2) ∪ {δ1,3 }, X1 (3) = X (2), X2 (3) = {δ1,3 , δ4,1 , δ3,4 , δ4,2 }.
X (4) = X (3) ∪ {δ2,4 }, X1 (4) = X (3), X2 (4) = {δ2,4 , δ1,3 , δ4,1 , δ1,2 , δ3,1 }.
X (5) = X (4) ∪ {δ1,4 }, X1 (5) = X (4), X2 (5) = {δ1,4 , δ4,2 , δ2,3 , δ3,1 }.
X (6) = X (5) ∪ {δ2,1 }, X1 (6) = X (5), X2 (6) = {δ2,1 , δ1,3 , δ3,4 , δ4,2 , δ1,4 }.
X (7) = X (6) ∪ {δ3,2 }, X1 (7) = X (6), X2 (7) = {δ3,2 , δ2,4 , δ4,1 , δ1,3 , δ2,1 , δ1,4 }.
X (8) = X (7) ∪ {δ4,3 } = S, X1 (8) = X (7), X2 (8) = {δ4,3 , δ3,2 , δ2,4 , δ1,2 , δ3,1 , δ1,4 , δ2,1 }.
Let k ∈ {1, . . . , 8}. Note that X (k) = X1 (k) ∪ X2 (k). The Coxeter graph VΓ4 (X (k)) is depicted
in Figure 2.3. In this figure the elements of X1 (k) are represented by punctures, while the
elements of X2 (k) are represented by small circles.
8
VΓ4 (X (1))
VΓ4 (X (5))
VΓ4 (X (2))
VΓ4 (X (3))
VΓ4 (X (6))
VΓ4 (X (4))
VΓ4 (X (7))
VΓ4 (X (8))
Figure 2.3. Coxeter graph VΓ4 (X (k)).
We solve the word problem for KB4 (X (k)) successively for k = 1, 2, . . . , 8, thanks to the following
observations. Since X (8) = S, this will provide a solution to the word problem for KB4 .
(1) Let k ∈ {1, . . . , 8}. Set X0 (k) = X1 (k) ∩ X2 (k). Observe that there is no edge in VΓ4
connecting an element of X1 (k) \ X0 (k) to an element of X2 (k) \ X0 (k). Hence, we can solve
with Algorithm D the word problem for KB4 (X (k)) from solutions to the word problem
for KB4 (X1 (k)) and for KB4 (X2 (k)).
(2) The subsets X1 (1) and X2 (1) are full, hence we can solve the word problem for KB4 (X1 (1))
and for KB4 (X2 (1)) with Algorithm B.
(3) Let k ≥ 2. On the one hand, we have X1 (k) = X (k−1). On the other hand, it is easily seen
that there is an injective morphism VΓ4 (X2 (k)) ,→ VΓ4 (X (k − 1)). Hence, by the remark
given at the beginning of the subsection, we can solve the word problem for KB4 (X1 (k))
and for KB4 (X2 (k)) from a solution to the word problem for KB4 (X (k − 1)).
3
Proof of Proposition 2.3
Recall that Fn = F (x1 , . . . , xn ) denotes the free group of rank n freely generated by x1 , . . . , xn ,
and that we have a representation ϕ : KBn → Aut(Fn ) which sends δi,j to ϕi,j , where
ϕi,j (xi ) = xi xj x−1
i , ϕi,j (xj ) = xi , and ϕi,j (xk ) = xk for k 6∈ {i, j} .
For X ⊂ S, we denote by ϕX : KBn (X ) → Aut(Fn ) the restriction of ϕ to KBn (X ). In this
section we prove that ϕX is faithful if X is a full subset of S.
9
Consider the following groups.
σi σj σi = σj σi σj if |i − j| = 1
,
Bn = σ1 , . . . , σn−1
σi σj = σj σi
if |i − j| ≥ 2
σi σj σi = σj σi σj if i ≡ j ± 1 mod n
,
B̃n = σ1 , . . . , σn
σi σj = σj σi
if i 6= j and i 6≡ j ± 1 mod n
n ≥ 3.
The group Bn is the classical braid group, and B̃n is the affine braid group.
We define representations ψn : Bn → Aut(Fn ) and ψ̃n : B̃n → Aut(Fn ) in the same way as ϕ as
follows.
ψn (σi )(xi ) = xi xi+1 x−1
i , ψn (σi )(xi+1 ) = xi , ψn (σi )(xk ) = xk if k 6∈ {i, i + 1}
ψ̃n (σi )(xi ) = xi xi+1 x−1
i , ψ̃n (σi )(xi+1 ) = xi , ψ̃n (σi )(xk ) = xk if k 6∈ {i, i + 1} , for i < n
ψ̃n (σn )(xn ) = xn x1 x−1
n , ψ̃n (σn )(x1 ) = xn , ψ̃n (σn )(xk ) = xk if k 6∈ {1, n}
The key of the proof of Proposition 2.3 is the following.
Theorem 3.1 (Artin [1], Bellingeri, Bodin [5]). The representations ψn : Bn → Aut(Fn ) and
ψ̃n : B̃n → Aut(Fn ) are faithful.
The support of a generator δi,j is defined to be supp(δi,j ) = {i, j}. The support of a subset X
of S is supp(X ) = ∪s∈X supp(s). We say that two subsets X1 and X2 of S are perpendicular 1 if
supp(X1 ) ∩ supp(X2 ) = ∅. Note that this condition implies that X1 ∩ X2 = ∅. More generally,
we say that a family X1 , . . . , X` of subsets of S is perpendicular if supp(Xi ) ∩ supp(Xj ) = ∅ for
all i 6= j. In that case we write X1 ∪ · · · ∪ X` = X1 · · · X` . We say that a subset X of S
is indecomposable if it is not the union of two perpendicular nonempty subsets. The following
observations will be of importance in what follows.
Remark. Let X1 and X2 be two perpendicular subsets of S, and let X = X1 X2 .
(1) X is a full subset if and only if X1 and X2 are both full subsets.
(2) KBn (X ) = KBn (X1 ) × KBn (X2 ).
Indeed, if δi,j ∈ X1 and δk,` ∈ X2 , then i, j, k, ` are distinct, and therefore δi,j and δk,` are
connected by an edge labelled by 2, and δi,j δk,` = δk,` δi,j .
Lemma 3.2. Let X1 and X2 be two perpendicular subsets of S, and let X = X1 X2 . Then
ϕX : KBn (X ) → Aut(Fn ) is faithful if and only if ϕX1 : KBn (X1 ) → Aut(Fn ) and ϕX2 :
KBn (X2 ) → Aut(Fn ) are both faithful.
Proof. For X ⊂ {x1 , . . . , xn }, we denote by F (X) the subgroup of Fn generated by X. There
is a natural embedding ιX : Aut(F (X)) ,→ Aut(Fn ) defined by
α(xi ) if xi ∈ X
ιX (α)(xi ) =
xi
otherwise
1
This terminology is derived from the theory of Coxeter groups.
10
Moreover, if X1 and X2 are disjoint subsets of {x1 , . . . , xn }, then the homomorphism
(ιX1 × ιX2 ) : Aut(F (X1 )) × Aut(F (X2 )) →
Aut(Fn )
(α1 , α2 )
7→ ιX1 (α1 ) ιX2 (α2 )
is well-defined and injective. From now on, we will assume Aut(F (X)) to be embedded in
Aut(Fn ) via ιX , for all X ⊂ {x1 , . . . , xn }.
By abuse of notation, for X ⊂ S, we will also denote by supp(X ) the set {xi | i ∈ supp(X )}. Set
X1 = supp(X1 ) and X2 = supp(X2 ). We have Im(ϕXi ) ⊂ Aut(F (Xi )) for i = 1, 2, X1 ∩ X2 = ∅,
and KBn (X ) = KBn (X1 )×KBn (X2 ). Hence, Lemma 3.2 follows from the following claim whose
proof is left to the reader.
Claim. Let f1 : G1 → H1 and f2 : G2 → H2 be two group homomorphisms. Let (f1 × f2 ) :
(G1 × G2 ) → (H1 × H2 ) be the homomorphism defined by (f1 × f2 )(u1 , u2 ) = (f1 (u1 ), f2 (u2 )).
Then (f1 × f2 ) is injective if and only if f1 and f2 are both injective.
For 2 ≤ m ≤ n we set
Zm = {δ1,2 , . . . , δm−1,m } , Z̃m = {δ1,2 , . . . , δm−1,m , δm,1 } .
Note that the map {σ1 , . . . , σm−1 } → Zm , σi 7→ δi,i+1 , induces an isomorphism fm : Bm →
KBn (Zm ). This follows from the presentation of KBn (Zm ) given in Step 2 of Section 2. Similarly, for m ≥ 3, the map {σ1 , . . . , σm } → Z̃m , σi 7→ δi,i+1 for 1 ≤ i ≤ m − 1, σm 7→ δm,1 , induces
an isomorphism f˜m : B̃m → KBn (Z̃m ).
Recall that the symmetric group Sn acts on S by g δi,j = δg(i),g(j) , and that this action induces
an action of Sn on KBn . On the other hand, there is a natural embedding Sn ,→ Aut(Fn ),
where g ∈ Sn sends xi to xg(i) for all i ∈ {1, . . . , n}, and this embedding induces by conjugation
an action of Sn on Aut(Fn ). It is easily seen that the homomorphism ϕ : KBn → Aut(Fn ) is
equivariant under these actions of Sn .
Lemma 3.3. If X is a full and indecomposable nonempty subset of S, then there exist g ∈ Sn
and m ∈ {2, . . . , n} such that either X = g Zm , or X = g Z̃m and m ≥ 3.
Proof. An oriented graph Υ is the data of two sets, V (Υ), called set of vertices, and E(Υ),
called set of arrows, together with two maps sou, tar : E(Υ) → V (Υ). We associate an oriented
graph ΥX to any subset X of S as follows. The set of vertices is V (ΥX ) = supp(X ), the set of
arrows is E(ΥX ) = X , and, for δi,j ∈ X , we set sou(δi,j ) = i and tar(δi,j ) = j. Assume that X
is a full and indecomposable nonempty subset of S. Since X is indecomposable, ΥX must be
connected. Since X is full, if s, t ∈ X are two different arrows of ΥX with a common vertex,
then there exist i, j, k ∈ {1, . . . , n} distinct such that either s = δj,i and t = δi,k , or s = δi,j
and t = δk,i . This implies that ΥX is either an oriented segment, or an oriented cycle with at
least 3 vertices (see Figure 3.1). If ΥX is an oriented segment, then there exist g ∈ Sn and
m ∈ {2, . . . , n} such that X = g Zm . If ΥX is an oriented cycle, then there exist g ∈ Sn and
m ∈ {3, . . . , n}, such that X = g Z̃m .
Proof of Proposition 2.3. Let X be a full nonempty subset of S. Write X = X1 · · · X` ,
where Xj is an indecomposable nonempty subset. As observed above, each Xj is also a full subset.
11
Figure 3.1. Oriented segment and oriented cycle.
Moreover, by Lemma 3.2, in order to show that ϕX is faithful, it suffices to show that ϕXj is
faithful for all j ∈ {1, . . . , `}. So, we can assume that X is a full and indecomposable nonempty
subset of S. By Lemma 3.3, there exist g ∈ Sn and m ∈ {2, . . . , n} such that either X = g Zm ,
or X = g Z̃m and m ≥ 3. Since ϕ is equivariant under the actions of Sn , upon conjugating by
g −1 , we can assume that either X = Zm , or X = Z̃m . Set Zm = {x1 , . . . , xm } = supp(Zm ) =
−1 and ϕ
˜−1
supp(Z̃m ), and identify Fm with F (Zm ). Then ϕZm = ψm ◦ fm
Z̃m = ψ̃m ◦ fm , hence
ϕX is faithful by Theorem 3.1.
4
Proof of Proposition 2.4
The proof of Proposition 2.4 is based on some general results on Coxeter groups and Artin
groups. Recall that the definitions of Coxeter diagram, Artin group and Coxeter group are
given at the beginning of Step 2 in Section 2. Recall also that, if Y is a subset of the set S of
vertices of Γ, then ΓY denotes the full subdiagram spanned by Y , AY denotes the subgroup of
A = A(Γ) generated by Y , and WY denotes the subgroup of W = W (Γ) generated by Y .
Let Γ be a Coxeter diagram, let S be its set of vertices, let A be the Artin group of Γ, and let
W be its Coxeter group. Since we have s2 = 1 in W for all s ∈ S, every element g in W can be
represented by a word over S. Such a word is called an expression of g. The minimal length of
an expression of g is called the length of g and is denoted by lg(g). An expression of g of length
lg(g) is a reduced expression of g. Let Y be a subset of S, and let g ∈ W . We say that g is
Y -minimal if it is of minimal length among the elements of the coset WY g. The first ingredient
in our proof of Proposition 2.4 is the following.
Proposition 4.1. (Bourbaki [6, Chap. IV, Exercise 3]). Let Y ⊂ S, and let g ∈ W . There
exists a unique Y -minimal element lying in the coset WY g. Moreover, the following conditions
are equivalent.
(a) g is Y -minimal,
(b) lg(sg) > lg(g) for all s ∈ Y ,
(c) lg(hg) = lg(h) + lg(g) for all h ∈ WY .
Remark. For g ∈ W and s ∈ S, we always have either lg(sg) = lg(g) + 1, or lg(sg) = lg(g) − 1.
This is a standard fact on Coxeter groups that can be found for instance in [6]. So, the inequality
lg(sg) > lg(g) means lg(sg) = lg(g)+1 and the inequality lg(sg) ≤ lg(g) means lg(sg) = lg(g)−1.
Let u be a word over S.
• Suppose that u is written in the form u1 ssu2 , where u1 , u2 are words over S and s is an
element of S. Then we say that u0 = u1 u2 is obtained from u by an M -operation of type I.
12
• Suppose that u is written in the form u = u1 hs, tims,t u2 , where u1 , u2 are words over S
and s, t are two elements of S connected by an edge labelled by ms,t . Then we say that
u0 = u1 ht, sims,t u2 is obtained from u by an M -operation of type II.
We say that a word u is M -reduced if its length cannot be shortened by M -operations of type
I, II. The second ingredient in our proof is the following.
Theorem 4.2 (Tits [17]). Let g ∈ W .
(1) An expression w of g is a reduced expression if and only if w is M -reduced.
(2) Any two reduced expressions w and w0 of g are connected by a finite sequence of M operations of type II.
Let Y be a subset of S. The third ingredient is a set-retraction ρY : A → AY to the inclusion
map ιY : AY → A, constructed in [10, 7]. This is defined as follows. Let α be an element of A.
• Choose a word α
b = s1 ε1 · · · s` ε` over S ±1 which represents α.
• Let i ∈ {0, 1, . . . , `}. Set gi = s1 s2 · · · si ∈ W , and write gi in the form gi = hi ki , where
hi ∈ WY and ki is Y -minimal.
−1
• Let i ∈ {1, . . . , `}. If εi = 1, set zi = ki−1 si ki−1
. If εi = −1, set zi = ki si ki−1 .
• Let i ∈ {1, . . . , `}. We set Ti = ziεi if zi ∈ Y . Otherwise we set Ti = 1.
• Set ρ̂Y (α) = T1 T2 · · · T` .
Proposition 4.3 (Godelle, Paris [10], Charney, Paris [7]). Let α ∈ A. The element ρY (α) ∈ AY
represented by the word ρ̂Y (α) defined above does not depend on the choice of the representative α
b
of α. Furthermore, the map ρY : A → AY is a set-retraction to the inclusion map ιY : AY ,→ A.
We turn now to apply these three ingredients to our group KBn and prove Proposition 2.4. Let
2 = 1, 1 ≤ i 6= j ≤ n. Note that KW is the
KWn denote the quotient of KBn by the relations δi,j
n
Coxeter group of the Coxeter diagram VΓn . For Y ⊂ X , we denote by KWn (Y) the subgroup
of KWn generated by Y.
Lemma 4.4. Let g ∈ KWn .
(1) An expression w of g is a reduced expression if and only if w is M -reduced.
(2) Any two reduced expressions w and w0 of g are connected by a finite sequence of M operations of type II(2) and II(3) .
(3) Let Y be a subset of S, and let w be a reduced expression of g. Then g is Y-minimal (in
the sense given above) if and only if w is MY -reduced.
Proof. Part (1) and Part (2) are Theorem 4.2 applied to KWn . So, we only need to prove Part
(3).
Suppose that g is not Y-minimal. By Proposition 4.1, there exists s ∈ Y such that lg(sg) ≤ lg(g),
that is, lg(sg) = lg(g) − 1. Let w0 be a reduced expression of sg. The word sw0 is an expression
of g and lg(sw0 ) = lg(w) = lg(g), hence sw0 is a reduced expression of g. By Theorem 4.2,
w and sw0 are connected by a finite sequence of M -operations of type II(2) and II(3) . On the
other hand, w0 is obtained from sw0 by an M -operation of type IIIY . So, w0 is obtained from
13
w by M -operations of type I, II(2) , II(3) and IIIY , and we have lg(w0 ) < lg(w), hence w is not
MY -reduced.
Suppose that w is not MY -reduced. Let w0 be an MY -reduction of w, and let g 0 be the element
of KWn represented by w0 . Since w0 is an MY -reduction of w, the element g 0 lies in the coset
KWn (Y) g. Moreover, lg(g 0 ) = lg(w0 ) < lg(w) = lg(g), hence g is not Y-minimal.
Proof of Proposition 2.4. Let Y be a subset of S. Consider the retraction ρY : KBn →
KBn (Y) constructed in Proposition 4.3. We shall prove that, if u is a word over S ±1 , then
πY (u) = ρY (ū). This will prove Proposition 2.4. Indeed, if ū = v̄, then πY (u) = ρY (ū) =
ρY (v̄) = πY (v). Moreover, since ρY : KBn → KBn (Y) is a retraction to the inclusion map
KBn (Y) ,→ KBn , we have ρY (ū) = ū if and only if ū ∈ KBn (Y), hence πY (u) = ū if and only
if ū ∈ KBn (Y).
Let u = sε11 · · · sε` ` be a word over S ±1 . Let α be the element of KBn represented by u.
• For i ∈ {0, 1, . . . , `}, we set u+
i = s1 · · · si , and we denote by gi the element of KWn
+
represented by ui .
• Let i ∈ {0, 1, . . . , `}. We write gi = hi ki , where hi ∈ KWn (Y), and ki is Y-minimal. Let
+
vi+ be an MY -reduction of u+
i . Then, by Lemma 4.4, vi is a reduced expression of ki .
−1
+
+
• Let i ∈ {1, . . . , `}. If εi = 1, we setzi = ki−1 si ki−1
and wi+ = vi−1
· si · op(vi−1
). If εi = −1,
−1
+
+
+
+
we set zi = ki si ki and wi = vi · si · op(vi ). Note that wi is an expression ofzi .
• Let i ∈ {1, . . . , `}. Let ri be an M -reduction of wi+ . By Lemma 4.4, ri is a reduced
expression of zi . Note that we have zi ∈ Y if and only if ri is of length 1 and ri ∈ Y.
• Let i ∈ {1, . . . , `}. If ri is of length 1 and ri ∈ Y, we set Ti = riεi . Otherwise we set Ti = 1.
• By construction, we have ρ̂Y (α) = πY (u) = T1 T2 · · · T` .
References
[1] E. Artin. Theory of braids. Ann. of Math. (2) 48 (1947), 101–126.
[2] D. Bar Natan, Z. Dancso. Finite Type Invariants of w-Knotted Objects I: w-Knots and the
Alexander Polynomial. arXiv:1405.1956.
[3] V. G. Bardakov. Virtual and welded links and their invariants. Siberian Electronic Mathematical
Reports, Volume: 2, page 196–199, 2005.
[4] V. G. Bardakov, P. Bellingeri. Combinatorial properties of virtual braids. Topology Appl. 156
(2009), no. 6, 1071–1082.
[5] P. Bellingeri, A. Bodin. The braid group of a necklace. To appear in Math. Z..
[6] N. Bourbaki. Eléments de mathématique. Fasc. XXXIV. Groupes et algèbres de Lie. Chapitre IV:
Groupes de Coxeter et systèmes de Tits. Chapitre V: Groupes engendrés par des réflexions. Chapitre
VI: Systèmes de racines. Actualités Scientifiques et Industrielles, No. 1337, Hermann, Paris, 1968.
[7] R. Charney, L. Paris. Convexity of parabolic subgroups in Artin groups. Bull. Lond. Math. Soc.
46 (2014), no. 6, 1248–1255.
14
[8] O. Chterental. Virtual braids and virtual curve diagrams. J. Knot Theory Ramifications 24 (2015),
no. 13, 1541001, 24 pp.
[9] Bruno A. Cisneros De La Cruz. Virtual braids from a topological viewpoint. J. Knot Theory
Ramifications 24 (2015), no. 6, 1550033, 36 pp.
[10] E. Godelle, L. Paris. K(π, 1) and word problems for infinite type Artin-Tits groups, and applications to virtual braid groups. Math. Z. 272 (2012), no. 3-4, 1339–1364.
[11] S. Kamada. Braid presentation of virtual knots and welded knots. Osaka J. Math. 44 (2007), no.
2, 441–458.
[12] L. H. Kauffman. Virtual knot theory. European J. Combin. 20 (1999), no. 7, 663–690.
[13] H. van der Lek. The homotopy type of complex hyperplane complements. Ph. D. Thesis, Nijmegen,
1983.
[14] V. O. Manturov. On the recognition of virtual braids. Zap. Nauchn. Sem. S. Peterburg. Otdel. Mat.
Inst. Steklov. (POMI) 299 (Geom. i Topol. 8) (2003) 267–286, 331–332 (in Russian). Translation in
J. Math. Sci. (N. Y.) 131 (2005) 5409–5419.
[15] L. Rabenda. Mémoire de DEA (Master thesis), Université de Bourgogne, 2003.
[16] J.-P. Serre. Arbres, amalgames, SL2 . Astérisque, No. 46. Société Mathématique de France, Paris,
1977.
[17] J. Tits. Le problème des mots dans les groupes de Coxeter. Symposia Mathematica (INDAM, Rome,
1967/68), Vol. 1, pp. 175–185, Academic Press, London, 1969.
[18] V. V. Vershinin. On homology of virtual braids and Burau representation. J. Knot Theory Ramifications 10 (2001), no. 5, 795–812.
Paolo Bellingeri,
LMNO UMR6139, CNRS, Université de Caen, F-14000 Caen, France.
E-mail: [email protected]
Bruno A. Cisneros de la Cruz,
Instituto de Matemáticas de la UNAM - Oaxaca, Oaxaca de Juárez, Oax. 68000, Mexico.
E-mail: [email protected]
Luis Paris,
IMB UMR5584, CNRS, Univ. Bourgogne Franche-Comté, F-21000 Dijon, France.
E-mail: [email protected]
15
| 4math.GR
|
arXiv:cs/0102003v1 [cs.CE] 2 Feb 2001
Fast Pricing of European Asian Options with
Provable Accuracy: Single-stock and Basket Options
Karhan Akcoglu†k, Ming-Yang Kao‡¶, and Shuba Raghavan§
† Department of Computer Science, Yale University, New Haven, CT 06520, USA.
‡ Department of Electrical Engineering and Computer Science, Tufts University,
Medford, MA 02155, USA.
§ Department of Economics, Yale University, New Haven, CT 06520, USA.
E-mail: [email protected], [email protected], [email protected]
Abstract. This paper develops three polynomial-time pricing techniques for European Asian options with provably small errors, where the stock prices follow
binomial trees or trees of higher-degree. The first technique is the first known Monte
Carlo algorithm with analytical error bounds suitable for pricing single-stock options
with meaningful confidence and speed. The second technique is a general recursive
bucketing-based scheme that can use the Aingworth-Motwani-Oldham aggregation
algorithm, Monte-Carlo simulation and possibly others as the base-case subroutine.
This scheme enables robust trade-offs between accuracy and time over subtrees of
different sizes. For long-term options or high-frequency price averaging, it can price
single-stock options with smaller errors in less time than the base-case algorithms
themselves. The third technique combines Fast Fourier Transform with bucketingbased schemes for pricing basket options. This technique takes polynomial time in the
number of days and the number of stocks, and does not add any errors to those already
incurred in the companion bucketing scheme. This technique assumes that the price
of each underlying stock moves independently.
k Supported in part by NSF Grant CCR-9896165.
¶ Supported in part by NSF Grants CCR-9531028 and CCR-9988376. Part of this work was performed
while this author was visiting Department of Computer Science, Yale University.
Fast Pricing of European Asian Options
2
1. Introduction
A call (respectively, put) option is a contract assigning its holder the right, but not
the obligation, to buy (respectively, sell) a security at some future time for a specified
strike price X [13]. If the holder exercises her right, the other party in the contract, the
writer , is obligated to assume the opposite side of the transaction. In exchange for this
right, the holder pays the writer an option price P . The security in this contract can
be any financial asset; for the purpose of this paper, we restrict it to a single stock or a
portfolio of stocks. An option in the latter case is commonly called a basket option; for
clarity, we call an option in the former case a single-stock option.
Options are popular financial instruments for a variety of trading strategies. For
example, options can be used to hedge risk. As protection from a potential price fall in
a stock price, one can purchase a put on the stock, thereby locking in a minimum sell
price. On the other hand, options can provide additional income for stockholders who
write calls on their holdings; of course, this strategy carries the risk of being forced to
sell the stock should the calls be exercised.
An option is valid until its expiry date. For a European option, the holder may
exercise it only on the expiry date. For an American option, the holder may exercise it
on any date up to and including the expiry date. The payoff of an option is the amount
of money its holder makes on the contract. A European call is worth exercising if and
only if S ≥ X, where S is the stock price on the expiry date. The payoff of the call is
(S − X)+ = max(S − X, 0). For an American call, S is set to the stock price at the
exercise time. An Asian option comes in European and American flavors, depending on
when it may be exercised. For an European Asian call, if A is the average stock price
over the entire life of the contract up to the expiry date, the payoff is (A − X)+ . For
an American Asian call, A is set to the average stock price up to the exercise date. The
payoffs of puts can be symmetrically defined.
The price P of an option is the discounted expected value of the payoff
with an appropriate martingale measure. Because of the popularity of options,
pricing techniques for computing P have been extensively researched [3, 5, 6, 13, 16, 19].
Generally, it is more difficult to price a basket option than a single-stock option. To
compute P , the price movement of each individual stock needs to be modeled. Typically,
it is modeled as Brownian motion with drift. Using a stochastic differential equation, P
can then be computed via a closed-form solution to the equation. When a closed-form
solution is not known, various approaches are used to find an approximate solution.
One class of approaches involves approximating the solution using numerical methods.
Other approaches approximate the Brownian motion model with a discrete model, and
use this model to approximate P . One such discrete model is the binomial tree model,
due to Cox, Ross, and Rubenstein [10]; see Section 2 for the definition of the model.
This paper develops three polynomial-time pricing techniques with provably small
errors. The remaining discussion makes the following assumptions:
(i) The option in question is an European Asian single-stock or basket call.
Fast Pricing of European Asian Options
3
(ii) Our task is to price the call at the start of its contract life.
(iii) The price of each underlying stock follows the binomial tree model.
(iv) In the case of a basket call, the price of each underlying stock moves independently.
Our results generalize easily for puts, for a later time point than the start of the contract
life, and for trees with higher degrees than two. The cases of American options and of
interdependent stocks remain open.
Monte Carlo simulation has been commonly used in the financial community.
Despite this popularity, most reported results on error bounds are experimental or
heuristic [1, 17]. Our first technique is the first known Monte Carlo algorithm that
has analytical error bounds suitable for pricing a European Asian call with meaningful
confidence and speed. As shown by Theorem 5, the number of simulations required is
polynomial in (P1) the logarithm of the inverse of the error probability and (P2) the
inverse of the price error relative to the strike price but is exponential in (E1) the square
root of the number of underlying stocks and (E2) the volatility of these stocks over the
call’s life. In particular, the algorithm is reasonably fast and accurate for a single-stock
European Asian call with reasonable volatility.
Monte Carlo simulation is a randomized technique, and thus there is always a
nonzero probability that the price obtained by polynomial-time Monte Carlo simulation
is not accurate enough. The aggregation algorithm of Aingworth, Motwani, and Oldham
(AMO) [1] is the first polynomial-time algorithm for pricing single-stock European
Asian calls and other path-dependent options with guaranteed worst-case price errors.
The AMO algorithm is based on a simple yet powerful idea called bucketing. Our
second technique is a general recursive bucketing-based scheme that can use the AMO
algorithm, Monte-Carlo simulation, and possibly others as the base-case subroutine.
This scheme enables robust trade-offs between accuracy and time over subtrees of
different sizes. For long-term options or high-frequency price averaging, it can price
single-stock European Asian calls with smaller error bounds in less time than the
base-case algorithms themselves. In particular, as implied by Theorem 14, given the
same runtime, this recursive scheme prices more accurately than the AMO algorithm;
similarly, given the same accuracy, the scheme runs faster than the AMO algorithm.
This recursive scheme works for calls written on a single stock. Our third technique
combines Fast Fourier Transform (FFT) [8] and bucketing-based schemes to price
basket calls and is applicable to European Asian calls as well as others. As shown
in Theorem 18, this technique takes polynomial time in the number of days and the
number of stocks, and does not add any errors to those already incurred in the companion
bucketing schemes.
The remainder of this paper is organized as follows. Section 2 reviews the binomial
tree model and basic definitions. Section 3 describes the new Monte Carlo algorithm.
Section 4 details the recursive scheme. Section 5 gives the FFT-based technique for
pricing basket calls. Section 6 concludes the paper with directions for further research.
Fast Pricing of European Asian Options
4
2. The Binomial Tree Model
A binomial tree T is a recombinant binary tree. If n is the depth of T , T has t + 1
nodes at depth t, for 0 ≤ t ≤ n. For 0 ≤ i ≤ t, let T [t, i] (or simply [t, i] if T is obvious
from context) be the i-th node at level t of T . For t > 0, T [t, 0] and T [t, t] have one
parent each, T [t − 1, 0] and T [t − 1, t − 1] respectively. For 0 < i < t, T [t, i] has two
.
parents, T [t − 1, i − 1] and T [t − 1, i]. The number of nodes in T is (n+1)(n+2)
2
Given a stock in the binomial tree model, the stock price is assumed to follow a
geometric random walk through T . Time is divided into n equal periods, with the
root T [0, 0] corresponding to time t = 0, when the option is priced, and the leaves
T [n, · ] corresponding to time t = n, the expiry date of the option. Let s(T [t, i]) (or
simply s(t, i)) be the stock price at node T [t, i]. At each time step, the stock price
s(t, i) rises to s(t + 1, i + 1) = u · s(t, i)—an uptick —with probability p or falls to
s(t + 1, i) = d · s(t, i)—a downtick —with probability q = 1 − p. Letting r denote the
risk-free interest rate, the√ parameters u and d satisfy 0 < d ≤ 1+r ≤ u and are typically
taken to be u = d1 = eσ/ n , where σ is the n-period volatility, or standard deviation, of
the stock price [13]. Although the probability p of an uptick is not known in general,
for the purposes of pricing options, we can use the risk-neutral probability model [13],
, where r is the risk-free interest rate for one period. This
which states that p = (1+r)−d
u−d
makes the expected return on the stock over one period, pu + (1 − p)d, equal to the
risk-free return, 1 + r.
Let Ω be the sample space of paths ω = (ω1 , . . . , ωn ) down T , where each
ωt ∈ {−1, 1}, with −1 corresponding to a downtick and 1 corresponding to an uptick.
Given ω ∈ Ω and 0 ≤ t ≤ n, let T [t, ω] be the unique node at level t that ω passes
through. Similar to the notation introduced above, we let s(T [t, ω]) = s(t, ω) be the
price at node T [t, ω].
We define the random variables Y1 , . . . , Yn on Ω by Yt (ω) = ωt , the t-th component
of ω. We define the probability measure Π on Ω to be the unique measure for which the
random variables Y1 , . . . , Yn are independent, identically distributed (i.i.d.) with P (Yi =
1) = p and P (Yi = −1) = q. We start with an initial fixed stock price S0 = s(TP[0, 0]).
t
For 1 ≤ t ≤ n, the stock price St is a random variable defined by St = S0 u i=1 Yi .
From the structure of the binomial tree, we have Pr(St = s(t, i)) = ti pi (1 − p)t−i ,
P
where 0 ≤ i ≤ t. The running total of the stock prices is defined by Tt =P ti=0 Si
t
and the running average is At = Tt /(t + 1). For ω ∈ Ω, we let St (ω) = S0 u i=1 Yi (ω) ,
P
Tt (ω) = ti=0 Si (ω), and At (ω) = Tt (ω)/(t + 1).
Recall that X is the strike price of a European Asian call. Using the above notation,
the price of this call is
1
1
E((Tn − (n + 1)X)+ ) =
E( max(Tn − (n + 1)X, 0)).
E((An − X)+ ) =
n+1
n+1
For large n, it is not known if this quantity can be computed exactly because the
stock price can follow exponentially many paths down the tree. Below, we show how to
estimate E((Tn −(n+1)X)+), from which the price of the option can be easily computed.
Fast Pricing of European Asian Options
5
3. A New Monte Carlo Algorithm
Monte Carlo simulation methods for asset pricing were introduced to finance by Boyle
[4]. They are very popular in pricing complex instruments, particularly path-dependent
European-style options. These methods involve randomly sampling paths ω ∈ Ω
according to the distribution Π and computing the payoff (An (ω) −X)+ on each sample.
Suppose N samples ω 1 , . . . , ω N are taken from Ω. The price estimate of the call is
N
1 X
(An (ω i ) − X)+ .
µ=
N i=1
The accuracy of this estimate depends on the number of simulations N and the variance
τ 2 of √the payoff: the error bound typically guaranteed by Monte Carlo methods is
O(τ / N ). Generally τ 2 is not known, and itself must be estimated to determine the
error bound of µ.
A number of techniques are used to reduce the error of µ. For example, [7, 18] use
the control variate technique, which ties the price of the option to the price of another
instrument (the control variate) for which an analytically tractable solution is known.
The simulation then estimates the difference between the option price and the control
variate, which can be determined with greater accuracy than the option price itself. The
antithetic variate method [9, 12] follows not only randomly sampled paths ω down the
binomial tree, but also the “mirror images” of each ω. None of these techniques have
known analytical error bounds for µ.
Below we use concentration of measure results in conjunction with Monte Carlo
simulation to estimate the price of a European Asian call and derive analytical error
bounds for this estimate. The error bounds are in terms of the strike price X of the
option and the maximum volatility σmax of the underlying stocks.
3.1. Analytical Error Bounds for the Single-stock Case
Let C = eE(ln Tn ) . In this section, we show that if (n + 1)X/C is “small”, E((Tn − (n +
1)X)+ ) is close to E(Tn − (n + 1)X) = E(Tn ) − (n + 1)X (Theorem 3(1)). The option
is deep-in-the-money and will probably be exercised. Since a closed-form formula exists
for E(Tn ), as described in Lemma 6, E(Tn − (n + 1)X) can be computed exactly, and
our algorithm uses it as our estimate for E((Tn − (n + 1)X)+ ). On the other hand, if
(n + 1)X/C is not small, the variance of (Tn − (n + 1)X)+ can be bounded from above
(Theorem 3(2)) and our algorithm estimates its expectation with bounded error using
Monte Carlo simulation.
We first give some theoretical results, then show how these results can be used in
our Monte Carlo algorithm, BoundedMC. We begin by finding bounds A and B such
that Tn ∈ [A, B] with high probability. Our main tool for this is an inequality arising
from the theory of martingales, usually known as Azuma’s Inequality [2, 15]. We use
the inequality in a form taken from Frieze [11]. Suppose we have a random variable
U = U(V ), where V = (V1 , . . . , Vη ) and, for 1 ≤ i ≤ η, Vi is chosen independently from
6
Fast Pricing of European Asian Options
the probability space Ωi . For V, W ∈ Ω = Ω1 × · · · × Ωη , write V ≅ W if V and W
differ in exactly one component.
Theorem 1 (Azuma’s Inequality) Suppose that V ≅ W implies |U(V ) − U(W )| ≤
a, for some constant a > 0. Then, for any real b ≥ 0,
2
− 2b2
Pr(|U − E(U)| ≥ b) ≤ 2e
a η
.
Lemma 2 Let C = eE(ln Tn ) . For any λ > 0, we have
2 /2
Pr (Tn ≤ Ce−σλ or Tn ≥ Ceσλ ) ≤ 2e−λ
,
where σ is the volatility of the stock.
Proof:
We apply Azuma’s Inequality to the random variable U = ln Tn . Let
ωX , ωY ∈ Ω be two paths that differ in a single component, say on day t. Assume
without loss of generality that ωX has an uptick on day t while ωY has a downtick.
X)
X)
Then TTnn(ω
≤ u2 and | ln TTnn(ω
| = | ln Tn (ωX ) − ln Tn (ωY )| ≤ 2 ln u = 2 √σn .
(ωY )
(ωY )
Applying Azuma’s Inequality with a = 2 √σn and b = σλ, we have
2 /2
Pr(| ln Tn − E(ln Tn )| ≥ σλ) ≤ 2e−λ
.
The claimed result follows from the fact that
| ln Tn − E(ln Tn )| ≥ σλ ⇔ ln Tn − E(ln Tn ) ≥ σλ or ln Tn − E(ln Tn ) ≤ −σλ
⇔ eln Tn /C ≥ eσλ or eln Tn /C ≤ e−σλ .
Now, fix ε > 0 and choose λ0 =
q
2 ln 2ε . Then, by Lemma 2,
Pr (Tn ≤ Ce−σλ or Tn ≥ Ceσλ ) ≤ ε.
Theorem 3(1) says that if (n + 1)X < Ce−σλ0 , then E((Tn − (n + 1)X)+ ) is close to
E(Tn −(n+1)X). Otherwise, Theorem 3(2) says that Var((Tn −(n+1)X)+ ) is bounded.
Theorem 3 Let C = eE(ln Tn ) .
(i) If (n + 1)X < Ce−σλ0 , then
E((Tn − (n + 1)X)+ ) − E(Tn − (n + 1)X) ≤ ε(n + 1)X.
(ii) If (n + 1)X ≥ Ce−σλ0 , then
Var((Tn − (n + 1)X)+ ) ≤ (n + 1)2 X 2 e4σλ0
1 + 2σε
.
λ0 − 2σ
Proof: Let φ(t) denote the probability density function of Tn .
Statement 1. Note first that
Z ∞
Z ∞
+
+
E((Tn − (n + 1)X) ) =
(t − (n + 1)X) φ(t)dt =
(t − (n + 1)X)φ(t)dt
0
= E(Tn − (n + 1)X) −
(n+1)X
Z
0
(n+1)X
(t − (n + 1)X)φ(t)dt.
7
Fast Pricing of European Asian Options
Then
|E((Tn − (n + 1)X)+ ) − E(Tn − (n + 1)X)|
Z (n+1)X
=
(t − (n + 1)X)φ(t)dt
≤
0
= (n + 1)X Pr (Tn ≤ (n + 1)X)
≤ ε(n + 1)X,
≤
(n + 1)X
Z
(n+1)X
φ(t)dt
0
(n + 1)X Pr (Tn ≤ Ce−σλ0 )
where the second-last inequality follows from the assumption that (n + 1)X < Ce−σλ0
and the last inequality follows from Lemma 2 and our choice of λ0 .
Statement 2. Since (n + 1)X ≥ Ce−σλ0 ,
Z ∞
+
+ 2
=
Var((Tn − (n + 1)X) ) ≤ E ((Tn − (n + 1)X) )
(t − (n + 1)X)2φ(t)dt
(n+1)X
Z ∞
=2
(t − (n + 1)X) Pr(Tn ≥ t)dt
(n+1)X
Z ∞
≤2
(t − Ce−σλ0 ) Pr(Tn ≥ t)dt
Ce−σλ0
Ceσλ0
Z
−σλ0
(t − Ce
Ce−σλ0
Z
2 2σλ0
2
≤C e
+ 4C σ
=2
∞
) Pr(Tn ≥ t)dt + 2
2 /2
e2σλ e−λ
Z
∞
Ceσλ0
(t − Ce−σλ0 ) Pr(Tn ≥ t)dt
dλ
λ0
(substitute t = Ceσλ in the second term)
Z ∞
1
2
2 2σλ0
2
2σ2
=C e
+ 4C σe
e− 2 (λ−2σ) dλ
λ0
2σλ0 −λ20 /2
R∞ 1 2
e
1 2
(from the inequality t e− 2 x dx ≤ 1t e− 2 t )
λ − 2σ
0
2σε
2 2 4σλ0
1+
≤ (n + 1) X e
λ0 − 2σ
≤ C 2 e2σλ0 + 4C 2 σ
e
2
(since (n + 1)Xe2σλ0 ≥ Ceσλ0 and ε = 2e−λ0 /2 )
1 + 2σε
≤ (n + 1)2 X 2 e4σλ0
.
λ0 − 2σ
3.2. The BoundedMC Algorithm
We next use these results in our algorithm. One approach would be to estimate C = eln Tn
to determine whether we should apply Theorem 3(1) or 3(2). Our algorithm takes a
more direct approach. We begin by selecting N samples ω 1 , . . . , ω N ∈ Ω and computing
Tn (ω 1 ), . . . , Tn (ω N ). For 1 ≤ i ≤ N, define the random variable Zi as
(
1 if Tn (ω i ) ≤ (n + 1)X,
Zi =
0 otherwise.
P
Let Z = N
i=1 Zi .
8
Fast Pricing of European Asian Options
Lemma 4 Let 0 < δ < 1 be given. With N = Θ(log 1δ ) trials, the following statements
hold.
(i) If
(ii) If
Z
N
Z
N
≤ 2ε, then Pr (Tn ≤ (n + 1)X) ≤ 4ε with probability 1 − δ.
> 2ε, then Pr (Tn ≤ (n + 1)X) ≥ ε with probability 1 − δ.
Proof: This follows from the Chernoff bound [15, Theorems 4.1 and 4.2].
Theorem 5 Let 0 < δ < 1 be given. With N = Θ(log 1δ +
following statements hold.
1 4σλ0 1+2σε
e
)
ε2
λ0 −2σ
trials, the
Z
≤ 2ε, then, with probability 1 − δ, E(Tn − (n + 1)X) estimates E((Tn − (n +
(i) If N
1)X)+ ) with error at most 4ε(n + 1)X. Correspondingly, the price of the call is
estimated with error at most 4εX.
P
+
Z
i
> 2ε, then, with probability 1 − δ, N1 N
(ii) If N
i=1 (Tn (ω ) − (n + 1)X) estimates
E((Tn − (n + 1)X)+ ) with standard deviation at most ε(n + 1)X. Correspondingly,
the price of the call is estimated with standard deviation at most εX.
Proof: The statements are proved as follows.
Statement 1. With probability 1 − δ, we have
|E((Tn − (n + 1)X)+ ) − E(Tn − (n + 1)X)| ≤ (n + 1)X Pr(Tn ≤ (n + 1)X)
≤ 4ε(n + 1)X,
where the first inequality follows from the proof of Theorem 3(1) and the second
1
E((Tn − (n +
inequality follows from Lemma 4(1). Since the price of the call is just n+1
+
1)X) ), the error in our estimate for the price is at most 4εX. We give a closed-form
formula for E(Tn − (n + 1)X) = E(Tn ) − (n + 1)X in Lemma 6 below.
Z
Statement 2. If N
> 2ε, then by Lemma 4(2), with probability 1 − δ, Pr(Tn ≤
(n + 1)X) ≥ ε. This implies (n + 1)X ≥ Ce−σλ0 . Otherwise, if (n + 1)X < Ce−σλ0 , we
would have
ε ≤ Pr(Tn ≤ (n + 1)X) ≤ Pr(Tn ≤ Ce−σλ0 ),
which contradicts our choice of λ0 above. Since (n + 1)X ≥ Ce−σλ0 , we can
P
i
+
for
apply Theorem 3(2). The variance of our estimate N1 N
i=1 (Tn (ω ) − (n + 1)X)
1
+
+
E((Tn − (n + 1)X) ) is N Var((Tn − (n + 1)X) ). Since
1 + 2σε
Var (Tn − (n + 1)X)+ ≤ (n + 1)2 X 2 e4σλ0
,
λ0 − 2σ
and its
the variance of our estimate is at most ε2 (n + 1)2 X 2 for N ≥ ε12 e4σλ0 λ1+2σε
0 −2σ
1
standard deviation is ε(n+1)X. Since the price of the call is just n+1 E((Tn −(n+1)X)+ ),
the expected error in our estimate for the payoff is at most εX.
We now summarize our algorithm.
9
Fast Pricing of European Asian Options
Algorithm 1 BoundedMC(δ, ε)
) paths, ω 1, . . . , ω N ;
generate N = Θ(log 1δ + ε12 e4σλ0 λ1+2σε
0 −2σ
let Z be the number of paths ω i such that Tn (ω i ) ≤ (n + 1)X;
if Z/N ≤ 2ε
1
1
E(Tn − (n + 1)X) = n+1
(E(Tn ) − (n + 1)X);
return n+1
else
P
i
+
return N1 N
i=1 (An (ω ) − X) .
For the completeness of Theorem 5, we conclude this section with a formula for
E(Tn ) [1].
Lemma 6 Let S0 be the initial stock price and let r be the interest rate. Then
(
(n + 1)S0 ,
if r = 0,
E(Tn ) =
(1+r)n+1 −1
S0 , if r > 0.
r
Proof:
Recall that s(t, ℓ) is the stock price at node T [t, ℓ], u and d are the uptick/downtick factors, and p and q are the risk neutral probabilities. We have
X
X
E(Tn ) =
s(t, ℓ) · Pr(St = s(t, ℓ))
level 0 ≤ t ≤ n node 0 ≤ ℓ ≤ t
n X
t
X
t ℓ
ℓ t−ℓ
p (1 − p)t−ℓ
S0 u d ·
=
ℓ
t=0 ℓ=0
n X
t
X
t
(pu)ℓ ((1 − p)d)t−ℓ
= S0
ℓ
t=0 ℓ=0
= S0
n
X
t=0
= S0
n
X
(pu + (1 − p)d)t
(1 + r)t
(using the binomial theorem)
(assuming risk-neutral rate of return)
t=0
=
(
(n + 1)S0 ,
if r = 0,
(1+r)n+1 −1
S0 ,
r
if r > 0,
as claimed.
3.3. Pricing Basket Options using BoundedMC
The results derived above are applicable for European Asian basket calls as well.
Suppose we have a basket of m stocks. Let Tn denote the running total of the
stocks in the basket up to day n, let Ωi be the sample space of paths the binomial
tree corresponding to stock i and let Πi be the corresponding measure on Ωi . Let
Ω = Ω1 × · · · × Ωm and Π = Π1 × · · · × Πm . Lemma 2 is again applicable, but this time
10
Fast Pricing of European Asian Options
our random variable U = ln Tn has mn components. Letting σmax denote the maximum
volatility among the m stocks, we can show, as in Lemma 2, that
√
Pr (Tn ≤ Ce−σmax λ
m
√
or Tn ≥ Ceσmax λ
m
2 /2
) ≤ 2e−λ
,
for any λ > q
0. The results of the single stock case are now applicable. Given ε > 0,
√
√
choose λ0 = 2 ln 2ε , so that Pr (Tn ≤ Ce−σmax λ m or Tn ≥ Ceσmax λ m ) ≤ ε. We sample
N paths ω 1 , . . . , ω N in Ω according to Π. Let Z be be number of paths ω i such that
Tn (ω i ) ≤ (n + 1)X, where X is the strike price of the basket call.
Theorem 7 Let 0 < δ < 1 be given. With N = Θ(log 1δ +
trials, the following statements hold.
√
√
1 4σmax λ0 m 1+2σmax ε √m
)
e
ε2
λ0 −2σmax m
Z
(i) If N
≤ 2ε, then, with probability 1 − δ, E(Tn − (n + 1)X) estimates E((Tn − (n +
1)X)+ ) with error at most 4ε(n + 1)X. Correspondingly, the price of the call is
estimated with error at most 4εX.
P
+
Z
i
> 2ε, then, with probability 1 − δ, N1 N
(ii) If N
i=1 (Tn (ω ) − (n + 1)X) estimates
E((Tn − (n + 1)X)+ ) with standard deviation at most ε(n + 1)X. Correspondingly,
the price of the call is estimated with standard deviation at most εX.
Proof: This theorem is analogous to Theorem 5 and can be proven using similar
techniques.
4. A Recursive Bucketing-Based Scheme
The AMO algorithm takes O(kn2 ) time to produce a price estimate in the range
, P , where P is the exact price of the call and k is any natural number. As in
P − nX
k
Section 3, this algorithm estimates E((Tn −(n+1)X)+ ), from which the price of the call,
E((An − X)+ ), can be easily estimated. Our recursive scheme is a generalization of a
variant of the AMO algorithm. Below we first describe this variant, called Bucketed Tree
Traversal (BTT) and then detail our scheme, called Recursive Bucketed Tree Traversal
(RecBTT).
Given binomial tree T of depth n, and numbers t, i, m such that 0 ≤ t ≤ n,
[t,i]
0 ≤ i ≤ t, and m ≤ n − t, let Tm be the subtree of depth m rooted at node T [t, i].
Given 0 ≤ t ≤ n, let ω|t be the prefix of ω up to level t of T . Note that ω|n = ω. Given
ψ, ω ∈ Ω, we say that ψ is an extension of ω|m if, for 0 ≤ t ≤ m, we have ψt = ωt . Given
another binomial tree U of depth n we say that ψ ∈ Ω(U) is isomorphic to ω ∈ Ω(T ),
if, for all 0 ≤ t ≤ n, ψt = ωt .
Like the AMO algorithm, BTT is based on the following simple observation.
Suppose that the running total Tm (ω) = Tm (ω|m) of the stock prices on path ω ∈ Ω
exceeds the barrier B = (n + 1)X. Then, for any extension ψ of ω|m, Tn (ψ) also exceeds
B and the call will be exercised. If we know the call will be exercised on all extensions
of ω|m , it is easy to compute the payoff of the call on these extensions, as described in
Lemma 6.
Fast Pricing of European Asian Options
11
As we travel down a path ω, once the running total Tm (ω) exceeds B, we can just
keep track of the running total on extensions ψ of ω|m weighted by Π(ψ), from which
the value of the option can be computed. Hence, we only need to individually keep
track of path prefixes ω|m that have running totals Tm (ω|m ) less than B.
Unfortunately, there may be exponentially many such ω|m. However, the running
totals Tm (ω|m ) are in the bounded range [0, B). Rather than trying to keep track of
each running total individually, we instead group the running totals terminating at each
node into buckets that subdivide this interval. This introduces some round-off error.
Suppose we use k buckets to divide [0, B) into equal-length subintervals and we use the
left endpoint of each interval as the representative value of the running totals contained
in that bucket. At each step down the tree, when we put a running total into a bucket,
an error of at most Bk is introduced. Traveling down n levels of the tree, the total Tn (ω)
and the average An (ω) is underestimated
of a path ω is underestimated by at most nB
k
B
by at most k .
BTT is detailed in Algorithm 2. At each node v = T [t, i] of T , create k + 1
buckets to store partial sums of path prefixes terminating at v. There will be k core
buckets and one overflow bucket. This overflow bucket is the only difference between
BTT and the AMO algorithm. For 0 ≤ j < k, core bucket bj (v) stores the probability
mass bj (v).mass of path prefixes that terminate at node v and have running totals in its
range range(j) = [j Bk , (j+1) Bk ). The representative value of partial sums in the bucket is
denoted by bj (v).value = j Bk . The overflow bucket bk (v) stores the probability-weighted
running total estimates of path prefixes that have estimated running totals exceeding
B. This quantity is denoted by bk (v).value. The probability mass of these path prefixes
is denoted by bk (v).mass. BTT iterates through each of the k + 1 buckets of each of
nodes in T , for a total runtime of O(kn2 ).
the (n+1)(n+2)
2
Algorithm 2 BTT(T , k, B)
for each node v ∈ T and each bucket bj (v)
set bj (v).mass ← 0;
take j such that initial price s(T [0, 0]) ∈ range(j); bj (T [0, 0]).mass ← 1; % assume
s(T [0, 0]) < B
for t = 0, . . . , (n − 1) % iterate through each level
for i = 0, . . . , t % iterate through each node at level t
let v = T [t, i]; % shorthand notation for node T [t, i]
for w ∈ {T [t + 1, i], T [t + 1, i + 1]} % for each child of node v
let p′ ∈ {p, q} be the probability of going from node v to w;
for bj (v) ∈ {b0 (v), . . . , bk (v)} % for each bucket at node v
let V ← bj (v).value + s(w);
let M ← bj (v).mass × p′ ;
if V < B
take ℓ such that V ∈ range(ℓ);
bℓ (w).mass ← bℓ (w).mass + M;
12
Fast Pricing of European Asian Options
else % in overflow bucket
k (w).value+M ×V
;
bk (w).value ← bk (w).mass×b
bk (w).mass+M
b (w).mass ← bk (w).mass + M;
Pn k
return i=0 bk (n, i).mass × (bk (n, i).value − B).
% return option price estimated from overflow buckets at leaves
We propose RecBTT, a recursive extension of BTT. Consider some level t in our
binomial tree T and assume that the weights of all path prefixes terminating at level t
have been put into the appropriate buckets. BTT uses these weights to compute the
bucket weights of nodes at level t + 1. In contrast, RecBTT recursively solves the
[t,i]
problem for subtrees Tm , 0 ≤ i ≤ t, of some depth m < n − t rooted at node T [t, i].+
As each recursive call is complete, RecBTT Merges the bucket weights at the leaves
[t,i]
of Tm into the corresponding nodes at level t+m of T . The advantages of the recursive
calls are twofold.
(i) They use finer bucket granularity, resulting in improved accuracy.
[t,i]
(ii) The results of a single recursive call on a particular subtree Tm are used to
[t,j]
Estimate the results of other recursive calls to other subtrees Tm , where j > i,
[t,j]
as long as the node prices in Tm are “sufficiently close” to the corresponding node
[t,i]
prices in Tm . This improves the runtime, since we do not need to make all t + 1
of the recursive calls, so there are portions of T that we do not directly traverse.
4.1. The Merge Procedure
[t ,i ]
Consider a recursive call on the subtree T1 = Tn10 0 of depth n1 rooted at node
v0 = T [t0 , i0 ]. A leaf v1 = T1 [n1 , i1 ] of T1 (0 ≤ i1 ≤ n1 ) corresponds to the node
v2 = T [t0 + n1 , i0 + i1 ] of T . Merge incorporates the bucket weights at v1 into the
bucket weights at v2 . Recall that the recursive call on T1 is made with finer bucket
granularity. Assume we use k1 = h1 k0 core buckets instead of just k0 = k. We first
combine each group of h1 buckets at v1 into a single bucket, so that we are left with
k0 buckets to merge into v2 . When we refer to a bucket bj1 (v1 ) below, 0 ≤ j1 < k0 , we
mean one of these k0 combined buckets.
The Merge procedure is described in Algorithm 3. Consider first the core buckets.
Let 0 ≤ j0 , j1 , j2 < k0 denote core bucket indices. Bucket bj0 (v0 ) contains the mass of
path prefixes in T terminating at v0 whose running total estimates fall into the interval
range(j0 ) = [j0 kB0 , (j0 + 1) kB0 ). Bucket bj1 (v1 ) contains the mass of full paths in T1
terminating at node v1 whose running total estimates fall into the interval range(j1 ).
This is equal to the mass of partial paths in T starting at node v0 and terminating
at node v2 whose running total estimates fall into the interval range(j1 ). Merging T1
into T involves merging each leaf v1 of T1 into the corresponding node v2 of T . Once
+
[t,i]
Actually, the trees on which we recursively solve the problem are not exactly the Tm . Each is
[t,i]
identical to the respective Tm , except the price at the root is changed from s(T [t, i]) to 0. The reason
for this is explained in Remark 1.
Fast Pricing of European Asian Options
13
the merging procedure is done, bj2 (v2 ).mass is updated to contain the weight of path
prefixes in T passing through node v0 and terminating at node v2 that have running
total estimates in the interval range(j2 ). The overflow buckets are handled similarly.
[t ,i ]
Algorithm 3 Merge(T , T1 = Tn10 0 )
let v0 ← T [t0 , i0 ];
for i1 = 0, . . . , n1 % for each leaf of T1
let v1 ← T1 [n1 , i1 ], v2 ← T [t0 + n1 , i0 + i1 ];
for j0 = 0, . . . , k0 % buckets in v0
for j1 = 0, . . . , k0 % buckets in v1
let V ← bj0 (v0 ).value + bj1 (v1 ).value;
let M ← bj0 (v0 ).mass × bj1 (v1 ).mass;
if V < B
take j2 such that V ∈ range(j2 );
bj2 (v2 ).mass ← bj2 (v2 ).mass + M;
else
k (v2 ).value+M ×V
;
bk (v2 ).value ← bk (v2 ).mass×b
bk (v2 ).mass+M
bk (v2 ).mass ← bk (v2 ).mass + M.
Remark 1 Notice that the price at v0 is counted twice: once in the path prefix from the
root T [0, 0] to v0 and once in the partial path between v0 and v2 . To address this issue,
when we recursively solve the problem on the subtree T1 , we set the price s(T1 [0, 0]) at
the root to be 0, ensuring that this price is counted once. This modification does not
change our algorithms.
Lemma 8 For an arbitrary node v, let E(v) be the maximum amount by which running
totals terminating at node v are underestimated by the bucket values. Using the Merge
algorithm, we have E(v2 ) ≤ E(v0 ) + E(v1 ) + Bk .
Proof: Consider first the core buckets. Bucket bj0 (v0 ) covers running totals in the
range [j0 Bk , (j0 + 1) Bk ) and bj1 (v1 ) covers running totals in the range [(j1 ) Bk , (j1 + 1) Bk ).
The concatenation of paths in these buckets covers running totals in the range
[(j0 +j1 ) Bk , (j0 +j1 +2) Bk ). Assuming j0 +j1 < k, we put these paths in bucket bj0 +j1 (v2 ),
and the running totals are underestimated by the original error of E(v0 ) + E(v1 ), plus
an additional Bk , since some partial paths that fall in the range of bucket bj0 +j1 +1 (v2 ) are
actually put in bj0 +j1 (v2 ). If j0 + j1 ≥ k, then these paths are put in the overflow bucket.
In this case, no additional error is introduced. The original error of E(v0 ) + E(v1 ) is
carried over. Similar arguments can be made when we are merging overflow buckets.
Lemma 9 Merge can be made to run in O(n1 k log k) time.
14
Fast Pricing of European Asian Options
Proof:
In the implementation of Algorithm 3, Merge runs in O(n1 k 2 ) time.
However, the core buckets can be merged with a faster technique. Let a0 (x) =
Pk−1
Pk−1
j1
j0
j1 =0 bj1 (v1 ).mass·x be polynomial representations
j0 =0 bj0 (v0 ).mass·x and a1 (x) =
of the bucket masses at nodes v0 and v1 . Let a2 (x) = a0 (x) · a1 (x). This product can be
computed in O(k log k) time with the Fast Fourier Transform (FFT) [8]. For 0 ≤ j2 < k,
the coefficient of the xj2 term of a2 (x) is the probability mass that should be added to
bucket bj2 (v2 ). For j2 > k, this coefficient should be added to the overflow bucket bk (v2 ).
So the core buckets can be merged in O(k log k) time per node. Merging the overflow
buckets takes O(k) time, so the total runtime per node is O(k log k).
4.2. The Estimate Procedure
[t ,i ]
[t ,i ]
Let T1 = Tm 0 1 and T2 = Tm 0 2 be two subtrees of T , where i2 > i1 We now describe
the Estimate(T2 , T1 ) procedure, which estimates the weights in the leaf buckets of T2
from the weights in the leaf buckets of T1 . This saves us the work of recursively solving
the problem on T2 .
Estimate is described in Algorithm 4. It uses the following fact. Given any node
v1 = T1 [t, i] in T1 , let v2 = T2 [t, i] be the corresponding node in T2 . Notice that there
is a constant α > 1 such that for all (v1 , v2 ) pairs, s(v2 ) = αs(v1 ). Hence, for any path
ψ ∈ Ω(T2 ), we have Tm (ψ) = αTm (ω), where ω ∈ Ω(T1 ) is isomorphic to ψ.
[t ,i ]
[t ,i ]
Algorithm 4 Estimate(T2 = Tm 0 2 , T1 = Tm 0 1 )
for i = 0, . . . , m % go through the leaf buckets of T1 and T2
let v1 ← T1 [m, i], v2 ← T2 [m, i];
for j1 = 0, . . . , k % go through each bucket at v1
let V ← αbj1 (v1 ).value;
let M ← bj1 (v1 ).mass;
if V < B
take j2 such that V ∈ range(j2 );
bj2 (v2 ).mass ← bj2 (v2 ).mass + M;
else
k (v2 ).value+M ×V
;
bk (v2 ).value ← bk (v2 ).mass×b
bk (v2 ).mass+M
bk (v2 ).mass ← bk (v2 ).mass + M;
Lemma 10 Suppose that α ≤ 2 and assume that the total path sums in T1 are
underestimated by our bucketing scheme by at most E. Estimate underestimates
the total path sums in T2 by at most 2E + 2 Bk .
Proof: Take any path ω ∈ Ω(T1 ) and let Um (ω) be the estimate for Tm (ω) made by the
bucketing scheme. By assumption, Tm (ω) − Um (ω) ≤ E. Let ψ ∈ Ω(T2 ) be isomorphic
to ω. Using αUm (ω) as an estimate for Tm (ψ) = αTm (ω), we underestimate Tm (ψ) by
at most αE ≤ 2E. This accounts for the first error term.
15
Fast Pricing of European Asian Options
Now, consider core bucket bj1 (v1 ) of some leaf v1 of T1 . The range of running totals
covered by this bucket is range(j1 ) = [j1 Bk , (j1 +1) Bk ). When we map this bucket to T2 , it
covers path sums in the range [αj1 Bk , α(j1 + 1) Bk ), which falls into at most three buckets
in the corresponding node v2 of T2 , since α ≤ 2. Call these buckets bj2 (v2 ), bj2 +1 (v2 ),
and bj2 +2 (v2 ) (assume these are all core buckets). Putting the entire mass into bj2 (v2 )
causes the portion of partial sums that fall into bucket bj2 +2 (v2 ) to be underestimated
by additional 2 Bk , which accounts for the second error term.
Similar errors occur if the range [αj1 Bk , α(j1 + 1) Bk ) spills into the overflow bucket
bk (v2 ) and when mapping the overflow bucket bk (v1 ) in v1 to T2 .
Lemma 11 Suppose that we would like to determine the leaf bucket weights of the
[t,i]
subtrees
Tm , where 0 ≤ i ≤ t. We need only call RecBTT at most once for every
√
Θ( σn ) subtrees, and use the Estimate procedure to estimate the leaf bucket weights
of the other subtrees with bounded error.
[t,0]
Proof: We begin by recursively solving the problem on the subtree Tm . Notice that
s(T [t, i]) = u2i s(T [t, 0]). By Lemma 10, as long as u2i ≤ 2, we can use Estimate to
[t,i]
compute the leaf bucket
weights of Tm , without introducing significant error. Recalling
√
the stock, we see that we can use
that we take u = eσ/ n , where σ is the√ volatility of
√
[t,i]
[t,0]
n ln 2
n
Estimate(Tm , Tm ) for i as large as 2σ = Θ( σ ) while keeping the error bounded.
This proves the lemma.
4.3. Error and Runtime Analysis
We now derive recursive expressions for the error and runtime of RecBTT. Suppose
there are n0 trading periods and we use a total of k0 buckets per node. For i > 0, let
ni be the number of trading periods and ki the number of buckets per node we use in
the i-th subproblem (at the i-th level into the recursion). Let Ei = E(ni , ki ) be the
error bound made in the i-th subproblem. The error is the amount by which RecBTT
underestimates the running totals. Let Ti = T (ni , ki ) be the runtime of RecBTT on
the i-th subproblem, including all recursive calls. We recursively call RecBTT on
binomial trees of decreasing depth (decreasing ni ) with an increasing number of buckets
(increasing ki ) until we reach a tree of some small depth.
Lemma 12 When i recursive calls are made, the error made by RecBTT is
E0 = 5Bn0
i
X
j=0
2i+1 n0
2j
Ẽi+1 ,
+
kj nj+1
ni+1
where Ẽi+1 is the error in solving the (i + 1)-st subproblem.
Proof: Consider the error in the i-th subproblem, Ei = E(ni , ki ). Each recursive
call has an error of Ei+1 . When the ki+1 smaller granularity buckets in the (i + 1)-st
problem are grouped into the ki < ki+1 buckets of the i-th problem, an error of one
16
Fast Pricing of European Asian Options
bucket size, kBi , is introduced. A call to Estimate underestimates the price sums by
at most 2(Ei+1 + kBi ) + 2 kBi = 4 kBi + 2Ei+1 , by Lemma 10. This error is made at most
ni
i
times, once at each of levels ni+1 , 2ni+1 , . . . , nni+1
ni+1 . At each of these levels, an
ni+1
additional error of kBi is introduced by the Merge procedure, as described in Lemma 8.
The total error in the i-th subproblem is
2ni
5Bni
+
Ei+1 .
Ei =
ki ni+1 ni+1
Unraveling i levels of the recursion on the original problem yields the result.
For the runtime of RecBTT, we must analyze how many times the (i + 1)-st
subproblem needs to be solved when solving the i-th subproblem. At each of levels
i
0, ni+1 , . . . , ( nni+1
− 1)ni+1 in the binomial subtree of the i-th subproblem, suppose the
(i + 1)-st subproblem is called at most ci times.
Lemma 13 When i recursive calls are made, the runtime of RecBTT is
!
Q
j−1
i
Y
X
n0 ik=0 ck
T̃i+1 ,
T0 = O
ck +
n0 nj kj log(kj )
ni+1
j=0
k=0
where T̃i+1 is the time it takes to solve the (i + 1)-st subproblem.
Proof:
Consider the runtime of the i-th subproblem, Ti = T (ni , ki ). Each
recursive call takes time Ti+1 and at most ci recursive calls are made at each of levels
i
− 1)ni+1 . The leaf bucket weights of the other at most ni subtrees
0, ni+1 , . . . , ( nni+1
rooted at the nodes at each of these levels can be determined using Estimate, which
i
levels t, each of the at
takes O(ki ni+1 ) time. For the Merge procedure, at each of nni+1
[t,j]
most ni nodes at level t must be merged with ni+1 leaves from the subtrees Tni+1 to get
the bucket weights at level t + ni+1 . The merging of each node takes O(ki log ki ) time.
The total runtime of the i-th subproblem is
ni
ni ci
Ti =
(ci Ti+1 + O(ni ki ni+1 ) + O(ni ni+1 ki log ki )) =
Ti+1 + O(n2i ki log ki ).
ni+1
ni+1
Unraveling i levels of the recursion on the original problem yields the lemma.
Theorem 14 Given integer R > 2, let γ = R1 , and for i > 0, let ni = ( nσ20 )1/2−iγ
and ki = 4i k0 ( nσ20 )iγ , where σ is the volatility of the stock. RecBTT underestimates
E((Tn −(n+1)X)+ ) by at most O(
1/2+γ
Bn0
k0
σ1−2γ
) and takes time O(21/γ n20 k0 ( γ1 +log kσ0 n2 0 )).
1
Proof: After K = 2γ
levels of recursion, we have nK = ( σn20 )0 = 1. The problem
consists of a single node and can be solved in constant time T̃K with an error of
ẼK = kBK = K σB 1/2 . ¿From Lemma 12, the error is
4 k0 n0
K
K
X
X
1
2K n0 σB
2K n0
2i−1
+
+
ẼK = 5Bn0
E0 = 5Bn0
n
k n
nK
2i−1 k0 ( σ20 )1/2−γ 4K k0 n1/2
0
i=1
i=1 i−1 i
!
1/2+γ
σ 1−2γ Bn0
.
=O
k0
17
Fast Pricing of European Asian Options
For the runtime analysis, by Lemma 11, note that c0 =
ni <
√
n
,
σ
√n
n/σ
√
= σ n. For i > 0, since
ci = 1. Combining this fact with Lemma 13 yields the runtime
!
K−1
X
T0 = n20 k0 log k0 + O
n0 ni c0 ki log ki + O(n0 c0 )
= n20 k0 log k0 + O
=
O(21/γ n20 k0 ( γ1
+
i=1
K−1
X
n20 4i k0 ( γ1 + log kσ0 n2 0 )
i=1
log kσ0 n2 0 )).
!
Corollary 15 Given integer R > 2, let γ = R1 and choose ni and ki as in Theorem 14.
In time O(21/γ n20 k0 ( γ1 + log kσ0 n2 0 )), RecBTT returns an option price in the range
[P −
1/2+γ 1−2γ
σ
Xn0
k0
, P ], where P is the exact price of the option.
Proof: This follows from Theorem 14 and the fact that
1
E((An − (n + 1)X)+ ) =
E((Tn − (n + 1)X)+ ).
n+1
Remark 2 By comparison, the AMO algorithm [1] has error O(n0 X/k0 ) and runtime
O(k0 n20 ).
4.4. Customization of the RecBTT Scheme
A useful feature of the RecBTT scheme is that the number of recursive calls, the
size of each recursive call, and the pricing method by which we solve the base case
of the recursion (the final recursive call) can be custom tailored for the application.
The base case of the recursion can be solved with any option pricing scheme, including
BoundedMC, other variants of the Monte Carlo algorithm, the AMO algorithm, or
exhaustive path traversal. For example, consider a 3-month (90-day) Asian option with
the average computed on a daily basis (n = 90). Pricing this option exactly requires
√
traversing 290 paths, which is infeasible. Using our RecBTT scheme, since n < 10, a
single recursive call can reduce this problem to several subproblems each with at most
210 paths.
In practice, larger values of n appear in several applications. Long-term options
contracts, LEAPS [13], are negotiated for exercise dates several years in advance.
Companies also offer stock options to employees over periods of up to five years or
more. Finally, in Istanbul options contracts [14], rather than computing the average
based on daily prices, the average is computed over prices taken at higher frequencies.
In each of these cases, n is sufficiently large that several recursive calls are required to
reduce the problem to a manageable size.
Fast Pricing of European Asian Options
18
5. Pricing Basket Options
The bucketed tree structure created by BTT and RecBTT can be used to price various
kinds of European basket options as well. Here, we describe here how to price European
Asian basket options. A basket option [13] is composed of m stocks, z1 , . . . , zm . For each
stock zi , we construct a binomial tree (according to the respective stock’s volatility), as
described in Section 2. For 1 ≤ t ≤ n and 1 ≤ i ≤ m, let Sti be the random variable
Pm i
denoting the price of stock zi on day t and define St =
i=1 St to be the random
variable denoting the total of the stock prices on day t. Recall that the payoff of a
Pn
1
European basket call with strike price X is E((Sn − X)+ ). Letting An = n+1
t=0 St
be the average total stock price, the payoff of a European Asian basket call with strike
price X is E((An − X)+ ). There is additional complexity with pricing basket options
that does not appear in their single-stock counterparts: the number of paths that the
total basket price can follow is exponential, not only in the number of trading periods n
but also in that of stocks m. Basket options are usually priced using traditional Monte
Carlo methods. The scheme we describe here is the first polynomial time, in both the
number of stocks and trading periods, pricing scheme for any kind of basket option with
provably small error bounds.
Our European Asian basket call pricing algorithm, BasketBTT, is described in
Algoirthm 5. Let B = (n+1)X, where X is the strike price of the basket option. For each
stock zi , 1 ≤ i ≤ m, use RecBTT to construct the bucketed binomial tree structure T i
described in Section 4, this time using B as the barrier; should the running total of any zi
exceed B, the basket option will always be exercised, regardless of what the other stocks
do. For each stock zi , we construct k + 1 superbuckets βji , 0 ≤ j ≤ k, where βji is the
combination of buckets bj (v) for all leaves v ∈ T i . For the core buckets βji , 0 ≤ j < k, let
P
βji .value = j Bk and βji .mass = nℓ=0 bj (T i [n, ℓ]).mass, where this summation ranges over
P
all leaves T i [n, ℓ] of T i . For the overflow bucket βki , let βki .mass = nℓ=0 bj (T i [n, ℓ]).mass
and
Pn
bk (T i [n, ℓ]).value × bj (T i [n, ℓ]).mass
i
βk .value = ℓ=0
.
βki .mass
Handling overflow superbuckets If the running total of a stock zi reaches the overflow
superbucket βki , the option will be exercised regardless of what the other stocks do.
Given this, we can determine the value of the option exactly, since
E((Tn − (n + 1)X)+ ) = E(Tn − (n + 1)X) = E(Tn ) − (n + 1)X
X
′
= βki .value +
E(Tni ) − (n + 1)X,
i′ 6=i
′
where Tni is the random variable denoting the running total of stock zi′ up to day n.
′
E(Tni ) can be computed exactly using Lemma 6.
Handling core superbuckets Consider now the core superbuckets βji , 0 ≤ j < k. Let
Pk−1 i
fi (x) = j=0
βj .mass · xj be the polynomial representation of the core bucket masses
Fast Pricing of European Asian Options
19
Q
of stock zi and let f (x) = m
i=1 fi (x). This product can be computed efficiently, as
described in Lemma 16. Notice that f (x) has the form
f (x) = b0 x0 + b1 x1 + · · · + bm(k−1) xm(k−1) .
¿From the definition of f (x), observe that bj is just the probability that the sum (over
all stocks zi ) of running totals Tni from the core buckets falls in the range [j Bk , (j + 1) Bk ).
That is,
!
m
X
bj = Pr
Tni ∈ [j Bk , (j + 1) Bk ) | Tni < B for all i .
i=1
Hence, the contribution to the option price from the core buckets can be estimated by
m(k−1)
X
j=k
bj (j Bk − (n + 1)X).
Pricing the option Combining the above results for the overflow and the core
superbuckets, we see that E((Tn − (n + 1)X)+ ) can be estimated by
m
X
i=1
m(k−1)
βki .mass(βki .value
+
i′
i′ 6=i E(Tn )
P
− (n + 1)X) +
X
j=k
bj (j Bk − (n + 1)X).
Algorithm 5 BasketBTT(z1 , . . . , zm , B = (n + 1)X)
for i = 1, . . . , n % for each stock zi
compute E(Tni ) using Lemma 6;
run RecBTT on stock zi with barrier B and k buckets;
for j = 0, . . . , k − 1 % construct core superbuckets
βji .value ← j Bk ;
P
βji .mass ← nℓ=0 bj (T i [n, ℓ]).mass;
P
βki .mass ← nℓ=0 bj (T i [n, ℓ]).mass; % construct overflow superbucket
Pn
1
i
i
βki .value ← β i .mass
ℓ=0 bk (T [n, ℓ]).value × bj (T [n, ℓ]).mass;
k
P
k−1 i
let fi (x) ← j=0
β .mass xj ;
Qm j
compute f (x) ← i=1 fi (x) as described in Lemma 16;
let f (x) ← b0 x0 + b1 x1 + · · · + bm(k−1) xm(k−1) ; % for some b0 , . . . , bm(k−1)
P
P
Pm(k−1)
i
i
i′
return m
bj (j Bk −(n+1)X);
j=k
i=1 βk .mass(βk .value+
i′ 6=i E(Tn )−(n+1)X)+
Lemma 16 Let σmin be the minimum volatility among the stocks in the basket. The
runtime of BasketBTT is
O(mT (σmin , n, k) + mk log m log k + mk log2 m),
where T (σmin , n, k) is the runtime of RecBTT on a binomial tree of size n when k
buckets are used and the volatility of the underlying stock is σmin .
Fast Pricing of European Asian Options
20
Proof: Computing the leaf bucket weights for each stock zi (using RecBTT) takes
O(m · T (σmin , n, k)) time. The runtime of the rest of the computation is dominated by
Q
the time to compute m
i=1 fi (x). We now describe an efficient way to do this. Assume
r
that m = 2 , for some r. The general case is handled similarly. We conduct the
multiplication by successively multiplying consecutive polynomials together log m = r
times until we are left with a single polynomial. For 1 ≤ i ≤ m, let fi0 (x) = fi (x).
j−1
(x)f2ij−1 (x). The answer that we
For 1 ≤ j ≤ r and 1 ≤ i ≤ 2mi , let fij (x) = f2i−1
Qm
are looking for is just i=1 fi (x) = f1r (x). At stage j, we multiply together 2mj pairs
of polynomials, each of degree 2j−1k. Using FFT, this takes O( 2mj 2j−1k log(2j−1k)) =
O(mk log(2j−1k)) time. The total runtime to compute the product, over all stages, is
P
O(mk(r log k + rj=1 log 2j−1 )) = O(mk(log m log k + log2 m)), from which the claimed
result follows.
Our definition of the error made by BasketBTT is symmetric to the definition of
error made by RecBTT; i.e., it is the maximum amount by which BasketBTT can
P Pn
i
i
i
i
underestimate m
t=0 St (ω ), for paths ω ∈ T .
i=1
Lemma 17 Let σmax be the maximum volatility among the stocks in the basket. The
error made by BasketBTT is at most mE(σmax , n, k), where E(σmax , n, k) is the error
made by RecBTT on a single stock with volatility σmax .
Proof: For 1 ≤ i ≤ m, let ωi be any path down the binomial tree corresponding
to stock zi and let Tni (ωi ) be the total price down ωi . When we run RecBTT on
zi , each Tni (ωi ) is underestimated by at most E(σmax , n, k). Hence for any ω1 , . . . , ωm ,
Pm i
i=1 Tn (ωi ) is underestimated by at most mE(σmax , n, k), as claimed.
Theorem 18 Given n, m, k, R > 2, γ = R1 , σmin and σmax , if we apply RecBTT
as described in Theorem 14 to construct the bucketed binomial tree for each stock,
1/2+γ 1−2γ
BasketBTT has an error of O(m Bn k σmax ) and runs in time
O(21/γ n2 mk( γ1 + log σkn
) + mk log m log k + mk log2 m).
2
min
Proof: This follows directly from Lemmas 16 and 17 and Theorem 14.
Corollary 19 Given n, m, k, R > 2, γ = R1 , σmin and σmax , BasketBTT
1/2+γ 1−2γ
underestimates the price of a European Asian basket call by at most O(m Xn k σmax )
and runs in time
) + mk log m log k + mk log2 m).
O(21/γ n2 mk( γ1 + log σkn
2
min
Proof: This follows from Theorem 18, which gives the amount by which the total
prices of the stocks may be underestimated. The amount by which the sum of their
1
.
daily averages is underestimated is smaller by a factor of n+1
Fast Pricing of European Asian Options
21
6. Further Research
This paper focuses on theoretical aspects of BoundedMC and RecBTT. In order
to optimize the performance of RecBTT in practice, we must consider three issues
in its implementation: (1) The number of recursive calls and the size of the recursive
subproblems will depend on the number of periods (see Section 4.4). (2) For options
with a high number of periods, nontrivial data structures may be required. (3) We have
a choice of several algorithms to use at the base-case of our recursion (see Section 4.4).
The implementation of our algorithms and empirical comparisons between them and
others will be reported in a subsequent paper.
This paper opens up several directions for further theoretical research. An
immediate open problem is whether similar Monte Carlo and bucketing techniques can
be used for pricing American Asian single-stock and basket options.
With regards to our algorithms, it would be an important result if the runtime of
BoundedMC on basket options can be bounded by a polynomial in the number of
stocks m; or if the recursive calls in RecBTT can be structured differently to yield
improved runtime or error bounds.
Finally, it would be interesting to see whether our Monte Carlo analysis or recursive
bucketing techniques can be used to price other types of options or applied to similar
problems.
Acknowledgments
We would like to thank Kyusik Chung, David Goldenberg, and Samuel Ieong for helpful
discussions; and Stan Eisenstat for pointers to numerical analysis software.
References
[1] D. Aingworth, R. Motwani, and J. D. Oldham. Accurate approximations for Asian options. In
Proceedings of the 11th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 891–900,
2000.
[2] K. Azuma. Weighted sums of certain random variables. Tokuku Mathematics Journal, 19:357–367,
1967.
[3] M. Baxter and A. Rennie. Financial Calculus: an Introduction to Derivative Pricing. Cambridge
University Press, Cambridge, United Kingdom, 1996.
[4] P. P. Boyle. Options: A Monte Carlo approach. Journal of Financial Economics, 4:323–338, 1977.
[5] P. Chalasani, S. Jha, and I. Saias. Approximate option pricing. Algorithmica, 25(1):2–21, 1999.
[6] P. Chalasani, S. Jha, and A. Varikooty. Accurate approximations for European-Asian options.
Journal of Computational Finance, 1(4):11–29, 1999.
[7] L. Clewlow and A. Carverhill. On the simulation of contingent claim. The Journal of Derivatives,
pages 66–74, Winter 1994.
[8] T. H. Cormen, C. E. Leiserson, and R. L. Rivest. Introduction to Algorithms. The MIT Press,
Cambridge, MA, 1999.
[9] C. Costantini. Variance reduction by antithetic random numbers of Monte Carlo methods for
unrestricted and reflecting diffusions. Mathematics and Computers in Simulation, 51(1-2):1–17,
1999.
Fast Pricing of European Asian Options
22
[10] J. C. Cox, S. A. Ross, and M. Rubenstein. Option pricing: A simplified approach. Journal of
Financial Economics, 7(3):229–263, 1979.
[11] A. Frieze. On the length of the longest monotone subsequence in a random permutation. The
Annals of Applied Probability, 1(2):301–305, May 1991.
[12] J. M. Hammersley and K. W. Morton. A new Monte Carlo technique: antithetic variates.
Proceedings of the Cambridge Philosophical Society, 52:449–475, 1956.
[13] J. C. Hull. Options, Futures, and Other Derivatives. Prentice Hall, Upper Saddle River, NJ, 4
edition, 2000.
[14] M. Jacques. The Istanbul option: Where the standard European option becomes Asian. Insurance:
Mathematics and Economics, 21(2):139–152, November 1997.
[15] R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press, New York,
NY, 1995.
[16] S. N. Neftci. An Introduction to the Mathematics of Financial Derivatives. Academic Press, New
York, NY, 1996.
[17] C. P. Robert and G. Casella. Monte Carlo statistical methods. Springer-Verlag, New York, 1999.
[18] F. J. Vazquez-Abad and D. Dufresne. Accelerated simulation for pricing asian options. In
Proceedings of the 1998 Winter Simulation Conference, pages 1493–1500, 1998.
[19] P. Wilmott, S. Howison, and J. Dewynne. The Mathematics of Financial Derivatives. Cambridge
University Press, Cambridge, United Kingdom, 1995.
| 5cs.CE
|
Applying deep learning techniques on medical
corpora from the World Wide Web: a prototypical
system and evaluation
Jose Antonio Miñarro-Giménez
1
1,3
, Oscar Marín-Alonso 1,2, Matthias Samwald 1*
Section for Medical Expert and Knowledge-Based Systems, Center for Medical Statistics, Informatics, and
Intelligent Systems, Medical University of Vienna, Vienna, Austria.
2
Dept. of Computer Technology, University of Alicante, Alicante, Spain.
3
Institute of Medical Informatics, Statistics, and Documentation, Medical University of Graz, Graz, Austria.
*
Corresponding author
E-mail: matthias.samwald [at] meduniwien.ac.at (MS)
1
Abstract
Background: The amount of biomedical literature is rapidly growing and it is becoming increasingly difficult
to keep manually curated knowledge bases and ontologies up-to-date. In this study we applied the word2vec
deep learning toolkit to medical corpora to test its potential for identifying relationships from unstructured text.
We evaluated the efficiency of word2vec in identifying properties of pharmaceuticals based on mid-sized,
unstructured medical text corpora available on the web. Properties included relationships to diseases (‘may
treat’) or physiological processes (‘has physiological effect’). We compared the relationships identified by
word2vec with manually curated information from the National Drug File – Reference Terminology (NDF-RT)
ontology as a gold standard. We used different word2vec parameter settings and models to compare their
impact on result quality.
Results: Our results revealed a maximum accuracy of 49.28% which suggests a limited ability of word2vec to
capture linguistic regularities on the collected medical corpora compared with other published results. We were
able to document the influence of different parameter settings on result accuracy and found and unexpected
trade-off between ranking quality and accuracy. Pre-processing corpora to reduce syntactic variability proved
to be a good strategy for increasing the utility of the trained vector models.
Conclusions: Word2vec is a very efficient implementation for computing vector representations and for its
ability to identify relationships in textual data without any prior domain knowledge. We found that the ranking
and retrieved results generated by word2vec were not of sufficient quality for automatic population of
knowledge bases and ontologies, but could serve as a starting point for further manual curation. Future
research should focus on how to combine word2vec tools with knowledge-based resources such as biomedical
ontologies to create hybrid systems with greater accuracy and flexibility than either approach on its own.
Keywords:deep learning; machine learning; text extraction; medicine; information retrieval; ontologies;
software evaluation
Introduction
The large amount of biomedical information in databases such as PubMed [1] is a valuable source for
automated information extraction [2] that facilitates the development of more efficient biomedical information
retrieval systems. The concept of ‘deep learning’ has recently gained a lot of attention. It refers to unsupervised
learning algorithms which automatically discover data without the need of supplying specific domain
knowledge [3]. This approach usually has higher performance rates than supervised and informed methods
when processing large unstructured corpora. However, the utility of these algorithms applied to realistic,
domain-specific use-cases still needs further evaluation.
2
Word2vec [4] implements an efficient deep learning algorithm for computing high-dimensional vector
representations of words and their relationships [5] based on unstructured text data. Once a vector model is
created from a text corpus, word2vec provides two basic tools to use these models: distance and analogy. The
distance tool provides a list of words closely related to a particular word from the vector model. These results
also contain the corresponding cosine similarity of each related word that indicates how close the words are in
the vector space model. The analogy tool, on the other hand, is able to query for textual regularities captured in
the vector model through simple vector subtraction and addition.
For example, let us assume that we use word2vec to create a vector model of the words appearing in a large
corpus of news articles. If the resulting vector space representation of cities and countries is projected to a twodimensional representation, we can observe patterns such as those sketched in Figure 1. Not only are ‘similar’
(e.g., bordering) countries close to each other in vector space, we also see that their capital cities are arranged
at predictable distances from their countries, because the deep learning algorithm was able to capture a notion
of the ‘capital city’ relation between two entities from unstructured text sources. Because of this, the vector
operation Paris – France + Berlin would result in a position in the vector space model that is close to the
position of the word Germany, i.e., the regularities in the vector representation can be used to search for words
related through a certain relationship to a query word. Because queries are simply defined ‘by example’,
word2vec allows querying for poorly formalized relationships, which is of special interest in complex and
evolving knowledge domains.
Germany
x
Berlin
x
France
x
Paris
x
Spain
x
Portugal
x
Madrid
x
Lisbon
x
Figure 1. Simplified, two-dimensional example of the regularities in the vector space representation
generated by word2vec.
3
These characteristics make word2vec of potential interest for improving the accessibility of unstructured
medical content. For example, word2vec could assist the curation of structure knowledge bases and ontologies,
and could help in refining information retrieval algorithms. However, deep learning algorithms such as
word2vec are known to require very large amounts of training data to provide good results, and the amount of
accessible, high-quality literature in specific domains (such as medicine) is often restricted, potentially
decreasing the practical utility of the approach.
While similar approaches such as GloVe [6] were recently claimed to outperform word2vec tools for the
unsupervised learning of word representations and word analogy, we utilized word2vec because it has been
widely studied [6-8] and, therefore, our results can be easily compared with others.
In this paper we report on our evaluation of word2vec for clinically relevant medical content based on diverse,
openly available, mid-sized medical text corpora. We compare the word relationships learned by word2vec
with curated medical relationship encoded in the National Drug File – Reference Terminology (NDF-RT)
ontology [9] to evaluate the results. The results of this exploratory work are intended to serve as an initial
guidance that informs more in-depth work on applying word2vec in the medical domain.
Material and Methods
Word2vec
Word2vec is an efficient implementation of deep learning techniques based on two architectures, continuous
bag-of-words (CBOW) and skip-gram (SG) [5], for computing continuous distributed vector representation of
words from large datasets (up to hundreds of billions of words).
Word2vec requires training the corpora using one of these architectures. The training tool provides the
following options: (i) type of architecture: continuous bag-of-words or skip-gram; (ii) the dimensionality of the
vector space; (iii) the size of the context window in number of words; (iv) the training algorithm: hierarchical
softmax and / or negative sampling; (v) the threshold for downsampling the frequent words; (vi) the number of
threads to use; and (vii) the format of the output word vector file (an example command line is shown in Table
1).
4
Table 1. An example of the parameters used by the word2vec training tool to generate the vector model
of the medical corpora.
Command line parameters
word2vec -train corpora.txt -output vector-model.bin -cbow 0 -size 200 -window 5
-negative 0 -hs 1 -sample 1e-3 -threads 12 -binary 1
In this example, the training tool uses the corpus file “corpora.txt” to generate the vector model and serialize it to the file “vectormodel.bin”. It uses the skip-gram architecture (-cbow 0) with a vector dimension of 200 (-size 200) and a window size of 5 (window 5). Besides, the vector model was generated using only the hierarchical softmax training algorithm (-hs 1) with a
threshold of 0,001 for downsampling the frequent words (-sample 1e-3). The training tool uses 12 execution threads to generate
the vector model as a binary file.
Medical text corpora
We assembled a collection of openly available text repositories relevant to clinical medicine (excluding
veterinary medicine) for use in this evaluation.
Two corpora were derived from PubMed. The first corpus was made up of PubMed abstracts with clinical
relevance. To select abstracts of clinical relevance, a PubMed query was assembled by merging the lists of
journals screened by the evidence-based medicine repositories DynaMed [10] and EvidenceUpdates [11]. The
list was further manually edited, and additional constraints were added (e.g., excluding articles published
before January 2005, excluding editorials) to create the final selection of PubMed abstracts. From this corpus
of PubMed abstracts we derived another corpus by extracting the conclusion sections of the abstracts. The
conclusion sections were further processed by expanding locally defined abbreviations in each abstract. This
resulted in a smaller corpus made up of very high-quality, self-contained key assertions made in the clinically
relevant research literature.
We created a corpus of medically relevant content from Wikipedia by selecting all articles that were associated
with Wikiproject Medicine or Wikiproject Pharmacology [12] through manual curation of Wikipedia editors.
This produced a corpus of Wikipedia articles with a good coverage of all major clinically relevant topics.
We also included two popular publicly available websites with content for medical professionals: Medscape
[13] and Merck Manual [14]. We created a script for crawling medical content from these websites based on
the PHPcrawl open-source library [15]. We also created scripts for stripping non-relevant portions of web
pages (such as headers and footers) that would have significantly degraded the quality of the corpora. HTML
markup was removed from the source data to yield raw text representations of the page contents.
Statistics on word counts and vocabulary sizes of the corpora generated in this way are summarized in Table 2.
Table 2. Corpora used in the experiment.
5
Corpus
Word count
Vocabulary size
Clinically relevant subset of PubMed, full abstracts
161,428,286
204,096
Clinically relevant subset of PubMed, extracted key assertions
(‘pubmed_key_assertions’)
17,342,158
47,703
Merck Manual
12,667,064
49,174
Medscape
25,854,998
63,600
Clinically relevant subset of Wikipedia (‘Wikipedia’)
10,945,677
65,875
Combined corpus (including all corpora above, ‘combined’)
236,835,672
261,353
Word counts refer to the final corpora that were derived from source datasets after all processing steps. Vocabulary sizes refer to
the number of distinct words found in each corpus. Underscored corpora were used for evaluating word2vec.
National Drug File – Reference Terminology (NDF-RT) ontology
The NDF-RT is a formal representation of knowledge about drugs and is maintained by the US Department of
Veterans Affairs. We chose the NDF-RT ontology as a reference for evaluating the results produced by the
word2vec algorithms because it is one of the richest manually curated and openly available knowledge bases
on medical drugs available. Several relationships between entities - such as ‘may treat’ or ‘has mechanism of
action’ - were extracted from NDR-RT through SPARQL queries. The relationships that were extracted in this
manner are described in Table 3.
Table 3. Relationships between entities extracted from NDF-RT.
NDF-RT relationship
Number of distinct drugs annotated
with at least one such relationship
Number of relationships
2,016
11,999
781
1,909
‘has_MoA’
(drug X has mechanism of
action Y)
2,436
9,126
‘has_PE’
(drug X has physiological
effect Y)
2,250
13,769
‘may_treat’
(drug X may treat disease /
symptom Y)
‘may_prevent’
(drug X may prevent disease
/ condition Y
6
In the text below we refer to relationships captured by NDF-RT as ‘correct’ relationships. Of course, this is a
simplification of reality, it might well be possible that not all relationships found in NDF-RT are factually
correct, and it is also likely that there are factually correct relationships missing from the NDF-RT ontology.
Evaluation
In order to conduct the pilot evaluation of word2vec tools for medical text, we defined the following workflow
for analyzing and evaluating the results produced by the word2vec tools: (1) gathering and processing openly
available medical text corpora; (2) training vector space representations of these corpora with word2vec; (3)
comparing results from the word2vec distance and analogy tools to assess how well the models captured the
medical relationships such as ‘may treat’ and ‘may prevent’ from the medical reference ontology ; (4)
calculating statistics to evaluate the impact of different parameter configurations.
Gathering and processing medical corpora
The texts gathered from the selected sources needed to be pre-processed before they could be used for training
vector models, since word2vec has no built-in functionalities for term normalisation or dealing with
punctuation. We found that unprocessed corpora contained an abundance of basic syntactic variations and
punctuation that had a negative impact on how word2vec indexes the terms and, therefore, the quality of the
resulting vector space models.
The processing of corpora involved:
1. Removing all punctuation signs and superfluous whitespace, e.g., to avoid that “diabetes.” (with a full
stop) and “diabetes” would be erroneously indexed as two separate terms.
2. Transforming all words to lower-case, e.g., to avoid that “Diabetes”, “diabetes” and “DIABETES”
would be indexed as three separate terms.
3. Merging multiword terms, e.g. The term “glucose metabolism disorder” was transformed to
“glucose_metabolism_disorder”, which could then be processed like a single word by the word2vec
algorithm. As word2vec processes every word in the corpus as a term, multiword terms cannot be
represented in vector space. Therefore, we created a dictionary of relevant medical multiword terms
from the NDF-RT ontology. When one term was found in the corpus, the words of the term were
concatenated using underscores
The processing workflow is depicted in Figure 2.
7
Figure 2. Workflow for gathering and pre-processing the content of the corpora.
Training word2vec vector space models
The pre-processed corpora described above were used to build vector models with the word2vec training tool.
The results of word2vec do not depend exclusively on the corpora but also on the parameters used. In order to
test the impact of these parameter settings we compared the results of vector models trained with different
parameter settings. We trained the corpora using (a) CBOW and SG vector model architectures; (b) 200, 300,
500 and 800 dimensionality of vector space; and (c) a word windows size of 5, 10 and 20. Other parameters,
such as the training algorithm or the threshold for downsampling frequent words, were not varied to keep the
complexity of results manageable.
Assessment system
To assess the vector models of the trained corpora, the results of the word2vec distance and analogy tools were
compared to the curated content of the NDF-RT ontology as a gold standard. To perform this evaluation, we
developed a system that automatically queried a trained vector model using the distance and analogy tools of
word2vec and matched the resulting list of words with the content of the NDF-RT ontology.
Figure 3 shows the software architecture of the assessment system. The “RESTful server” module made the
word2vec analogy and distance tools accessible through RESTful services. RESTful services were deployed
through the Java-based Jersey framework [16] and facilitated the access of the vector space models from
external applications. Both services returned forty words with their cosine similarity values by default.
8
Figure 3. The software architecture of the assessment system of word2vec tools using NDF-RT ontology.
The “RESTful client” module (Figure 3) was responsible for gathering required information from the NDF-RT
ontology, preparing the queries to consult the analogy and distance services, processing the responses and
matching the retrieved terms with the information from NDF-RT to obtain the evaluation results. A batch
process was also defined to automatically execute these tasks and to obtain the corresponding evaluation result
for each trained corpus. The “Query module” collected subject-predicate-object triples for particular predicates
from the NDF-RT ontology. Then, it created a list of unique subjects from all collected triples and used it to
call the analogy and distance services. The “Matching module” received the results from the services and
checked which words from the retrieved vector matched their corresponding words from the object values of
the collected triples. Finally, the results of the word2vec tools evaluation were processed to calculate their
accuracy values.
Results and Discussion
The main results of this paper are: (1) the definition of a methodology to evaluate deep learning techniques in
medical corpora; (2) the development of the batch system to automatically run the word2vec tools and match
the results against the content of the NDF-RT ontology as a gold standard; and (3) the assessment of the results
of the batch system for the evaluation of word2vec on medical corpora. The source code of the software for
running these experiments is freely available for download at [17]; the software can be easily adapted to other
use-cases, corpora and ontologies.
Removing sources of syntactic variability in the corpora proved to be a good strategy for increasing the utility
of the trained vector models. As an example, Table 4 shows an excerpt of the result list when searching for
words related to “aspirin” in the combined corpus with the distance tool. The results from the raw corpus
9
include terms such as “[coumadin]” or “aspirin’s” which contain irrelevant symbols. Furthermore, the lack of
multiword terms in the raw corpus reduced the quality of retrieved terms, e.g., the term “acetylsalicylic_acid”
(the active ingredient of Aspirin) is the closest to “aspirin” in the pre-processed corpus with a cosine similarity
value of 0.83, whereas in the raw corpus the term “acetylsalicylic” is only found at the 5th position with a
cosine similarity of 0.68.
Table 4. Results of the Distance tool for the term “aspirin” when querying the pre-processed and raw
versions of the combined corpus).
Pre-processed combined corpus
Raw combined corpus
Rank
Word
Cosine similarity
Word
Cosine similarity
1
acetylsalicylic_acid
0.83
clopidogrel
0.79
2
clopidogrel
0.76
ticlopidine
0.72
3
ticlopidine
0.74
thienopyridine
0.69
4
warfarin
0.70
warfarin
0.69
5
clopidogrel_bisulfate
0.67
acetylsalicylic
0.68
27
cilostazol
0.60
[coumadin]
0.60
30
antithrombotics
0.60
aspirin’s
0.59
The vector model based on the pre-processed corpus yields better results than the vector model based on the raw corpus, e.g., the
term ‘acetylsalicylic_acid’ is the closest term to ‘aspirin’.
The analogy tool aims to find textual regularities captured in the vector model to search for words related to a
query word through a specific kind of relationship, while the distance tool simply returns words that are
similar or generically related to a query word. To test whether the vector model actually captured textual
regularities as expected, we tested if the analogy tool was superior in retrieving the right words from a specific
relation. We calculated the accuracy of both tools for the three selected corpora: combined,
pubmed_key_assertions and Wikipedia (see Table 2); and the relationships: may_treat, may_prevent, has_PE
and has_MoA (see Table 3) from the reference ontology NDF-RT as a gold standard. Table 5 shows that the
analogy tool indeed provided better results than the unselective distance tool, which underlines the ability of
the analogy tool to recognize textual regularities beyond generic notions of relatedness. Models trained on the
combined corpus, the largest corpus among the test corpora, produced best results, which confirms the
hypothesis that word2vec model quality increases as corpus size increases. The number of correct results for
the may_treat relationship with the combined corpus and the analogy tool was the best among all test cases
with an accuracy 38.78%. Consequently, we choose the best configuration to run our evaluations regarding the
influence of window-size and vector dimensionality parameters on the accuracy of the resulting vector models.
The optimisation of window-size and vector dimensionality parameters is described by the developers of
word2vec as the most important ones for achieving good results. The window-size parameter corresponds to
the span of words in the text that is taken into account, backwards and forwards, when iterating through the
words during model training.
10
Table 5. Accuracy values of the word2vec tools.
Corpus
combined
pubmed_key_ass
ertions
wikipedia
Tool
may_treat
may_prevent
has_PE
has_MoA
Analogy
38.78%
13.82%
2.54%
8.14%
Distance
4.36%
8.68%
0.67%
8.89%
Analogy
22.66%
5.6%
0.93%
1.59%
Distance
2.66%
5.85%
0.37%
3.94%
Analogy
20.95%
6.46%
2.22%
2.91%
Distance
1.39%
3.56%
0.32%
3.56%
Accuracy values of the word2vec tools using combined, pubmed_key_assertions and Wikipedia corpora, with analogy and
distance tools, and the may_treat, may_prevent, has_PE and has_MoA relationships. The best accuracy was obtained with the
combined corpus and analogy tool when querying may_treat information.
Figure 4 shows the accuracy when querying the resulting vector models generated with SG and CBOW
architectures and with three different window sizes, 5, 10 and 20. Larger window sizes increase the time
required to train vector models, so choosing the optimal window size can reduce computing time. Our findings
show that the window-size parameter is also relevant for improving the accuracy of the gathered medical
corpora. With our corpora, the optimal value is a window size of 10.
Figure 4. Comparison between the accuracy values using skip-gram (SG) and continuous bag-of-words
(CBOW) architectures. Comparisons were conducted for window sizes of 5, 10 and 20. Results are based
on a 200-dimensional vector model, the combined corpus and the may_treat relationship from NDF-RT
ontology as the gold-standard for querying the word2vec analogy tool.
11
The influence of the vector dimensionality parameter was evaluated using vector models generated with 200,
300, 500 and 800 vector dimensions, and the SG and CBOW architectures. Figure 5 shows the comparison
between the accuracy values with various parameter configurations. In our experiments, the best ranking
results were obtained with the SG architecture and a vector dimensionality of 300, while worst ranking results
were observed with the CBOW architecture and a vector dimensionality of 800. The results suggest that the SG
architecture consistently yielded better result ranking than the CBOW architecture. This is also consistent with
other evaluations such as [6] which obtained an accuracy of 61% with SG, a dimension vector of 300 and 1B
word corpus. We observed a U-shaped relationship between the dimensionality of the model and the accuracy
of results with SG and CBOW architecture, so a higher vector dimensionality will not provide better results
than the stated ones. A higher dimension of the vectors implies a bigger size of the resulting vector model.
Therefore, having a vector model with a dimension of 300 requires less memory space and provides better
results than a vector model with 500 or 800 vector dimensions.
Figure 5. Accuracy of the vector models generated using the combined corpus with skip-gram (SG) and
Continuous bag-of-words (CBOW) architectures. Comparisons were conducted for vector
dimensionalities of 200, 300, 500 and 800. The may_treat information from NDF-RT ontology is used as
the gold standard for querying word2vec analogy tool.
12
Conclusions
The ability of word2vec to retrieve expected terms from the size-restricted corpora we used is not suitable for
applications requiring high precision since we only obtained an accuracy of 49.28%. These modest results
could be explained by the restricted size of gathered medical corpora as well as the complexity of the medical
knowledge domain. It could be of interest for future research to test this tool with larger, commercially
available medical corpora.
Due to the complexity of medical terminology, we found pre-processing of corpora necessary to decrease
syntactic variability. We also found that many relevant medical terms of interest were composed of multiple
words, and that ontology-based pre-processing of these terms led to a marked improvement of the results from
the word2vec tools.
As expected, the analogy tool produced better results for identifying related entities for a specific type of
relationship than the distance tool, and the largest corpus provided better results than smaller sub-corpora. As a
consequence of our results, we recommend to use SG architecture rather than CBOW to train other medical
datasets because such architecture always produced the best accuracy values. Moreover, the combination of a
10 window-size with a 300 vector dimension produced the best results among all tested parameters
configurations. We also conclude that a vector dimension greater than 800 and a window size greater than 20
are not recommended due to the observed quick deterioration of accuracy values.
Regarding the low matching rate observed in our evaluation with mid-sized medical corpora, our future
research will focus on how to improve accuracy values by combining word2vec analogy tool with knowledgebased resources such as ontologies to create hybrid systems and also running the evaluations with larger,
commercially available corpora.
Acknowledgments
We thank the word2vec team for providing assistance with tuning parameters of the word2vec tools.
References
1.
PubMed home. Available: http://www.ncbi.nlm.nih.gov/pubmed/. Accessed 14 February 2010.
2.
Kim J-D, Ohta T, Tsujii J (2008) Corpus annotation for mining biomedical events from literature. BMC
Bioinformatics 9: 10. doi:10.1186/1471-2105-9-10.
13
3.
Bengio Y, Courville A, Vincent P (2013) Representation Learning: A Review and New Perspectives. IEEE
Trans Pattern Anal Mach Intell 35: 1798–1828. doi:10.1109/TPAMI.2013.50.
4.
word2vec - Tool for computing continuous distributed representations of words. - Google Project Hosting.
Available: https://code.google.com/p/word2vec/. Accessed 22 August 2014.
5.
Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient Estimation of Word Representations in Vector
Space. ArXiv13013781 Cs. Available: http://arxiv.org/abs/1301.3781. Accessed 22 August 2014.
6.
Pennington J, Socher R, Manning CD (2014). Glove: Global vectors for word representation. Proceedings
of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP).
7.
Levy O, Goldberg Y (2014). Linguistic Regularities in Sparse and Explicit Word Representations.
Proceedings of the Eighteenth Conference on Computational Natural Language Learning, Baltimore,
Maryland, USA.
8.
Frome A, Corrado GS, Shlens J, Bengio S, Dean J, Mikolov T (2013). Devise: A deep visual-semantic
embedding model. In Advances in Neural Information Processing Systems. pp. 2121-2129.
9.
National Drug File - Reference Terminology - Summary | NCBO BioPortal. Available:
http://bioportal.bioontology.org/ontologies/NDFRT. Accessed 24 August 2014.
10. Evidence Based Content, Up to Date Content, Clinical Reference | DynaMed. Available:
https://dynamed.ebscohost.com/. Accessed 16 October 2013.
11. EvidenceUpdates. Available: http://plus.mcmaster.ca/evidenceupdates/. Accessed 16 October 2013.
12. Heilman JM, Kemmann E, Bonert M, Chatterjee A, Ragar B, et al. (2011) Wikipedia: a key tool for global
public health promotion. J Med Internet Res 13: e14. doi:10.2196/jmir.1589.
13. Diseases & Conditions - Medscape Reference. Available: http://emedicine.medscape.com/. Accessed 15
October 2013.
14. THE MERCK MANUALS - Trusted Medical and Scientific Information. Available:
http://www.merckmanuals.com/. Accessed 15 October 2013.
15. PHPCrawl webcrawler/webspider library for PHP - About. Available: http://phpcrawl.cuab.de/. Accessed
16 October 2013.
16. Jersey. Available: https://jersey.java.net/index.html. Accessed 22 August 2014.
17. biomedical-text-exploring-tools - Deep learning on medical text corpora on the World Wide World Google Project Hosting. Available: https://code.google.com/p/biomedical-text-exploring-tools/. Accessed
22 August 2014.
14
| 9cs.NE
|
arXiv:1504.02925v5 [math.AC] 14 Mar 2016
ON THE VANISHING OF LOCAL COHOMOLOGY OF
THE ABSOLUTE INTEGRAL CLOSURE IN POSITIVE CHARACTERISTIC
PHAM HUNG QUY
Abstract. The aim of this paper is to extend the main result of C. Huneke and G. Lyubeznik
in [Adv. Math. 210 (2007), 498–504] to the class of rings that are images of Cohen-Macaulay
local rings. Namely, let R be a local Noetherian domain of positive characteristic that is an image
of a Cohen-Macaulay local ring. We prove that all local cohomology of R (below the dimension)
maps to zero in a finite extension of the ring. As a direct consequence we obtain that the absolute
integral closure of R is a big Cohen-Macaulay algebra. Since every excellent local ring is an image
of a Cohen-Macaulay local ring, this result is a generalization of the main result of M. Hochster
and Huneke in [Ann. of Math. 135 (1992), 45–79] with a simpler proof.
1. Introduction
Let (R, m) be a commutative Noetherian local domain with fraction field K. The absolute integral closure of R, denoted R+ , is the integral closure of R in a fixed algebraic closure K of K.
A famous result of M. Hochster and C. Huneke says that if (R, m) is an excellent local Noetherian
domain of positive characteristic p > 0, then R+ is a (balanced) big Cohen-Macaulay algebra, i.e.
every system of parameters in R becomes a regular sequence in R+ (cf. [7]). Furthermore, K.E.
Smith in [15] proved that the tight closure of an ideal generated by parameters is the contraction
of its extension in R+ : I ∗ = IR+ ∩ R. This property is not true for every ideal I in an excellent
Noetherian domain since tight closure does not commute with localization (cf. [1]).
As mentioned above, Hmi (R+ ) = 0 for all i < dim R provided R is an excellent local Noetherian
domain of positive characteristic. Hence, the natural homomorphism Hmi (R) → Hmi (R+ ), induced
from the inclusion R → R+ , is the zero map for all i < dim R. In the case R is an image of a
Gorenstein (not necessarily excellent) local ring, as the main result of [8], Huneke and G. Lyubeznik
proved a stronger conclusion that one can find a finite extension ring S, R ⊆ S ⊆ R+ , such that
the natural map Hmi (R) → Hmi (S) is zero for all i < dim R. Therefore, they obtained a simpler
proof for the result of Hochster and Huneke in the cases where the assumptions overlap, e.g., for
complete Noetherian local domain. The techniques used in [8] are the Frobenius action on the local
cohomology, (modified) equation lemma (cf. [7], [15], [8]) and the local duality theorem (This is
the reason of the assumption that R is an image of a Gorenstein local ring). The motivation of the
present paper is our belief: If a result was shown by the local duality theorem, then it can be proven
under the assumption that the ring is an image of a Cohen-Macaulay local ring (for example, see
[12]). The main result of this paper extends Huneke-Lyubeznik’s result to the class of rings that
are images of Cohen-Macaulay local rings. Namely, we prove the following.
Key words and phrases. Absolute integral closure; Local cohomology; Big Cohen-Macaulay; Characteristics p.
2010 Mathematics Subject Classification: 13A35; 13D45; 13B40; 13D22; 13H10; 14B15.
This work is partially supported by a fund of Vietnam National Foundation for Science and Technology Development
(NAFOSTED) under grant number 101.04-2014.25.
1
2
PHAM HUNG QUY
Theorem 1.1. Let (R, m) be a commutative Noetherian local domain containing a filed of positive
characteristic p. Let K be the fraction field of R and K an algebraic closure of K. Assume that R
is an image of a Cohen-Macaulay local ring. Let R′ be an R-subalgebra of K (i.e. R ⊆ R′ ⊆ K)
that is a finite R-module. Then there is an R′ -subalgebra R′′ of K (i.e. R′ ⊆ R′′ ⊆ K) that is finite
as an R-module such that the natural map Hmi (R′ ) → Hmi (R′′ ) is the zero map for all i < dim R.
As a direct application of Theorem 1.1 we obtain that the absolute integral closure R+ is a big
Cohen-Macaulay algebra (cf. Corollary 3.2). It worth be noted that every excellent local ring is an
image of a Cohen-Macaulay excellent local ring by T. Kawasaki (cf. [9, Corollary 1.2]). Therefore,
our results also extend the original result of Hochster and Huneke with a simpler proof. The main
results will be proven in the last section. In the next section we recall the theory of attached primes
of Artinian (local cohomology) modules.
2. Preliminaries
Throughout this section (R, m) be a commutative Noetherian local ring. We recall the main
result of [12] which is an illustration for our belief (mentioned in the introduction).
I.G. Macdonald, in [10], introduced the theory of secondary representation for Artinian modules,
which is in some sense dual to the theory of primary decomposition for Noetherian modules. Let
A 6= 0 be an Artinian R-module. We say that A is secondary ifp
the multiplication by x on A is
surjective or nilpotent for every x ∈ R. In this case, the set p := (AnnR A) is a prime ideal of R
and we say that A is p-secondary. Note that every Artinian R-module A has a minimal secondary
representation A = A1 +. . .+An , where Ai is pi -secondary, each Ai is not redundant and pi 6= pj for
all i 6= j. The set {p1 , . . . , pn } is independent of the choice of the minimal secondary representation
of A. This set is called the set of attached primes of A and denoted by AttR A. Notice that if R is
complete we have the Matlis dual D(A) of A is Noetherian and AttR A = AssR D(A).
For each ideal I of R, we denote by V ar(I) the set of all prime ideals of R containing I. The
following is easy to understand from the theory of associated primes.
Lemma 2.1 ([10]). Let A be an Artinian R-module. The following statements are true.
(i) A 6= 0 if and only if AttR A 6= ∅.
(ii) A 6= 0 has finite length if and only if AttR A 6= {m}.
(iii) min AttR A = min V ar(AnnR A). In particular,
dim(R/AnnR A) = max{dim(R/p) | p ∈ AttR A}.
(iv) If 0 →
A′
→ A → A′′ → 0 is an exact sequence of Artinian R-modules then
AttR A′′ ⊆ AttR A ⊆ AttR A′ ∪ AttR A′′ .
b be the m-adic complete of R. Note that every Artinan R-module A has a natural structure
Let R
b
as an R-module and with this structure, each subset of A is an R-submodule if and only if it is an
b
b
R-submodule.
Therefore A is an Artinian R-module.
So, the set of attached primes AttRb A of A
b
over R is well defined.
Lemma 2.2. ([2, 8.2.4, 8.2.5]). AttR A = P ∩ R | P ∈ AttRb A .
Let M be a finitely generated R-module. It is well known that the local cohomology module
Hmi (M ) is Artinian for all i ≥ 0 (cf. [2, Theorem 7.1.3]). Suppose that R is an image of a Gorenstein
local ring. R.Y. Sharp, in [14], used the local duality theorem to prove the following relation
i−dim(R/p)
AttRp HpRp
(Mp ) = qRp | q ∈ AttR (Hmi (M )), q ⊆ p
ON THE ABSOLUTE INTEGRAL CLOSURE IN POSITIVE CHARACTERISTIC
3
for all p ∈ Supp(M ) and all i ≥ 0. Based on the study of splitting of local cohomology (cf. [4], [5]),
L.T. Nhan and the author showed that the above relation holds true on the category of finitely
generated R-modules if and only if R is an image of a Cohen-Macaulay local ring (cf. [12]). It
worth be noted that R is an image of a Cohen-Macaulay local ring if and only if R is universally
catenary and all its formal fibers are Cohen-Macaulay by Kawasaki (cf. [9, Corollary 1.2]). More
precisely, we proved the following.
Theorem 2.3. The following statements are equivalent:
(i) R is an image of a Cohen-Macaulay local ring;
i−dim(R/p)
(ii) AttRp HpRp
(Mp ) = qRp | q ∈ AttR (Hmi (M )), q ⊆ p for every finitely generated
R-module M , integer i[
≥ 0 and prime ideal p of R;
i
b R)
b for every finitely generated R-module M and inAssRb (R/p
(iii) AttRb (Hm (M )) =
i (M ))
p∈AttR (Hm
teger i ≥ 0.
The above Theorem says that the attached primes of local cohomology modules have good
behaviors with completion and localization when R is an image of a Cohen-Macaulay local ring.
This will be very useful in the next section.
3. Proof the main result
Throughout this section, let (R, m, k) be a commutative Noetherian local ring that is an image
of a Cohen-Macaulay local ring. The following plays the key role in our proof of the main result.
Proposition 3.1. Let M and N be finitely generated R-modules and ϕ : M → N a homomorphism.
For each i ≥ 0, ϕ induces the homomorphism ϕi : Hmi (M ) → Hmi (N ). Suppose for all p ∈
AttR (Hmi (M )) and p 6= m, the map ϕp : Mp → Np induces the zero map
i−tp
ϕp
i−t
i−t
: HpRpp (Mp ) → HpRpp (Np ),
where tp = dim R/p. Then Im(ϕi ) has finite length.
Proof. Suppose Im(ϕi ) has not finite length. By Lemma 2.1 there exists m 6= p ∈ AttR (Im(ϕi )). So
b
p ∈ AttR (Hmi (M )) by Lemma 2.1 (iv). Consider Im(ϕi ) as an Artinian R-module.
By Lemma 2.2,
i
there exists P ∈ AttRb (Im(ϕ )) such that P ∩ R = p. Hence we have P ∈ AttRb (Hmi (M )) by Lemma
2.1 (iv) again. Since R is an image of a Cohen-Macaulay local ring, Theorem 2.3 (iii) implies that
b R).
b Therefore dim R/P
b
b is complete,
= dim R/p by [3, Theorem 2.1.15]. We have R
P ∈ AssRb (R/p
so it is an image of a Gorenstein local ring S (of dimension n). By local duality we have
c) (∼
c, S)) ∼
b∼
D(ExtSn−i (M
= Hmi (M ) ⊗R R
= Hmbi (M
= Hmi (M )),
b is
where D = HomRb (−, ERb (k)) is the Matlis duality functor (cf. [2, Theorem 11.2.6]). Since R
c, S) ∼
complete we have ExtSn−i (M
= D(Hmi (M )).
i
i
We write the map ϕ : Hm (M ) → Hmi (N ) as the composition of two maps
Hmi (M ) → I = Im(ϕi ) → Hmi (N ),
where the first of which is surjective and the second injective. Applying the Matlis duality functor
c, S) is the composition of two maps
b , S) → Extn−i (M
D we get the map D(ϕi ) : ExtSn−i (N
S
c, S)
b , S) → D(I) → Extn−i (M
ExtSn−i (N
S
4
PHAM HUNG QUY
with the first of which is surjective and the second injective. We have D(I) is a finitely generated
b
R-module
and P ∈ AssRb D(I) (= AttRb I). Let P ′ be the pre-image of P in S. Localization at P ′
the above composition we get the composition
cP , SP ′ )
bP , SP ′ ) → (D(I))P → Extn−i (M
ExtSn−i′ (N
S ′
P
P
with the first of which is surjective and the second injective. Since (D(I))P 6= 0, we have the map
cP , SP ′ )
bP , SP ′ ) → Extn−i (M
D(ϕi )P : ExtSn−i′ (N
S ′
P
P
is a non-zero map. Notice that dim SP ′ = n − tp . Applying local duality (for SP ′ ) we have the map
i−tp
ϕ
bP
:H
i−tp c
bP (MP )
PR
→H
i−tp b
bP (NP ),
PR
c→N
b , is a non-zero map. Recalling our assumption that the map
induced from the map ϕ
b:M
i−tp
ϕp
i−t
i−t
: HpRpp (Mp ) → HpRpp (Np ),
induced from ϕ : M → N , is zero.
b m)
b induces the
On the other hand, the faithfully flat homomorphism of local rings (R, m) → (R,
bP , P R
bP ) by [11, Theorem 7.1]. It is a local homomorphism
flat homomorphism (Rp , pRp ) → (R
q
so we have a faithfully flat homomorphism. It should be noted that
following commutative diagram of flat homomorphisms
R
/ Rp
/R
bP
b
R
one can check that
H
i−tp c
bP (MP )
PR
bP = P R
bP . Using the
pR
i−t
∼
bP
= HpRpp (Mp ) ⊗Rp R
i−t
i−t
i−t
bP . Therefore the maps ϕpi−tp and ϕ
bP p are either zero or non-zero,
and ϕ
bP p is just ϕp p ⊗Rp R
simultaneously. This is a contradiction. The proof is complete.
We are ready to prove the main result of this paper. In the rest of this section, (R, m) is a local
domain of positive characteristic p that is an image of a Cohen-Macaulay local ring. Let I be an
ideal of R and R′ an R-algebra. Notice that the local cohomology, HIi (−), can be computed via
the Čech co-complex of the generators of I. The Frobenius ring homomorphism
f : R′ −→R′ ; r 7→ r p
induces a natural map f∗ : HIi (R′ ) → HIi (R′ ) on all i ≥ 0. It is called the (natural) action of
Frobenius on HIi (R′ ).
Proof of Theorem 1.1. We proceed by induction on d = dim R. There is nothing to prove when
d = 0. Assume that d > 0 and the theorem is proven for all smaller dimension. For each i < d and
ei,p that is finite as
p ∈ AttR Hmi (R), p 6= m, by the inductive hypothesis there is an Rp′ -subalgebra R
Rp -module such that the natural map
i−t
i−t ei,p
HpRpp (Rp′ ) → HpRpp (R
)
ON THE ABSOLUTE INTEGRAL CLOSURE IN POSITIVE CHARACTERISTIC
5
ei,p = R′ [z1 , ..., zk ], where z1 , . . . , zk ∈ K are integral
is the zero map, where tp = dim R/p. Let R
p
over Rp . Multiplying, if necessary, some suitable element of R \ p, we can assume that each zj is
integral over R. Set Ri,p = R′ [z1 , . . . , zk ]. Clearly, Ri,p is an R′ -subalgebra of K that is finite as
R-module.
Since the sets {i | 0 ≤ i < d} and AttR (Hmi (R)) are finite, the following is a finite extension of R
R∗ = R[Ri,p | i < d, p ∈ AttR (Hmi (R)) \ {m}].
We have R∗ is an Ri,p -subalgebra of K for all i < d and all p ∈ AttR (Hmi (R)) \ {m}. The inclusions
R′ → Ri,p → R∗ induce the natural maps
i−t
i−t ei,p
i−t
HpRpp (Rp′ ) → HpRpp (R
) → HpRpp (Rp∗ ).
ei,p we have the natural map
By the construction of R
i−t
i−t
HpRpp (Rp′ ) → HpRpp (Rp∗ )
is the zero map for all i < d and all p ∈ AttR Hmi (R) \ {m}. By Proposition 3.1 we have the natural
map
ϕi : Hmi (R′ ) → Hmi (R∗ ),
induced from the inclusion ϕ : R′ → R∗ , has ℓ(Im(ϕi )) < ∞ for all i < d.
Since the natural inclusion ϕ : R′ → R∗ is compatible with the Frobenius homomorphism on R′
and R∗ , we have ϕi is compatible with the Frobenius action f∗ on Hmi (R′ ) and Hmi (R∗ ). Therefore
Imϕi is an f∗ -stable R-submodule of Hmi (R∗ ), i.e. f∗ (α) ∈ Imϕi for every α ∈ Imϕi . Since Imϕi
has finite length, every α ∈ Imϕi satisfies the ”equation lemma” of Huneke-Lyubeznik (cf. [8,
Lemma 2.2]). Hence each element of Imϕi will map to the zero in local cohomology of a certainly
finite extension of R∗ . Since Imϕi is a finitely generated R-module for all i < d, there is an
R∗ -subalgebra R′′ of K that is finite as R-module such that the composition of the natural maps
Hmi (R′ ) → Hmi (R∗ ) → Hmi (R′′ ) is zero for all i < d. The proof is complete.
Similar to [8, Corollary 2.3] we have the following.
Corollary 3.2. Let (R, m) be a commutative Noetherian local domain containing a field of positive
characteristic p and R+ the absolute integral closure of R in K. Assume that R is an image of a
Cohen-Macaulay local ring. Then the following hold:
(i) Hmi (R+ ) = 0 for all i < dim R.
(ii) Every system of parameters of R is a regular sequence on R+ , i.e. R+ is a big Cohen-Macaulay
algebra.
We close this paper with the following.
Remark 3.3. (i) In [13], A. Sannai and A.K. Singh showed that the finite extension in HunekeLyubeznik’s result can be chosen as a generically Galois extension. It is not difficult to see
that our method also works for Sannai-Singh’s paper. Hence Theorem 1.3 (2) and Corollary
3.3 of [13] hold true when the ring is an image of a Cohen-Macaulay local ring.
(ii) Since an excellent local ring is an image of a Cohen-Macaulay excellent local ring ([9, Corollary
1.2]), Corollary 3.3 is a generalization of the original result of Hochster and Huneke in [7]
with a simpler proof. Thus our results cover all previous results. On the other hand, R.C.
Heitmann constructed examples of universally catenarian local domains with the absolute
integral closures are not Cohen-Macaulay (cf. [6, Corollary 1.8]). Comparing with the result
6
PHAM HUNG QUY
of Kawasaki ([9, Corollary 1.2]), the condition that R is an image of a Cohen-Macaulay local
ring is seem to be the most general case for R+ is big Cohen-Macaulay.
Acknowledgement . The author is grateful to Professor Kei-ichi Watanabe for his helpful recommendations. He is also grateful to Kazuma Shimomoto for his interest and many discussions.
References
[1] H. Brenner and P. Monsky, Tight closure does not commute with localization, Ann. of Math. 171 (2010),
571–588.
[2] M. Brodmann and R.Y. Sharp, Local cohomology: an algebraic introduction with geometric applications, Cambridge University Press, 1998.
[3] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge University Press, 1998.
[4] N.T. Cuong and P.H. Quy, A splitting theorem for local cohomology and its applications, J. Algebra 331 (2011),
512–522.
[5] N.T. Cuong and P.H. Quy, On the splitting of local cohomology and structure of finitely generated modules in
local rings, Preprint.
[6] R.C. Heitmann, Tight closure in non-excellent local rings, J. Algebra 324 (2010), 430–441.
[7] M. Hochster and C. Huneke, Infinite integral extensions and big Cohen-Macaulay algebras, Ann. of Math. 135
(1992), 45–79.
[8] C. Huneke and G. Lyubeznik, Absolute integrak closure in positive characteristic, Adv. Math. 210 (2007), 498–
504.
[9] T. Kawasaki, On arithmetic Macaulayfication of Noetherian rings, Trans. Amer. Math. Soc. 354 (2001), 123–149.
[10] I.G. Macdonald, Secondary representation of modules over a commutative ring, Symp. Math. 11 (1973), 23–43.
[11] H. Matsumura, Commutative ring theory, Cambridge University Press, 1986.
[12] L.T. Nhan and P.H. Quy, Attached primes of local cohomology modules under localization and completion, J.
Algebra 420 (2014), 475–485.
[13] A. Sannai and A.K. Singh, Galois extensions, plus closure, and maps on local cohomology, Adv. Math. 229
(2012), 1847–1861.
[14] R.Y. Sharp, Some results on the vanishing of local cohomology modules, Proc. London Math. Soc. 30 (1975),
177–195.
[15] K.E. Smith, Tight closure of parameter ideals, Invent. Math. 115 (1994), 41–60.
Department of Mathematics, FPT University, Hoa Lac Hi-Tech Park, Ha Noi, Viet Nam
E-mail address: [email protected]
| 0math.AC
|
Context Attentive Bandits:
Contextual Bandit with Restricted Context
arXiv:1705.03821v2 [cs.AI] 7 Jun 2017
Djallel Bouneffouf1 , Irina Rish2 , Guillermo A. Cecchi3 , Raphaël Féraud4
1,2,3
IBM Thomas J. Watson Research Center, Yorktown Heights, NY USA
4
Orange Labs, 2 av. Pierre Marzin, 22300 Lannion (France)
{dbouneffouf, Irish, gcecchi }@us.ibm.com
[email protected]
Abstract
We consider a novel formulation of the multiarmed bandit model, which we call the contextual bandit with restricted context, where
only a limited number of features can be accessed by the learner at every iteration. This
novel formulation is motivated by different online problems arising in clinical trials, recommender systems and attention modeling.
Herein, we adapt the standard multi-armed
bandit algorithm known as Thompson Sampling to take advantage of our restricted context setting, and propose two novel algorithms,
called the Thompson Sampling with Restricted
Context (TSRC) and the Windows Thompson
Sampling with Restricted Context (WTSRC),
for handling stationary and nonstationary environments, respectively. Our empirical results
demonstrate advantages of the proposed approaches on several real-life datasets.
1
Introduction
In sequential decision problems encountered in various
applications, from clinical trials [Villar et al., 2015] and
recommender systems [Mary et al., 2015] to visual attention models [Mnih et al., 2014], a decision-making algorithm must choose among several actions at each time
point. The actions are typically associated with a side
information, or a context (e.g., a user’s profile), and the
reward feedback is limited to the chosen option. For example, in image processing with attention models [Mnih
et al., 2014], the context is an input image, the action
is classifying the image into one of the given categories,
and the reward is 1 if classification is correct and 0 otherwise. A different example involves clinical trials [Villar
et al., 2015], where the context is the patient’s medical
record (e.g. health condition, family history, etc.), the
actions correspond to the treatment options being compared, and the reward represents the outcome of the proposed treatment (e.g., success or failure). In this setting,
we are looking for a good trade-off between the exploration (e.g., of the new drug) and the exploitation (of the
known drug).
This inherent exploration vs. exploitation trade-off exists in many sequential decision problems, and is traditionally formulated as the multi-armed bandit (MAB)
problem, stated as follows: given K “arms”, or possible actions, each associated with a fixed but unknown
reward probability distribution [Lai and Robbins, 1985;
Auer et al., 2002], an agent selects an arm to play at each
iteration, and receives a reward, drawn according to the
selected arm’s distribution, independently from the previous actions. A particularly useful version of MAB is
the contextual multi-arm bandit (CMAB), or simply contextual bandit problem, where at each iteration, before
choosing an arm, the agent observes an N -dimensional
context, or feature vector. The agent uses this context,
along with the rewards of the arms played in the past, to
choose which arm to play in the current iteration. Over
time, the agent’s aim is to collect enough information
about the relationship between the context vectors and
rewards, so that it can predict the next best arm to play
by looking at the corresponding contexts [Langford and
Zhang, 2008; Agrawal and Goyal, 2013].
We introduce here a novel and practically important
special case of the contextual bandit problem, called the
contextual bandit with restricted context (CBRC), where
observing the full feature vector at each iteration is too
expensive or impossible for some reasons, and the player
can only request to observe a limited number of those
features; the upper bound (budget) on the feature subset is fixed for all iteration, but within the budget, the
player can choose any feature subset of the given size.
The problem is to select the best feature subset so that
the overall reward is maximized, which involves exploring both the feature space as well as the arms space.
For instance, in [Tekin et al., 2015], the analysis of
clinical trials shows that a doctor can ask a patient only
a limited number of question before deciding on drug
prescription. In the visual attention models involved in
visual pattern recognition [Mnih et al., 2014], a retina-
algorithm based on the LASSO estimator is presented.
Still, the above work, unlike ours, assumes full observability of the context variables.
In classical online learning (non-bandit) setting, where
the actual label of a mislabeled sample is revealed to the
classifier (unlike 0 reward for any wrong classification
decision in the bandit setting), the authors of [Wang et
al., 2014] investigate the problem of Online Feature Selection, where the aim is to make accurate predictions
using only a small number of active features.
Finally, [Durand and Gagné, 2014] tackles the online
feature selection problem by addressing the combinatorial optimization problem in the stochastic bandit setting
with bandit feedback, utilizing the Thompson Sampling
algorithm. Note that none of the previous approaches addresses the problem of context restriction (variable selection) in the contextual bandit setting, which is the main
focus of this work.
like representation is used, where at each moment only a
small region of an image is observed at high resolution,
and image classification is performed based on such partial information about the image. Furthermore, in recommender system setting [Hu and Ogihara, 2011], recommending an advertisement depends on user’s profile, but
usually only a limited aspects of such profile are available. The above examples can be modeled within the
proposed framework, assuming that only a limited number of features from the full context can be selected and
observed before choosing an arm (action) at each iteration.
Overall, the main contributions of this paper include
(1) a new formulation of a bandit problem with restricted
context, motivated by practical applications with a limited budget on information access, (2) two new algorithms, both for stationary and non-stationary settings
of the restricted-context contextual bandit problem, and
(3) empirical evaluation demonstrating advantages of the
proposed methods over a range of datasets and parameter
settings.
This paper is organized as follows. Section 2 reviews
related works. Section 3 introduces some background
concepts. Section 4 introduces the contextual bandit
model with restricted context, and the proposed algorithms for both stationary and non-stationary environments. Experimental evaluation on several datasets, for
varying parameter settings, is presented in Section 5. Finally, the last section concludes the paper and points out
possible directions for future works.
2
3
Background
This section introduces some background concepts our
approach builds upon, such as contextual bandit, combinatorial bandit, and Thompson Sampling.
The contextual bandit problem. Following [Langford
and Zhang, 2008], this problem is defined as follows.
At each time point (iteration) t ∈ {1, ..., T }, a player
is presented with a context (feature vector) c(t) ∈ RN
before choosing an arm k ∈ A = {1, ..., K}. We will
denote by C = {C1 , ..., CN } the set of features (variables) defining the context. Let r = (r1 (t), ..., rK (t))
denote a reward vector, where rk (t) ∈ [0, 1] is a reward
at time t associated with the arm k ∈ A. Herein, we
will primarily focus on the Bernoulli bandit with binary
reward, i.e. rk (t) ∈ {0, 1}. Let π : C → A denote a
policy. Also, Dc,r denotes a joint distribution (c, r). We
will assume that the expected reward is a linear function
of the context, i.e. E[rk (t)|c(t)] = µTk c(t), where µk is
an unknown weight vector (to be learned from the data)
associated with the arm k.
Related Work
The multi-armed bandit problem has been extensively
studied. Different solutions have been proposed using a stochastic formulation [Lai and Robbins, 1985;
Auer et al., 2002; Bouneffouf and Féraud, 2016] and a
Bayesian formulation [Agrawal and Goyal, 2012]; however, these approaches did not take into account the context.
In LINUCB [Li et al., 2010], Neural Bandit [Allesiardo et al., 2014] and in Contextual Thompson Sampling (CTS) [Agrawal and Goyal, 2013], a linear dependency is assumed between the expected reward of an action and its context; the representation space is modeled
using a set of linear predictors. However, the context is
assumed to be fully observable, unlike in this paper.
Motivated by dimensionality reduction task, [AbbasiYadkori et al., 2012] studied a sparse variant of stochastic
linear bandits, where only a relatively small (unknown)
subset of features is relevant to a multivariate function
optimization. Similarly, [Carpentier and Munos, 2012]
also considered the high-dimensional stochastic linear
bandits with sparsity, combining the ideas from compressed sensing and bandit theory. In [Bastani and Bayati, 2015], the problem is formulated as a MAB with
high-dimensional covariates, and a new efficient bandit
Thompson Sampling (TS). The TS [Thompson, 1933],
also known as Basyesian posterior sampling, is a
classical approach to multi-arm bandit problem, where
the reward rk (t) for choosing an arm k at time t is
assumed to follow a distribution P r(rt |µ̃) with the
parameter µ̃. Given a prior P r(µ̃) on these parameters,
their posterior distribution is given by the Bayes rule,
P r(µ̃|rt ) ∝ P r(rt |µ̃)P r(µ̃). A particular case of the
Thomson Sampling approach assumes a Bernoulli bandit
problem, with rewards being 0 or 1, and the parameters
following the Beta prior. TS initially assumes arm k to
have prior Beta(1, 1) on µk (the probability of success).
At time t, having observed Sk (t) successes (reward
= 1) and Fk (t) failures (reward = 0), the algorithm
updates the distribution on µk as Beta(Sk (t), Fk (t)).
The algorithm then generates independent samples θk (t)
2
A, π(c) = π̂(s(c))} of compound-function policies,
where the function s : RN → RN
C d maps each c(t) to
cd (t), for a given subset C d , and the function π̂ : RN
Cd →
A maps cd (t) to an action in A.
As mentioned before, in our setting, the rewards are
binary rk (t) ∈ {0, 1}. The objective of a contextual
bandit algorithm would be learn a hypothesis π over T
iterations maximizing the total reward.
from these posterior distributions of the µk , and selects
the arm with the largest sample value. For more details,
see, for example, [Agrawal and Goyal, 2012].
Combinatorial Bandit. Our feature subset selection
aproach will build upon the combinatorial bandit problem [Durand and Gagné, 2014], specified as follows:
Each arm k ∈ {1, ..., K} is associated with the corresponding variable xk (t) ∈ R, which indicates the reward obtained when choosing the k-th arm at time t, for
t > 1. Let us consider a constrained set of arm subsets
S ⊆ P (K), where P (K) is the power set of K, associated with a set of variables {rM (t)}, for all M ∈ S
and t > 1. Variable rM (t) ∈ R indicates the reward
associated with selecting a subset of arms M at time t,
where rM (t) = h(xk (t)), k ∈ M , for some function
h(·). The combinatorial bandit setting can be viewed as
a game where a player sequentially selects subsets in S
and observes rewards corresponding to the played subsets. Here we will define the reward function h(·) used
to compute rM (t) as aP
sum of the outcomes of the arms
in M , i.e. rM (t) =
k∈M xk (t), although more sophisticated nonlinear rewards are also possible. The objective of the combinatorial bandit problem is to maximize the reward over time. We consider here a stochastic
model, where xi (t) observed for an arm k are random
variables independent and distributed according to some
unknown distribution with unknown expectation µk . The
outcomes distribution can be different for each arm. The
global rewards rM (t) are also random variables independent and distributed according to some unknown distribution with unknown expectation µM .
4
Algorithm 1 The CBRC Problem Setting
1: Repeat
2: (c(t), r(t)) is drawn according to distribution Dc,r
3: The player chooses a subset C d ⊆ C
4: The values of cd (t) of features in C d are revealed
5: The player chooses an arm k(t) = π̂(cd (t))
6: The reward rk(t) is revealed
7: The player updates its policy π
8: t = t + 1
9: Until t=T
Thompson Sampling with Restricted Context
We now propose a method for solving the stationary CBRC problem, called Thompson Sampling with Restricted Context (TSRC), and summarize it in Algorithm
2 (see section 3 for background on Thompson Sampling);
here the combinatorial task of selecting the best subset of features is approached as a combinatorial bandit
problem, following the approach of [Durand and Gagné,
2014], and the subsequent decision-making (action selection) task as a contextual bandit problem solved by
Thompson Sampling [Agrawal and Goyal, 2013], respectively.
Let ni (t) be the number of times the i-th feature has
been selected so far, let rif (t) be the cumulative reward
associated with the feature i, and let rk (t) be the reward
associated with the arm k at time t.
The algorithm takes as an input the desired number of
features d, as well as the initial values of the Beta distribution parameters in TS. At each iteration t, we update
the values of those parameters, Si and Fi (steps 5 and
6), to represent the current total number of successes and
failures, respectively, and then sample the ”probability
of success” parameter θi from the corresponding Beta
distribution, separately for each feature i to estimate µi ,
which is the mean reward
the use of the
P conditioned to
1
d
variable i : µi = K
k E[rk .1{i ∈ C }] (step 7). We
select the best subsetP
of features, C d ⊆ C, such that
d
C = arg maxC d ∈C i∈C d θi (step 9). So the goal of
the combinatorialP
bandit in TSRC algorithm is to maximize: E[rC d ] = i∈C d µi ; note that implementing this
step does not actually require combinatorial search1 .
Problem Setting
In this section, we define a new type of a bandit problem,
the contextual bandit with restricted context (CBRC); the
combinatorial task of feature subset (i.e., restricted context) selection as treated as a combinatorial bandit problem [Durand and Gagné, 2014], and our approach will be
based on the Thompson Sampling [Agrawal and Goyal,
2012].
4.1
Contextual Bandit with Restricted Context
in Stationary Environment (CBRC)
Let c(t) ∈ RN denote a value assignments to the vector
of random variables, or features, (C1 , ..., CN ) at time t,
and let C = {1, ..., N } be the set of their indexes. Furthermore, let cd (t) denote a sparse vector of assignments
to only d ≤ N out of N features, with indexes from a
subset C d ⊆ C, |C d | = d, and with zeros assigned to all
features with indexes outside of C d .
Formally, we denote the set of all such vectors as
d
N
d
RN
/ C d }. In the
C d = {c (t) ∈ R | ci = 0 ∀i ∈
future, we will always use C d to denote a feature subset of size d, and by cd the corresponding sparse vector. We will consider a set Πd = ∪C d ⊆C {π : RN →
1
Since the individual rewards θi are non-negative (recall
that they follow Beta-distribution), we can simply select the
set C d of d arms with the highest individual rewards θi (t).
3
with R > 0, ∈]0, 1], γ ∈]0, 1] constants, and µ̂ =
Pt−1
B(t)−1 ( τ =1 c(τ )c(τ )).
At each time point t, and for each arm, we sample a ddimensional µ˜k from N (µˆk (t), v 2 Bk (t)−1 ), and choose
an arm maximizing cd (t)> µ˜k (step 14 in the algorithm),
obtain the reward (step 15), and update the parameters
of the distributions for each µ˜k (steps 16-21). Finally,
the reward rk (t) for choosing an arm k is observed, and
relevant parameters are updated.
Algorithm 2 Thompson Sampling with Restricted Context (TSRC)
1: Require: The size d of the feature subset, the initial values Si (0) and Fi (0) of the Beta distribution
parameters.
2: Initialize: ∀k ∈ {1, ..., K}, Bk = Id , µˆk =
0d , gk = 0d , and ∀i ∈ {1, ..., N }, ni (0) = 0,
rif (0) = 0.
3: Foreach t = 1, 2, ..., T do
4:
Foreach context feature i = 1, ..., N do
5:
Si (t) = Si (0) + rif (t − 1)
6:
Fi (t) = Fi (0) + ni (t − 1) − rif (t − 1)
7:
Sample θi from Beta(Si (t), Fi (t)) distribution
8:
9:
10:
11:
12:
13:
14:
4.2
In a stationary environment, the context vectors and the
rewards are drawn from fixed probability distributions;
the objective is to identify a subset of features allowing
for the optimal context-to-arm mapping. However, the
objective changes when the environment becomes nonstationary.
In the non-stationary CBRC setting, we will assume
that the rewards distribution can change only at certain times, and remain stationary between such changes.
Given the non-stationarity of the reward, the player
should continue looking for feature subsets which allow
for the optimal context-arm mapping, rather than converge to a fixed subset.
Similarly to stationary CBRC problem described earlier, at each iteration, a context c(t) ∈ RN describes
the environment, the player chooses a subset C d ⊆ C of
the feature set C, and observes the values of those features as a sparse vector cd (t), where all features outside
of C d are set to zero. Given cd (t), the player chooses
an arm k(t). The reward rk (t) of the played arm is revealed. The reward distribution is non-stationary as described above, and the (stationary) reward distributions
between the change points are unknown. We will make
a very specific simplifying assumptions that the change
points occur at every v time points, i.e. that all windows
of stationarity have a fixed size.
End do
P
Select C d (t) = argmaxC d ⊆C i∈C d θi
Obtain sparse vector cd (t) of feature assignments
in C d , where ci = 0 ∀i ∈
/ Cd
Foreach arm k = 1, ..., K do
Sample µ˜k from N (µˆk , v 2 Bk−1 ) distribution.
End do
Select arm k(t) = argmax cd (t)> µ˜k
k⊂{1,...,K}
Observe rk (t)
if rk (t) = 1 then
Bk = Bk + cd (t)cd (t)T
gk = gk + cd (t)rk (t)
µˆk = Bk−1 gk
20:
∀i ∈ C d (t), ni (t) = ni (t − 1) + 1 and rif (t) =
f
ri (t − 1) + 1.
21:
End if
22: End do
15:
16:
17:
18:
19:
Once a subset of features is selected using the combinatorial bandit approach, we switch to the contextual
bandit setting in steps 10-13, to choose an arm based on
the context consisting now of a subset of features.
We will assume that the expected reward is a linear function of a restricted context, E[rk (t)|cd (t)] =
µTk cd (t); note that this assumption is different from the
linear reward assumption of [Agrawal and Goyal, 2013]
where single parameter µ was considered for all arms,
there was no restricted context, and for each arm, a separate context vector ck (t) was assumed.
Besides this difference, we will follow the approach
of [Agrawal and Goyal, 2013] for the contextual Thompson Sampling. We assume that reward ri (t) for choosing
arm i at time t follows a parametric likelihood function
P r(r(t)|µ˜k ), and that the posterior distribution at time
t+1, P r(µ̃|r(t)) ∝ P r(r(t)|µ̃)P r(µ̃) is given by a multivariate Gaussian distribution N (µˆk (t + 1), v 2 Bk (t +
Pt−1
>
1)−1 ), where Bk (t) = Id + τ =1 c(τ )c(τ
q) with d
the size of the context vectors c, v = R
Contextual Bandit with Restricted Context
in Non-stationary Environments
Windows TSRC Algorithm
Similarly to the TSRC algorithm proposed earlier for the
stationary CBRC problem, our algorithm for the nonstationary CBRC uses Thompson Sampling (TS) to find
the best d features of the context. The two algorithms
are practically identical, except for the following important detail: instead of updating the Beta distribution with
the number of successes and failures accumulated from
the beginning of the game, only the successes and failures within a given stationarity window are counted. The
resulting approach is called the Window Thompson Sampling with restricted Context, or WTSRC.
Note that v, the true size of the stationary window assumed above, is not known by the algorithm, and is replaced by some approximate window size parameter w.
In this paper, we assumed a fixed w, and experiment with
24
1
dln( γ )
4
Table 1: Datasets
UCI Datasets
Covertype
CNAE-9
Internet Advertisements
Poker Hand
Instances
581 012
1080
3279
1 025 010
Features
95
857
1558
11
the misclassification error, computed as the total number
of misclassified samples over the number of iterations.
This average errors for each algorithm were computed
using 10 cyclical iterations over each dataset, and over
the four different sparsity levels mentioned above.
As expected, the CTS algorihtm with the full set of
features (Fullfeatures) achieved the best performance as
compared with the rest of the algorithms, underscoring
the importance of the amount of context observed in an
on-line learning setting.
However, when the context is limited, is in the CBRC
problem considered in this work, our TSRC approach
shows superior performance (shown in bold in Table 2)
when compared to the rest of the algorithms, except for
the Fullfeatures, confirming the importance of efficient
feature selection in the CBRC setting.
Overall, based on their mean error rate, the top three
algorithms were TSRC (mean error 49.01%), Randomfix (mean error 55.46%), and MAB (mean error 57.98%),
respectively, suggesting that using a fixed randomly selected feature subset may still be a better strategy than
not considering any context at all, as in MAB. However,
as it turns out, ignoring the context in MAB may still be
a better approach than randomly changing the choice of
feature at each iteration in Random-EI; the latter resulted
into the worst mean error of 61.18%.
Classes
7
9
2
9
several values of it; however, in the future, can be also
adjusted using a bandit approach.
5
Empirical Evaluation
Empirical evaluation of the proposed methods was based
on four datasets from the UCI Machine Learning Repository 2 : Covertype, CNAE-9, Internet Advertisements and
Poker Hand (for details of each dataset, see Table 1).
In order to simulate a data stream, we draw samples
from the dataset sequentially, starting from the beginning
each time we draw the last sample. At each round, the algorithm receives the reward 1 if the instance is classified
correctly, and 0 otherwise. We compute the total number
of classification errors as a performance metric.
We compare our algorithms with the following competing methods:
• Multi-arm Bandit (MAB): this is the standard Thomspon Sampling approach to (non-contextual) multiarm bandit setting.
• Fullfeatures: this algorithm uses the contextual
Thomspon Sampling (CTS) with the full set of features.
• Random-EI: this algorithm selects a Random subset
of features of the specified size d at Each Iteration
(thus, Random-EI), and then invokes the contextual
bandit algorithm (CTS).
• Random-fix: similarly to Random-EI, this algorithm
invokes CTS on a random subset of d features; however, this subset is selected once prior to seeing any
data samples, and remains fixed.
We ran the above algorithms and our proposed TSRC
and WTSRC methods, in stationary and non-stationary
settings, respectively, for different feature subset sizes,
such as 5%, 25%, 50% and 75% of the total number of
features.
In the Figures presented later in this section, we used
the parameter, called sparsity, to denote the percent of
features that were not selected, resulting into the sparsity
levels of 95%, 75%, 50% and 25%, respectively. In the
following sections, we will present our results first for
the stationary and then for the non-stationary settings.
5.1
Detailed analysis on Covertype dataset
Figure 1 provides a more detailed evaluation of the algorithms for different levels of sparsity, on a specific
dataset. Ignoring the Fullfeatures which, as expected,
outperforms all methods since it considers the full context, we observe that:
95 % sparsity: TSRC has the lowest error out of all
methods, followed tightly by MAB, suggesting that, at a
very high sparsity level, ignoring the context (MAB) is
similar to considering only a very small (5%) fraction of
it. Also, as mentioned above, sticking to the same fixed
subset of randomly selected features appears to be better
than changing the random subset at each iteration.
75 % sparsity: we observe that Random-EI has a
lower error than the Random-fix.
50 % sparsity: TSRC has the lowest error, followed
closely by the Random-fix, which implies that, at some
levels of sparsity, random subset selection can perform
almost as good as our optimization strategy. Again,
we observe that Random-fix performs better that the
Randon-EI, where the performance of the latter is close
to the multi-arm bandit without any context.
25 % sparsity: we observe that TSRC perform practically as good as Fullfeatures, which implies that at this
sparsity level, our approach was able to select the optimal
feature subset.
Stationary case
Table 2 summarizes our results for the stationary CBRC
setting; it represents the average classification error, i.e.
2
5.2
Non-stationary Case
In this setting, for each dataset, we run the experiments for 3,000,000 iterations, where we change the la-
https://archive.ics.uci.edu/ml/datasets.html
5
Table 2: Stationary Environment
Datasets
Covertype
CNAE-9
Internet Advertisements
Poker Hand
MAB
Fullfeatures
TSRC
Random-fix
Random-EI
70.54 ± 0.30
79.85 ± 0.35
19.22 ± 0.20
62.29 ± 0.21
35.27 ± 0.32
52.01 ± 0.20
14.33 ± 0.22
58.57 ± 0.12
53.54 ± 1.75
68.47 ± 0.95
15.21 ± 1.20
58.82 ± 0.71
72.29 ± 2.38
68.50 ± 3.49
21.21 ± 1.93
59.83 ± 2.57
82.69 ± 1.87
80.02 ± 0.23
23.53 ± 1.64
58.49 ± 0.81
Figure 1: Stationary Environment (Covertype dataset)
95 % sparsity: MAB has the lowest error, as compared with the Random-EI and Random-fix, which implies that, at high sparsity levels, ignoring the context can
be better then even considering a small random subset of
it.
75 % sparsity: TSRC yields the best result, implying
that even in a non-stationary environment, a stationary
strategy for best subset selection can be beneficial.
50 % sparsity: both Random-Fix and Randon-EI
yield the worst results, implying that random selection is
not a good strategy at this sparsity level in non-stationary
environment.
25 % sparsity: our nonstationary method, WTSRC,
outperforms all other algorithms, demonstrating the advantages of a dynamic feature-selection strategy in a
non-stationary environment at relatively low sparsity levels.
bel of class at each 500,000 iteration to simulate the
non-stationarity. We evaluate our WTSRC algorithm for
the nonstationary CBRC problem against our stationarysetting TSRC method, and against the same five baseline
methods we presented before.
Similarly to the stationary case evaluation, the Table
3 reports the mean error over all iterations and over the
same for level of sparsity as before. Overall, we observed
that the WTSRC performs the best, confirming the effectiveness of using our time-windows approach in a nonstationary on-line learning setting.
Our top three performing algorithms were WTSRC,
Fullfeatures and TSRC, in that order, which underscores
the importance of the size of the observed context even
in the non-stationary on-line learning setting.
Detailed analysis on Covertype dataset
Figure 2 provides a more detailed evaluation of the algorithms for different levels of sparsity, on a specific
dataset. The Fullfeatures, as expected, has the same performance on different level of sparsity, since it has the
access to all features. We observe that:
6
Conclusions
We have introduced a new formulation of MAB, motivated by several real-world applications including visual
attention modeling, medical diagnosis and RS. In this
6
Table 3: Non-Stationary Environment
Datasets
Covertype
CNAE-9
Internet Advertisements
Poker Hand
MAB
Fullfeatures
TSRC
Random-fix
Random-EI
WTSRC
69.72 ± 4.30
74.34 ± 5.39
43.99 ± 3.85
82.90 ± 4.03
68.54 ± 2.10
69.21 ± 2.12
40.21 ± 1.87
82.44 ± 0.43
76.12 ± 2.86
71.87 ± 4.5
42.01 ± 1.79
78.86 ± 1.25
80.96 ± 2.19
79.87 ± 4.5
40.04 ± 4.52
79.99 ± 1.48
80.71 ± 2.22
76.21 ± 1.90
40.56 ± 1.19
79.81 ± 0.48
60.56
65.56
38.06
77.56
± 2.95
± 1.05
± 0.85
± 1.79
Figure 2: Non-Stationary Environment (Covertype dataset)
setting, which we refer to as contextual bandit with restricted context (CBRC), a set of features, or a context,
is used to describe the current state of world; however,
the agent can only choose a limited-size subset of those
features to observe, and thus needs to explore the feature space simultaneously with exploring the arm space,
in order to find the best feature subset. We proposed
two novel algorithms based on Thompson Sampling for
solving the CBRC problem in both stationary and nonstationary environments. Empirical evaluation on several datasets demonstrates advantages of the proposed
approaches.
7
stochastic bandits. In Proceedings of the Fifteenth
International Conference on Artificial Intelligence
and Statistics, AISTATS 2012, pages 1–9, 2012.
[Agrawal and Goyal, 2012] Shipra Agrawal and Navin
Goyal. Analysis of thompson sampling for the multiarmed bandit problem. In COLT 2012 - The 25th
Annual Conference on Learning Theory, June 25-27,
2012, Edinburgh, Scotland, pages 39.1–39.26, 2012.
[Agrawal and Goyal, 2013] Shipra Agrawal and Navin
Goyal. Thompson sampling for contextual bandits
with linear payoffs. In ICML (3), pages 127–135,
2013.
Acknowledgments
[Allesiardo et al., 2014] Robin Allesiardo, Raphaël
Féraud, and Djallel Bouneffouf. A neural networks
committee for the contextual bandit problem. In
Neural Information Processing - 21st International
Conference, ICONIP 2014, Kuching, Malaysia,
November 3-6, 2014. Proceedings, Part I, pages
374–381, 2014.
The authors thank Dr. Alina Beygelzimer and Dr.
Karthikeyan Shanmugam for the critical reading of the
draft manuscript.
References
[Abbasi-Yadkori et al., 2012] Yasin
Abbasi-Yadkori,
Dávid Pál, and Csaba Szepesvári.
Online-toconfidence-set conversions and application to sparse
[Auer et al., 2002] Peter Auer, Nicolò Cesa-Bianchi,
and Paul Fischer. Finite-time analysis of the mul7
[Villar et al., 2015] Sofı́a S Villar, Jack Bowden, and
James Wason. Multi-armed bandit models for the optimal design of clinical trials: benefits and challenges.
Statistical science: a review journal of the Institute of
Mathematical Statistics, 30(2):199, 2015.
[Wang et al., 2014] Jialei Wang, Peilin Zhao, Steven CH
Hoi, and Rong Jin. Online feature selection and its
applications. IEEE Transactions on Knowledge and
Data Engineering, 26(3):698–710, 2014.
tiarmed bandit problem. Machine Learning, 47(23):235–256, 2002.
[Bastani and Bayati, 2015] Hamsa
Bastani
and
Mohsen Bayati.
Online decision-making with
high-dimensional covariates.
Available at SSRN
2661896, 2015.
[Bouneffouf and Féraud, 2016] Djallel Bouneffouf and
Raphaël Féraud. Multi-armed bandit problem with
known trend. Neurocomputing, 205:16–21, 2016.
[Carpentier and Munos, 2012] Alexandra
Carpentier
and Rémi Munos. Bandit theory meets compressed
sensing for high dimensional stochastic linear bandit.
In Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, AISTATS
2012, pages 190–198, 2012.
[Durand and Gagné, 2014] Audrey Durand and Christian Gagné. Thompson sampling for combinatorial
bandits and its application to online feature selection.
2014.
[Hu and Ogihara, 2011] Yajie Hu and Mitsunori Ogihara. Nextone player: A music recommendation system based on user behavior. In ISMIR, pages 103–108,
2011.
[Lai and Robbins, 1985] T. L. Lai and Herbert Robbins.
Asymptotically efficient adaptive allocation rules. Advances in Applied Mathematics, 6(1):4–22, 1985.
[Langford and Zhang, 2008] John Langford and Tong
Zhang. The epoch-greedy algorithm for multi-armed
bandits with side information. In Advances in neural information processing systems, pages 817–824,
2008.
[Li et al., 2010] Lihong Li, Wei Chu, John Langford,
and Robert E. Schapire. A contextual-bandit approach
to personalized news article recommendation. CoRR,
2010.
[Mary et al., 2015] Jérémie Mary, Romaric Gaudel, and
Philippe Preux. Bandits and recommender systems. In
Machine Learning, Optimization, and Big Data - First
International Workshop, MOD 2015, pages 325–336,
2015.
[Mnih et al., 2014] Volodymyr Mnih, Nicolas Heess,
Alex Graves, et al. Recurrent models of visual attention. In Advances in Neural Information Processing
Systems, pages 2204–2212, 2014.
[Tekin et al., 2015] Cem Tekin, Onur Atan, and Mihaela
van der Schaar. Discover the expert: Context-adaptive
expert selection for medical diagnosis. IEEE Trans.
Emerging Topics Comput., 3(2):220–234, 2015.
[Thompson, 1933] William R Thompson. On the likelihood that one unknown probability exceeds another
in view of the evidence of two samples. Biometrika,
25(3/4):285–294, 1933.
8
| 2cs.AI
|
Stock Trading via Feedback Control:
Stochastic Model Predictive or Genetic?
arXiv:1708.08857v2 [cs.CE] 4 Oct 2017
Mogens Graf Plessen and Alberto Bemporad
Abstract— We seek a discussion about the most suitable feedback control structure for stock trading under the consideration
of proportional transaction costs. Suitability refers to robustness
and performance capability. Both are tested by considering
different one-step ahead prediction qualities, including the ideal
case, correct prediction of the direction of change in daily
stock prices and the worst-case. Feedback control structures
are partitioned into two general classes: stochastic model
predictive control (SMPC) and genetic. For the former class
three controllers are discussed, whereby it is distinguished
between two Markowitz- and one dynamic hedging-inspired
SMPC formulation. For the latter class five trading algorithms
are disucssed, whereby it is distinguished between two different
moving average (MA) based, two trading range (TR) based, and
one strategy based on historical optimal (HistOpt) trajectories.
This paper also gives a preliminary discussion about how
modified dynamic hedging-inspired SMPC formulations may
serve as alternatives to Markowitz portfolio optimization. The
combinations of all of the eight controllers with five different
one-step ahead prediction methods are backtested for daily
trading of the 30 components of the German stock market
index DAX for the time period between November 27, 2015
and November 25, 2016.
We compare eight different stock trading algorithms that
can be partitioned into the two aforementioned classes.
For scenario generation, on which we rely all controllers,
we assume five different one-step ahead price prediction
methods. Their quality ranges from ideal (perfect price-ahead
prediction) to totally off (wrong price rate sign-prediction at
all sampling intervals). It is stressed that we explicitly do
not consider multi-asset portfolio optimization, but instead
concentrate on the trading of separate single assets for a
given period of time. The used real-world data is drawn from
the 30 components of the German stock market index DAX.
The remainder of this paper is organized as follows.
System dynamics are described in Section II. Section III
introduces two different stochastic model predictive stock
trading frameworks. Genetic stock trading algorithms are
outlined in Section IV. Simulation experiments are reported
in Section V before concluding with Section VI.
I. I NTRODUCTION
Let time index t be associated with sampling time Ts such
that all time instances of interest can be described as tTs ,
whereby, throughout this papers, we have Ts = 1 day. Let
us define the system state by
T
Z(t) = I(t) M (t) N (t) W (t) ,
(1)
Within the context of performance-related asset trading,
we distinuish between three general tasks: system identification (finding of cause and effect relations), scenario
generation (future asset price predictions) and trade decision
taking (control logic). This paper focuses on the third task.
For low-level trading mechanics and feedback control thereof
we refer to [1]. For interesting recent control theory-related
research problems in finance associated with the control of
order book dynamics, see [2]. This work is founded on
[3] and [4]. The motivation for this paper is the intention
to extend a stochastic model predictive control approach
to multiple-asset portfolio optimization for profit- and riskrelated objectives. However, the suitability of SMPC needs
first to be compared to alternative control strategies. Such
a comparison is provided below. The main contribution of
this paper is thus analysis to find the most suitable general
feedback control structure for stock trading out of two
general and large classes: stochastic model predictive control
(SMPC) and genetic algorithms. In this paper, we here refer
to a genetic algorithm as any customized control method of
arbitrary structure whose parameters are optimized through
simulation using real-world data.
MGP and AB are with IMT School for Advanced Studies Lucca,
Piazza S. Francesco 19, 55100 Lucca, Italy, {mogens.plessen,
alberto.bemporad}@imtlucca.it. Presented as a poster at the
XVIII Workshop on Quantitative Finance 2017 (QFW2017).
II. T RANSITION DYNAMICS M ODELING
with I(t) ∈ {0, 1} indicating a cash- or stock-investment,
respectively, M (t) ∈ R the current cash position (measured
in currency e ), N (t) ∈ Z+ the number of shares held,
and W (t) ∈ R+ the current portfolio wealth. Thus, to
analyze a suitable stock trading algorithm, we assume no
fractional investments being possible, but either entirely
cash- or stock-investment. Transition dynamics can then
be modeled as a Markov decision process (MDP). Control
variable J(t) ∈ {0, 1} decides upon investment positions
according to Figure 1. In general, we model transaction costs
as non-convex with a fixed charge for any nonzero trade
(fixed transaction costs) and a linear term scaling with the
quantity traded (proportional transaction costs). Thus, at time
t − 1, the purchase of N (t − 1) shares of an asset results
J =1
J =0
cash
I(t)=0
stock
I(t)=1
J =1
J =0
Fig. 1. Visualization of the Markov decision process (MDP) when trading
cash and a stock only.
in M (t) = M (t − 1) − N (t − 1)s(t − 1) (1 + ǫbuy ) − βbuy ,
with s(t) denoting asset (closing) price at time t. Similarly,
for the selling of N (t − 1) assets we have M (t) = M (t −
1) + N (t − 1)s(t − 1) (1 − ǫsell ) − βsell . For the remainder
of this paper we assume no fixed transaction costs, i.e.,
βsell = βbuy = 0. This simplification is done to directly
adapt the convex problem formulations from [3] and for the
dynamic hedging-inspired formulation proposed in Section
III. Fixed transaction costs, that render the problem nonconvex, can be approached by iterative relaxation methods
[5] or hybrid nsystem theory. For wealth
dynamics, we
o
have W (t) ∈ W (t − 1), M (t), W̃ (t) , whereby W̃ (t) =
M̃ (t)+ Ñ (t)s(t) with M̃ (t) denoting the optimizer and Ñ (t)
the optimal objective function value of
M (t − 1) − βbuy − M (t)
.
max N (t) ∈ Z+ : N (t) =
s(t − 1) (1 + ǫbuy )
M(t)≥0
Thus, given M (t − 1), we find the largest possible positive
integer number of assets we can purchase under consideration of transaction costs. The smallest possible cash residual
is M̃ (t) = 0.
Within this paper, the focus is on unconstrained trading
frequency of one asset, i.e., trading is permitted on any
two consecutive trading days, and confining cash and asset
to be based on the same currency. More general is the
treatment of multiple assets, multiple foreign exchange rates
(forex), and various constraints such as a bound on the total
number of admissable trades, a waiting period in between
two consecutive trades or diversification constraints, which
require a state-space extension but or not subject of this paper
explicitly.
To summarize, at every trading interval t we conduct
following algorithm:
1) Read current s(t) to update W (t) and thus Z(t).
2) Decide on J(t) ∈ {0, 1}.
3) Rebalance the portfolio according to J(t).
All of the following two sections are concerned about the
decision on J(t) ∈ {0, 1} with fundamental objective profit
maximization.
III. S TOCHASTIC M ODEL P REDICTIVE S TOCK T RADING
Let us discuss a relation between portfolio optimization
and dynamic hedging using stochastic model predictive control (SMPC). For a financial institution, hedging a derivative
contract implies to dynamically rebalance a so-called replicating portfolio of underlying assets at periodic intervals so
that, at the expiration date of the contract, the value of the
portfolio is as close as possible to the payoff value to pay
to the customer. For the multiple asset replicating portfolio
case, wealth dynamics w(t) of the replicating portfolio can
be defined as
! n
n
X
X
bi (t)ui (t) (2)
hi (t) +
w(t+1) = (1+r) w(t) −
i=1
i=1
where ui (t) is the fractional quantity of asset i held at time
t, bi (t) = si (t + 1) − (1 + r)si (t) is the excess return, i.e.,
how much the asset gains (or loses) with respect to the riskfree rate r over interval Ts , and transaction costs hi (t) are
assumed to be proportional to the traded quantity of stock,
i.e.,
hi (t) = ǫi si (t)|ui (t) − ui (t − 1)|,
(3)
with fixed quantity ǫi ≥ 0 depending on commissions on
trading asset i, i = 1, . . . , n (we assume no costs are applied
on trasacting the risk-free asset). A standard option contract
typically covers 100 shares. Thus, ui (t) = 1 implies a
portfolio such that at the end of the rebalancing interval 100
shares of underlying asset i are held. For our setting, we here
assume one asset only and drop subscripts correspondingly.
Furthermore, we set ǫ = ǫbuy and ǫsell = ǫbuy . For the
formulation of convex optimization problems, we introduce
a virtual portfolio, constrain1 u(t) ∈ {0, 1}, and then relate
I(t) = u(t).
A. Two Markowitz-inspired SMPCs
With regard of portfolio optimization, Markowitz [3]
trades-off the mean (performance) and variance (risk) of the
return. For our setting, this objective can be formulated as
α
max
E[w(t + 1)] − Var[w(t + 1)],
(4)
2
u(t)∈{0,1}
where α denotes the trade-off parameter. The decision of
selecting u(t) is largely dependent on s(t + 1), which is unknown at time t. Employing a SMPC approach, we therefore
generate M scenarios for possible future prices sj (t+1) with
corresponding probabilities π j , j = 1,
P.M. . , M . Accordingly,
we obtain wj (t + 1), E[w(t + 1)] = j=1 π j wj (t + 1) and
Var[w(t + 1)] = E[w2 (t + 1)] − E2 [w(t + 1)]. We generate
scenarios as
sj (t + 1) = ŝ(t + 1) + σpert η j (t), η j (t) ∼ N (0, 1),
(5)
where ŝ(t + 1) denotes our mean estimate of s(t + 1) and
σpert ≥ 0 is a tuning parameter to add perturbation noise.
For final experiments we assume M = 100. Thus, our first
SMPC-based controller solves (4) with scenario generation
according (5). It is referred to as SMPC-M100. For the
fractional case relevant for dynamic option hedging (and
specific for ∆-hedging [6]), the corresponding to (4) can
be cast into a quadratic program (QP) by the introduction
of two slack variables. For our case, we just evaluate the
objective function for both u(t) ∈ {0, 1} and therefore do
not require a QP-solver.
In a second setting we assume M = 1 which implies
Var[w(t + 1)] = 0. In order to still maintain a possible knob
to trade-off performance and risk, we define
max
u(t)∈{0,1}
E[w(t + 1)] −
β
(u(t) − u(t − 1))2 σ,
2
(6)
where β is a tuning parameter and σ was introduced to relate
to data. Assuming the log-normal stock model, it is typically
estimated as the maximum likelihood (ML) from T + 1
1 The unconstrained case admitting u(t) < 0 would imply the possibility
of shortselling stocks.
30
25
8
25
8
s(T ) − s(0)
p̄(t)
w(t)
[e ]
6
4
4
2
2
0
1
0
1
u(t) [-]
[e ]
30
s(T ) − s(0)
p̄(t)
w(t)
6
u(t) [-]
35
stock price [e ]
stock price [e ]
35
0.5
0
0.5
0
0
50
100
150
200
250
t
0
50
100
150
200
250
t
Fig. 2. Visualization of the dynamic hedging-inspired SMPC formulation for stock trading. A noncausal setting with perfect s(t + 1) knowledge is
assumed for illustration. Parameters for scenario generation and perturation noise are (M, σpert ) = (100, 0.3). (Left) Dynamic hedging result. (Right)
Result of SMPC-DH. In both
cases, (7) is solved. The difference is the generation of references pj (t + 1), j = 1, . . . , M , see Section III-B. The average
1 PM
j
is denoted by p̄(t) = M
j=1 p (t). The underlying stock (top frame) is of Vonovia SE (November 27, 2015 until November 25, 2016).
s(t)
+1)
past stock prices using {ln( s(t−T
s(t−T ) ), . . . , ln( s(t−1) )}. The
higher β, the less frequent u(t) is varied. Again, we solve
(6) by evaluation for both u(t) ∈ {0, 1} instead of solving
the more general QP. We refer to the resulting controller as
SMPC-E+, eventhough, strictly it is not stochastic anymore
since M = 1.
B. Dynamic hedging-inspired SMPC
For option hedging, the objective is to typically minimize
the so-called hedging error e(T ) = w(T ) − p(T ), where
w(T ) and p(T ) denote replicating portfolio wealth and option price at expiration date T , respectively. Guiding notion
of dynamic option hedging is to minimize the “tracking
error” e(t) = w(t) − p(t), ∀t = 0, . . . , T for all possible
asset price realization. In order to minimize under transaction
costs, three different stochastic measures of the predicted
hedging error are discussed in [4]. We here focus on the
LP-MinMax formulation, minimizing the maximal hedging
error resulting from scenario generation, i.e.,
min
max
u(t)∈{0,1} j=1,...,M
|wj (t + 1) − pj (t + 1)|.
(7)
Its benefit over the other two stochastic measures is independence from any trade-off parameter. Let us discuss how
the framework (7) can serve for stock trading. First, for
hedging of a European call option, the analytical scheme
to generatio option price scenarios is pj (t + 1) = (1 +
r)−(T −(t+1)) max(sj (T ) − Ks , 0), j = 1, . . . , M . The
option strike price is denoted by Ks . For our stock trading
objective, we modify these “references”. For (7), we therefore propose
pj (t + 1) = (1 + r)−(T −(t+1)) max(sj (t + 1) − s(0), l(t)),
(8)
with
(
max(w(t), 0), if w(t) > l(t),
l(t) =
l(t − 1),
otherwise,
and initialize l(0) = 0. The corresponding dynamic hedginginspired controller shall be referred to as SMPC-DH. Both
the difference with respect to dynamic option hedging and
the motivation for employing (8) for stock trading are
visualized in Figure 2. Using (8) in combination with (7)
can be interpreted as a trailing stop-loss strategy. Finally,
we remark that the other two stochastic measures (QP-Var
and LP-CVaR) from [4] can be employed likewise using (8).
IV. G ENETIC S TOCK T RADING
We discuss five genetic stock trading methods.
A. Two moving average-based controllers
Let us define the
average (MA) of a stock price as
Pmoving
pMA −1
1
s̃(t+1−τ
), with s̃(t+1) = ŝ(t+1)
sMA (t+1) = pMA
τ =0
and s̃(t + 1 − τ ) = s(t + 1 − τ ), ∀τ ≥ 1, and where
pMA is the moving average length parameter. The first MAbased controller, referred to as MA-Cross in the following,
(K)
performance [%]
100
50
0
−50
0
100
200
300
400
500
trading days
Fig. 3. Normalized two-year evolution of all 30 components held in the
DAX between November 28, 2014, and November 25, 2016. The data is
partitioned by the black-dashed vertical line into training and evaluation
data, respectively. Thus, t = 0 is initialized at trading day 261 (November
27, 2015).
triggers a buy-signal (J(t) = 1) in case of a short-term MA
coming from below crossing a long-term MA. Similarly, a
sell-signal (J(t) = 0) is generated in case of the shortterm MA crossing the long-term MA from above. The two
parameters defining MA-lengths shall be denoted by pMA,s
and pMA,l .
The second MA-based controller, below referred to as
MA-Sign, takes as input parameters pMA and TMA . It then
computes ∆sMA (t − τ ) = sMA (t + 1 − τ ) − sMA (t −
τ ), ∀τ = 0, 1, . . . , TMA − 2. A buy-signal is generated if
sign (∆sMA (t − τ )) > 0, ∀τ = 0, 1, . . . , TMA − 2, a sellsignal otherwise, and where sign(·) denotes the sign-operator.
The guiding idea is to exploit price trends using constantsign MA-slope rates for the past TMA − 1 intervals. We refer
to this second MA-based controller as MA-Sign.
B. Two trading range-based controllers
Let us select a time window [t − Twin , t] and partition it
such that
Twin = KpTR + δ,
(9)
where pTR ∈ Z++ is a parameter, K ∈ Z++ , and δ ≥ 0 a
corresponding residual of time-instances. We define intervalwise local maxima by
s(k)
max =
max
τ ∈[t−Twin +(k−1)K,t−Twin +kK]
s(τ ),
(k)
(10)
and the corresponding time arguments by tmax , ∀k =
(k)
(k)
1, . . . , K. Similarly we derive local minima smin and tmin .
Local minima and maxima are suitable to generate trading
ranges (TR). We refer to our first TR-based controller as
TR-Inside. It triggers trading signals as follows:
max (t+1)|
< ǫTR ,
if |ŝ(t+1)−ŷ
ŷmax (t+1)
0,
|ŝ(t+1)−ŷmin (t+1)|
J(t) = 1,
if
< ǫTR ,
ŷmin (t+1)
J(t − 1), otherwise,
(K)
where ŷmax (t + 1) = (t + 1 − tmax )qmax (t + 1) + smax ,
(K)
(K−1)
(K)
(K−1)
qmax (t + 1) = (smax − smax )/(tmax − tmax ), and analogously for ŷmin (t + 1) and qmin (t + 1). Thus, buy(sell)signals are triggered upon reaching the lower(upper) trading
range affinely constructed based upon the last two local
minima(maxima).
Our second TR-based controller is referred to as TROutside. It triggers trading signals according to:
max (t+1)
> ǫTR ,
if ŝ(t+1)−ŷ
1,
ŷmax (t+1)
ŝ(t+1)−ŷmin (t+1)
J(t) = 0,
if
< −ǫTR,
ŷmin (t+1)
J(t − 1), otherwise.
Thus, buy(sell)-signals are triggered upon outbraking the
upper(lower) trading corridor affinely constructed based upon
the last two local maxima(minima).
Note that for the final trading rules of both TR-Inside and
TR-Outside, we only employ the last two local maxima and
(k)
minima, eventhough we derived smax in (10) (and similarly
(k)
(k)
(k)
tmax , smin and tmin ) for all k = 1, . . . , K. This has the reason
that a partition according (9) can be constructed either with
uniform spacings starting at at time t − Twin and partitioning
proceeding forward in time until t, or, alternatively, starting
at time t and partitioning going backward in time. Interestingly, when testing both methods we found the former
method to almost always perform better. We attribute this to
the residual δ that typically enlarges the final time-window
for k = K.
C. Historical optimal-based causal controller
Given past stock prices historical optimal (HistOpt) trading
trajectory can be reconstructed a posteriori with hindsight.
This can be done efficiently by graph generation and evaluation. A valid question is whether such optimal trajectories
generated up until time t + 1 with predicted ŝ(t + 1) as final
stock price can also be exploited for real-time (RT) stock
trading. Thus, trading signals are
(
˜
˜ − τ ) = J(t),
˜
J(t),
if J(t
∀τ = 1, . . . , THO − 1,
J(t) =
0,
otherwise,
(11)
˜ denotes the historical optimal trading trajectory
where J(t)
at time t. Tuning parameter THO determines the number of
past consecutive identical trading signals necessary to trigger
a buy/sell signal. We refer to this controller as HistOpt-RT.
D. Final remark
All trading controllers discussed within this section were
designed to rely on an estimated step-ahead stock price
ŝ(t + 1). Naturally, price data can arbitrarily be shifted
by one sampling interval to the past, thereby making the
controllers independent of ŝ(t + 1) while not prohibiting
their applicability. The formulation using ŝ(t + 1) admits
for exploitation of any potential good estimate of step-ahead
stock prices. Furthermore, it allows a better comparison with
the SMPC-based methods, whose core is an estimate of onestep ahead stock prices. As will be shown in the subsequent
section, this is crucial for both performance and robustness.
V. S IMULATION
EXPERIMENTS
TABLE I. Overfitting illustrated by means of MA-Cross when optimizing
parameters on past data. The parameters are (pMA,l , pMA,s ). They are
optimized on the training data (November 28, 2014 until November 26,
2015) and then validated for November 27, 2015 until November 25, 2016.
The corresponding return performances in percent are denoted by ftrain and
fval . The 30 DAX components are ordered according to increasing absolute
performance for the 2-year time period. Buy-and-hold performances fval,B&H
are given for comparison. The final row indicates average returns.
Stock
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
Avg.
Parameters
(1, 95)
(20, 74)
(9, 38)
(7, 32)
(21, 37)
(18, 11)
(7, 14)
(20, 29)
(9, 18)
(17, 11)
(17, 10)
(14, 26)
(18, 25)
(6, 26)
(20, 27)
(4, 21)
(17, 27)
(19, 156)
(16, 29)
(12, 39)
(5, 22)
(12, 20)
(11, 21)
(4, 9)
(14, 27)
(17, 10)
(9, 23)
(21, 16)
(17, 43)
(13, 19)
–
ftrain,MA
0.0
2.2
14.4
11.4
35.4
15.6
10.5
54.8
24.4
20.5
35.9
13.7
41.0
29.8
33.9
26.8
27.0
6.9
34.1
11.0
40.8
12.7
27.9
29.0
38.2
21.2
32.3
37.5
46.5
48.8
26.1
fval,MA
61.7
-29.4
-18.2
-28.1
-19.1
-31.1
-9.3
-20.5
-30.7
-16.5
-30.5
-9.9
-4.4
-0.6
-27.2
-30.5
-17.3
12.1
1.8
1.3
7.6
-8.4
-24.8
-29.2
1.6
-12.2
-29.6
-9.5
-8.4
15.9
-11.4
fval,B&H
7.0
-25.5
-37.3
-39.5
-5.8
-30.4
-7.9
-20.6
-35.4
-24.4
-28.7
4.7
-8.7
-9.8
-16.2
-12.8
-11.0
4.9
1.7
8.3
-2.3
2.5
-10.9
-6.0
10.5
2.9
7.3
-4.4
13.1
50.5
-7.5
TABLE II. Parameter selections for final simulation experiments.
Controller
QP-E+
SMPC-M100
SMPC-DH
MA-Cross
MA-Sign
TR-Inside
TR-Outside
HistOpt-RT
Parameters
(M, α)
(M, α)
M
(pMA,l , pMA,s )
(TMA , pMA )
(Twin , pTR , ǫTR )
(Twin , pTR , ǫTR )
THO
Values
(1, 1)
(100, 10)
100
(50, 1)
(10, 100)
(261, 100, 0.01)
(261, 20, 0.03)
1
For simulation experiments, we employ the stock prices
of the 30 components of the German stock market index
DAX between November 28, 2014 and November 25, 2016.
For closed-loop trading we only considered the past year
and initialized t = 0 on November 27, 2015. Nevertheless,
previous price data was still relevant for the generation
TABLE III. Results for the one-year trading period between November
27, 2015 and November 25, 2016. The average total number of trades per
year is denoted by N̄tr . The minimum number of trading days between
any two trades is tmin . The average, minimum and maximum performance
(all measured in %) are f¯, fmin and fmax . The total percentage of positive
returns is Fpos .
1. Perfect: ŝ(t + 1) = s(t + 1)
N̄tr
Controller
tmin
fmin
fmax
Fpos
f¯
QP-E+
56
1
150.7
40.8
534.5
100
SMPC-M100
28
1
45.7
0
220.9
100
SMPC-DH
13
1
8.2
0
68.7
100
MA-Cross
21
1
27.7
-3.9
95.0
93.3
MA-Sign
4
26
-8.7
-32.2
50.5
23.3
TR-Inside
2
43
-2.5
-27.2
15.3
60.0
TR-Outside
5
19
7.6
-31.4
55.5
70.0
HistOpt-RT
75
1
133.3
24.1
506.7
100
2. Indifferent: ŝ(t + 1) = s(t)
f¯
N̄tr
Controller
tmin
fmin
fmax
Fpos
QP-E+
0
0
0
0
0
100
SMPC-M100
0
0
0
0
0
100
SMPC-DH
2
1
-4
-43.6
2.0
76.7
MA-Cross
1
10
0.7
-19.7
28.9
86.7
MA-Sign
4
26
-8.7
-32.2
50.5
23.3
TR-Inside
2
42
-0.5
-26.5
18.9
60.0
TR-Outside
5
21
-4.2
-42.5
50.8
36.7
HistOpt-RT
75
1
-52.5
-73.8
-30.5
0
Pt
3. Random: ŝ(t + 1) = s(t) + η(t) 1t
τ =0 |s(τ ) − s(τ − 1)|
N̄tr
f¯
Controller
tmin
fmin
fmax
Fpos
QP-E+
57
1
-40.1
-60.7
0.4
3.3
SMPC-M100
20
1
-15.6
-55.3
0.0
26.7
SMPC-DH
25
1
-22.4
-54.8
0.0
23.3
MA-Cross
9
11
-5.0
-36.5
68.6
23.3
MA-Sign
4
26
-8.7
-32.3
50.5
23.3
TR-Inside
2
42
-0.2
-26.4
26.4
56.7
TR-Outside
7
11
-6.7
-39.1
29.7
30.0
HistOpt-RT
116.4
1
-67.2
-80.3
-34.7
0
4. Correct Sign: ŝ(t + 1) = s(t) + 10ξ(t)sign(s(t + 1) − s(t))
Controller
tmin
fmin
fmax
Fpos
N̄tr
f¯
QP-E+
117
1
74.7
-25.9
400.4
86.7
SMPC-M100
119
1
57.8
-29.6
338.6
90.0
SMPC-DH
65
1
3
-34.1
74.5
60.0
MA-Cross
21
1
22.0
-17.1
171.8
73.3
MA-Sign
4
26
-8.7
-32.3
50.5
23.3
TR-Inside
6
21
-10.9
-32.2
16.2
23.3
TR-Outside
51
1
28.3
-32.7
188.1
70.0
HistOpt-RT
125
1
69.3
-28.9
380.7
83.3
5. Wrong Sign: ŝ(t + 1) = s(t) − 10ξ(t)sign(s(t + 1) − s(t))
f¯
N̄tr
Controller
tmin
fmin
fmax
Fpos
QP-E+
117.4
1
-93.5
-98.7
-87.6
0
SMPC-M100
119
1
-93.4
-98.6
-88.5
0
SMPC-DH
120
1
-93.7
-98.7
-88.8
0
MA-Cross
11
3
-16.9
-40.8
2.7
3.3
MA-Sign
4
26
-8.7
-32.3
50.5
23.3
TR-Inside
5
28
-3.1
-28.7
26.1
40.0
TR-Outside
56
1
-64.3
-97.1
-19.3
0
HistOpt-RT
136
1
-94.9
-98.7
-91.4
0
Global Optimum (Trading w/ Hindsight)/Buy-and-Hold
N̄tr
Controller
tmin
fmin
fmax
Fpos
f¯
HistOpt
42
1
192.6
64.7
609.5
100
Buy-and-Hold
1
0
-7.5
-39.5
50.5
36.7
of measures such as moving averages at t = 0. All data
was drawn from finance.yahoo.com, see Figure 3 for
visualization. Throughout, proportional transaction costs of
1% are assumed. All simulations were run on a laptop
running Ubuntu 14.04 equipped with an Intel Core i7 CPU
@2.80GHz×8, 15.6GB of memory and using Python 2.7.
A. Closed-loop trading results
Each stock is traded separately and the portfolio with
transition dynamics according
to Section II is initialized with
Z(0) = 0 M0 0 M0 where M0 = 100000e . We
compare eight different controllers and five different methods
that we use for the prediction of ŝ(t+1). In addition we state
the results for a buy-and-hold strategy (investing maximally
into the stock at t = 0 and consequently holding the investments throughout), and for the global optimal trading result
(trading with hindsight). We assume proportional transaction
costs ǫ = 0.01 identical for both buying and selling of stocks.
0
100 with T the
Performance is defined by f = W (TM)−M
0
final trading date.
Regarding parameter selections for the genetic algorithms,
we tested three settings. First, we optimized parameters on
training data (November 28, 2014 until November 26, 2015)
and then validated for November 27, 2015 until November
25, 2016. Thus, for each stock and for each controller
individual parameters were selected . Second, we recursively
updated parameters. Thus, every 100 days (we also tested 20
and 50 days) we recomputed parameters optimized on past
data of one year at that time. Third, we arbitrarily chose
a fixed parameter set for each controller and used this for
the trading of all 30 stocks. For both the first and the second
approach strong overfitting could be observed, see Table I for
illustration. We therefore opted for fixed parameter selections
for the trading of all stocks. More conservative parameter
selections, such as, e.g., larger MA-windows pMA,l , performend on average better. For HistOpt-RT we intentionally
chose THO = 1, which is the most aggressive but least robust
choice as outlined in the following. The parameter selections
employed for final simulation experiments are summarized
in Table II.
Closed-loop trading results are summarized in Table III.
Because of the importance of step-ahead predictions and for
robustness considerations, we compare five different versions
for ŝ(t + 1). Ideally (but unrealistically in general), s(t + 1)
is estimated perfectly, i.e., ŝ(t + 1) = s(t + 1). This case
is noncausal. Nevertheless, it serves as an important benchmark. Case 4 and 5 (“correct” and “wrong sign prediction”)
are likewise noncausal since s(t + 1) is not known at time
t. Guiding notion for their introduction was to analyze
influence of correct trend prediction (up or down) for the
price one time-step ahead but without knowledge of exact
level of price rise or fall. We therefore add multiplicative
time-varying perturbation 10ξ(t) with ξ(t) ∼ U(0, 1) uniformly distributed. Case 2 (“indifferent”) uses the current
stock price as the estimate for the ext time-step ahead. Case
3 (“random”) randomly perturbs s(t) as the estimate for
ŝ(t + 1), whereby η(t) ∼ N (0, 1) normally distributed.
Results are discussed in the next section. Importantly, we
remark that only 36.7% of the 30 DAX components rose,
i.e., s(T ) > s(0), for the one-year trading period considered
between November 27, 2015 and November 25, 2016.
B. Discussion
Several observations can be made from Table III. Let us
first discuss the results for SMPC-based stock trading and
HistOpt-RT. For the ideal case of perfect s(t+1) knowledge
truely excellent results can be obtained. Both for QP-E+
and HistOpt-RT. Performances are even in range with the
global optimum (HistOpt) despite only one-step ahead price
knowledge. By reduction of α from 10 to 1, quasi identical
performance to QP-E+ is achieved by SMPC-M100, i.e.,
(f¯, fmin , fmax ) = (139.4, 42.7, 533.7). We selected α = 10 to
illustrate its role in adding robustness. This becomes apparent
for the random ŝ(t+1) prediction method (case 3): while f¯ =
−40.1% for QP-E+, it is f¯ = −15.6% for SMPC-M100,
and additionally yielding 26.7% positive returns overall. For
the indifferent prediction ŝ(t + 1) = s(t), QP-E+ and
SMPC-M100 never enter a trade. This was expected. Both
optimization problem formulations essentially rely on the
mean difference between s(t) and ŝ(t+1). Characteristically
they do not consider any past data points except the current
price s(t). This is in contrast to HistOpt-RT where all
available past data up until t is searched for the optimal traing
trajectory. SMPC-DH was found to not be competitive with
the other two SMPC-based controllers (QP-E+ and SMPCM100), neither with respect to performance nor robustness.
Nevertheless, its framework is favorable in that more (and
better) heuristics can easily be incorporated by adjusting
reference scenarios pj (t + 1), j = 1, . . . , M .
Of great relevance to SMPC-based trading strategies and
HistOpt-RT are the experiments for correct and wrong sign
predictions (case 4 and 5). Importantly, they indicate that
perfect one-step ahead sign predictions of price changes s(t+
1)− s(t) are sufficient for excellent results. Thus, the precise
level of increase or decrease in stock prices is not necessarily
required. For illustration, consider the average gain of 74.7%
per stock for QP-E+ despite the fact that only 36.7% of
all 30 DAX components actually rose during the past year
and moreover on average yielding minus 7.5% (see the Buyand-Hold strategy). Even more important are the results for
case 5, i.e., when at every trading instant wrongly predicting
the direction of change in stock prices. For all SMPC-based
trading methods and HistOpt-RT, after just one year, at least
93.5% of all initial wealth is lost.
Finally, note that for the SMPC-based methods and
HistOpt-RT the minimum time-span between any two trades
is always 1 day (except for case 2 when there are no trades
at all). Furthermore, the average number of trades per year,
N̄tr , is considerably larger in comparison to the genetic
algorithms.
Throughout experiments, more robust performance could
be observed for the genetic algorithms. For the given parameter choices, MA-Cross appeared to be best suited to
exploit potential knowledge of future stock prices (case 1).
Encouraging are the returns for MA-Cross and TR-Inside
for the causal prediction case, i.e., ŝ(t + 1) = s(t). Despite
the fact that only 36.7% of all 30 DAX components were
actually rising since last year, the two controllers yielded
positive returns for 86.7% and 60%, respectively. For random
price-ahead predictions (case 3), the performance of all
four MA- and TR-based controllers was comparable to the
Buy-and-Hold method; with TR-Inside performing best on
average and with respect to worst-case losses.
Before concluding, let us remark some realistic success
ratios reported in the literature for correct sign predictions
of step-ahead price difference s(t + 1) − s(t). In [7], support
vector machines (SVM) in combination with 12 technical
indicators (such as Williams %R, stochastic %K, disparity,
etc.) are used to predict the direction of change in the daily
Korea composite stock price index (KOSPI). For validation
data and their best tuning parameter choices, they report a
prediction performance between 50.1% and 57.8%. The same
author mentioned similar results in earlier work [8].
VI. C ONCLUSION
For stock trading, the general class of genetic algorithms
appears more suitable than methods based on stochastic
model predictive control. The former class is signficantly
more robust. A SMPC-approach is justifiable only for consistently perfect prediction of direction of price changes. This
is not achievable in practice. The relations and differences
between using SMPC for dynamic hedging and stock trading
were discussed.
Findings motivate the following:
1) A detailed analysis of scenarios when MA- and TRbased algorithms fail and succeed, respectively.
2) An artificial and automated generation of genetic trading algorithms to further improve performance and
robustness [9].
3) The usage of options for their predictable worst-case
loss [10].
Subject of future research are the application of genetic
trading algorithms to both multi-asset portfolio optimization
and dynamic option hedging.
R EFERENCES
[1] B. R. Barmish and J. A. Primbs, “Stock trading via feedback control,”
Encyclopedia of Systems and Control, pp. 1357–1364, 2015.
[2] B. R. Barmish, S. Condie, D. Materassi, J. A. Primbs, and S. Warnick,
“On nasdaq order book dynamics: New problems for the control field,”
in IEEE American Control Conference, pp. 5671–5672, American
Automatic Control Council (AACC), 2016.
[3] H. Markowitz, “Portfolio selection,” The Journal of Finance, vol. 7,
no. 1, pp. 77–91, 1952.
[4] A. Bemporad, L. Puglia, and T. Gabbriellini, “A stochastic model predictive control approach to dynamic option hedging with transaction
costs,” in IEEE American Control Conference, pp. 3862–3867, 2011.
[5] M. S. Lobo, M. Fazel, and S. Boyd, “Portfolio optimization with linear
and fixed transaction costs,” Annals of Operations Research, vol. 152,
no. 1, pp. 341–365, 2007.
[6] F. Black and M. Scholes, “Pricing of options and corporate liabilities,”
Journal of Political Economy, vol. 81, no. 3, pp. 637–654, 1973.
[7] K.-j. Kim, “Financial time series forecasting using support vector
machines,” Neurocomputing, vol. 55, no. 1, pp. 307–319, 2003.
[8] K.-j. Kim and I. Han, “Genetic algorithms approach to feature discretization in artificial neural networks for the prediction of stock price
index,” Expert Systems with Applications, vol. 19, no. 2, pp. 125–132,
2000.
[9] F. Allen and R. Karjalainen, “Using genetic algorithms to find technical trading rules,” Journal of Financial Economics, vol. 51, no. 2,
pp. 245–271, 1999.
[10] M. Graf Plessen and A. Bemporad, “Parallel investments in multiple
call and put options for the tracking of desired profit profiles,” in IEEE
American Control Conference, pp. 1091–1096, 2017.
| 5cs.CE
|
arXiv:1602.03287v2 [math.SG] 15 Feb 2016
ON THE AUTONOMOUS NORM ON THE GROUP OF
HAMILTONIAN DIFFEOMORPHISMS OF THE TORUS
MICHAEL BRANDENBURSKY, JAREK KĘDRA, AND EGOR SHELUKHIN
Abstract. We prove that the autonomous norm on the group
of Hamiltonian diffeomorphisms of the two-dimensional torus is
unbounded. We provide explicit examples of Hamiltonian diffeomorphisms with arbitrarily large autonomous norm. For the proofs
we construct quasimorphisms on Ham(T2 ) and some of them are
Calabi.
1. Introduction
2. Preliminaries
3. The injectivity theorem
4. Quasimorphisms with vanishing properties
5. Vanishing on autonomous diffeomorphisms
6. Further results
References
1
6
11
15
17
23
26
1. Introduction
Let M be a smooth manifold and let X : M → T M be a compactly
supported vector field with the flow ΨX : R → Diff(M). The timeone map ΨX (1) of the flow is called the autonomous diffeomorphism
associated with the vector field X. The subset Aut(M) ⊂ Diff 0 (M)
of autonomous diffeomorphisms is conjugation invariant and, since the
group of diffeomorphisms isotopic to the identity is simple, it generates
Diff 0 (M). In other words, a compactly supported diffeomorphism of
M isotopic to the identity is a finite product of autonomous ones. One
may ask for a minimal decomposition and this question leads to the
concept of the autonomous norm which is defined by
kf kAut := min{n ∈ N | f = a1 · · · an , ai ∈ Aut(M)}.
It is the word norm associated with the generating set Aut(M). Since
this set is conjugation invariant, so is the autonomous norm. It follows
2000 Mathematics Subject Classification. Primary 53; Secondary 57.
Key words and phrases. Hamiltonian diffeomorphisms, autonomous norm, quasimorphisms, braid groups.
1
2
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
from the work of Burago-Ivanov-Polterovich [5] and Tsuboi [18, 19]
that for many manifolds all conjugation invariant norms on Diff 0 (M)
are bounded. Hence the autonomous norm is bounded in those cases.
The situation is different for the groups of area preserving diffeomorphisms of surfaces. For example, the autonomous norm on the group
Diff 0 (D2 , area) of compactly supported area preserving diffeomorphisms
of the open disc is unbounded [3]. The same is true for the group
Ham(Σ) of Hamiltonian diffeomorphisms of closed oriented surfaces
different from the torus [8, 2, 4]. The present paper deals with the
remaining case of the torus:
Theorem 1.1. The autonomous norm on the group Ham(T2 ) of Hamiltonian diffeomorphism of the torus is unbounded.
One way to prove unboundedness of a conjugation invariant norm on a
group G is to construct an unbounded quasimorphism ψ : G → R which
is Lipschitz with respect to this norm. If such a norm is a word norm
then it suffices to construct a quasimorphism which is bounded on the
generating set which implies that it is Lipschitz. If G = Diff 0 (M, vol)
is the group of volume preserving diffeomorphisms of a manifold M
then nontrivial quasimorphisms on G can be obtained from nontrivial
quasimorphisms on the fundamental group of M as follows.
Let z ∈ M be the basepoint and let g be an auxiliary Riemannian
metric on M. For every point x ∈ M chose a path γx : [0, 1] → M from
z to x by choosing a measurable section of the map π : P → M, where
P = {γ : [0, 1] → M | γ(0) = z, γ(1) = x and γ is a geodesic of g}.
Let f ∈ Diff 0 (M, vol) and let {ft } be an isotopy from the identity
to f . For every x ∈ M the isotopy {ft } defines a loop based at x by
γ(f, x) = γx ft (x)γf (x) , where the bar denotes the path in the reverse
direction. This loop is well defined up to homotopy of loops based
at z provided that evaluating loops of diffeomorphisms of M at the
basepoint produces homotopically trivial loops in M. This holds, for
example, if the center of the fundamental group of M is trivial or if
{ft } is a Hamiltonian isotopy in a symplectic manifold.
Let ψ : π1 (M, z) → R be a quasimorphism and let f ∈ Diff 0 (M, vol) be
a compactly supported diffeomorphism isotopic to the identity. Then,
given that the volume of M is finite, the map Ψ : Diff 0 (M, vol) → R
defined by
Z
Ψ(f ) =
ψ(γ(f, x))dx
M
ON THE AUTONOMOUS NORM ON Ham(T2 )
3
is a well defined quasimorphism. This construction and the argument
are due to Polterovich [13]. Notice that the construction can be performed for an action G → Diff 0 (M, vol) of a group G on M. For example, if M is simply connected then G = Diff 0 (M) can act on another
manifold which is not simply connected. Concretely, if Σ is a surface
then Diff 0 (Σ, area) acts on the configuration space M = Xn (Σ). The
fundamental group of this configuration space is (by definition) the
pure braid group on n-strings on the surface Σ. Geometrically, this
construction generalizes the above one in the sense that an isotopy and
a configuration of points defines a pure braid γ(f, x1 , . . . , xn ) rather
than a single loop (up to homotopy). We provide more details in Section 2.4. Historically the braid approach was the first original idea due
to Gambaudo and Ghys [8] applied to diffeomorphisms of the disc and
the sphere. It was later generalized by the first named author to other
surfaces [2]. To sum up, the construction gives a linear map
G : Q(Pn (Σ)) → Q(Ham(Σ)),
from the space of homogeneous quasimorphisms on the pure braid
group to the space of homogeneous quasimorphisms on the group of
Hamiltonian diffeomorphisms of the surface.
There are two main problems in proving the unboundedness of the
autonomous norm. The first, which is a general one, is to show that the
above construction yields nontrivial quasimorphisms. The second is to
show that among these nontrivial quasimorphisms there are ones which
are bounded on the set of autonomous diffeomorphisms. These are the
main objectives of the present paper as well as the earlier ones [2, 3, 4].
The solution of the first problem has two parts. The first one, which
is essentially the same for all surfaces, is the claim that the kernel of
the composition Q(Bn (Σ)) → Q(Pn (Σ)) → Q(Ham(Σ)) consists of
homomorphisms. The idea of the proof is due to Ishida who did it in
the case of the disc and the sphere [9] and his argument was generalized to all surfaces in [3]. The second part is to construct nontrivial
quasimorphisms on the full braid group. Here, the solution depends on
the genus.
The problem of identifying quasimorphisms on braid groups which yield
quasimorphisms vanishing on autonomous diffeomorphisms is the main
problem in all the cases and, again, solutions depend on the genus.
In what follows we provide a proof of Theorem 1.1 by reducing the
argument to several results which are then proved in the rest of the
paper.
4
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Proof of Theorem 1.1. The structure of the proof is presented in
the following composition of linear maps.
Q(F2 ; Z/2 × Z/2)
/
Q(F2 )
π∗ /
Q(P2 (T2 ))
G
/
Q(Ham(T2 ))
Here, Q(G) denotes a space of homogeneous quasimorphisms on a
group G and Q(F2 ; Z/2 × Z/2) ⊂ Q(F2 ) is the subspace of quasimorphisms invariant under the action generated by inverting generators.
• The construction of Gambaudo and Ghys (Section 2.4), provides a
linear map G : Q(Pn (T2 )) → Q(Ham(T2 )) from the space of homogeneous quasimorphisms of the pure braid group to the space of
homogeneous quasimorphisms of the group Ham(T2 ) of Hamiltonian diffeomorphisms of the torus. This map has a nontrivial kernel
in general and the goal is to construct a suitable quasimorphism ψ
on the pure braid group such that its image G(ψ) is a nontrivial
quasimorphism bounded on the set of autonomous diffeomorphisms.
• In our proof, we specify the braid group to two strings. There is
an isomorphism P2 (T2 ) ∼
= F2 × Z2 (Lemma 2.6). We construct a
suitable quasimorphism on the pure braid group by constructing a
quasimorphism ψ : F2 → R on the free group and composing it with
the projection π : P2 (T2 ) → F2 . The free group here is the quotient
of the braid group by its center and hence the projection is canonical
(i.e. every automorphism of P2 (T2 ) descends to an automorphism
of the quotient F2 ). Thus the refined goal is to construct a quasimorphism ψ : F2 → R such that the image G(ψ ◦ π) is nontrivial and
bounded on the set of autonomous elements.
• Let F2 = ha, bi and let σa , σb ∈ Aut(F2 ) be automorphisms defined by
σa (a) = a−1 , σa (b) = b, σb (a) = a and σb (b) = b−1 . They generate an
action of Z/2 ×Z/2 on F2 . Let Q(F2 ; Z/2 ×Z/2) denote the space of
homogeneous quasimorphisms which are invariant under this action.
The composition
π∗
G
Q(F2 ; Z/2 × Z/2) → Q(P2 (T2 )) → Q(Ham(T2 ))
is injective (Proposition 3.1). Moreover, the space Q(F2 ; Z/2 × Z/2)
is infinite dimensional (Proposition 4.3). We obtain this way an infinite dimensional space of quasimorphisms on the group Ham(T2 ) and
the next step is to prove that it contains quasimorphisms bounded
on autonomous diffeomorphisms.
• We prove in Lemma 5.1 that if a quasimorphism ψ ∈ Q(F2 ) vanishes
on primitive elements and on the commutator [a, b] of the generators
ON THE AUTONOMOUS NORM ON Ham(T2 )
5
then the quasimorphism G(π ∗ ψ) vanishes on autonomous elements.
This reduces our task to showing that the space Q(F2 ; Z/2 × Z/2)
contains quasimorphisms vanishing on primitive elements and the
commutator of the generators. Observe that the second condition
is automatic. Indeed, if ψ ∈ Q(F2 ; Z/2 × Z/2) then we have the
following computation in which we use invariance under σa and homogeneity (which implies conjugation invariance).
ψ[a, b] = ψ aba−1 b−1
= ψ σa a−1 bab−1
= ψ a−1 bab−1
= ψ a−1 bab−1 a−1 a
= ψ[b, a] = −ψ[a, b].
It follows that ψ[a, b] = 0.
• Let σ = σa ◦ σb ∈ Aut(F2 ) be the automorphism acting on a word by
inverting all its letters. If a quasimorphism ψ ∈ Q(F2 ) is invariant
under the action of Z/2 × Z/2 then it is invariant under the action
generated by σ. This, in turn, implies that ψ vanishes on palindromes
(see the proof of Corollary 5.5). It has been observed by Bardakov,
Shpilrain and Tolstykh [1] that a primitive element of the free group
F2 is a product of two palindromes. Hence every quasimorphism
from Q(F2 ; Z/2 × Z/2) vanishes on primitive elements. This finishes
the proof.
Remark 1.2. In the proof Proposition 3.1 which claims the injectivity
of the homomorphism Q(F2 ; Z/2 × Z/2) → Q(Ham(T2 )), we provide
explicit examples of Hamiltonian diffeomorphisms on which quasimorphisms of the form G(π ∗ ψ) evaluate nontrivially (Example 3.3). Such
examples are quite standard and have been considered, for example,
by Khanevsky [10] and Polterovich-Shelukhin [14].
Remark 1.3. Another side result is concerned with the Calabi property
and continuity of the quasimorphisms we construct in the paper, see
Section 6.6 for a discussion of the Calabi property, and a new example of a Calabi-type quasimorphism. More precisely, if ψ ∈ Q(F2 ) is
a nontrivial quasimorphism vanishing on palindromes then G(π ∗ ψ) is
nontrivial (see Corollary 6.4 and the discussion that follows it). Moreover, as proven in [15] and [7, Proposition 4.1],
• if ψ[a, b] 6= 0 then G(π ∗ ψ) has the Calabi property;
• if ψ[a, b] = 0 then G(π ∗ ψ) is continuous in C 0 -topology.
6
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
It follows that the quasimorphisms constructed in the proof of Proposition 4.3 are nontrivial and C 0 -continuous. Since by Lemma 5.1 such
quasimorphisms vanish on autonomous diffeomorphisms, we can strengthen Theorem 1.1 to the following statement:
Theorem 1.4. The group Ham(T2 ) equipped with the word norm associated with the C 0 -closure of the set of autonomous diffeomorphisms
has infinite diameter.
2. Preliminaries
In this section we provide necessary definitions, review in detail the
construction of Gambaudo-Ghys and state some known results which
we need for the proof.
Definition 2.1. Let G be a group. A function k · k : G → [0, ∞) is
called a conjugation invariant norm on G if it satisfies the following
conditions:
(1) kf k = 0 if and only if f = 1,
(2) kf −1 k = kf k,
(3) kf gk ≤ kf k + kgk,
(4) kgf g −1k = kf k.
Definition 2.2. A function ψ : G → R is called a quasimorphism if
there exist Dψ ≥ 0 such that the inequality
|ψ(f ) − ψ(f g) + ψ(g)| ≤ Dψ
holds for all f, g ∈ G. A quasimorphism ψ is called homogeneous if
ψ (f n ) = nψ(f ),
for all f ∈ G and n ∈ Z. The space of all homogeneous quasimorphisms
on a group G is denoted by Q(G). Let S ⊂ G. We denote by Q(G; S)
the space of homogeneous quasimorphism which vanish on S.
If ψ : G → R is a quasimorphism then its homogenization ψ : G →
R, the unique homogeneous quasimorphism that differs from ψ by a
bounded function, satisfies
ψ(g n )
.
ψ(g) = lim
n→∞
n
Moreover, the homogenization behaves well with respect to group actions.
ON THE AUTONOMOUS NORM ON Ham(T2 )
7
Lemma 2.3. Let ψ : G → R be a quasimorphism and let α : H →
Aut(G) be an action of a group H on the group G. If ψ is invariant
under the action α then so is its homogenization.
Proof. The invariance of ψ under the action α means that ψ(α(h)(g)) =
ψ(g) for every h ∈ H and every g ∈ G. The statement is a consequence
of the following straightforward computation.
ψ ((α(h)(g))n)
n→∞
n
ψ (α(h)(g n ))
= lim
n→∞
n
n
ψ (g )
= lim
n→∞
n
= ψ(g).
ψ(α(h)(g)) = lim
2.4. The Gambaudo-Ghys construction. Let Σ be an oriented surface and let Xn (Σ) denote the space of configurations of ordered ntuples of points in Σ. Its quotient by the n-th symmetric group is the
space of unordered configurations and it is denoted by Cn (Σ). The
fundamental groups
Pn (Σ) := π1 (Xn (Σ))
Bn (Σ) := π1 (Cn (Σ))
are called the pure braid group and the (full) braid group of the surface
Σ respectively.
Let (z1 , . . . , zn ) ∈ Xn (Σ) be an n-tuple of distinct points which is the
basepoint in the configuration space. We fix an auxiliary Riemannian
metric on Σ and for every point x ∈ Σ we fix a geodesic γi,x of minimal
length from zi to x. We denote by γ i,x the reversed geodesic.
Let ht ∈ Ham(Σ) be a Hamiltonian isotopy from the identity to a
diffeomorphism h = h1 ∈ Ham(Σ) and let (x1 , . . . , xn ) ∈ Xn (Σ) be a
point in the configuration space. Let γ(h, x1 , . . . , xn ) ∈ Pn (Σ) be the
braid represented by the loop [0, 3] → Xn (Σ) defined by
for 0 ≤ s ≤ 1
(γ1,x1 (s), . . . , γn,xn (s))
s 7→ (hs−1 (x1 ), . . . , hs−1(xn ))
for 1 ≤ s ≤ 2
(γ
for 2 ≤ s ≤ 3.
1,h(x1 ) (s − 2), . . . , γ n,h(xn ) (s − 2))
8
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
This braid is only well defined on a set of points (x1 , . . . , xn ) of full
measure.
Let ψ : Pn (Σ) → R be a homogeneous quasimorphism and let
G : Q(Pn (Σ)) → Q(Ham(Σ))
be defined by
1
G(ψ)(h) := lim
p→∞ p
Z
ψ(γ(hp , x1 , . . . , xn )) dx1 ∧ · · · ∧ dxn .
Xn (Σ)
The fact that the value G(ψ) is a homogeneous quasimorphism, when
ψ is a signature quasimorphism, was first proved by Gambaudo and
Ghys [8] for the case of the disc and the sphere and later extended to
all ψ and all surfaces by Brandenbursky [3]. The map G is linear and,
in general, has a nontrivial kernel. In Section 3 we prove that G is
injective on a certain subspace of Q(P2 (T2 )).
2.5. Braid groups on two strings. We use the following presentations of the braid groups the free group:
B2 (T2 ) = ha1 , a2 , b1 , b2 , σ | Relations i
P2 (T2 ) = ha1 , a2 , b1 , b2 , σ 2 | Relations i
F2 = ha, bi.
We omit the relations because they are quite complicated and we don’t
need them in our discussions. They can be found in [17, Theorem 1.3
and 1.4]. The generators are presented in Figure 2.1, which should
be understood as follows. For example, the generator a1 is a braid in
a1
z1
z2
a2
b1
b2
z1
z1
z1
z2
z2
z2
σ
z1
z2
Figure 2.1. Generators of the braid group B2 (T2 ).
which the first basepoint traces the horizontal loop going once around
the torus and the second basepoint remains still.
Lemma 2.6. The map Φ : X2 (T2 ) → T2 \ {0} × T2 defined by
Φ(x, y) := (x − y, y)
ON THE AUTONOMOUS NORM ON Ham(T2 )
9
is a diffeomorphism. It induces an isomorphism
Φ∗ : P2 (T) → F2 × Z2 ,
which on the generators is given by
a1 7→ (a, (0, 0))
a2 7→ (a−1 , (1, 0))
b1 7→ (b, (0, 0))
−1
σ2
7→ ([a, b], (0, 0))
b2 7→ (b , (0, 1)).
Proof. The fact that Φ is a diffeomorphism is straightforward. Let
π : P2 (T2 ) → F2 denote the projection onto the free factor. The following figures describe the value of π on a generator.
x = x0
xt
x−y
y
xt − y
Figure 2.2. The image π(b1 ) = b ∈ F2 .
x
yt
x−y
y
x − yt
Figure 2.3. The image π(b2 ) = b−1 ∈ F2 .
The left hand side of the figure represents the image of the generator b1
(blue) with respect to the projection onto the torus (the black square).
The generator b1 moves the point x along the meridian of the torus and
keeps the point y fixed. The generator b2 keeps the point x fixed and
moves the point y along the meridian of the torus. The right hand sides
of the figures present the free part of Φ∗ (bi ) as loops on the punctured
torus. The abelian parts are straightforward to see. The values on the
generators ai are computed analogously.
It follows from the above proposition that the quotient of the pure
braid group P2 (T2 ) by its center is isomorphic to the free group F2 .
10
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Let π : P2 (T2 ) → F2 denote the projection. It induces the linear map
π ∗ : Q(F2 ) → Q(P2 (T2 )).
In the second part of the paper we will need quasimorphisms on the
full braid group. In what follows we identify those quasimorphisms ψ
on the free groups such that π ∗ ψ extends to the full braid group.
Definition 2.7. A word w in F2 = ha, bi is called a palindrome if w is
equal to itself read from right to left. Let PAL ⊂ F2 denote the set of
all palindromes.
Proposition 2.8. A quasimorphism ψ ∈ Q(F2 ) vanishes on palindromes if and only if the quasimorphism π ∗ ψ extends to B2 (T2 ). In
particular, we get a linear map
Q(F2 , PAL) → Q(B2 (T2 )).
Proof. The pure braid group is a normal subgroup of finite index in
the full braid group. According to [11, Lemma 4.2], a homogeneous
quasimorphism ψ : H → R on a finite index normal subgroup H ⊳ G
extends to the group G if and only if for every h ∈ H and every g ∈ G
we have that ψ(ghg −1) = ψ(h).
It follows that a quasimorphism on the pure braid group extends if and
only if it is invariant under the automorphism defined by the conjugation by σ. Since F2 is the quotient of the pure braid group by the
center, the conjugation by σ descends to an automorphism of the free
group. By abuse of notation we denote it by σ ∈ Aut(F2 ). Observe
that σ is defined by specifying its values on generators as σ(a) = a−1
and σ(b) = b−1 .
We conclude that if ψ ∈ Q(F2 ) then the quasimorphism π ∗ ψ extends
to the full braid group if and only if ψ is invariant under σ. That is,
ψ(σ(g)) = ψ(g) for every g ∈ F2 .
Observe that σ(g) = g −1 if and only if g is a palindrome. In particular,
every element of the form σ(g)g −1 is a palindrome.
If ψ vanishes on palindromes then ψ(σ(g)g −1) = 0 for every g. The
following computation shows that ψ is invariant under σ. Let w ∈ F2
be any element.
1
|ψ(σ(w)) − ψ(w)| = |ψ(σ(w n )) − ψ(w n )|
n
Dψ
1
≤
|ψ(σ(w n )w −n )| + Dψ =
.
n
n
ON THE AUTONOMOUS NORM ON Ham(T2 )
11
Conversely, if ψ is invariant with respect to σ then if g ∈ F2 is a
palindrome we get that
ψ(g −1 ) = ψ(σ(g)) = ψ(g)
and by homogeneity we obtain that ψ(g) = 0.
3. The injectivity theorem
Let σa , σb ∈ Aut(F2 ) be automorphisms defined by
σa (a) = a−1 ,
σa (b) = b
σb (a) = a,
σb (b) = b−1 .
They generate an action of Z/2 × Z/2 on the free group F2 .
Proposition 3.1. Let Q(F2 ; Z/2 × Z/2) ⊂ Q(F2 ) be the space of homogeneous quasimorphisms which are invariant under the above action.
The composition
π∗
G
Q(F2 , Z/2 × Z/2) → Q(P2 (T2 )) → Q(Ham(T2 ))
is injective.
Lemma 3.2. Let h ∈ Ham(T2 ) be a Hamiltonian diffeomorphism and
let x, y ∈ T2 be two points. If h(x) = x and h(y) = y then
γ(hp , x, y) = γ(h, x, y)p
Proof. Immediate from the definition of γ(−, −, −).
Proof. Let ψ ∈ Q(F2 , Z/2 × Z/2). We shall prove that G(π ∗ ψ) 6= 0 in
Q(Ham(T2 )) by constructing explicit examples of Hamiltonian diffeomorphisms on which G(π ∗ ψ) evaluates nontrivially.
Let s ∈ (0, 14 ) and let 0 < ǫ < 10−3 s. Let Fs : [0, 1] → R be a smooth
function with the following properties:
(1) Fs (x) = 0 for x ∈ 0, 14 − s − ǫ ∪ 14 + s + ǫ, 1 ,
(2) Fs′ (x) = 1 for x ∈ 14 − s, 41 − ǫ ,
(3) Fs 14 − x = Fs 41 + x for x ∈ 0, 41 ,
see Figure 3.1.
12
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Figure 3.1. The function Fs .
Let H, V : T2 → R be defined by H(x, y) = Fs (1 − y) and V (x, y) =
Fs (x) respectively. Let ht , vt ∈ Ham(T2 ) be the corresponding Hamiltonian flows and let t0 > 0 be a real number
chosen so that the restric
tion of vt0 to the annulus 41 − s, 14 − ǫ ×S1 is the identity. Let h := ht0
and v := vt0 . The support of v is marked green and the support of h
is marked blue in Figure 3.2 below. The isotopy {vt } is supported between the green lines and the support of the isotopy {ht } is between
the blue lines.
S1 z
1
S4
S3
S2
z2
Figure 3.2. Diffeomorphisms h and v.
Define the following pairwise disjoint open subsets of the torus:
• S1 := 14 − s, 14 + s × 43 − s, 34 + s – the red square,
ON THE AUTONOMOUS NORM ON Ham(T2 )
13
• S2 := T2 \ 14 − s, 41 + s × [0, 1] ∪ [0, 1] × 34 − s, 43 + s – the
complement of the union of the blue and green annuli,
• S3 := 14 − s, 14 + s × [0, 1] \ S1 – the green annulus minus the
closure of the red square,
• S4 := [0, 1] × 14 − s, 41 + s \ S1 – the blue annulus minus the
closure of the red square.
Let ψ : F2 → R be a nontrivial homogeneous quasimorphism invariant
under the action of Z/2 × Z/2 and let w(a, b) ∈ F2 be an element such
that ψ(w(a, b)) > 0. Let g = w(v, h) ∈ Ham(T2 ). Now we investigate
the value of the integral
Z
ψ(π(γ(g, x, y))) dx ∧ dy
X2 (T2 )
by decomposing it into a sum of integral over subsets of the configuration space. First observe that the subset
[
[
X2 (Si ) ⊂ X2 (T2 )
Si × Sj ∪
i6=j
i
is open and dense so we have
Z
XZ
ψ(π(γ(g, x, y))) dx ∧ dy =
X2 (T2 )
+
ψ(π(γ(g, x, y))) dx ∧ dy
i6=j
Si ×Sj
i
X2 (Si )
XZ
ψ(π(γ(g, x, y))) dx ∧ dy
It will be useful to know the volumes of the sets Si × Sj . They are as
follows:
• vol(S1 × S1 ) = 16s4 ,
• vol(S1 × S2 ) = 4s2 (1 − 2s)2 = 4s2 − 16s3 + 16s4 ,
• vol(S1 × S3 ) = vol(S1 × S4 ) = 4s2 (1 − 2s)2s = 8s3 − 16s4 ,
• vol(S2 × S2 ) = (1 − 2s)2 = 1 − 4s + 4s2 ,
• vol(S2 × S3 ) = vol(S2 × S4 ) = (1 − 2s)2 (1 − 2s)2s = 2s − 12s2 +
24s3 − 16s4 ,
• vol(S3 × S3 ) = vol(S4 × S4 ) = vol(S3 × S4 ) = 4s2 (1 − 2s)2 = 4s2 −
16s3 + 16s4 .
The volumes are polynomials of s and what will be important below is
their degrees. Let us consider the element π(γ(w(v, h), x, y)) ∈ F2 for
various configurations:
14
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
• (x, y) ∈ S1 × S2 ; depending on the position of x in the red square we
obtain:
– (top left) π(γ(w(v, h), x, y)) = w(a, b),
– (top right) π(γ(w(v, h), x, y)) = w(a−1 , b),
– (bottom left) π(γ(w(v, h), x, y)) = w(a, b−1 ),
– (bottom right) π(γ(w(v, h), x, y)) = w(a−1 , b−1 ),
Since the quasimorphism ψ is invariant under inverting generators
we get that ψ(π(γ(g, x, y))) = ψ(w(a, b)) 6= 0. Thus
Z
ψ(π(γ(g, x, y))) dx ∧ dy = vol(S1 × S2 )ψ(w(a, b))
S1 ×S2
= (4s2 − 16s3 + 16s4 )ψ(w(a, b)).
• (x, y) ∈ S1 ×(S3 ∪S4 ); for fixed g the braid γ(g, x, y) can attain finitely
many values in this case and we let C1 := max |ψ(π(γ(g, x, y)))|.
Z
ψ(π(γ(g, x, y))) dx ∧ dy ≤ vol(S1 × S2 )C1
S1 ×S3
= (6s3 − 16s4 )C1 .
• (x, y) ∈ S2 × (S3 ∪ S4 ); we get that π(γ(g, x, y)) is either a power of
a or a power of b so ψ(π(γ(g, x, y))) = 0.
• (x, y) ∈ S3 × S4 ; here the situation is similar to the first case and the
value of π(γ(g, x, y)) depends on the positions x and y in the strips
and we obtain that ψ(π(γ(g, x, y))) = ψ(w(a, b)) 6= 0. We get
Z
ψ(π(γ(g, x, y))) dx ∧ dy = vol(S3 × S4 )ψ(w(a, b))
S3 ×S4
= (4s2 − 16s3 + 16s4 )ψ(w(a, b)).
• (x, y) ∈ X2 (S1 ); for fixed g the braid γ(g, x, y) can attain finitely
many values and let C2 := max |ψ(π(γ(g, x, y)))|. We have that
Z
ψ(π(γ(g, x, y))) dx ∧ dy ≤ vol(S1 × S1 )C1
X2 (S1 )
= 16s4C2 .
• (x, y) ∈ X2 (S3 ) ∪ X2 (S4 ); in this case the braid γ(g, x, y) is equal to
n
m n
either am
1 a2 or b1 b2 and hence π(γ(g, x, y)) is equal to a power of a
generator and ψ(π(γ(g, x, y))) = 0.
ON THE AUTONOMOUS NORM ON Ham(T2 )
15
Chose s ∈ 0, 14 small enough so that
(6s3 − 16s4)C1 + 16s4 C2 < 2(4s2 − 16s3 + 16s4 )ψ(w(a, b)).
For such an s we obtain that
Z
ψ(π(γ(g, x, y))) dx ∧ dy 6= 0.
X2 (T2 )
Since g(x) = x and g(y) = y for (x, y) outside the subset of arbitrarily
small measure (depending on the number ǫ), we have that γ(g p , x, y) =
γ(g, x, y)p, for (x, y) in the set of measure which is arbitrarily close to
full, according to Lemma 3.2. This implies that ψ(π(γ(g p , x, y))) =
pψ(π(γ(g, x, y))) and finally we get that
Z
1
ψ(π(γ(g p , x, y))) dx ∧ dy 6= 0.
lim
p→∞ X (T2 ) p
2
which finishes the proof.
Example 3.3 (Eggbeater). Let g = h2m v 2m−1 · · · h2 v ∈ Ham(T2 ),
where h, v ∈ Ham(T2 ) are Hamiltonian diffeomorphisms defined in the
above proof. It follows from Proposition 4.3 below and the above proof
that the quasimorphism G(ψcm ◦π) is unbounded on the cyclic subgroup
of Ham(T2 ) generated by g. Here ψcm : F2 → R is the quasimorphism
associated with the function cm : Zm → Z given by
cm (i1 , . . . , im ) = sgn(|i1 | − |im |),
see Lemma 4.1. Since the quasimorphism G(ψcm ◦ π) vanishes on autonomous elements, we get that the cyclic subgroup generated by g is
unbounded with respect to the autonomous norm.
♦
4. Quasimorphisms with vanishing properties
In this section we prove that the space Q(F2 ; Z/2 × Z/2) of quasimorphisms on the free group invariant under the action of the Klein
group is infinite dimensional. It is done by constructing explicit examples. Our construction is inspired by the example from the proof of
Theorem 1.1 in [1].
Let w ∈ Fn be a reduced word. A syllable in w is a maximal power
of a generator occurring in w. The exponent of a syllable s is denoted
by e(s). For example, the commutator [a, b] ∈ F2 has four syllables all
with exponents equal to 1. Let w = s1 s2 . . . sk be a reduced word, where
si are syllables. It defines a k-tuple of integers (e(s1 ), e(s2 ), . . . , e(sk )).
16
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Let c : Zm → Z be a bounded function which satisfies the identity:
(4.1)
c(i1 , . . . , im ) = −c(−im , . . . , −i1 ).
Let ψc : Fn → R be defined as follows. Let w = s1 . . . sk ∈ Fn . If
k < m then ψc (w) = 0. If k ≥ m then
ψc (s1 s2 . . . sk ) :=
k−m+1
X
c(e(si ), . . . , e(si+m−1 )).
i=1
Lemma 4.1. Let c : Zm → Z∩[−B, B] be a bounded function satisfying
the identity (4.1). Then the function ψc is a quasimorphism with defect
bounded by 3(m + 1)B.
Proof. Let s = s1 · · · sk and t = t1 · · · tl be reduced words such that
st is also reduced. We have the following expression for the value of
ψc (st):
ψc (s) + ψc (t)
Pm−1
+
i=1 c(e(sk−m+1+i ), . . . , e(sk ), e(t1 ), . . . , e(ti ))
if last letter of s is different from the first letter of t, or
ψc (st) = ψc (s) + ψc (t)
P
+ m−1
i=1 c(e(sk−m+1+i ), . . . , e(sk t1 ), . . . , e(ti ))
−c(e(sk−m+1 ), . . . , e(sk )) − c(e(t1 ), . . . , e(tm )),
otherwise.
It follows that in this special case we have the following estimate:
|ψc (s) − ψc (st) + ψc (t)| ≤ (m + 1)B.
Let us now consider the general case. Let s = s1 · · · sk , t = t1 · · · tl , u =
u1 · · · up ∈ Fn be reduced words, where si , ti and ui are syllables. Suppose that su, u−1 t and st are reduced. Using the previous inequality
we obtain the following estimate of the defect:
ψc (su) − ψc (st) + ψc (u−1 t) ≤
≤ ψc (s) + ψc (u) − ψc (s) − ψc (t) + ψc (u−1 ) + ψc (t) + 3(m + 1)B
= ψc (u) + ψc (u−1) + 3(m + 1)B
≤ 3(m + 1)B.
The fact that ψc (u−1 ) = −ψc (u) follows from the identity (4.1). This
proves that ψc is a quasimorphism with defect 3(m + 1)B.
ON THE AUTONOMOUS NORM ON Ham(T2 )
17
Example 4.2. Let c : Z2 → Z be defined by c(m, n) = sgn(|m| − |n|).
The associated quasimorphism ψc : F2 → Z is clearly invariant under
the action of Z/2×Z/2. To see that it is unbounded consider the cyclic
subgroup generated by a4 b3 a2 b. We have that
ψc a4 b3 a2 b)n = 2n + 1.
which implies that the homogenization of this quasimorphism is nontrivial. It follows from Lemma 2.3 that the homogenization is also
invariant.
Proposition 4.3. The subspace Q(F2 ; Z/2 × Z/2) ⊂ Q(F2 ) of homogeneous quasimorphism invariant under the action of Z/2 × Z/2 is
infinite dimensional.
Proof. Let cm : Zm → Z for m ≥ 2 be a function defined by
cm (i1 , . . . , im ) := sgn(|i1 | − |im |).
Consider the sequence ψcm of quasimorphisms defined in the beginning
of this section. Since the function cm depends only on the absolute values, the quasimorphisms ψcm are invariant under inverting generators.
Let wm = a2m b2m−1 · · · a2 b. We get that
ψcm (w n ) = 2n + m − 1.
Let k ∈ N be a positive integer. Consider the square k × k-matrix with
entries aij = ψc3i (w3j ) and observe that it is upper triangular with positive entries on the diagonal and hence it has a positive determinant. It
implies that the functions ψc3i for i = 1, . . . , k are linearly independent
for any k ∈ N. This shows that there exists an infinite dimensional
subspace of quasimorphisms invariant under inverting generators. It
follows from Lemma 2.3 that dim Q(F2 ; Z/2 × Z/2) = ∞.
5. Vanishing on autonomous diffeomorphisms
Lemma 5.1. Let ψ : F2 = ha, bi → R be a homogeneous quasimorphism. If it vanishes on primitive elements and on the commutator
[a, b] then the quasimorphism
G(ψ ◦ π) : Ham(T2 ) → R
vanishes on the set of autonomous diffeomorphisms.
18
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Proof. Let H : T2 → R be a smooth function and let ht ∈ Ham(T2 )
be the autonomous flow induced by H. Let
O(ht , x) := {ht (x) ∈ T2 | t ∈ R }
denote the orbit of the point x with respect to the flow ht . Such an
orbit is either periodic (including constant) or it is an interval between
one (homoclinic) or two (heteroclinic) fixed points.
Let z1 , z2 ∈ T2 be basepoints and let x, y ∈ T2 . In what follows we
analyze the braid γ(hp , x, y) for p ∈ N. We break it down into cases
depending on the form of the orbits O(ht , x) and O(ht , y). We consider
the following cases:
(1) O(ht , x) = {x}.
(a) If O(ht , y) = {y} then γ(hp , x, y) is trivial.
(b) If O(ht , y) is a contractible periodic orbit bounding a disc containing the point x then γ(hp , x, y) is an integer power of σ 2 .
(c) If O(ht , y) is a contractible periodic orbit bounding a disc not
containing the point x then γ(hp , x, y) is either trivial or equal
to σ 2 .
(d) If O(ht , y) is a homotopically nontrivial periodic orbit then
its image is a simple closed curve. There exists a symplectic diffeomorphism of the torus f ∈ Symp(T2 ) preserving the
basepoints z1 and z2 such that the image f (O(ht , y)) of the
orbit represents the standard generator (1, 0) ∈ Z2 = π1 (T2 )
disjoint from f (x). Thus the braid γ(f hp f −1 , f (x), f (y))) =
(F, A), where both F ∈ F2 and A ∈ Z2 are powers of primitive
elements. Observe that γ(f hp f −1 , f (x), f (y))) = f∗ (γ(hp , x, y)),
where f∗ : P2 (T2 ) → P2 (T2 ) is the automorphism induced
by f . Since f induces an automorphism of the quotient F2 we
get that π(γ(hp , x, y)) is a power of a primitive element.
(e) If the orbit O(ht , y) is nonperiodic then there exists p0 ∈ N
such that
#{γ(hp , x, y) ∈ P2 (T2 ) | p ≥ p0 } ≤ 2.
Indeed, let y+ := limt→∞ ht (y) be the limit point and let ǫ > 0.
There exists p0 such that |hp (y) − y+ | < ǫ for every p ≥ p0 .
Depending on a relative position of the points z1 , z2 , x and
hp (y) the braids γ(hp , x, y) and γ(hq , x, y) for p, q ≥ p0 may
differ by at most one crossing arising when the endpoints hp (y)
or hq (y) and x are joined to the basepoints.
ON THE AUTONOMOUS NORM ON Ham(T2 )
19
(2) The orbit O(ht , x) is nonperiodic. Let x+ := limt→∞ ht (x).
(a) If O(ht , y) is either constant or nonperiodic then
#{γ(hp , x, y) ∈ P2 (T2 ) | p ≥ p0 } ≤ 2
as in the previous case. The only difference is that for given
ǫ one has to choose p0 such that both |hp (x) − x+ | < ǫ and
|hp (y) − y+ | < ǫ for all p ≥ p0 .
(b) If O(ht , y) is a contractible periodic orbit such that O(ht , x) is
contained in the disc bounded by O(ht , y) then γ(hp , x, y) is a
power of σ 2 .
(c) If O(ht , y) is a contractible periodic orbit such that O(ht , x) is
not contained in the disc bounded by O(ht , y) then γ(hp , x, y)
is either trivial or equal to σ 2 .
(d) If O(ht , y) is a homotopically nontrivial periodic orbit then, as
in the case (1)(d) above, we get that γ(hp , x, y) is a power of
a primitive element.
(3) The orbit O(ht , x) is contractible periodic.
(a) The case when O(ht , y) is either constant or nonperiodic has
been dealt with above.
(b) If the orbits O(ht , x) and O(ht , y) are concentric then γ(hp , x, y)
is a power of σ 2 .
(c) If the orbits O(ht , x) and O(ht , y) are contractible and not
concentric then γ(hp , x, y) is either trivial or equal to σ 2 .
(d) If O(ht , y) is periodic and homotopically nontrivial then the
braid γ(hp , x, y) is a power of a primitive element and the
argument is the same as in the analogous cases above.
(4) If O(ht , x) is periodic and homotopically nontrivial then the only
case which has not been done above is when the orbit O(ht , y) is
periodic and not contractible. In this case the images of both orbits
are disjoint simple closed curves and thus there exists a symplectic diffeomorphism f ∈ Symp(T2 ) preserving basepoints z1 and
z2 such that both f (O(ht , x)) and f (O(ht , y)) are disjoint simple
closed curves representing the generator (1, 0) ∈ H1 (T2 ; Z) (recall that the intersection form on H1 (T2 ; Z) is non-degenerate and
20
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
anti-symmetric). In this case we have
(
γ(f hp f −1 , f (x), f (y)) =
n
bm
i bj
n 2
bm
j bi σ ,
for some m, n ∈ Z, where i, j ∈ {1, 2} are distinct. To see this
recall that the conjugation by σ swaps the generators b1 and b2 . It
may also be useful to use the following computation
n
m n −1 2
m n 2
σbm
i bj σ = σbi bj σ σ = bj bi σ .
Thus the image of the above braid in the free group is equal to
either bm−n or bm−n [a, b]. Thus the image of the braid γ(hp , x, y) in
the free group is a product of a power of a primitive element and
a commutator of two primitive elements.
According to a theorem of Nielsen [12], the commutator of two primitive
elements is conjugate to [a, b]±1 .
As a conclusion we obtain that the projection π(γ(hp , x, y)) is equal to
either one of the following:
• an integer power of the commutator [a, b],
• an integer power of a primitive element,
• a product of a power of a primitive element and a conjugate of
the commutator [a, b] or its inverse,
or there exists p0 ∈ N such that
#{γ(hp , x, y) ∈ P2 (T2 ) | p ≥ p0 } ≤ 2.
Let ψ : F2 → R be a homogeneous quasimorphism vanishing on primitive elements and on the commutator [a, b]. If γ(hp , x, y) attains finitely
many values for p ≥ p0 then
1
ψ(γ(hp , x, y)) = 0
p→∞ p
lim
and hence G(ψ)(h) = 0. If π(γ(hp , x, y)) is a power of either a primitive
element or a conjugate of the commutator [a, b] then ψ(π(γ(hp , x, y))) =
0 by the hypothesis and we also get that G(ψ)(h) = 0. Finally, if
π(γ(hp , x, y)) is a product of a power of a primitive element and a
power of a conjugate of the commutator [a, b] then
|G(ψ)(h)| ≤ Dψ .
ON THE AUTONOMOUS NORM ON Ham(T2 )
21
Since for every autonomous diffeomorphism h and n ∈ N, hn is also
autonomous we get
Dψ
G(ψ)(hn )
≤
.
n
n
|G(ψ)(h)| =
This concludes the proof of the vanishing property of G(ψ) on the set
of autonomous diffeomorphisms.
5.2. Palindromes and primitive elements. The following observation and its proof are due to Bardakov, Shpilrain and Tolstykh [1].
Lemma 5.3. Every primitive element w ∈ F2 of the free group of rank
2 is a product of up to two palindromes.
Proof. Let σ : F2 → F2 be an automorphism defined on generators by
σ(a) = a−1 and σ(b) = b−1 . Consider the extension
ι
π
F2 → Aut(F2 ) → Out(F2 ) = GL(2, Z),
where the quotient is identified with the automorphism group of the
abelianisation Z2 of the free group F2 .
Let θ ∈ Aut(F2 ) be an automorphism. Since the image π(σ) ∈ Out(F2 )
is equal to − Id, the image π[σ, θ] of the commutator is trivial. This
implies that [σ, θ] = Ip(θ) is an inner automorphism for some p(θ) ∈ F2 .
For example, if τ ∈ Aut(F2 ) is defined by τ (a) = ab and τ (b) = b then
p(τ ) = a.
The following computation proves that σ(p(θ)) = p(θ)−1 which means
that p(θ) is a palindrome (notice that σ is an involution).
Iσ(p(θ)) = σ Ip(θ) σ −1 = σ[σ, θ]σ −1
= [θ, σ] = [σ, θ]−1 = Ip(θ)−1 .
The second observation is that
p(θξ) = p(θ)θ(p(ξ))
for any θ, ξ ∈ Aut(F2 ). Indeed,
Ip(θξ) = [σ, θξ]
= σθσθ−1 · θσξσξ −1 θ−1
= [σ, θ]θ[σ, ξ]θ−1
= Ip(θ) θIp(ξ) θ−1
= Ip(θ) Iθ(p(ξ)) = Ip(θ)θ(p(ξ)) .
22
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Evaluating this identity on the automorphism τ defined above we get
that
p(θτ ) = p(θ)θ(p(τ ))
p(θτ ) = p(θ)θ(a)
p(θ)−1 p(θτ ) = θ(a),
which shows that any primitive element θ(a) is a product of two palindromes.
Corollary 5.4. If ψ : F2 → R is a homogeneous quasimorphism which
vanishes on palindromes then it vanishes on primitive elements.
Proof. Let w ∈ F2 be a primitive element. It follows from Lemma 5.3
that w = uv, where u, v ∈ F2 are palindromes. If n ∈ N is a positive
integer then
w n = (uv)n = (uv)n−1u v,
which shows that w n is a product of two palindromes. It implies that
|ψ(w n )| ≤ Dψ . Since ψ is homogeneous we get that it vanishes on
primitive elements.
Corollary 5.5. If ψ ∈ Q(F2 ; Z/2 × Z/2) then the quasimorphism
G(π ∗ ψ) ∈ Q(Ham(T2 ) vanishes on autonomous diffeomorphisms.
Proof. Let σ = σa σb ∈ Aut(F2 ). It acts on words by inverting all
letters. In particular, an element w ∈ F2 is a palindrome if and only if
σ(w) = w −1 . If ψ ∈ Q(F2 ; Z/2 × Z/2) then ψ is invariant under the
action of σ:
ψ(σ(w)) = ψ(w).
If w ∈ F2 is a palindrome then ψ(w −1) = ψ(σ(w)) = ψ(w) and it
follows from the homogeneity of ψ that it vanishes on palindromes. It
is then a consequence of Corollary 5.4 that ψ vanishes on primitive
elements.
As explained in the computation on page 5, ψ vanishes on the commutator [a, b] of the generators of F2 and hence, according to Lemma 5.1, the
quasimorphism G(π ∗ ψ) vanishes on autonomous diffeomorphisms.
ON THE AUTONOMOUS NORM ON Ham(T2 )
23
6. Further results
6.1. A lift to symplectic diffeomorphisms. The braid γ(h, x, y) is
not well defined for h ∈ Symp0 (T2 ). That is, it depends on the isotopy
from the identity to h. For example, the isotopy defined ht (u, v) =
(u + t, v) is a loop based at the identity and γ(ht , x, y) = a1 a2 .
Lemma 6.2. Let ℓ : [0, 1] → Symp0 (T2 ) be a loop based at the identity.
Then for every x, y ∈ T2 the braid γ(ℓ, x, y) is central.
Proof. Since the inclusion T2 → Symp0 (T2 ) (where the torus acts on
itself by translations) is a homotopy equivalence, the loop ℓ is isotopic
to a concatenation of loops ha : (u, v) 7→ (u + t, v) and hb : (u, v) 7→
(u, v + s). This implies that
m n n
2
γ(ℓ, x, y) = am
1 a2 b1 b2 ∈ P2 (T ),
for some m, n ∈ Z. Observe that the center of P2 (T2 ) is isomorphic to
Z2 and is generated by a1 a2 and b1 b2 (see Lemma 2.6.
It follows from the lemma that if ψ ∈ Q(P2 (T2 )) vanishes on the center
then ψ(γ(h, x, y)) is well defined for h ∈ Symp0 (T2 ). We thus have a
lift of the Gambaudo-Ghys homomorphism
Gb : Q(P2 (T2 ), CENTER) → Q(Symp0 (T2 )).
Lemma 6.3. The composition
Gb
ι∗
Q(B2 (T2 ), CENTER) → Q(P2 (T2 ), CENTER) → Q(Symp0 (T2 ))
is injective.
The proof of this lemma is essentially the same as the proof of Theorem
2 in [2]. If ψ ∈ Q(F2 ) then the composition ψ ◦ π vanishes on the
center and hence every quasimorphism on F2 yields a quasimorphism
on the group of symplectic diffeomorphisms. In order to ensure that
it is nontrivial we require that ψ ◦ π extends to the full braid group
and this holds if the quasimorphism ψ vanishes on palindromes (see
Proposition 2.8).
Corollary 6.4. The composition
π∗
Gb
Q(F2 , PAL) → Q(P2 (T2 )) → Q(Symp0 (T2 ))
is injective.
24
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
Since Ham(T2 ) is equal to the commutator subgroup of Symp0 (T2 ) the
kernel of the homomorphism Q(Symp0 (T2 )) → Q(Ham(T2 )) induced
by the inclusion consists of homomorphisms. Since Q(F2 , PAL) contains
no homomorphism we get that the composition
Q(F2 , PAL) → Q(Ham(T2 ))
is injective.
Remark 6.5. This is a slightly stronger statement than Proposition 3.1
and it could serve as an alternative part of the proof of the main Theorem 1.1. We chose a more direct approach in order to have a complete
proof which makes the paper selfcontained and also because the proof
of Proposition 3.1 provides explicit examples of diffeomorphisms. On
the other hand, the above arguments allow us to provide examples of
Calabi quasimorphisms which are presented next.
6.6. The Calabi property and continuity. Let (M, ω) be a symplectic manifold and let B ⊂ M is a displaceable symplectic ball. A
homogeneous quasimorphism Ψ : Ham(M, ω) → R is called Calabi (or
has the Calabi property) if its restriction to a subgroup Ham(B) coincides with the Calabi homomorphism. The definition is due to Entov
and Polterovich [6] who constructed first examples of Calabi quasimorphisms using quantum homology. Their examples include the sphere
S2 and they asked whether there were Calabi quasimorphisms for other
surfaces. Pierre Py gave a positive answer to this question in [16, 15].
Here we provide a Calabi quasimorphism by producing a slightly modified example.
Example 6.7 (The snake quasimorphism). Let w ∈ F2 be an element.
It defines a path on the plane starting at the origin, consisting of horizontal and vertical segments of integer length with turning points on
the integer lattice. See Figure 6.1 for an example.
Let ξ : F2 → Z be defined by ξ(w) := L(w) − R(w), where L(w) and
R(w) denote the number of the left and right turns of the path defined
by w. Thus the value at an element drawn in Figure 6.1 is ξ(w) =
5 − 4 = 1.
If w ∈ F2 is a palindrome then the induced path is symmetric with
respect to the half turn about its mid point and hence the initial turns
become the opposite terminal turns and hence they cancel. Thus ξ
vanishes on palindromes and hence its homogenization vanishes both
on palindromes and primitive elements. Since ξ ([a, b]n ) = 4n − 1 we
ON THE AUTONOMOUS NORM ON Ham(T2 )
25
Figure 6.1. The element w = a2 b−1 a2 b5 a−3 b−6 a−3 b4 a4 b
see that ξ is unbounded. Hence its homogenization ξb is nontrivial and
b b]) = 4. Thus the quasimorphism G(π ∗ ξ) is nontrivial and has the
ξ([a,
Calabi property.
♦
Acknowledgments. We would like to thank Luis Haug for fruitful
discussions and Pierre Py for comments on the first version of the paper.
We thank the CRM and ISM who supported the visit of Kędra in
Montreal. Kędra also thanks the Max Planck Institute for Mathematics
for supporting his visit in Bonn.
Part of this work has been done during Brandenbursky’s stay at IHES
and CRM. We wish to express his gratitude to both institutes. Brandenbursky was supported by CRM-ISM fellowship and NSF grant No.
1002477
This work has been done during Shelukhin’s stay in CRM, ICJ Lyon 1,
Institut Mittag Leffler, and IAS. He thanks these institutions for their
warm hospitality. He was partially supported by CRM-ISM fellowship,
ERC Grant RealUMan, Mittag Leffler fellowship, and NSF grant No.
DMS-1128155.
26
BRANDENBURSKY, KĘDRA, AND SHELUKHIN
References
[1] Valery Bardakov, Vladimir Shpilrain, and Vladimir Tolstykh. On the palindromic and primitive widths of a free group. J. Algebra, 285(2):574–585, 2005.
[2] Michael Brandenbursky. Bi-invariant metrics and quasi-morphisms on groups
of Hamiltonian diffeomorphisms of surfaces. Internat. J. Math., 26(9):1550066,
29 pages, 2015.
[3] Michael Brandenbursky and Jarek Kedra. On the autonomous metric on the
‘
group of area-preserving diffeomorphisms of the 2-disc. Algebr. Geom. Topol.,
13(2):795–816, 2013.
[4] Michael Brandenbursky and Egor Shelukhin. On the Lp -geometry of autonomous Hamiltonian diffeomorphisms of surfaces. To appear in Math. Res.
Lett.
[5] Dmitri Burago, Sergei Ivanov, and Leonid Polterovich. Conjugation-invariant
norms on groups of geometric origin. In Groups of diffeomorphisms, volume 52
of Adv. Stud. Pure Math., pages 221–250. Math. Soc. Japan, Tokyo, 2008.
[6] Michael Entov and Leonid Polterovich. Calabi quasimorphism and quantum
homology. Int. Math. Res. Not., (30):1635–1676, 2003.
[7] Michael Entov, Leonid Polterovich, and Pierre Py. On continuity of quasimorphisms for symplectic maps. In Perspectives in analysis, geometry, and
topology, volume 296 of Progr. Math., pages 169–197. Birkhäuser/Springer,
New York, 2012. With an appendix by Michael Khanevsky.
[8] Jean-Marc Gambaudo and Étienne Ghys. Commutators and diffeomorphisms
of surfaces. Ergodic Theory Dynam. Systems, 24(5):1591–1617, 2004.
[9] Tomohiko Ishida. Quasi-morphisms on the group of area-preserving diffeomorphisms of the 2-disk via braid groups. Proc. Amer. Math. Soc. Ser. B, 1:43–51,
2014.
[10] Michael Khanevsky. Hamiltonian commutators with large Hofer norm. Sep
2014. Available at http://arxiv.org/abs/1409.7420v2.
[11] A. V. Malyutin. Pseudocharacters of braid groups and the simplicity of links.
Algebra i Analiz, 21(2):113–135, 2009.
[12] J. Nielsen. Die Isomorphismen der allgemeinen, unendlichen Gruppe mit zwei
Erzeugenden. Math. Ann., 78(1):385–397, 1917.
[13] Leonid Polterovich. Floer homology, dynamics and groups. In Morse theoretic
methods in nonlinear analysis and in symplectic topology, volume 217 of NATO
Sci. Ser. II Math. Phys. Chem., pages 417–438. Springer, Dordrecht, 2006.
[14] Leonid Polterovich and Egor Shelukhin. Autonomous Hamiltonian flows,
Hofer’s geometry and persistence modules. Selecta Math. (N.S.), 22(1):227–
296, 2016.
[15] Pierre Py. Quasi-morphismes de Calabi et graphe de Reeb sur le tore. C. R.
Math. Acad. Sci. Paris, 343(5):323–328, 2006.
[16] Pierre Py. Quasi-morphismes et invariant de Calabi. Ann. Sci. École Norm.
Sup. (4), 39(1):177–195, 2006.
[17] G. P. Scott. Braid groups and the group of homeomorphisms of a surface. Proc.
Cambridge Philos. Soc., 68:605–617, 1970.
[18] Takashi Tsuboi. On the uniform simplicity of diffeomorphism groups. In Differential geometry, pages 43–55. World Sci. Publ., Hackensack, NJ, 2009.
[19] Takashi Tsuboi. On the uniform perfectness of the groups of diffeomorphisms
of even-dimensional manifolds. Comment. Math. Helv., 87(1):141–185, 2012.
ON THE AUTONOMOUS NORM ON Ham(T2 )
Ben-Gurion University, Israel
E-mail address: [email protected]
University of Aberdeen and University of Szczecin
E-mail address: [email protected]
IAS, Princeton
E-mail address: [email protected]
27
| 4math.GR
|
A REMARK ON TORSION GROWTH IN HOMOLOGY
AND VOLUME OF 3-MANIFOLDS
arXiv:1802.09244v1 [math.GR] 26 Feb 2018
HOLGER KAMMEYER
Abstract. We show that Lück’s conjecture on torsion growth in homology implies that two 3-manifolds have equal volume if the fundamental
groups have the same set of finite quotients.
The purpose of this note is to relate two well-known open problems which
both deal with a residually finite fundamental group Γ of an odd-dimensional
aspherical manifold. The first one [11, Conjecture 1.12(2)] predicts that the
ℓ2 -torsion ρ(2) (Γ) determines the exponential rate at which torsion in middledegree homology grows along a chain of finite index normal subgroups.
Conjecture A. Let M be an aspherical closed manifold of dimension 2d+1.
Suppose that Γ = π1 M is residually finite and letTΓ = Γ0 ≥ Γ1 ≥ · · · be any
chain of finite index normal subgroups of Γ with ∞
n=0 Γn = {1}. Then
lim
n→∞
log |Hd (Γn )tors |
= (−1)d ρ(2) (Γ).
[Γ : Γn ]
The term |Hd (Γn )tors | denotes the order of the torsion subgroup of Hd (Γn ).
The ℓ2 -torsion ρ(2) (Γ) is the ℓ2 -counterpart to Reidemeister torsion as surveyed in [12] and [7]. The second conjecture says that volume of 3-manifolds
can be recovered from the finite quotients of the fundamental group.
Conjecture B. Let Γ and Λ be infinite fundamental groups of connected,
b∼
b Then
closed, orientable, irreducible 3-manifolds and suppose that Γ
= Λ.
vol(Γ) = vol(Λ).
b of Γ is the projective limit over all finite
Here the profinite completion Γ
quotients of Γ. Two groups have isomorphic profinite completions if and only
if they have the same set of finite quotients [18, Corollary 3.2.8]. If Γ = π1 M
for a 3-manifold M with the stated properties, then Thurston geometrization
applies to M : there is a minimal choice of finitely many disjointly embedded
incompressible tori in M , unique up to isotopy, which cut M into pieces such
that each piece carries one out of eight geometries. The sum of the volumes
of the hyperbolic pieces gives the well-defined quantity vol(Γ). Conjecture B
is often stated as a question [2, Question 3.18]. But we dare to promote it
to a conjecture in view of the following result.
Theorem 1. Conjecture A implies Conjecture B.
The theorem seems to be folklore among the experts in the field but I
could not find a proof in the literature so that this note is meant as a service
to the community.
2010 Mathematics Subject Classification. 20E18, 57M27.
1
2
HOLGER KAMMEYER
The contrapositive of Theorem 1 says that constructing two profinitely isomorphic 3-manifold groups with differing covolume would disprove Conjecture A. Funar [4] and Hempel [6] constructed examples of closed 3-manifolds
with non-isomorphic but profinitely isomorphic fundamental groups. These
examples carry Sol and H2 ×R geometry, respectively, and thus all have zero
volume by definition. Wilkes [19] showed that Hempel’s examples are the
only ones among Seifert-fiber spaces. It seems to be open whether there exist
such examples with H3 -geometry. As a first step in the negative direction,
Bridson and Reid [3] showed that the figure eight knot group is determined
among 3-manifold groups by the profinite completion.
The paper at hand is divided into two sections. Section 1 presents the
proof of Theorem 1. As a complement, Section 2 discusses how the related
asymptotic volume conjecture and the Bergeron–Venkatesh conjecture fit into
the picture.
I wish to thank S. Kionke and J. Raimbault for helpful discussions during
the junior trimester program “Topology” at the Hausdorff Research Institute
for Mathematics in Bonn.
1. Proof of Theorem 1
For the moment, let Γ and Λ be any two finitely generated, residually
finite groups. To prepare the proof of Theorem 1, we collect a couple of
propositions from the survey article [17] and include more detailed proofs
for the sake of a self-contained treatment. We first recall that the open
b are precisely the subgroups of finite index. One direction is
subgroups of Γ
b
easy: Γ is compact and the cosets of an open subgroup form a disjoint open
cover. The converse is a deep theorem due to Nikolov and Segal [15] that
crucially relies on the assumption that Γ is finitely generated. The proof
moreover invokes the classification of finite simple groups.
The assumption that Γ is residually finite says precisely that the canonical
b is an embedding. If Q is a finite group, the universal property of
map Γ → Γ
b
b Q) → Hom(Γ, Q) is a surjection. By
Γ says that the restriction map Hom(Γ,
ϕ
b−
the above, the kernel of any homomorphism Γ
→ Q is open which implies
that ϕ is continuous and is thus determined by the values on the dense subset
b Thus Hom(Γ,
b Q) → Hom(Γ, Q) is in fact a bijection which clearly
Γ ⊂ Γ.
b Q) → Epi(Γ, Q) of surjective homomorphisms.
restricts to a bijection Epi(Γ,
This has the following consequence.
b then there is an epimorphism
Proposition 2. If Λ embeds densely into Γ,
H1 (Λ) → H1 (Γ).
Proof. Let p be a prime number which does not divide the group order
|H1 (Λ)tors | and let us set r = dimQ H1 (Γ; Q). It is apparent that we have
an epimorphism Γ → (Z/pZ)r ⊕ H1 (Γ)tors . By the above remarks, this
b → (Z/pZ)r ⊕ H1 (Γ)tors .
epimorphism extends uniquely to an epimorphism Γ
b the latter map restricts to an epimorphism
Since Λ embeds densely into Γ,
r
Λ → (Z/pZ) ⊕ H1 (Γ)tors . This epimorphism must lift to an epimorphism
Λ → Zr ⊕ H1 (Γ)tors ∼
= H1 (Γ) because p is coprime to |H1 (Λ)tors |. Of course
this last epimorphism factors through the abelianization H1 (Λ).
A REMARK ON TORSION GROWTH IN HOMOLOGY
3
b ∼
b then
Corollary 3. The abelianization is a profinite invariant: if Γ
= Λ,
∼
H1 (Γ) = H1 (Λ).
Proof. Since we have surjections in both directions the groups H1 (Γ) and
H1 (Λ) have the same free abelian rank. Thus either surjection restricts to
an isomorphism of the free parts and thus induces a surjection of the finite
torsion quotients—which then must be a bijection.
b called the profinite
Let us now endow Γ with the subspace topology of Γ,
topology of Γ. For the open subgroups of Γ we have the same situation as
b
we observed for Γ.
Proposition 4. A subgroup H ≤ Γ is open in the profinite topology if and
only if H has finite index in Γ.
b carries the coarsest topology under which the projecProof. Recall that Γ
b
tions Γ → Γ/Γi for finite index normal subgroups Γi E Γ are continuous.
b → Γ/Γi are the canonical projections, it
Since the compositions Γ → Γ
b is given by the
follows that a subbase for the subspace topology of Γ ⊂ Γ
cosets of finite index normal subgroups of Γ.
T
If H has finite index in Γ, then so does the normal core N = g∈Γ gHg−1
because N is precisely the kernel of the permutation representation
S of Γ on
the homogeneous set Γ/H defined by left translation. Thus H = h∈H hN
is open. Conversely, let H ≤ Γ be open. Then H is a union of finite
intersections of finite index normal subgroups of Γ. In particular H contains
a finite index subgroup, whence has finite index itself.
b defines a 1-1–correspondence
Proposition 5. Taking closure H 7→ H in Γ
from the open (or finite index) subgroups of Γ to the open (or finite index)
b The inverse is given by intersection H 7→ H ∩ Γ with Γ.
subgroups of Γ.
This correspondence preserves the index, sends a normal subgroup N E Γ to
∼
b and in the latter case we have Γ/N
b
a normal subgroup N E Γ,
= Γ/N .
The proof is given in [18, Prop. 3.2.2, p. 84]. Here is an easy consequence.
Corollary 6. For H1 , H2 ≤ Γ of finite index we have H1 ∩ H2 = H1 ∩ H2 .
b and we get
Proof. By the proposition H1 ∩ H2 has finite index in Γ
(H1 ∩ H2 ) ∩ Γ = (H1 ∩ Γ) ∩ (H2 ∩ Γ) = H1 ∩ H2 .
Applying the proposition again yields H1 ∩ H2 = H1 ∩ H2 .
Note that for a finitely generated, residually finite group Γ there is a
canonical choice of a chain
Γ = M1 ≥ M2 ≥ M3 ≥ · · ·
T
of finite index normal subgroups Mn E Γ satisfying ∞
n=1 Mn = {1}. Simply
define Mn to be the intersection of the (finitely many!) normal subgroups
of index at most n. By the last two results, Mn is the intersection of all
b with index at most n.
normal subgroups of Γ
T
Proposition 7. The intersection ∞
n=1 Mn is trivial.
4
HOLGER KAMMEYER
T
b
Proof. Let g ∈ ∞
n=1 Mn ⊂ Γ. Since Γ is finitely generated, it has only countably many subgroups of finite index. Therefore the description of the topolb given above shows that Γ
b is second and thus first countable. Hence
ogy of Γ
b with limi→∞ gi = g.
we can pick a sequence (gi ) from the dense subset Γ ⊂ Γ
b
b
Let pn : Γ → Γ/Mn and pbn : Γ → Γ/Mn denote the canonical projections.
Since pbn is continuous, we have
Mn = pbn (g) = lim pbn (gi )
i→∞
∼
b
and hence limi→∞ pn (gi ) = Mn ∈ Γ/Mn because Γ/M
n = Γ/Mn by Proposition 5. As Γ/Mn is discrete, the sequence pn (gi ) is eventually constant.
This means that for all n ≥ 1 there is N ≥ 1 such that for all i ≥ N we
have pn (gi ) = Mn , or equivalently gi ∈ Mn . But the open sets Mn form a
neighborhood basis of 1 ∈ Γ as follows from the description of the profinite
topology of Γ given in the proof of Proposition 4. So the last statement gives
b (and hence Γ) is Hausdorff, we conclude g = 1.
limi→∞ gi = 1. Since Γ
b is residually finite as an abstract group. Before we give
It follows that Γ
the proof of Theorem 1, we put down one more observation. If H ≤ Γ is any
b is a profinite group so that the universal
subgroup, then the closure H in Γ
b
b → H which restricts
property of H gives a canonical homomorphism η : H
to the identity on H. This is always an epimorphism because the image is
dense, as it contains H, and closed because it is compact and H is Hausdorff.
However, in general we cannot expect that η is injective, not even if H is
finitely generated. Nevertheless:
b →
Proposition 8. If H ≤ Γ has finite index, then the canonical map η : H
H is an isomorphism.
Proof. Let h ∈ ker η. The group H is finitely generated because it is a finite
b is second and hence
index subgroup of Γ. As above we conclude that H
b we can thus pick a sequence of
first countable. Since H lies densely in H,
elements hi ∈ H such that limi→∞ hi = h. By continuity of η, we obtain
limi→∞ η(hi ) = η(h) = 1 and thus limi→∞ hi = 1 in the topology of H. A
neighborhood basis of 1 ∈ H is given by the sets Mn ∩ H where Mn are the
b from above. It follows that for all n ≥ 1
finite index normal subgroups of Γ
there exists N ≥ 1 such that for all i ≥ N we have hi ∈ Mn ∩H. Since H has
finite index in Γ, it follows that any finite index normal subgroup K E H
has also finite index as a subgroup of Γ. Thus there exists n ≥ 1 such that
Mn lies in the normal core of K as a subgroup of Γ. Hence for all K E H
of finite index there exists N ≥ 1 such that for all i ≥ N we have hi ∈ K.
But the finite index normal subgroups K E H form a neighborhood basis
of 1 ∈ H in the profinite topology of H. Hence we have limi→∞ hi = 1 in
b Since H
b is Hausdorff, we conclude h = 1.
the topology of H.
Proof of Theorem 1. Note that Γ and Λ are finitely generated and residually
b∼
b
finite, as a consequence of geometrization [5]. We fix an isomorphism Γ
= Λ.
Again, let Mn ≤ Γ be the intersection of all normal subgroups of Γ of
index at most n. By Proposition 5 it follows that Ln = Λ ∩ Mn is the
intersection of all normal subgroups of Λ of index at most n and [Γ : Mn ] =
A REMARK ON TORSION GROWTH IN HOMOLOGY
5
T
T
[Λ : Ln ]. By Proposition 7 we have n Mn = {1} so that n Ln = {1}. From
dn ∼
cn so that Corollary 3 implies |H1 (Mn )tors | =
Proposition 8 we get M
=L
|H1 (Ln )tors |. A theorem of Lück and Schick [13, Theorem 0.7] conjectured
in Lott and Lück [9, Conjecture 7.7] shows that ρ(2) (Γ) = − vol(Γ)/6π and
similarly for Λ, see also [12, Theorem 4.3, p. 216]. If Conjecture A holds
true, this implies
log |H1 (Ln )tors |
log |H1 (Mn )tors |
= 6π lim
= vol(Λ).
n→∞
n→∞
[Γ : Mn ]
[Λ : Ln ]
vol(Γ) = 6π lim
2. Related conjectures
One can find companion conjectures to Conjecture A in the literature
which likewise predict an exponential rate of torsion growth in homology
proportional to volume. However, these conjectures restrict the aspherical
manifolds under consideration in one way or another. Specifically dealing
with 3-manifolds is Lê’s asymptotic volume conjecture.
Conjecture C. Let Γ be the fundamental group of a connected, orientable,
irreducible, compact 3-manifold whose boundary is either empty or a collection of tori. Then
lim sup
Γn →{1}
vol(Γ)
log |H1 (Γn )tors |
=
.
[Γ : Γn ]
6π
The conjecture appears in [8, Conjecture 1 (a)]. The volume vol(Γ) is defined by a geometric decomposition as before which also exists for toroidal
boundary. The lim sup on the left hand side is defined as the lowest upper
bound of all lim sups along sequences (Γn ) of (not necessarily nested!) finite index normal subgroups of Γ with lim supn Γn = {1}. Recall that by
definition
\ [
lim supn Γn =
Γn
N ≥0 n≥N
so that the condition lim supn Γn = {1} is actually equivalent to requiring
(
1 if g = e,
lim trC[Γ/Γn ] (gΓn ) = trC[Γ] (g) =
n→∞
0 otherwise,
for all g ∈ Γ where the traces are the usual traces of group algebras given
by the unit coefficient.
Question 9. Does Conjecture C imply Conjecture B?
The proof of Theorem 1 does not immediately carry over to Question 9 as
lim supn Γn = {1} for some sequence (Γn ) does not imply lim sup Λn = {1}
for the groups Λn = Λ ∩ Γn . Here is an example.
Example 10. Let Γ = Z×Z with (nested) chain of subgroups Γn = 2n Z×3nZ.
Q
b=Z
b × Z.
b From the description Z
b∼
Clearly, we have Γ
= p Zp it is apparent
b : N Z]
b = N for any N ≥ 1. Since N Z is the only subgroup of index N
that [Z
6
HOLGER KAMMEYER
b Thus we have Γn = 2n Z
b × 3n Z.
b
in Z, Proposition 5 implies that N Z = N Z.
It follows that
∞
\
Y
Y
b × Z.
b
Γn ∼
Zp × Z2 × {0} ×
Zq ≤ Z
= {0} ×
n=1
p>2
q>3
b be the subgroup generated by the two elements
So if we let Λ ≤ Γ
((0, 1, 1, . . .), (1, 0, 0, 0, . . .)) and ((1, 0, 0, 0 . . .), (0, 1, 1, 1, . . .))
Q
b × Z,
b then clearly Λ ∼
b so that
in p Zp × p Zp ∼
=Z
= Z × Z is dense in Γ
b→Λ=Γ
b is a surjective homomorphism of isomorphic
the canonical map Λ
finitely generated profinite groups. Hence it T
must be an isomorphism [18,
Proposition 2.5.2, p. 46]. However, we have ∞
n=1 Λn 6= {0} even though
T
∞
Γ
=
{0}.
n
n=1
Q
We remark that Lê has proven the inequality “≤” of Conjecture C, even
if the subgroups are not required to be normal. Another conjecture, which
leaves both the realm of 3-manifolds and of normal subgroups, is due to
Bergeron and Venkatesh [1, Conjecture 1.3]. It does however assume a
somewhat rigorous arithmetic setting. This is what we want to present
next.
Let G be a semisimple algebraic group, defined and anisotropic over Q.
Let Γ ≤ G(Q) be a congruence subgroup. This means that for some (and
then for any) Q-embedding ρ : G → GLn there is k ≥ 1 such that the group
ρ(Γ) contains the kernel of ρ(G) ∩ GLn (Z) → GLn (Z/kZ) as a subgroup of
finite index. Fix an algebraic representation of G on a finite-dimensional
Q-vector space W and let M ⊂ W be a Γ-invariant Z-lattice, which always
exists according to [16, Remark, p. T
173]. Let Γ = Γ0 ≥ Γ1 ≥ · · · be a
chain of congruence subgroups with n Γn = {1}. For a maximal compact
subgroup K of G = G(R), we denote by X = G/K the symmetric space
associated with G. Let g and k be the Lie algebras of G and K and let
δ(G) = rankC g ⊗ C − rankC k ⊗ C be the deficiency of G, sometimes also
known as the fundamental rank δ(X) of X.
Conjecture D. For each d ≥ 1 there is a constant cG,M,d ≥ 0 such that
log |Hd (Γn ; M )tors |
= cG,M,d vol(Γ)
n→∞
[Γ : Γn ]
lim
and cG,M,d > 0 if and only if δ(G) = 1 and dim X = 2d + 1.
In this case the volume vol(Γ) is the volume of the closed locally symmetric
space Γ\X which is defined by means of a Haar measure on G and as such
only unique up to scaling. But any rescaling of this measure would also
rescale the constant cG,M,d by the reciprocal value so that the product is
well-defined. To make sure that cG,M,d really only depends on G, M , and
d, we agree upon the following normalization of the Haar measure. The
Killing form on g restricts to a positive definite form on the subspace p
in the orthogonal Cartan decomposition g = k ⊕ p. Identifying p with the
tangent space TK X, we obtain a G-invariant metric on X by translation.
We require that the volume of Γ\X determined by Haar measure be equal
to the volume of Γ\X as Riemannian manifold.
A REMARK ON TORSION GROWTH IN HOMOLOGY
7
To relate Conjecture D to Conjecture B, we need to restrict our attention to arithmetic hyperbolic 3-manifolds. These are quotients of hyperbolic
3-space H3 by arithmetic Kleinian groups. A Kleinian group is a discrete
subgroup Γ ≤ PSL(2, C) ∼
= Isom+ (H3 ) such that vol(Γ) = vol(Γ\H3 ) < ∞.
A Kleinian group Γ ≤ PSL(2, C) is called arithmetic if there exists a semisimple linear algebraic Q-group H ≤ GLn and an epimorphism of Lie groups
φ : H(R)0 → PSL(2, C) with compact kernel such that Γ is commensurable
with φ(H(Z) ∩ H(R)0 ). Here H(R)0 denotes the unit component and two
subgroups of a third group are called commensurable if their intersection has
finite index in both subgroups. Note that we consider PSL(2, C) as a real
Lie group so that the complexified lie algebra is sl(2, C) ⊕ sl(2, C) and hence
δ(PSL(2, C)) = 1. There is an alternative and equivalent approach to the
definition of arithmetic Kleinian groups via orders in quaternion algebras
over number fields [14].
b = Λ.
b
Question 11. Let Γ and Λ be arithmetic Kleinian groups such that Γ
Suppose Conjecture D holds true. Can we conclude that vol(Γ) = vol(Λ)?
Again, various problems arise when trying to adapt the proof of Theorem 1
to settle this question in the affirmative. To be more concrete, a direct translation fails for the following reason. Let Mn be the intersection of all normal
subgroups of index at most n in the arithmetic group H(Z) corresponding
to Γ as above. Then Mn will not consist of congruence subgroups. In fact,
H(Z) has the congruence subgroup property if and only if all the groups Mn
are congruence subgroups. But the congruence subgroup property is well
known to fail for all arithmetic Kleinian groups [10]. Instead, one could try
to start with a chain of congruence subgroups Γn of Γ but then it seems
unclear if or under what circumstances the chain Λn = Γn ∩ Λ consists of
congruence subgroups in Λ.
We remark that for the trivial coefficient system Z ⊂ Q, Conjecture D
is wide open. However, in our relevant case of δ(G) = 1, Bergeron and
Venkatesh construct strongly acyclic coefficient modules M with the property that the spectrum of the Laplacian acting on M ⊗Z C-valued p-forms
on Γn \X is bounded away from zero for all p and n. In the special case
G = SL(2, C), they show that Conjecture D holds true for any strongly
acyclic M .
References
[1] N. Bergeron and A. Venkatesh, The asymptotic growth of torsion homology for arithmetic groups, J. Inst. Math. Jussieu 12 (2013), no. 2, 391–447. MR 3028790 ↑6
[2] M. Boileau and S. Friedl, Profinite completions and 3-manifold groups, RIMS
Kokyuroku 1991 (2016), 54–68. http://hdl.handle.net/2433/224629. ↑1
[3] M. Bridson and A. Reid, Profinite rigidity, fibering, and the figure-eight knot (2015).
e-print. arXiv:1505.07886 ↑2
[4] L. Funar, Torus bundles not distinguished by TQFT invariants, Geom. Topol. 17
(2013), no. 4, 2289–2344. With an appendix by Funar and Andrei Rapinchuk.
MR 3109869
↑2
[5] J. Hempel, Residual finiteness for 3-manifolds, Combinatorial group theory and topology (Alta, Utah, 1984), Ann. of Math. Stud., vol. 111, Princeton Univ. Press, Princeton, NJ, 1987, pp. 379–396. MR 895623 ↑4
[6]
, Some 3-manifold groups with the same finite quotients (2014). e-print.
arXiv:1409.3509 ↑2
8
HOLGER KAMMEYER
[7] H. Kammeyer, Introduction to ℓ2 -invariants, lecture notes, 2018. Available for download at http://topology.math.kit.edu/21_679.php. ↑1
[8] T. T. Q. Lê, Growth of homology torsion in finite coverings and hyperbolic volume
(2014). e-print. arXiv:1412.7758 ↑5
[9] J. Lott and W. Lück, L2 -topological invariants of 3-manifolds, Invent. Math. 120
(1995), no. 1, 15–60. MR 1323981 ↑5
[10] A. Lubotzky, Group presentation, p-adic analytic groups and lattices in SL2 (C), Ann.
of Math. (2) 118 (1983), no. 1, 115–130. MR 707163 ↑7
[11] W. Lück, Approximating L2 -invariants and homology growth, Geom. Funct. Anal. 23
(2013), no. 2, 622–663. MR 3053758 ↑1
[12]
, L2 -invariants: theory and applications to geometry and K-theory, Ergebnisse
der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in
Mathematics, vol. 44, Springer-Verlag, Berlin, 2002. MR 1926649 ↑1, 5
[13] W. Lück and T. Schick, L2 -torsion of hyperbolic manifolds of finite volume, Geom.
Funct. Anal. 9 (1999), no. 3, 518–567. MR 1708444 ↑5
[14] C. Maclachlan and A. W. Reid, The arithmetic of hyperbolic 3-manifolds, Graduate
Texts in Mathematics, vol. 219, Springer-Verlag, New York, 2003. MR 1937957 ↑7
[15] N. Nikolov and D. Segal, On finitely generated profinite groups. I. Strong completeness
and uniform bounds, Ann. of Math. (2) 165 (2007), no. 1, 171–238. MR 2276769 ↑2
[16] V. Platonov and A. Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics, vol. 139, Academic Press, Inc., Boston, MA, 1994. Translated
from the 1991 Russian original by Rachel Rowen. MR 1278263 ↑6
[17] A. W. Reid, Profinite properties of discrete groups, Groups St Andrews 2013, London
Math. Soc. Lecture Note Ser., vol. 422, Cambridge Univ. Press, Cambridge, 2015,
pp. 73–104. MR 3445488 ↑2
[18] L. Ribes and P. Zalesskii, Profinite groups, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics, vol. 40, SpringerVerlag, Berlin, 2000. MR 1775104 ↑1, 3, 6
[19] G. Wilkes, Profinite rigidity for Seifert fibre spaces, Geom. Dedicata 188 (2017), 141–
163. MR 3639628 ↑2
Karlsruhe Institute of Technology, Institute for Algebra and Geometry,
Germany
E-mail address: [email protected]
URL: www.math.kit.edu/iag7/~kammeyer/
| 4math.GR
|
Optimization–Based Decoding Algorithms
for LDPC Convolutional Codes in
Communication Systems
arXiv:1708.05618v1 [cs.IT] 18 Aug 2017
Banu Kabakulak∗1 , Z. Caner Taşkın1 , and Ali Emre Pusane2
1
2
Department of Industrial Engineering, Boğaziçi University, İstanbul, Turkey
Department of Electrical and Electronics Engineering, Boğaziçi University, İstanbul, Turkey
In a digital communication system, information is sent from one place to another
over a noisy communication channel. It may be possible to detect and correct
errors that occur during the transmission if one encodes the original information
by adding redundant bits. Low–density parity–check (LDPC) convolutional codes,
a member of the LDPC code family, encode the original information to improve
error correction capability. In practice these codes are used to decode very long
information sequences, where the information arrives in subsequent packets over
time, such as video streams. We consider the problem of decoding the received
information with minimum error from an optimization point of view and investigate
integer programming–based exact and heuristic decoding algorithms for its solution.
In particular, we consider relax–and–fix heuristics that decode information in small
windows. Computational results indicate that our approaches identify near–optimal
solutions significantly faster than a commercial solver in high channel error rates.
Our proposed algorithms can find higher quality solutions compared with commonly
used iterative decoding heuristics.
Keywords: Telecommunications, integer programming, relax–and–fix heuristic.
∗
Corresponding author.
E-mail addresses:
[email protected]
[email protected] (Z. C. Taşkın), [email protected] (A. E. Pusane).
(B.
Kabakulak),
1 Introduction and Literature Review
A digital communication system represents digital information flow from a source to sink over
an unreliable environment, such as air or space. Daily communication with digital cellular
phones (CDMA, GSM), high speed data modems (V. 32, V. 34), computer networks such as
Internet, TV broadcasting or weather forecasting through digital satellites, image and data
transmission to a space craft traveling in deep space as in the case of NASA’s Pluto mission
[1], optical recording in CD-ROMs are some examples of digital communication systems.
Since communication environments are unreliable in nature, errors may be introduced during
transmission. In order to minimize the effects of these transmission errors, encoder applies
certain techniques known as channel coding to add redundant bits to original information.
When information reaches the receiver, decoder makes use of these redundant bits to detect
and correct the errors in the received vector to obtain the original information. Work on
channel coding, which started in the 1950s, has focused on turbo codes (obtained by parallel
concatenation of two convolutional codes with an interleaver) and LDPC codes (described by
low–density parity–check matrices).
LDPC codes find wide application areas such as the wireless network standard (IEEE
802.11n), WiMax (IEEE 802.16e) and digital video broadcasting standard (DVB-S2) due to
their high error detection and correction capabilities. LDPC code family, first proposed by
Gallager in 1962, has sparse parity–check matrix representations [2]. In the following years,
LDPC codes were represented by Tanner graphs, which belong to a special type of bipartite
graphs that are intensively studied in graph theory [3, 4]. Sparsity property of the parity–check
matrix gives rise to the development of iterative message–passing decoding algorithms (such
as beilef propagation, Gallager A and B algorithms) on Tanner graph with low complexity
[5] – [8]. Ease of the application of iterative message–passing decoding algorithms brings the
advantage of low decoding latency.
Maximum likelihood (ML) decoding is the optimal decoding algorithm in terms of minimizing
error probability. Since ML decoding problem is known to be NP–hard, iterative message–
passing decoding algorithms for LDPC codes are preferred in practice [9]. However, these
2
heuristic decoding algorithms do not guarantee optimality of the decoded vector and they may
fail to decode correctly when the graph representing an LDPC code includes cycles. Feldman
et al. use optimization methods and they develop linear relaxation based maximum likelihood
decoding algorithms for LDPC and turbo codes in [10, 11]. However, the proposed models do
not allow decoding in an acceptable amount of time for codes with practical lengths.
Convolutional codes, first introduced by Elias in 1955, differ from block codes in that the
encoder contains memory and the encoder outputs, at any time unit, depend both on the
current inputs and on the previous input blocks [12]. Convolutional codes find application
areas such as deep–space and satellite communication starting from early 1970s. They can
be decoded with Viterbi algorithm, which provides maximum–likelihood decoding by dynamic
programming, by dividing the received vector into smaller blocks of bits. Although Viterbi
algorithm has a high decoding complexity for convolutional codes with long block lengths, it
can easily implemented on hardware due to its highly repetitive nature [13, 14]. For long block
lengths, sequential decoding algorithms such as Fano algorithm [15] and later stack algorithm
that is developed by Zigangirov [16] and independently by Jelinek [17] fit well. While Viterbi
algorithm finds the best codeword, sequential decoding is suboptimal since it focuses on a
certain number of likely codewords [18].
LDPC Convolutional (LDPC–C) codes, introduced by J. Feltström and Zigangirov in 1999,
are preferred to LDPC block codes in decoding for the cases where information is obtained
continuously. They can be decoded by sliding window decoders which implement iterative
decoding algorithms (such as belief propagation and density evaluation) at each window [19].
Although LDPC–C codes provide short–delay and low–complexity in decoding, they are not
in communication standards such as WiMax and DVB-S2 yet [20].
In this study, we consider LDPC–C codes and propose optimization based sliding window
decoders that can give a near optimal decoded codeword for a received vector of practical length
(approximately n = 4000) in an acceptable amount of time. The mathematical formulation
and proposed decoding algorithms are explained in Section 3. Our proposed decoders can
be used in a real–time reliable communication system since they have low decoding latency.
Besides, they are applicable in settings such as deep–space communication system due to their
3
high error correction capability.
The rest paper is organized as follows: we define the problem in more detail in the next
section. Section 3 explains the proposed decoding techniques. We give the corresponding
computational results in Section 4. Some concluding remarks and comments on future work
appear in Section 5.
2 Problem Definition
Digital communication systems transmit information from a sender to a receiver over a communication channel. Communication channels are unreliable environments, such as air, that
many sender–receiver pairs share. Hence, during transmission some of the transmitted symbols
can be lost or their values can change. In coding theory, information is encoded in order to
overcome the occurance of such errors during the transmission. Let the information to be sent
be represented by a k–bits long sequence u = u1 u2 ...uk (ui ∈ {0, 1}). In order to test whether
the information is sent correctly or not, parity bits are added by the encoder. This is done with
a k × n generator matrix G through the operation v = uG (mod 2). As a result, an n–bits
long (n ≥ k) codeword v = v1 v2 ...vn (vi ∈ {0, 1}) is obtained. Without loss of generality, we
can assume that the first k bits of the codeword are information bits, and the remaining n − k
bits are parity bits.
As shown in Figure 1, when the encoded information reaches receiver as an n–bits long
vector r, the correctness of the vector is tested at the decoder using the parity bits. If r is
detected to be erroneous, the decoder attempts to determine the locations of the errors and
fix them [21, 22]. Hence, the information u sent from the source is estimated as û at the sink.
4
Generator Matrix
G
Digital Source
u
Encoder
v
Noise
r
û
Digital Sink
Coding Channel
Decoder
Parity-Check Matrix
H
Figure 1: Digital communication system diagram
There are several models used to model the noisy communication channels. In our study,
we employ binary symmetric channel (BSC) model for noisy channel. As shown in Figure 2, a
transmitted bit is received correctly with probability 1 − p or an error occurs with probability
p [23].
0
1−p
0
p
p
1
1
1−p
Figure 2: Binary symmetric channel
In BSC, the received vector r includes both correct and incorrect bits. Although we do
not know which bits are erroneously received, flipping the bit fixes the error when the error
location is known. Hence, the aim of the decoder is to determine the error locations in BSC.
As explained above, original information u is encoded with k × n generator matrix G. Received vector r is decoded with a parity–check matrix H of dimension (n−k)×n. (J, K)−regular
5
LDPC codes are member of linear block codes that can be represented by a parity–check matrix H with J−many ones at each column and K−many ones at each row. An example of a
parity–check matrix from (3, 6)−regular LDPC code family is given in Figure 3.
1
0
H=
1
0
1
1 1 1 1 0 0 0 1 0
1
1
0
1
0 1 0 0 1 1 1 1
0 0 0 0 1 1 1 1
1 0 1 1 1 1 1 0
1 1 1 1 0 0 0 0
Figure 3: A parity–check matrix from (3, 6)−regular LDPC code family
Vectors v that satisfy the equation vHT = 0 (mod 2) are codewords. For any original
information u, encoded vector v = uG (mod 2) is a codeword. The channel decoder concludes
that whether the received vector r has changed or not by checking the value of expression rHT
is equal to vector 0 in (mod 2) or not [21].
LDPC codes can also be represented using Tanner graphs [8]. On one side of this bipartite
graph, there are n variable nodes standing for n codeword symbols of the code and on the
other side of the bipartite graph there are (n − k) check nodes corresponding to (n − k) parity–
check equations defined by each row of the H matrix. Here, H matrix is the bi–adjacency
matrix of Tanner graph. This representation of LDPC codes brings the advantage of applying
the iterative decoding and other decoding algorithms easily. Figure 4 shows Tanner graph
representation of H matrix defined in Figure 3.
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
c1
c2
c3
c4
c5
Figure 4: Tanner graph representation of the parity–check matrix given in Figure 3
6
In our study, we focus on LDPC–C codes. LDPC–C codes divide the original information
into smaller blocks and decode each block by considering the previous blocks [19]. In the
code, the nonzero elements are located on the diagonal as a ribbon and the code has infinite
dimension. As given in Figure 5 below, an LDPC–C code consists of ms –many small parity–
check matrices at each column, where ms parameter represents the width of the ribbon. The
diagonal pattern is obtained by shifting the columns down as the dimension increases.
H0 (1)
H1 (1)
H0 (2)
.
..
..
.
H1 (2)
..
..
H = Hms (1)
.
.
.
Hms (2) . .
H0 (L)
..
.
H1 (L)
..
..
.
.
.
Hms (L) . .
Figure 5: Generic structure of an LDPC–C code
These codes find application areas such as satellite communication and video streams where
the information is received continuously. Finite dimension LDPC–C codes, namely terminated
LDPC–C codes, can be obtained by limiting the dimension of LDPC–C code by specifying a
finite row or column size [24].
In Figure 6, an example of (3, 6)–regular terminated LDPC–C code obtained by limiting
the row size is given.
7
H=
010000000000000000000000000000000000
101000000000000000000000000000000000
010101000000000000000000000000000000
101010010000000000000000000000000000
011001100100000000000000000000000000
100110011001000000000000000000000000
000101101010010000000000000000000000
000010010101101000000000000000000000
000000100110010101000000000000000000
000000001001101010010000000000000000
000000000010011001100100000000000000
000000000000100110011001000000000000
000000000000001011010100010000000000
000000000000000010011001101000000000
000000000000000000100110010101000000
000000000000000000001001101010010000
000000000000000000000010011001100100
000000000000000000000000100110011001
Figure 6: A (3, 6)–regular terminated LDPC–C code
(3, 6)–regular structure cannot be seen for the first and the last parts of the code. For
example, the number of ones for the first five rows of the (3, 6)–regular code in Figure 6 is less
than 6. Similarly, number of ones is less than 3 in the last nine columns of the code. One can
observe the (3, 6)–regular structure for the intermediary rows and columns.
The repeating structure of LDPC–C codes allow the application of sliding window decoding
approaches which use iterative decoding algorithms (such as belief propagation, density evaluation, Gallager A and B) at each window [24]. Although iterative decoding algorithms are
easily applicable, they cannot guarantee that the solution is near optimal. They may even fail
to decode if the received vector includes errors.
Our goal in this study is to develop algorithms to decode a finite length received vector with
terminated LDPC–C codes on BSC. Then we generalize these decoding algorithms to decode
practically infinite length received vectors with LDPC–C codes. Our proposed decoders can
give a near optimal feasible decoding for any real sized received vector in acceptable amount
of time.
8
3 Solution Methods
We propose three different sliding window decoders for terminated LDPC–C codes and a
sliding window decoder for LDPC–C codes. The terminology used in this paper is summarized
in Table 1.
Table 1: List of symbols
Parameters
k
length of the original information
G
generator matrix
H
parity-check matrix
ŷ
received vector
n
length of the encoded information, # of columns in H
p
error probability in BSC
m
# of rows in base permutation matrix
ms
width of the ribbon of an LDPC–C code
C
set of check nodes
V
set of variable nodes
w
height of the window
hs
horizontal step size
vs
vertical step size
r
hs /vs ratio
Decision Variables
fi
ith bit of the decoded vector
kj
an auxiliary integer variable
3.1 Mathematical Formulation
The decoding problem of a terminated LDPC–C code can be represented with Exact Model
(EM) which is given in [25]. The columns and rows of a (n − k) × n parity-check matrix H of
a binary linear code can be represented with index sets V = {1, ..., n} and C = {1, ..., n − k},
respectively. In EM, Hij is the (i, j)−entry of parity-check matrix H, fi is a binary variable
denoting the value of the ith code bit and kj is an integer variable. Here, ŷ is the received
vector.
9
Exact Model:
min
X
(1 − fi ) +
i:ŷi =1
s.t.
X
i∈V
X
fi
(1)
i:ŷi =0
Hij fi = 2kj , ∀j ∈ C
(2)
fi ∈ {0, 1}, ∀i ∈ V,
(3)
kj ≥ 0, kj ∈ Z, ∀j ∈ C.
(4)
Constraints (2) guarantee that the decoded vector f satisfies the equality fHT = 0 (mod 2).
The objective (1) minimizes the Hamming distance between the decoded vector f and the
received vector ŷ. That is, the aim is to find the nearest codeword to the received vector.
Constraints (3) and (4) set the binary and integrality restrictions on decision variables f and
k, respectively.
Since EM is an integer programming formulation, it is not practical to obtain an optimal
decoding using commercial solver for real–sized (approximately n = 4000) terminated LDPC–
C codes. This can be seen from the computational experiments in Section 4. Instead, we will
look at terminated LDPC–C code in small windows and solve limited models at each window.
Note that LDPC–C code decoding problem cannot be represented as a compact mathematical formulation since this would require infinite number of decision variables fi and constraints.
3.2 LDPC–C Code Generation
We implement the terminated LDPC–C code generation scheme given in [26] which is also
explained in this section. We generate a terminated LDPC–C code with the help of a base
permutation matrix. As shown in Figure 7, by randomly permuting the columns of an s × s
identity matrix Is , we can obtain a (5, 10)–regular base permutation matrix of dimension
(m, 2m) where m = 5 × s. Regularity of the matrix is provided through augmenting identity
matrices 10 times at each row and 5 times at each column. In Figure 7, Iis represents the ith
randomly permuted identity matrix.
10
A
Hbase
I1s
11
Is
21
=
Is
31
Is
I41
s
I2s
I3s
I4s
I5s
I6s
I12
s
I13
s
I14
s
I15
s
I16
s
I7s
I8s
I9s
I10
s
I17
s
I18
s
I19
s
I20
s
B
I22
I23
I24
I25
I26
I27
I28
I29
I30
s
s
s
s
s
s
s
s
s
33 I34 I35 I36 I37 I38 I39 I40
I32
I
s
s
s
s
s
s
s
s
s
43 I44 I45 I46 I47 I48 I49 I50
I42
I
s
s
s
s
s
s
s
s
s
A
B
Figure 7: (5, 10)–regular base permutation matrix
Then, we split the base permutation matrix into two matrices, namely lower triangular A
and upper triangular B as shown in Figure 8.
hs
vs
=
Figure 8: A and B matrices
We divide Hbase with a horizontal step length hs and a vertical step length vs . One can
=
observe that when Hbase is (5, 10)–regular,
its dimension is (m, 2m) for some m. Then,
r = hs /vs = 2, since there is a number c such that hs c = 2m and vs c = m.
A
B
A
B
A
hs
Figure 9: (5, 10)–regular terminated
LDPC–C code
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
vs
11
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Horizontal Step
Verti
010000000000000000000000000000000000
101000000000000000000000000000000000
010000000000000000000000000000000000
101000000000000000000000000000000000
Then, these A and B matrices are repeatedly located until the desired terminated LDPC–C
code size is obtained. After t–many repetitions, terminated LDPC–C code has size (tm, 2tm)
as shown in Figure 9. The ribbon size is ms = m + vs for such a code.
3.3 Sliding Window Decoders
Sliding window decoders in practical applications make use of special structure of the LDPC–C
codes [27, 28]. As explained in Section 2, LDPC–C codes have all nonzero entries on a ribbon,
with width ms , that lies on the diagonal. Then, one can consider a window on the LDPC–C
code with height w and decode the received vector partially. Decoding of the received vector
proceeds iteratively by sliding the window hs units horizontally and vs units vertically.
In sliding window decoders, we can pick window row size w > ms and column size larger or
equal to rw where r = hs /vs . For the rows of the LDPC–C code corresponding to the window,
all entries in the columns after the window are zero with this window dimension selection.
Algorithm 1 explains the main steps of a generic sliding window decoder. Part of the received
vector corresponding to the current window is decoded with an algorithm. Hence, performance
of a sliding window decoder depends on how fast and correctly the windows are decoded. As
we mention in Section 2, if we implement Gallager A or B algorithm for windows, decoded
vector may not close to the original information. We investigate the performance of Gallager
A and B algorithms in sliding window decoder with computational experiments in Section 4.
Algorithm 1: (Generic Sliding Window)
Input: Received vector ŷ, Binary code H
1. Decode the current window with an algorithm
2. Move the window hs units horizontally, vs units vertically
3. Fix the decoded values of hs –many leaving bits
4. If all bits decoded, Then STOP, Else go to Step 1
Output: A decoded vector/codeword
12
In our approach, we solve each window with EM formulation that is written for the decision
variables and constraints within the window. At each iteration, hs –many bits and vs –many
constraints leave the window. Exiting bits are decoded in the previous window and can be
fixed to their decoded values in the proceeding iterations. The decoded bits will affect the
upcoming bits by appearing as a constant in the constraints (2). Our sliding window decoding
algorithm has main steps that are given in Algorithm 2.
Algorithm 2: (Sliding Window)
Input: Received vector ŷ, Binary code H
1. Solve EM for the current window
2. Move the window hs units horizontally, vs units vertically
3. Fix the decoded values of hs –many leaving bits
4. Update constraints (2) with the fixed bits
5. If all bits decoded, Then STOP, Else go to Step 1
Output: A decoded codeword
It is possible to apply different strategies in window dimension selection and window solution
generation. This gives rise to our four different sliding window decoders, i.e. complete window,
finite window and repeating windows decoders for terminated LDPC–C codes and an LDPC–C
code decoder, that are explained in the next sections.
3.3.1 Complete Window (CW) Decoder
Complete window (CW) decoder requires that binary code has finite dimension. Hence, it is
applicable only for terminated LDPC–C codes. In CW, the window height is w and width is
n (the length of the received vector ŷ). This means in a window we have w–many constraints
and n–many bits as fi decision variables.
We consider two diffrent ways in window decoding. In the first approach, i.e. Some Binary
CW (SBCW), we restrict the first undecoded hs bits of the window to be binary and relax the
bits coming after those as continuous variables. As an example, when we solve the first window
13
000000000000000000000000100110011001
000000000000000000000000100110011001
0100000000000
1010000000000
0101010000000
1010100100000
0110011001000
1001100110010
0001011010100
0000100101011
0000001001100
0000000010011
0000000000100
0000000000001
0 0 0 0 0 s0 0 0 0 0 0 0 0
0000000000000
0000000000000
0000000000000
0000000000000
0000000000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
0 0s0 0 0 0 0 0 0
000000000
000000000
000000000
100000000
001000000
0 1s 0 0 1 0 0 0 0
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
0 0 s0 0 0 0
000000
000000
010000
100100
011001
of the code in Figure 10, first h bits (corresponding to the dotted rectangle) are binary and we
relax all the remaining bits as continuous. When we move to the next window by shifting the
window v units down, first h bits have been fixed to their decoded values, the next h bits
are set to be binary and the bits coming after are continuous variables. The decoder proceeds
in this fashion.
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Figure 10: Sliding window in CW decoder
One can see that the dashed rectangle in Figure 10 covers all nonzero entries in the window.
From this observation as a second approach, i.e. All Binary CW (ABCW), we consider to force
the first undecoded (rw)–many bits (corresponding to the dashed rectangle) of the window to
be binary and the ones after these are continuous. As we move to the next window, hs –many
bits are fixed and the dashed rectangle shift to right hs units. Moving from one window to the
other requires removing first vs –many constraints and including new vs –many constraints.
The method of fixing some of the decision variables and relaxing some others is known as
Relax–and–Fix heuristic in the literature [29, 30]. In general, fixing the values of the variables
may lead to infeasibility in the next iterations. However, we do not observe such a situation
in our computational experiments when we pick the window that is sufficiently large to cover
all nonzero entries for the undecoded bits in the corresponding rows. We can observe that a
window of size w × (rw) (dashed rectangle) can cover the undecoded nonzero entries.
14
A
B
3.3.2 Finite Window (FW) Decoder
A
In finite window (FW) decoder, we have smaller window of size w × (rw). That is we have
B
w–many constraints and (rw)–many fi decision variables. At each iteration, after solving EM
model for the window, we fix first hs –many bits and slide the window. In Some Binary FW
A
(SBFW) decoder, we restrict first hs –many bits to be binary and relax the rest as continuous.
For All Binary FW (ABFW) method, all (rw)–many bits are binary variables.
hs
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
vs
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Horizontal Step
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
0 0i 0 0 0 0 0 0
00000000
00000000
00000000
00000000
0 1 in
0 0 0FW
0000
000000000000000000000000000
Figure
decoder
0 0 0 0 011:
0 0 0 0Sliding
0 0 0 0 0 0window
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Vertical Step
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
s
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
The window position can be seen in Figure 11 as the window slides. The previous decoded
bits appear as a constant in constraints (2) of EM formulation for the current window. In
FW, we store only one window model. This means we are storing w–many constraints and
(rw)–many f decision variables in the memory at a time.
As we move from one window to the other, we remove h –many decision variables and
introduce hs –many new ones. Also, we remove vs –many constraints and add vs –many new
constraints.
3.3.3 Repeating Windows (RW) Decoder
As explained in Section 3.2, a terminated LDPC–C code is obtained by repetitively locating A
and B matrices. As can be seen in Figure 12, a window will come out again after m–iterations,
where m is the number of rows in Hbase .
15
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
vs
hs
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Horizontal Step
00000000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
Figure
0 1 0in
0 0RW
0 0 0 0decoder
000000000000000000000000000
0 0 0 0 0 012:
0 0 0 Sliding
0 0 0 0 0 0 window
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
Vertical Step
01000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
This means that there are m–many different windows. However, the first and the (m + 1)st
windows still differ from each other in terms of their EM formulation. That is the constant
term in constraints (2) and the objective function coefficients change but the coefficients of the
decision variables stay the same. Hence, we store m–many window models and when its turn
comes we solve the window after updating the constant term and the objective function.
Assuming that a window is of size w×(rw), having m–many window models requires to store
(mw)–many
variables in the memory. However, we
0 1 0 0 0 0 0constraints
0 0 0 0 0 0 0 0 0 0and
0 0 0 0(mrw)–many
0 0 0 0 0 0 0 0 0 0 0 0f0i 0 decision
0
m
s
do
101000000000000000000000000000000000
010101000000000000000000000000000000
not
need to add or remove constraints and decision variables. FW decoder has the burden
101010010000000000000000000000000000
011001100100000000000000000000000000
1 0 0 1 1 0 0 1 1 operations
0 0 1 0 0 0 0 0 0 0and
0 0 0 0the
0 0 0advantage
0 0 0 0 0 0 0 0 0 of
0 low memory usage. On the other hand, RW
of add/remove
w
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
0 0 0 0s0 0
000000
000000
000000
010000
100100
011001
decoder directly calls the window models on the expense of memory.
In Some Binary RW (SBRW) only first h –many bits are binary, whereas All Binary RW
(ABRW) has all (rw)–many bits as binary variables.
3.3.4 LDPC Convolutional Code (CC) Decoder
hs
0 1 0 0 0 0 0 0CW,
0 0 0 0 0FW
0 0 0 0and
0000
0 0 0 assume
0 0 0 0 0 0 0that
0 0 0 0 we
0
The
RW
are given a finite dimensional code that can
vs decoders
10100
01010
10101
01100
10011
00010
00001
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000
00000000
10000000
00100000
11001000
00110010
11010100
00101011
01001100
00010011
00000100
00000001
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000000
10000000
01000000
10101000
01010010
11001100
00110011
00101101
00010010
00000100
00000001
00000000
00000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
000000000
100000000
001000000
010010000
101101000
110010101
001101010
010011001
000100110
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
000000
010000
100100
011001
be represented by a H matrix. Hence, they are applicable for terminated LDPC–C codes.
However, as explained in Section 2, LDPC–C codes are practically infinite dimensional codes
and cannot be represented by a compact H matrix on computer. On the other hand, they are
generated from A and B matrices. Therefore, we can store a part of LDPC–C code as given
16
in Figure 13 that includes the required information.
A
B
Figure 13: A part of LDPC–C code
With this part of the LDPC–C code, we can represent the (i, j)th entry of the code with a
function. Hence, we can represent the current window model using this small matrix. This
allows the application of FW and RW decoders to LDPC–C codes. Note that our CW decoder
is not applicable to CC, since it takes into account all bits of the received vector.
4 Computational Results
The computations have been carried out on a computer with 2.6 GHz Intel Core i5-3230M
processor and 4 GB of RAM working under Windows 10 Professional.
In our computational experiments, we evaluate the performance of our sliding window decoders. In our decoders, the number of the constraints and decision variables in EM formulation
limited with the size of the window. We make use of CPLEX 12.6.0 to solve EM for the current window (see Step 1 of Algorithm 2). We compare the performance of our sliding window
decoders with Exact Model Decoder (EMD). In EMD, EM formulation includes all constraints
and decision variables corresponding to terminated LDPC–C code. That is, for a terminated
LDPC–C code of size (n/2, n) we have n/2−many constraints (2) and n−many fi decision
variables in EM. We again utilize CPLEX for solving EM of EMD.
17
Table 2: Summary of methods
Method
# CPLEX
Models
Move to
Next Window
Window
Size
# Vars
kj
fi
CW
1
w×n
w
n
FW
1
w × rw
w
rw
RW
m
w × rw
mw
mrw
EMD
1
Delete/add vs –const.
Update const. (2)
Delete/add vs –const., hs –vars
Update const. (2)
Update obj. func. coeffs
Update const. (2)
—
(n/2) × n
(n/2)
n
# Vars
int./binary/cont.
SB
AB
SB
AB
SB
AB
—
w/hs /(n − hs )
w/rw/(n − rw)
w/hs /(rw − hs )
w/rw/0
mw/mhs /m(rw − hs )
mw/mrw/0
(n/2)/n/0
We summarize the solution methods in Table 2. “# CPLEX Models” gives the number of
CPLEX models stored in the memory. One needs to carry out operations given in “Move to
Next Window” column when sliding the window of size “Window Size”. In “# Vars” columns,
we list the number of kj and fi decision variables and also give the number of integer, binary
and continuous decision variables stored in the memory for SB and AB approaches of the
methods. CC decoder is not listed in Table 2, since it is the application of FW and RW
decoders to practically infinite dimensional codes.
Table 3: List of computational parameters
Parameters
n
1200, 3600, 6000, 8400, 12000
p
0.02 (low), 0.05 (high)
m
150
w
m + 1 (small),
hs
2
vs
1
3m
2
+ 1 (large)
A summary of the parameters that are used in the computational experiments are given in
Table 3. We generate a base permutation matrix of size (m, 2m) = (150, 300). We obtain a (5,
10)–regular terminated LDPC–C code H of desired dimensions from this base permutation matrix. In our experiments, we consider four different code length, i.e. n = 1200, 3600, 6000, 8400
for terminated LDPC–C codes. In order to test the algorithms for LDPC–C codes, we consider
a larger code length n = 12000. For each code length n, we experiment 10 random instances
and report the average values. We investigate two levels of error rate, i.e. low error p = 0.02
and high error p = 0.05. There are two alternatives for the window sizes, namely small window
18
w = m + 1 and large window w =
3m
2
+ 1.
In our sliding window algorithms, we solve the window models with CPLEX within 1 minute
time limit. On the other hand, we set a time limit of 4000 seconds to EMD for solving a
terminated LDPC–C code instance. Since we are testing a larger code length, i.e. n = 12000,
for LDPC–C codes, we set a time limit of 5000 seconds to EMD to find a solution.
Table 4: Performance of EMD with p = 0.02 and 0.05
p
0.02
0.05
n
z
CPU
Gap (%)
# OPT
z
CPU
Gap (%)
# OPT
1200
23.9
0.16
0
10
56.3
221.42
0
10
3600
72.7
0.23
0
10
736.0
1797.72
35.44
6
6000
121.0
0.32
0
10
1358.3
3890.06
43.37
4
8400
169.9
0.54
0
10
3623.7
4049.98
80.45
1
12000
238.6
0.85
0
10
4300.6
4457.34
70.84
2
Table 4 gives the performance of EMD under low and high error rates. The column “z”
shows the objection function value of the best known solution found within the time limitation.
“CPU” is the computational time in terms of seconds. “Gap (%)” is the relative difference
between the best lower and upper bounds. “# OPT” is the number of instances that are
solved to optimality among 10 trials. The first four rows in Table 4 are average results for
terminated LDPC–C codes. The last row is the average result for LDPC–C code. As the error
rate increases, EMD has difficulty in finding optimal solutions. A similar pattern is observed
when the length of the received vector n increases. That is, the optimality gap increases when
the code gets longer as expected.
4.1 Terminated LDPC–C Code Results
In this section, we discuss the results of the computational experiments of n = 1200, 3600, 6000, 8400
for error probabilities 0.02 and 0.05 and two levels of window size, i.e., small and large.
19
Table 5: Performances of SBCW and ABCW
p
SB
0.02
0.05
AB
0.02
0.05
w
n
z
CPU
small
Gap (%)
# SOLVED
z
CPU
large
Gap (%)
# SOLVED
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
23.9
72.7
121.0
169.9
56.3
181.8
564.2
2243.75
23.9
72.7
121.0
169.9
58.3
181.8
392.9
533.5
5.48
31.89
83.77
168.97
26.70
224.30
1165.43
2188.73
6.33
34.05
88.51
177.23
17.85
67.58
537.07
642.54
0
0
0
0
0
2.83
14.92
48.27
0
0
0
0
2.59
2.83
16.81
17.50
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
23.9
72.7
121.0
169.9
107.3
1052.1
2504.2
4016.5
23.9
72.7
121.0
169.9
56.3
614.9
1279.3
3119.4
6.31
39.05
97.54
187.91
334.28
1056.68
1419.37
1346.33
7.65
43.86
107.59
201.58
57.01
494.62
941.14
761.03
0
0
0
0
9.03
53.46
80.25
89.48
0
0
0
0
0
26.82
36.05
67.13
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Table 5 summarizes the results for CW decoder explained in Section 3.3.1. “Gap (%)”
column represents the percent difference from the best known lower bound found by CPLEX
while obtaining the results in Table 4. “# SOLVED” column shows the number of instances
that can be decoded by the method.
When p = 0.02, CW decoder can find optimal solutions as EMD in Table 4. However, CW
completes decoding in longer time for both SB and AB variants and both window sizes. This
is since solving EM model with CPLEX (in EMD) under low error probability is easy and
decoding in small windows takes longer time in CW. When the error probability increases to
0.05 and window size is small, we can see that CW finds better feasible solutions in shorter
time than EMD (in Table 4) for SB and AB variants. As the window size gets larger, only AB
alternative gives better gap and time values compared with EMD.
In general, with high error probability AB takes shorter time and obtains better gaps than SB
(see results for p = 0.05 in Table 5, Table 6 and Table 7). Note that this is somewhat counter
intuitive since the number of binary variables in AB variant is larger than SB. However, note
that AB has the advantage of being able to use the integral solution of the previous window
as a starting solution of the new window. Hence, AB has more time to find a better solution
20
in the current window within the time limit compared with SB.
When p = 0.05, the performance of CW deteriorates as the window size gets larger. Solving
a larger model in a window decreases the quality of the solution obtained within the time limit.
Size of the window model also depends on the length of the received vector n. Hence, the gap
values increase as n increases.
Table 6: Performances of SBFW and ABFW
p
SB
0.02
0.05
AB
0.02
0.05
w
n
z
CPU
small
Gap (%)
# SOLVED
z
CPU
large
Gap (%)
# SOLVED
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
23.9
72.7
121.0
169.9
56.3
196.6
353.9
629.8
23.9
72.7
121.0
169.9
58.3
214.6
368.1
617.8
10.91
24.25
41.98
62.48
15.07
348.29
973.76
3561.43
6.59
22.95
39.22
56.45
21.48
387.94
653.55
1792.25
0
0
0
0
0
5.89
12.89
26.65
0
0
0
0
2.59
10.97
16.05
25.96
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
23.9
72.7
121.0
169.9
56.3
177.0
300.3
427.0
23.9
72.7
121.0
169.9
56.3
177.0
300.3
427.0
11.90
47.47
74.04
129.44
364.67
3581.46
6889.76
11445.70
15.80
65.66
114.59
165.71
72.78
999.93
2087.41
3061.82
0
0
0
0
0
0.78
0.86
1.03
0
0
0
0
0
0.78
0.86
1.03
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Results given in Table 6 shows that FW (see Section 3.3.2) can find optimal solution in all
cases when p = 0.02. With this error probability, FW needs more time to find the optimal
solution for SB and AB alternatives when the window size gets larger. The computational
times are larger than EMD for both alternatives.
However, as error probability gets higher, FW can find better solutions than EMD in shorter
time for SB and AB methods. AB method is faster than SB, since it can make use of integral
solution found in the previous window. Note that a similar pattern also appears in CW as
discussed before. FW takes more time than CW for both SB and AB alternatives, since it
needs to add and remove variables while moving to the next window position. On the other
hand, the size of the window model is independent from code length n, hence we can find
better solutions within the time limit. As a result, the gap values are better than CW decoder.
21
Table 7: Performances of SBRW and ABRW
p
SB
0.02
0.05
AB
0.02
0.05
w
n
z
CPU
small
Gap (%)
# SOLVED
z
CPU
large
Gap (%)
# SOLVED
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
1200
3600
6000
8400
23.9
72.7
168.9
234.2
75.2
269.9
628.5
999.0
23.9
72.7
121.0
169.9
56.3
217.6
369.1
616.2
6.52
24.42
766.64
1088.18
330.79
2094.05
7330.18
13140.1
6.25
24.78
42.41
61.40
10.24
674.08
845.82
2806.38
0
0
11.17
7.86
18.42
27.82
50.52
57.29
0
0
0
0
0
11.35
16.20
25.81
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
23.9
72.7
122.1
168.8
109.0
480.5
–
–
23.9
72.7
121.0
169.9
56.3
175.9
300.3
432.5
55.84
258.60
410.82
438.66
6542.20
74868.29
–
–
8.77
38.88
67.53
98.08
155.94
1498.80
5067.35
9445.67
0
0
0
0
35.10
62.15
–
–
0
0
0
0
0
0.68
0.86
1.23
10
10
9
7
10
5
0
0
10
10
10
10
10
8
10
8
We also observe that, at p = 0.05 increasing the window size improves the gap values in
contrast to CW decoder. In FW decoder, although the window size does not depend on n,
gap values still depend on n due to error accumulation during the iterations. That is, if a
window is not decoded optimally, this near optimal window solution will propagate to the
upcoming window decodings. As the code length n gets larger, this effect becomes more
apparent and the gap values increases. If the window size is larger, then we are considering
more information during the window decoding, which improves the gap values. This effect is
explained graphically in Figure 14.
From Table 7, we can see that RW cannot complete decoding at all cases. RW decoder stores
m–CPLEX models in memory and CPLEX needs additional memory for branch–and–bound
tree while solving the window model. Hence, when the window size gets larger, we see that
memory is not sufficient to complete the iterations for some instances.
Comparison of Tables 6 and 7 shows that ABFW and ABRW methods give similar gap
values as expected. However, ABRW method requires more time to manage window models.
As the window size gets larger, the computational time of ABRW is even worse than EMD (in
Table 3) with high error probability.
22
4.2 LDPC–C Code Results
We also investigate the performance of FW and RW decoders for very large code length. For
this purpose we take n = 12000 and consider high (p = 0.05) error probability, small and
large window sizes. CW method is inapplicable in practice for very large code lengths, since it
includes all the bits of the codeword as a decision variable to the window model. Performance
of EMD for n = 12000 is given in the last row of Table 4.
Table 8: Performances of FW and RW decoders
w
FW
RW
SB
AB
SB
AB
z
CPU
small
Gap (%)
# SOLVED
z
CPU
large
Gap (%)
# SOLVED
926.6
990.2
1434.1
907.0
5796.79
3686.91
20013.34
4537.75
30.93
34.03
58.21
27.74
10
10
10
10
597.6
597.6
–
596.4
14749.46
3999.73
–
5428.56
0.77
0.77
–
0.45
10
10
0
5
Table 8 summarizes the average results of 10 instances for FW and RW decoders with SB
and AB alternatives. When we have small window size, all methods can decode the received
vector. Among all, ABFW completed decoding within shortest time.
When the window size gets larger, RW decoder cannot solve all instances due to memory
limit. On the other hand, FW decoder can solve all instances with better gap values compared
with the small window size. ABFW takes less time by making use of integral starting solution
advantage over SBFW. Moreover, compared with the EMD (last row of Table 4), ABFW
finds near optimal solutions in shorter time for all instances. However, EMD can solve only 2
instances to optimality. For the 5 instances that ABRW can decode, ABFW and ABRW get
the same objective values. For these cases, ABFW is faster than ABRW as expected.
Considering the computational results for LDPC–C codes, we can see that ABFW is the
best alternative for decoding process in terms of both time and solution quality. We further
evaluate the performances of the methods by analyzing their decoding errors with respect to
the original vector as given in Figure 14.
23
14
ABFW, w small
ABFW, w large
ABRW, w small
ABRW, w large
12
Error (bits)
10
8
6
4
2
0
0
20
40
60
80
100
120
Section
Figure 14: Error accumulation in decoding
In this figure, the average decoding errors of the 10 instances for code length n = 12000
with error probability p = 0.05 are given. We divide n into 120 sections each include 100 bits.
For each section, average errors from the original code vector is plotted. When the window
size is small, average error gets larger as the iterations proceeds. That is when we make error
in decoding in early steps of the decoding process, this error will increase the probability that
we are decoding erroneously in the upcoming windows. On the other hand, when the window
size gets larger, we have more information about the LDPC–C code, which decreases the
error accumulation during the iterations. However, taking a large window size requires more
decoding time. As a result, one should take into account the trade off between computational
time and the solution quality when deciding on the window size.
The performance of decoding algorithms are interpreted with Bit Error Rate (BER) in
telecommunications literature. BER is the percentage of the decoded bits that are different
than the original vector [23].
24
Pn
BER =
i=0
| yio − yid |
× 100
n
(5)
BER can be calculated with the formula given in equation (5), where yo is the original and
yd is the decoded codeword.
Table 9: BER of Sliding Window Decoders
w
ABFW
ABRW
small
6.918
5.402
large
0.008
0.003
We can calculate the BER values for our decoding algorithms using the data of Figure 15.
The BER results given in Table 9 show that the error correction capability increases when we
have larger window. For example, among 100 bits of the codeword that is decoded by ABFW
method, approximately 7 bits (% 6.918) are different from the original codeword when window
size is small. As the window size gets larger, this difference drops to 8 bits among 100,000 bits
(% 0.008).
In our final experiment, our goal is to compare our proposed decoding algorithms with two
commonly used algorithms. In practical applications, decoding of a received vector is done
with iterative algorithms. Among these Gallager A and B algorithms are popular due to their
ease of application [31, 32]. The performance of our proposed decoding algorithm (ABFW) can
be tested against a sliding window decoder that uses Gallager A or B algorithm for decoding
windows (see Algorithm 1).
Gallager A and B algorithms are quite similar. For each bit of the received codeword ŷ,
the algorithm collects messages, which are the values of the parity–check equations, from each
check node. If the neighboring check node is unsatisfied, then this is considered as an indication
of an error in the corresponding bit. If most of the neighbors of a bit are unsatisfied, we have
a strong intuition that the bit is erroneous. Let di be the number of neighbors of variable node
i in the Tanner graph of the code.
As given in Algorithm 3, Gallager A algorithm prefers to flip the bit that has the maximum
number of unsatisfied checks. At each iteration of the algorithm, we flip only one bit which
guarantees that the number of unsatisfied check nodes will decrease at each iteration. Gallager
25
B algorithm decides whether to flip or not each bit at an iteration. For each bit, Gallager
B flips the bit if the number of unsatisfied check nodes is larger than the satisfied ones. In
Gallager B algorithm, decrease in the unsatisfied check nodes at each iteration is not for sure
since it applies multi–flip at an iteration.
Algorithm 3: (Gallager A and B)
Input: Received vector, ŷ
1. Calculate all parity–check equations
2. If all check nodes are satisfied, Then STOP.
3. Else Calculate the number of all unsatisfied parity–check
equations for all received bits, say ui for bit i.
4 - A. Let l = argmaxi {ui }. If ul > dl /2, Then flip bit l.
4 - B. If ui > di /2, Then flip bit i.
5. End If
6. If stopping is satisfied, Then STOP.
7. Else Go to Step 1.
8. End If
Output: A feasible decoded codeword, or no solution
We apply Gallager algorithm at each window of the sliding window algorithm instead of
solving window model with CPLEX. A known problem with these algorithms is that they may
get stuck when there is a cycle in the LDPC code [33]. In such a case, the algorithm may
terminate with no conclusion. To avoid such a situation, we take the stopping criterion as the
number of iterations and bound it with value 100. Note that this may result in ending with
an infeasible solution when the algorithm terminates.
Table 10: Performance of Gallager A
p
0.02
0.05
w
n
z
1200
3600
6000
8400
12000
1200
3600
6000
8400
12000
159.2
213.1
268.3
323.9
395.1
191.3
348.9
518.7
684.8
917.7
CPU
small
Gap (%)
# FEAS
BER
z
10.69
49.32
99.14
159.23
261.76
11.44
52.08
102.85
163.59
252.25
84.99
65.99
54.96
47.63
39.65
70.62
49.66
42.59
38.83
35.35
0
0
0
0
0
0
0
0
0
0
11.94
4.33
2.81
2.21
1.67
15.43
10.57
9.96
9.64
9.22
234.1
276.9
338.2
386.4
451.5
259.1
387.5
539.3
684.8
879.1
26
CPU
large
Gap (%)
# FEAS
BER
20.79
99.38
192.15
299.31
506.70
20.74
99.06
192.02
299.04
485.53
89.79
73.81
64.27
56.12
47.22
78.31
54.69
44.71
38.74
32.39
0
0
0
0
0
0
0
0
0
0
18.23
5.89
3.74
2.72
1.85
19.7
10.13
8.78
7.99
7.11
Table 10 shows the average of 10 instances with Gallager A algorithm when it is applied in
the windows of sliding window decoder. Gallager A algorithm cannot find a feasible solution
for any of the cases, as given in “# FEAS” column. That is the decoded vector does not satisfy
the equality vHT = 0 (mod 2). Besides, decoded vectors are far away from the best known
lower bounds (found by CPLEX while obtaining the results in Table 4) which can be seen from
the “Gap (%)” column.
“BER” column shows the percent difference from the original codeword. When the values
compared with the ones in Table 9 for n = 12000 and p = 0.05, our proposed ABFW algorithm
provides significantly higher quality solutions compared to Gallager A.
Table 11: Performance of Gallager B
p
0.02
0.05
w
n
z
CPU
small
Gap (%)
# FEAS
BER
z
CPU
large
Gap (%)
# FEAS
BER
1200
3600
6000
8400
12000
1200
3600
6000
8400
12000
174.2
781.1
1854.7
3037.8
4816.4
519.4
1704.5
2889.0
4073.6
5847.9
10.75
50.46
93.83
149.39
250.38
11.63
47.87
94.72
150.02
251.29
86.15
85.63
92.02
94.11
94.95
89.01
89.69
89.69
89.71
89.85
0
0
0
0
0
0
0
0
0
0
13.55
20.96
30.49
35.85
39.95
42.87
47.22
48.09
48.49
48.71
469.7
1645.2
2846.3
4056.3
5865.5
591.4
1791.3
2983.8
4184.2
5973.1
21.19
100.27
193.40
301.41
489.15
21.11
100.43
194.52
302.27
489.25
94.69
95.53
95.74
95.81
95.93
90.48
90.19
90.02
89.99
90.07
0
0
0
0
0
0
0
0
0
0
38.67
45.51
47.34
48.19
48.83
49.29
49.69
49.88
50.03
49.96
As summarized Table 11, BER values are high since on the contrary to Gallager A algorithm,
Gallager B does not guarantee to decrease the error as its iterations proceed. That is error
accumulation effect appears in BER results more dramatically for Gallager B. Both Gallager
A and B algorithms are faster than ABFW method. However, their solutions are usually not
feasible and are distant from the best known lower bound.
These results indicate that ABFW is a strong candidate for decoding problem in communication systems. Gallager A and B algorithms give quick but poor quality solutions. These
algorithms may be practical for TV broadcasting and video streams since fast decoding is
crucial for these applications. On the other hand, as in the case of NASA’s Mission Pluto,
we may have some received information that cannot be reobtained from the source. For such
cases high solution quality is the key issue instead of decoding speed. Hence, ABFW method
27
is more practical for these kind of communication systems.
5 Conclusions
We proposed optimization–based sliding window decoders for terminated LDPC–C codes,
namely complete window (CW), finite window (FW), repeating windows (RW) decoders.
We explained how one can utilize these algorithms to practically decode infinite dimensional
LDPC–C codes and introduce LDPC convolutional code (CC) decoder. The computational
results indicate that within the given time limit sliding window decoders find better feasible solutions in shorter time compared with exact model decoder (EMD). For each proposed
decoder, we implement some binary (SB) and all binary (AB) variants. Among the sliding
window decoders, AB approach is better than SB due to starting solution advantage.
For the decoding of convolutional codes, our proposed ABFW algorithm is the best among
all methods in terms of both computational time and solution quality. One can obtain better
solutions by increasing the window size in the expense of computational time.
Although, RW approach reveals worse performance than FW method, it can still be a nice
candidate to decode time invariant LDPC–C codes where all windows are same. In such a
case, one needs to store a single window model instead of m. This can decrease the memory
usage and improve the computational time.
Gallager A and B algorithms are popular in practical applications. Compared with ABFW
approach, these algorithms give poor quality solution in shorter time. Our proposed algorithm
ABFW can contribute to the communication system reliability by providing near optimal
decoded codewords. It is applicable in settings such as deep space communications where
obtaining a high–quality decoding within reasonable amount of time is crucial.
Acknowledgements
This research has been supported by the Turkish Scientific and Technological Research Council
with grant no 113M499.
28
References
[1] C. C. DeBoy et al., “The RF Telecommunications System for the New Horizons Mission
to Pluto”, in Proc. 2004 IEEE Aerospace Conf., pp. 1463–1478.
[2] R. G., Gallager, “Low-density parity-check codes,” IRE Trans. on Information Theory,
vol. 8, no. 1, pp. 21–28, January 1962.
[3] R., Diestel, Graph Theory. 4th ed. Berlin, Germany: Springer-Verlag, June 2010.
[4] J. P., van den Berg, and A. J. R. M., Gademann, “Optimal routing in an automated
storage/retrieval system with dedicated storage,” IIE Transactions, vol 31, pp. 407–415,
1999.
[5] V., Guruswami, “Iterative decoding of low–density parity–check codes”, Bulletin of the
European Association for Theoretical Computer Science (EATCS), Issue 90, October 2006.
[6] F. R., Kschischang, B. J. Frey, and H. A., Loeliger, “Factor graphs and the sum–product
algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 498–519, February 2001.
[7] J. S., Yedidia, W. T., Freeman, and Y., Weiss, “Understanding belief propagation and
its generalizations,” Mitsubishi Electric Res. Labs, Cambridge, Tech. Rep., TR2001-22,
November 2001.
[8] R. M., Tanner, “A recursive approach to low complexity codes,” IEEE Trans. on Inf.
Theory, vol IT-27, no. 5, pp. 533–547, September 1981.
[9] E. R., Berlekamp, R. J., McEliece, and H. C. A., van Tilborg, “On the inherent intractability of certain coding problems,” IEEE Trans. Inf. Theory, vol. 24, pp. 384–386,
May 1978.
[10] J., Feldman, and D. R., Karger, “Decoding turbo-like codes via linear programming,” J.
Computer and System Sciences, vol 68, pp. 733-752, 2004.
[11] J., Feldman, M. J., Wainwright, and D. R., Karger, “Using linear programming to decode
binary linear codes,” IEEE Trans. on Information Theory, vol 51, no. 3, pp. 954–972,
March 2005.
[12] P., Elias, “Coding for Noisy Channels”, MIT Res. Lab. of Electronics, Cambridge, MA,
IRE Convention Rec., Part 4, pp. 37–46, 1955.
29
[13] A. J., Viterbi, “Error bounds for convolutional codes and an asymptotically optimum
decoding algorithm”, IEEE Trans. on Information Theory, vol. 13, no. 2, pp. 260–269,
April 1967.
[14] P., Brice, W. Jiang, and G. Wan, “A Cluster-Based Context-Tree Model for Multivariate
Data Streams with Applications to Anomaly Detection”, INFORMS J. on Computing,
vol. 23, no. 3, pp. 364–376, September 2010.
[15] R. M., Fano, “A heuristic discussion of probabilistic decoding,” IEEE Trans. Inform.
Theory, vol. IT-9, no. 2, pp. 64–73, April 1963.
[16] K., Sh. Zigangirov, “Some sequential decoding procedures,” Probl. Peredachi Inf., 2, pp.
13–25, 1966.
[17] F., Jelinek, “A fast sequential decoding algorithm using a stack,” IBM J. Res. and Dev.,
13, pp. 675–685, November 1969.
[18] Y. S., Han, and P.-N., Chen, “Sequential decoding of convolutional codes” in Wiley Encyclopedia of Telecommunications, Wiley, 2003.
[19] A., Jimenez-Feltström and K., Sh. Zigangirov, “Time-varying periodic convolutional codes
with low-density parity-check matrix,” IEEE Trans. on Information Theory, vol. 45, pp.
2181–2191, 1999.
[20] I. E. Bocharova, B. D. Kudryashov, and R. Johannesson, “LDPC convolutional codes
versus QC LDPC block codes in communication standard scenarios”, in 2014 IEEE Int.
Symp. on Information Theory, pp. 2774–2778.
[21] T. K., Moon, Error correction coding: mathematical methods and algorithms. New Jersey:
Wiley, 2005.
[22] I., Ben-Gal, Y. T., Herer, and T., Raz, “Self-correcting inspection procedure under inspection errors”, IIE Transactions, vol. 34, pp. 529–540, 2002.
[23] D. J. C., MacKay, Information theory, inference, and learning algorithms. Cambridge,
United Kingdom: Cambridge Univ. Press, 2003.
[24] S., Kudekar, T. J., Richardson, and R. L., Urbanke, “Threshold saturation via spatial
coupling: why convolutional LDPC ensembles perform so well over the BEC,” IEEE
Trans. on Information Theory, vol 57, no. 2, pp. 803–834, February 2011.
30
[25] A. B., Keha, and T. M., Duman, “Minimum distance computation of LDPC codes using
branch and cut algorithm,” IEEE Trans. on Communications, vol 58, no. 4, pp. 1072–1079,
2010.
[26] R., Ashrafi, and A. E., Pusane, “Spatially-coupled communication system for the correlated erasure channel,” IET Communications, vol. 7, no. 8, pp. 755–765, May 2013.
[27] M., Lentmaier, A., Sridharan, D. J., Costello, Jr., and K., Sh. Zigangirov, “Iterative
decoding threshold analysis for LDPC convolutional codes,” IEEE Trans. on Information
Theory, vol. 56, pp. 5274–5289, 2010.
[28] D. J., Costello, Jr., A. E., Pusane, S., Bates, K., Sh. Zigangirov, “A comparison between
LDPC block and convolutional codes,” Proc. Inf. Theory and Applications Workshop, San
Diego, CA, USA, 2006.
[29] L. A., Wolsey, Integer programming. New Jersey: Wiley, 1998, pp. 215–216.
[30] N. K., Freeman,J., Mittenthal, and S. H., Melouk, “Parallel-machine scheduling to minimize overtime and waste costs,” IIE Transactions, vol 46, pp. 601–618, 2014.
[31] B. M. J., Leiner, “LDPC codes - a brief tutorial,” Wien Technical University, 2005.
[32] C. C., Moallemi, and B., Van Roy, “Resource Allocation via Message Passing,” INFORMS
J. on Computing, vol 23, no. 2, pp. 205–219, July 2010.
[33] A., Sarıduman, A. E., Pusane, and Z. C., Taşkın, “An integer programming-based search
technique for error-prone structures of LDPC codes,” AEU - Int. J. of Electron. and
Commun., vol. 68, no. 11, pp. 1097-1105, November 2014.
31
| 7cs.IT
|
Optimal Stochastic Package Delivery Planning with
Deadline: A Cardinality Minimization in Routing
arXiv:1803.02232v1 [cs.AI] 28 Feb 2018
∗
Suttinee Sawadsitang∗, Siwei Jiang†, Dusit Niyato∗ , Ping Wang∗
School of Computer Science and Engineering, Nanyang Technological University
† Singapore Institute of Manufacturing Technology (SIMTech) A*STAR
Abstract—Vehicle Routing Problem with Private fleet and
common Carrier (VRPPC) has been proposed to help a supplier
manage package delivery services from a single depot to multiple
customers. Most of the existing VRPPC works consider deterministic parameters which may not be practical and uncertainty
has to be taken into account. In this paper, we propose the
Optimal Stochastic Delivery Planning with Deadline (ODPD) to
help a supplier plan and optimize the package delivery. The aim
of ODPD is to service all customers within a given deadline
while considering the randomness in customer demands and
traveling time. We formulate the ODPD as a stochastic integer
programming, and use the cardinality minimization approach
for calculating the deadline violation probability. To accelerate
computation, the L-shaped decomposition method is adopted. We
conduct extensive performance evaluation based on real customer
locations and traveling time from Google Map.
Index Terms—Full-truckload, less-than-truckload, stochastic
programming, cardinality, decomposition, L-shaped.
I. I NTRODUCTION
To achieve a higher profit and better standard of customer
satisfaction, suppliers need to meticulously plan their delivery.
The suppliers usually use one of the two delivery modes,
which are (i) to deliver customer packages by themselves
known as full-truckload (FTL) mode and (ii) to outsource
customer package delivery to a third-party carrier known as
less-than-truckload (LTL) mode. For the FTL mode, suppliers
are required to pay a truck rental, driver stipend, fuel, and other
expenses. For the LTL mode, the carriers charge suppliers
based on actual package size or weight in an on-demand
basic. However, the FTL mode is more economical when the
utilization of the FTL trucks is high. The Vehicle Routing
Problem (VRP) that combines both FTL and LTL is usually
referred to as Vehicle Routing Problem with Private fleet and
common Carrier (VRPPC).
In VRPPC, suppliers need to reserve FTL trucks in advance
before the exact customer demand is observed. The under- and
over-reservation of FTL trucks may occur due to unknown
customer demand. Furthermore, the suppliers need to limit
their delivery time, i.e., setting the delivery deadline, to save
cost, e.g., overtime driver pay and other extra charges. The
uncertainty in road traffic, weather, road-construction, and
other factors may cause FTL trucks to miss their delivery
deadline. To address this issue, one approach is to use a large
set of historical customer demand and traveling time which
can improve the accuracy of the vehicle routing solution.
However, utilizing a large set of historical data incurs more
computational time and the problem becomes intractable in
practice.
In this paper, we aim to help a supplier to minimize their
payment of delivering packages from a single warehouse,
i.e., a depot, to multiple customer addresses by using either
FTL, LTL, or the combination. The delivery has a specific
deadline, and missing the deadline will incur a penalty cost
to the supplier. Therefore, we propose the Optimal Stochastic
Deliver Planning with Deadline (ODPD) to address the aforementioned challenges. The uncertainties in customer demand
and delivery time are estimated by historical data. We consider
the uncertainty in customer demand in the form of a traditional
stochastic integer programming. Alternatively, we consider the
uncertainty of traveling time in the form of the cardinality
minimization problem. Based on this problem representation,
the proposed ODPD does not rely on the strong assumptions,
which may not be practical in some cases. These strong
assumptions are commonly used in existing works. Such
assumptions are (i) the traveling time for each path must follow
the normal distribution, (ii) the traveling time of each path
must be independent to each other, and (iii) the deadline should
be longer than the total traveling time. Moreover, to reduce
the computational time of obtaining an optimal solution, we
decompose the ODPD into small sub-problems by L-shaped
decomposition [9]. Finally, the performance evaluation of the
ODPD is conducted based on the real dataset obtained from
real carrier company and traveling time from Google Map.
Compared with the baseline methods, the ODPD can achieve
lower deadline violation probability. The execution time of
solving the decomposed ODPD is significantly lower than that
of the original (nondecomposed) problem.
II. R ELATED W ORK
In VRP, a traveling time of a path is random because of road
traffic and other factors. The VRP with deterministic traveling
time may not be desirable in practical applications. Instead,
VRP with random traveling time can be modeled as the
stochastic shortest path problem with least expected traveling
time (LET) as in [3], [4], [5]. The authors in [3] assumed
that the weight of each path is independent and follows timeinvariant distribution. They proposed the Expected Value (EV)
algorithm and the Expected Lower Bound (ELB) algorithm
to solve the problem. From their assumptions, the problem
can be regarded as a deterministic shortest path by using the
expected values as path weights. Taking different approach, the
authors in [4] proposed a new criterion, which is to reduce the
probability of the late arrival instead of to minimize the expected traveling time. This criterion relaxes the problem while
keeping the problem practical when people actually plan their
journey. Later on, the authors in [5] adopted the shortest path
with deadline violation probability problem from [4] and formulated the problem as the cardinality minimization problem.
They transformed the cardinality minimization problem into
L1-norm minimization and solved it as a mix-integer linear
programming. By using cardinality minimization approach, the
proposed algorithm can avoid the major assumptions in [3]
i.e., the weight of each path is independent and follows timeinvariant distribution. However, the L1-norm minimization
does not guarantee to return the optimal result because L1norm minimization in [5] is to minimize the summation of
the exceeding time from all samples.
To enhance the quality of business strategies, many studies
extend the shortest path problem by combining it with business
planning constraints, e.g. VRPPC, VRP with time window,
multi-depot VRP, and other aspects [2]. The VRPPC has been
earlier introduced in [6]. Subsequently, the authors in [7]
proposed “Selection, Routing and Improvement Algorithms”
(SRI), which use λ-interchange procedure. The authors in [8]
proposed the “Iterated Density Estimation Evolutionary Algorithm (IDEA)”, which achieves better computational performance and more accurate results compared to the SRI
algorithms. Most of the VRPPC studies were dedicated to improving the quality of the heuristic solver and considered only
dynamic parameters. In our previous work [1], we considered
different aspect of VRPPC, which was to focus on modeling
and formulating the VRPPC problem. The Optimal Delivery
Planning (ODP) was proposed to handle stochastic customer
demand. However, the ODP problem becomes intractable
easily when the number of trucks, carriers, customers, and
especially demand scenarios increase.
For further improvement of the supplier planning, in this
paper, we propose the Optimal Delivery Planning with Deadline (ODPD) as an extension of ODP, where the supplier can
limit the delivery time. In practice, the uncertainty in traveling
time is also unavoidable. Thus, we reformulate the system
model as the cardinality minimization problem to handle
this randomness. Since the solution of L1-norm formulation
in [5] does not guarantee to be optimal, we directly minimize
the cardinality of all samples instead of minimize the total
exceeding time of all samples. Furthermore, to speed up the
computation, the L-shaped decomposition [9] is applied into
the ODPD.
III. S YSTEM M ODEL AND A SSUMPTIONS
This section presents the system model for the ODPD. The
ODPD is formulated as a two stage stochastic programming
model in (1) and (2). The first stage and second stage decisions
are made before and after observing the actual customer
demand, respectively.
First stage: The ODPD helps a supplier decide how to
reserve FTL trucks based on FTL truck information and
customer demand probability.
Second stage: The ODPD assigns all customers with
non-zero demand to either an outsource LTL carrier or
the reserved FTL truck. Additionally, the ODPD finds
the best delivery routing of the FTL trucks. The routing
is calculated based on both traveling distance and the
probability of serving all assigned customers before the
deadline.
The supplier has a set of customers, which is represented
as C = {C1 , C2 , . . . , Cn′ }, where n′ is the total number of
customers. Customer Ci may or may not have demand. The demand of the customer Ci is denoted as Di , where Di = 1 if the
customer i has demand, and Di = 0 otherwise. Let the demand
scenario be a collection of all customers’ demand represented
as ω = (D1 , D2 , . . . , Dn′ ). Let Ω = {ω1 , ω2 , . . . , ωq′ } be a set
of demand scenarios, where q ′ is the total number of scenarios.
The list of package weight of each customer is represented as
A = (A1 , A2 , . . . , An′ ).
A set of FTL trucks is represented as T = {T1 , T2 , . . . , Tt′ }
where t′ denotes the total number of FTL trucks. Each
truck has its own capacity, which is denoted as F =
(F1 , F2 , . . . , Ft′ ). Let T denote a deadline or an instant of time
that all FTL trucks aim to finish the delivery. The supplier
may outsource third-party LTL carriers to deliver packages.
Let R = {R1 , R2 , . . . , Rr′ } represent a set of LTL carriers,
where r′ is the total number of LTL carriers. Note that the
routing of FTL trucks is determined by the ODPD while LTL
carriers are responsible for finding the routing by themselves.
The traveling distance and traveling time from location u to
location v are represented as Ku,v and Qu,v , respectively.
Without loss of generality, let u and v be taken from the set of
customer locations and the supplier depot, which is denoted
as U. In this paper, both traveling distance and the probability
of deadline violation are used to calculate the best routing of
the FTL trucks.
There are four different payments involved in the ODPD,
bi,r is the service
i.e., C̄t is the initial cost of FTL truck t, C
charge of LTL carrier r to serve customer i, C̈u,v is the routing
cost of FTL truck t from location u to location v, and Ć is
the penalty cost that incurs when the delivery time exceeds
the deadline T.
IV. P ROBLEM F ORMULATION
In this section, we present the original ODPD problem
formulation. Then, we reformulate the original ODPD as a
cardinality minimization problem. Here, the total traveling
time is from when the FTL vehicle leaves the depot to when
it serves all customers and comes back to the depot again.
TABLE I: List of key notations
Symbols
Main Sets
T
C
U
R
Parameter
Ft
Ai
Di
Ku,v
Qu,v
T
Pricing
C̄t
bi,r
C
C̈u,v
Ć
Uncertainty
Ω
of FTL trucks, (ii) LTL carrier service charge, (iii) the cost
of FTL vehicle routing, and (iv) the penalty cost due to the
delivery time exceeding the deadline. Note that the last term in
(2) is to minimize the probability of arriving at the destination
after the deadline T. The penalty cost is calculated based
on this deadline violation probability. The deadline violation
probability
Pof truck
P t under demand scenario ω is expressed
as P rob( u∈U v∈U Qu,v Vu,v,t (ω) ≥ T). P rob(ω) denotes
the probability of demand scenario ω.
The constraints in (3), (4), and (5) are controlling the
assignment, limiting the capacity of each truck, and forcing
the objective function to include the initial cost of each truck,
respectively. Note that ∆ denotes a large scalar number. In
addition, (6) to (11) are the routing constraints.
X
X
Xi,t +
Yi,r (ω) ≥ Di (ω),
i ∈ C, ω ∈ Ω
(3)
Description
Set of private trucks, where t ∈ T is truck’s index
Set of customers, where i, j ∈ C are customers’ indexes
Set of routing locations C ∪ {depot}, where
u, v ∈ U are locations’ indexes
Set of carriers, where r ∈ R is carrier’s index
Capacity of truck t
Weight of the customer i’s package
Demand indicator of customer i
Distance between locations u and v in kilometer
Traveling time between locations u and v
Deadline
Initial cost of FTL truck t
Service charge of the carrier r that serves customer i
Cost of FTL truck t routing from location u to v
Penalty cost of the deadline violation probability
Set of demand scenarios, where ω ∈ Ω is the scenario’s
index
P rob(ω)
Probability of the scenario ω to happen
Decision Variables
Xi,t
A binary variable for allocating FTL truck in which Xi,t =
1 if customer i is served by FTL truck t, and Xi,t = 0
otherwise
Wt
A binary variable for representing the reservation of truck t
in which Wt = 1 if FTL truck t is reserved, and Wt = 0
otherwise
Yi,r (ω)
A binary variable for allocating LTL carrier in which Yi,r =
1 if customer i is served by LTL carrier r, and Yi,r = 0
otherwise
Vu,v,t (ω)
A binary variable that indicates routing path of FTL truck
in which Vu,v,t (ω) = 1 if FTL truck t uses the path from
location u to location v, and Vu,v,t (ω) = 0 otherwise
Si,t (ω)
An integer variable that assists the system to eliminate subtour from the routing solution
t∈T
X
r∈R
Ai Xi,t ≤ Ft ,
t∈T
(4)
Xi,t ≤ ∆Wt ,
t∈T
(5)
Vu,u,t (ω) = 0
X
Vu,0,t (ω) ≤ 1,
u ∈ U, t ∈ T , ω ∈ Ω
(6)
t ∈ T ,ω ∈ Ω
(7)
t ∈ T ,ω ∈ Ω
(8)
Vu,i,t (ω) = Xi,t Di (ω)
i ∈ C, t ∈ T , ω ∈ Ω
(9)
Vi,u,t (ω) = Xi,t Di (ω)
i ∈ C, t ∈ T , ω ∈ Ω
(10)
i∈C
X
i∈C
u∈U
X
V0,u,t (ω) ≤ 1,
u∈U
A. Original ODPD Formulation
u∈U
We formulate the ODPD problem as a stochastic integer
programming model. In the model, there are five decision variables, i.e., Xi,t , Wt , Yi,r (ω), and Vu,v,t (ω) which are binary
variables, and Si,t (ω) which takes an integer value between
one and the total number of customers. The description of the
decision variables is presented in Table I.
Minimize:
XX
Xi,t +
i∈C t∈T
where
X
C̄t Wt + E[Q(Xi,t (ω), Wt (ω))],
(1)
t∈T
XX X
i∈C r∈R ω∈Ω
XXXX
bi,r Yi,r (ω)+
P rob(ω)C
u∈U
Si,t (ω) − Sj,t (ω) + |C|Vi,j,t (ω) ≤ |C| − 1,
i, j ∈ C, t ∈ T , ω ∈ Ω
(11)
ĆP rob(ω)P rob(
Xi,t , Wt ∈ {0, 1},
i ∈ C, t ∈ T
(12)
Yi,r (ω), Vu,v,t (ω), ∈ {0, 1},
r ∈ R, u, v ∈ U, t ∈ T , ω ∈ Ω
(13)
Si,t,ω ∈ {0, 1, 2, 3, . . . , n′ }, i ∈ C, t ∈ T , ω ∈ Ω
(14)
P rob(ω)C̈u,v Vu,v,t (ω)+
B. Reformulate the system fomulation with cardinality
u∈U v∈U t∈T ω∈Ω
t∈T ω∈Ω
X
The last three constraints in (12)-(14) indicate the types and
bounds of the decision variables.
Q(Xi,t (ω), Wt (ω)) =
XX
X
XX
Qu,v Vu,v,t (ω) ≥ T)
u∈U v∈U
(2)
subject to: (3) - (14).
The objective function
P given
P in (1) and (2) includes five
terms. The first term
i∈C
t∈T Xi,t is to minimize the
allocation of FTL trucks. The second to fifth terms aim to
minimize the total payment which includes (i) the initial cost
While the probability of demand scenario is a given parameter, the deadline violation probability is calculated as expressed
in the cardinality minimization formulation. In particular, the
cardinality is the number of non-zero
elements in a vector
or a matrix. For example, for ~x = 0 3 4 0 8 1 , the
cardinality is Card(~x) = 4.
To reformulate the optimization as the cardinality minimization problem, a set of traveling time samples is required.
Let S = {S1 , S2 , . . . , Ss′ } denote a set of traveling time
samples, where s′ is the total number of samples. The set
of traveling time from location u to location v in the sample
1
2
s′
, Wu,v
, . . . , Wu,v
}. Let
{S1 , S2 , . . . , Ss′ } is denoted as {Wu,v
Gt (ω) = {g1,t (ω), g2,t (ω), g3,t (ω), . . . , gs′ ,t (ω)} be a set of
the time exceeding the deadline that truck t uses to deliver
packages in scenario ω. Note that Vu,v,t (ω) is the decision
variable. For traveling time sample s, we let gs,t (ω) equal
zero when the delivery is done before the deadline T, and the
formulation is expressed as follows:
!
!
XX
s
gs,t (ω) = max 0,
Wu,v Vu,v,t (ω) − T . (15)
u∈U v∈U
Therefore, the deadline violation probability can be calculated by the cardinality of the time exceeding the deadline
divided by the total number of samples as follows:
!
XX
Card(Gt (ω))
.
P rob
Qu,v Vu,v,t (ω) ≥ T =
s′
u∈U v∈U
(16)
Note that the denominator s′ is a constant, and we can simplify
the fraction expression by including the denominator in the
penalty cost Ć.
We introduce a new decision variable for reformulating
the ODPD as the cardinality minimization problem. Zt,s is
a variable indicating the deadline status of truck t associated
with traveling time sample s in which Zt,s = 1 if FTL truck t
will not be able to deliver packages to all assigned customers
before deadline T, and Zt,s = 0 otherwise. We introduce the
expression
X
Card(Gt (ω)) =
Zt,s (ω)
(17)
s∈S
as a part of the objective function. Correspondingly, the
constraint in (18) must be introduced together with (17). The
constraint in (18) is to force Zt,s (ω) = 1 in which truck t
takes longer time to deliver than the deadline T when using
traveling time sample s.
XX
s
Wu,v
Vu,v,t (ω) − ∆Zt,s (ω) ≤ T,
are proposed to solve the first stage and the second stage, respectively. Both master and sub-problems are solved iteratively
in a loop until the optimal solution is found.
A. Master Problem
The master problem is expressed as follows:
Minimize:
XX
X
Xi,t +
C̄t Wt + θ,
i∈C t∈T
(20)
t∈T
subject to: (3) - (4), (12), (21).
The objective function of the master problem is expressed
in (20), where θ is an additional positive continuous decision
variable. The constraints on the capacity limit and the initial
cost of FTL trucks are included in the master problem. The
optimal solution values obtained from the master problem are
denoted as X̄i,t , W̄t , and θ̄. In the traditional L-shaped method,
the additional constraints consist of feasibility cuts and optimal
cuts. In this paper, we present only the optimal cut constraint
because the feasibility cuts are not necessary. This is from
the fact that the first stage decision variables of the ODPD
are binary variables. The optimal cut constraint is given as
follows:
XX
Ei,t Xi,t + θ ≥ e,
(21)
i∈C t∈T
which requires a solution from the sub-problem, i.e., Ei,t
and e.
B. Sub-problem
The sub-problem is given as follows:
Minimize:
XX
XX
bi,r Yi,r +
P rob(ω)C
P rob(ω)ĆZt,s
i∈C r∈R
+
XXX
t∈T s∈S
(22)
P rob(ω)C̈u,v Vu,v,t
u∈U v∈U t∈T
(19)
subject to: (3), (6)-(11), (13)-(14), (18)-(19).
The objective function of the sub-problem is to minimize
the expected value of the second stage. In the sub-problem,
in constraint (3), the decision variable Xi,t becomes the
parameter denoted as X̄i,t , the value of which is obtained
from the master problem. Note that Wt does not appear in
any constraints in the sub-problem, and thus W̄t is not used
to solve the sub-problem.
When the number of demand scenarios is large, the optimization problem presented in Section IV becomes more
complicated and takes long time to be solved. To make
the optimization problem more computationally tractable, we
apply L-shaped decomposition method [9]. We first reformulate the optimization problem into a master problem and
a sub-problem. Since the ODPD is a two stage stochastic
optimization problem, the master problem and the sub-problem
C. Algorithm
The L-shaped decomposition algorithm for the ODPD is
presented in Algorithm 1, where k is the iteration index. In
the first iteration, the master problem is solved by ignoring the
optimal cut constraint, and θ̄ is set to a very small value. For
the later iterations, the optimal cut constraint is added, and θ
becomes a decision variable, the optimal solution of which is
denoted as θ̄. Next, the second stage solutions are solved for
all scenarios as the inner loop.
u∈U v∈U
t ∈ T , s ∈ S, ω ∈ Ω
(18)
The constraint in (19) indicates that Zt,s is a binary variable.
Zt,s (ω) ∈ {0, 1},
s ∈ S, t ∈ T , ω ∈ Ω
V. D ECOMPOSITION
At the end, the values of e and Ei,t are calculated by
adapting from the original L-shaped method because the
ODPD is a mix integer programming, which is not solvable by
simplex method. The value of e depends on the values of J(ω),
P (ω), and M (ω), which are the LTL cost for outsourcing
all package delivery, the penalty cost, and the routing cost,
respectively. To calculating Ei,t , we use similar idea from
J(ω) and M (ω), but the value is specific for each i ∈ C
and t ∈ T .
When the difference between θ̄ and B is less than ǫ,
i.e., error tolerance, the convergence condition is met. The
loop is terminated when the total cost H increases and the
convergence condition has been met before. The optimal
solution is achieved at the second latest iteration, which is
the iteration before the total cost increases.
XX
P (ω) = P rob(ω)
Ć Z̄t,s
(23)
t∈T s∈S
M (ω) = P rob(ω)
XX X
C̈u,i V̄u,i,t + C̈i,u V̄i,u,t
i∈C t∈T u∈U
J(ω) = P rob(ω)
X
i∈C
bi,r
Di (ω) min C
r∈R
bi,r
Ii,t (ω) = P rob(ω) Di (ω) min C
r∈R
X
− P rob(ω)
C̈u,i V̄u,i,t + C̈i,u V̄i,u,t
(24)
(25)
(26)
u∈U
VI. P ERFORMANCE E VALUATION
A. Parameter Setting
We consider the system model with one FTL truck and one
LTL carrier. The initial cost of a van, i.e., the FTL truck,
is C¯1 = S$280, and the cost is estimated based on the truck
renting price in Singapore [10]. The capacity limit of the van is
F1 = 1060 kilograms [11]. The routing cost between location
u to location v is C̈u,v = Ku,v × 0.105, where Ku,v is the
distance in kilometers from location u to location v [12].
In this study, we assume that all customer packages have
equal weight in which A = 30 kilograms, and the cost of
b = S$21 per package. The cost of the
the LTL carrier is C
LTL carrier is calculated based on the Speedpost service by
SingPost Company [13]. Note that the above parameter setting
is similar to that in [1].
Moreover, the penalty cost is Ć = 1, and the deadline is T
= 105 minutes. The error tolerance for terminating the decomposition loop ǫ is 0.001. The customer demands are random
in the cases of multiple demand scenarios, and all customers
have demand in the case of one demand sceanrio. Note that
some parameters will be varied for different experiment. The
ODPD system model is implemented in GAMS software [14],
the solver of which is CPLEX, to obtain an optimal solution.
All the experiments are conducted on the PC computer with a
3.70GHz Intel Xeon CPU E5-1630 v3 processor and 32 GB
installed memory (RAM).
Result: The optimal solution is achieved at the second
latest iteration
while (H k < H k−1 ) or (N 6= 1) do
if k = 0 then
//first iteration;
- θ̄k is set to a very small value;
- Solving master problem without the additional
cutting constraint;
else
- Generating the additional constraint as in (21)
k−1
by using ek−1 and Ei,t
;
- Solving master problem with the additional
cutting constraint;
end
k
- Obtaining all X̄i,t
, W̄tk and θ̄k ;
foreach scenario ω ∈ Ω do
k
- Solving the sub-problem by using X̄i,t
as
parameters;
k
k
k
- Obtaining Ȳi,r
, V̄u,v,t
, and Z̄t,s
;
- Calculating P (ω) , M (ω), and J(ω) from
Equation (23), (24), and (25), respectively;
- Calculating Ii,t (ω), ∀t ∈ T , ∀i ∈ C from (26);
- Calculating h(ω), which is the objective
k
k
k
cost (22) by using Ȳi,r
, V̄u,v,t
, and Z̄t,s
as
parameters;
end
P
- Calculating
ek = ω∈Ω (J(ω) − P (ω) − M (ω));
P
k
Ei,t
=
P ω∈Ω Ii,t (ω), ∀t ∈ T , ∀iP∈ C;P
k
k
X̄i,t
;
(ω)) − i∈C t∈T Ei,t
B= P
ω∈Ω (J(ω) − M
P
k
k
H = t∈T C̄t W̄t + ω∈Ω h(ω) ;
if |B − θ̄k | > ǫ then
N = 1 ; // convergence condition is
met
end
- Increasing iteration index by one k = k + 1;
end
Algorithm 1: Decomposition algorithms
In this paper, the input dataset is synthesized based on the
real data from a logistics private limited company in Singapore
and Google Map. We first use the geographical locations of
the depot and 24 customers from the real logistics dataset.
We implement a Java program to collect the distance and the
traveling time of each customer’s location from the Google
Map Distance Matrix API [15]. Additionally, to provide more
variety of tests, we use the synthesized traveling time dataset
which are generated from the normal distribution by using
Google traveling time as the mean value, and the variance is
set to be 100 seconds.
B. Routing Path and Payment
The vehicle routing path is presented in Figure 1 for the case
that all 24 customers have demand. We observe that all the
customers are served by the FTL truck. With the deadline of
295
1
Total Payment
FTL Payment
Payment (S$)
285
280
LTL Payment
Routing Payment
Penalty Payment
8
6
4
Probability of Deadline Violation
290
2
0
80
ODP with Deadline
The original ODP
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
90
100
110
120
130
140
150
160
170
0
80
180
90
100
110
Deadline (Minutes)
310
LTL Payment
Routing Payment
Penalty Payment
30
20
10
30
25
3
6
9
12
15
18
21
24
27
30
Penalty Cost (S$)
Fig. 4: Different value of the penalty
cost when the deadline = 105 minutes
170
310
300
290
280
1
2
3
4
5
6
7
20
1
15
10
0
0
160
320
ODPD (heuristic)
ODPD (w/o heuristic)
ODPD Decom (heuristic)
5
0
150
40
Probability of
Deadline Violation
Payment (S$)
280
Execution Time (Minutes)
Total Payment
FTL Payment
290
140
Fig. 3: Comparison between the original ODP and the ODPD on the deadline violation
35
300
130
Fig. 2: Payment breakdown of different deadline when penalty cost = 1
Total Payment (S$)
Fig. 1: Vehicle routing for Singapore
Road Map
120
Increase in Deadline (Seconds)
ODPD (heuristic)
ODPD (w/o heuristic)
ODPD Decom (heuristic)
0.5
0
1
2
3
4
5
6
7
1
Numner of Scenarios
Fig. 5: Comparison of Time performance before and after decomposing
T = 105 minutes and the penalty cost of Ć = 1, the probability
of deadline violation is 0.4. In other words, two out of five
samples violate the deadline. Note that the total payment is
S$286.19, S$280 of which is the initial cost of the FTL truck,
S$4.19 is the routing payment, and S$2 is the penalty of
violating deadline. Note that the penalty = S$2 because we
ignore the denominator in the constraint in (16).
C. Impact of Different Deadline
We vary the deadline for the following experiments while
the penalty cost is fixed as one (Ć = 1). For ease of
presentation, we only consider the scenario that all customers
have demand. When the deadline is set to be larger, there
will be more choices for the system to select an optimal path
without paying the penalty of deadline violation.
1) Total Payment: Figure 2 shows the payment breakdown.
When the deadline is between 80 and 85 minutes, no path can
meet the deadline for all samples. For the deadline from 90
to 95 minutes, the system can find the optimal path with four
samples which violate the deadline. When the deadline is set
to 100 and 105 minutes, three and two samples violate the
deadline, respectively. The penalty payment reaches zero when
2
3
4
5
6
7
Number of Scenarios
Fig. 6: Comparison of Total Payment
(top) and Deadline Violation Probability (bottom) before and after decomposing
the deadline is set to be 110 minutes, which means the system
can find the feasible path without violating the deadline for
all traveling time samples.
2) Comparison of Deadline Violation Probability: Figure 3
shows the deadline violation probabilities of the original ODP
and the ODPD. The probabilities are calculated from the
optimal routing solutions. The ODPD always achieves smaller
probability of deadline violation. This result clearly shows that
the ODPD yields a better solution than that of the ODP as the
ODPD explicitly incorporates the deadline violation into the
formulation.
D. Impact of Penalty Cost
We consider the case that all customers have demand and
the deadline is fixed (i.e., T = 105 minutes). Then, we vary the
value of the penalty cost. The payment breakdown is presented
in Figure 4. When the penalty cost Ć ≤ 9, all 24 customers are
delivered by the FTL truck and the penalty cost increases as
the penalty payment increases. However, when the penalty cost
Ć ≥ 12, which is too expensive compared with hiring the LTL
carrier. Therefore, the ODPD system allocates 23 customers
to the FTL truck and 1 customer to the LTL carrier. This
180
solution guarantees that all the customers will be served before
the deadline for all traveling time samples. Thus, the penalty
payment is zero.
E. Decomposition Evaluation
We next evaluate the performance in terms of total payment
and computation time of the ODPD with decomposition. We
perform the following experiments by two different solver setting options, i.e., CPLEX with heuristic and CPLEX without
a high speed feature or heuristic. The former may achieve a
local optimal solution with fast computation. On the contrary,
the latter is to obtain a global optimal solution. In general, the
former executes faster than the latter. Note that the ODPD with
decomposition is solved only using CPLEX with heuristic. The
following results are obtained by varying the size of demand
scenarios. We present only less than 7 scenarios in Figure 5
and Figure 6 because the ODPD with heuristic does not yield
the optimal solution for more than 7 scenarios. In particular,
the customers are served by the outsourcing carrier, which
can be considered as an unsolvable problem. Moreover, for
multiple scenarios, the execution time of the ODPD without
heuristic is too slow, i.e., taking more than 2 hours for 2
scenarios and 12 hours for 5 scenarios, which is unacceptable
in practice.
The execution time of the ODPD is presented in Figure 5.
Note that we omit the results of the ODPD without heuristic
for multiple scenarios in Figure 6 because the execution time
is too long. For only one demand scenario case, the ODPD
solved by CPLEX without heuristic achieves the best total
payment (Figure 6 (top)). However, the total payment is less
than S$5 different compared with the other methods.
On the other hand, the ODPD and the decomposed ODPD,
which are solved by CPLEX with heuristic, can be solved
with acceptable execution time. The time performances of
ODPD with and without decomposition are observed to be
comparable as indicated in Figure 5. However, the total
payment and the deadline violation probability of the ODPD
without decomposition, which is illustrated in Figure 6, are
significantly higher than those of the decomposed ODPD.
In summary, with the decomposition method, we can
achieve an optimal solution for more than 7 scenarios with
acceptable execution time. To illustrate, solving 25 scenarios
using the decomposed ODPD is still faster than solving 2
scenarios using the original (undecomposed) ODPD without
heuristic. The execution time of the decomposed ODPD increases according to number of scenarios. The decomposed
ODPD always achieves the better total payment than that
of the ODPD when solving with CPLEX with heuristic.
Moreover, the deadline violation probability of the original
(undecomposed) ODPD is always one or close to one which
might not be preferable for the customers.
VII. C ONCLUSION
We have proposed the Optimal Stochastic Delivery Planning
with Deadline (ODPD) to help supplier make the decision
whether to serve customers by full-truckload (FTL) trucks or
less-than-truckload (LTL) carriers before the actual customer
demand is observed. The ODPD, which is able to accommodate the traveling time limit requirement, enhances supplier
benefit and customer satisfaction. The uncertainties in both
customer demand and traveling time have been considered in
the ODPD. While we assume that the distribution probability
of the customer demand is known, the assumption of traveling
time is not required as we have adopted the cardinality
minimization approach in the ODPD routing. The L-shaped
decomposition method has been applied to the ODPD. The
experiments based on the real Singapore customer locations
have been carried out. The decision between FTL truck and
LTL carrier assignment has been optimized when the penalty
cost of FTL trucks exceeding the deadline is accounted. The
computational time of the decomposed ODPD is significantly
lower than that of the original (undecomposed) ODPD. For
the future work, multi-stage scenarios and FTL truck break
down events will be incorporated.
VIII. ACKNOWLEDGMENT
This work is partially supported by Singapore Institute of
Manufacturing Technology-Nanyang Technological University
(SIMTech-NTU) Joint Laboratory and Collaborative research
Programme on Complex Systems.
R EFERENCES
[1] S. Sawadsitang, R. Kaewpuang, S. Jiang, D. Niyato, and P. Wing.
“Optimal Stochastic Delivery Planning in Full-Truckload and Less-ThanTruckload Delivery”. In Proceedings of IEEE 85th Vehicular Technology
Conference (VTC2017-Spring), 4-7 June 2017.
[2] C. Lin, K. L. Choy, G. T. S. Ho, S. H. Chung, and H. Y. Lam. “Survey of
Green Vehicle Routing Problem: Past and future trends”. Expert System.
Appl. vol. 41, pp. 1118-1138, 2014.
[3] E. D. Miller-Hooks and H. S. Mahmassani .“Least Expected Time Paths
in Stochastic, Time-Varying Transportation Networks”. Transportation
Science, vol. 34, pp. 198-215, 2000.
[4] H. Frank, Shortest paths in probabilistic graphs, Operations Research,
vol. 17, no. 4, pp. 583599, 1969.
[5] Z. Cao, H. Guo, J. Zhang, D. Niyato, and U. Fastenrath. “Finding the
Shortest Path in Stochastic Vehicle Routing: A Cardinality Minimization
Approach”. IEEE Transportation on Intelligent Transportation Systems,
vol. 17, pp. 1688-1702, 2015
[6] Ball, M. O., Golden, A., Assad, A., & Bodin, L. D. (1983). “Planning
for truck fleet size in the presence of a common-carrier option”. Decision
Sciences, 14, pp. 103-120, 1983.
[7] M.-C. Bolduc, J. Renaud, and F. Boctor,“A heuristic for the routing and
carrier selection problem”, European Journal of Operational Research,
vol. 183, no. 2, pp. 926-932, 2007.
[8] Euchi J., Chabchoub H. “Iterated Density Estimation with 2-opt local
search for the vehicle routing problem with private fleet and common
carrier”. In Proceedings of International Conference on Computers &
Industrial Engineering (CIE), pp.1058-1063, 2009.
[9] J. R. Birge, and F. Louveaux, “Introduction to Stochastic Programming”.
New York: Springer.
[10] http://sgvehiclerent.com/
[11] https://rentalorrysingapore.com/lorry-dimensions/
[12] https://www.lta.gov.sg/content/dam/ltaweb/corp/PublicationsResearch/files/
FactsandFigures/Statistics in Brief 2009.pdf
[13] http://www.singpost.com/
[14] D. Chattopadhyay, “Application of General Algebraic Modeling System
to Power System Optimization”, IEEE Transactions on Power Systems,
vol.14 pp.15-22, 1999.
[15] https://developers.google.com/maps/documentation/distance-matrix/
| 2cs.AI
|
WNtags: A Web-Based Tool For Image Labeling And
Retrieval With Lexical Ontologies
Marko Horvat, Anton Grbin, Gordan Gledec
Faculty of Electrical Engineering and Computing, University of Zagreb,
Unska 3, HR-10000 Zagreb, Croatia
{Marko.Horvat2, Anton Grbin, Gordan.Gledec}@fer.hr
Cite as: Horvat, M., Grbin, A., & Gledec, G. (2013). Wntags: A web-based tool
for image labeling and retrieval with lexical ontologies. Advances in
Knowledge-Based and Intelligent Information and Engineering System, IOS
Press, 585-594. doi: 10.3233/978-1-61499-105-2-585
Abstract. Ever growing number of image documents available on the Internet
continuously motivates research in better annotation models and more efficient
retrieval methods. Formal knowledge representation of objects and events in
pictures, their interaction as well as context complexity becomes no longer an
option for a quality image repository, but a necessity. We present an ontologybased online image annotation tool WNtags and demonstrate its usefulness in
several typical multimedia retrieval tasks using International Affective Picture
System emotionally annotated image database. WNtags is built around WordNet lexical ontology but considers Suggested Upper Merged Ontology as the
preferred labeling formalism. WNtags uses sets of weighted WordNet synsets
as high-level image semantic descriptors and query matching is performed with
word stemming and node distance metrics. We also elaborate our near future
plans to expand image content description with induced affect as in stimuli for
research of human emotion and attention.
1 Introduction
Image repositories annotated with ground truth labels are welcome tools in many
areas of computer science which address the problems of visual data classification,
recognition or retrieval. Prevalence of visual media on the Web stresses practical uses
of such repositories. Image retrieval algorithms with high precision and recall, and
low fall-out are being increasingly sought as the number of indexed pictures on the
Internet is getting larger. But because building a large dataset of annotated images is a
costly and lengthy endeavor, it is important to create a flexible and modular system
with small footprint that reuses available off the shelve components as much as possible. These motives guided planning and implementation of the WNtags ontologybased image annotation tool.
The WNtags allows image tagging with weighted sets of concepts. In the current
version WNtags uses WordNet (hence the tool’s name) as the knowledge database but
its system architecture can accommodate concepts from other truly formal ontologies
as image semantics descriptors. Apart from providing user friendly Web interface for
collaborative image annotation, WNtags is also a tool for searching emotionally annotated real-life images. The image retrieval is based on semantic similarity between the
search query and stored image annotations. This is accomplished using circumplex
model of emotions [1], ontology-based image labels [2], corpus independent graphbased distance metrics [3] which, as the entire tool, has modular implementation and
can be optimized or replaced independently from other parts of the system.
WNtags tool is different from other similar tools for manual image annotation (per
example see [4][5]) in several respects. Firstly, WNtags uses International Affective
Picture System (IAPS) dataset with semantically and emotionally described photographs that induce negative, neutral or positive affective reactions in stimulated persons [6]. IAPS pictures are annotated sparsely and inadequately using free-text keywords [7]. WNtags can also be viewed as a tool that significantly improves functionality and usability of this and other similar databases. Secondly, WNtags’ modular
construction test enables experimentation with different algorithms in various stages
of image retrieval. Thirdly, WNtags uses WordNet [8] for description of image content and in the future envisages applying Suggested Upper Merged Ontology (SUMO)
[9] to formally schematize image concepts. Finally, WNtags uses weighted measures
and other heuristics to more accurately describe the most semantically important
concepts in a picture and to reconcile different measures of observer variability.
The remainder of this paper is organized as follows. Chapter 2 discusses the
WNtags as the collaborative web-based research tool in the area of image annotation,
knowledge representation and document retrieval and describes its architecture. Chapter 3 describes formal methods of image annotation. Chapter 4 focuses on image retrieval and discusses the results. Finally, chapter 5 gives insight into future work and
concludes the paper.
2 The WNtags collaborative web-based annotation tool
The WNtags is primarily a research tool in areas of image annotation, knowledge
representation and multimedia document retrieval on the Web. Its development was
motivated by a number of objectives and possible applications. In this section we
describe the goals behind WNtags inception, its system architecture and the images
used to test the tool’s annotation and retrieval capabilities. There are two groups of
motives or incentives for development of WNtags tool (Table 1). The goals are divided in two broad and complementary areas in computer science and artificial intelligence: knowledge representation and image retrieval.
Table 1. Goals motivating the WNtags tool development with possible applications.
Goals
Possible applications
Knowledge representation Image content description, annotation expressivity, semantic gap, affective dimensions, …
Image retrieval
Image repositories, emotionally annotated databases, web directories, …
The most important goal of this project was to develop a collaborative online tool
that could be modified and customized with minimal effort. In doing this, at least in
the first version of WNtags, the system requirements were confined to representation
of high-level semantics in static images. Some compromises had to be made since we
wanted semantics annotations to be formal, but at the same close to a natural language. With the aid of a reasoning engine, the semantics annotations should be able to
yield new knowledge about the image content. This is by no means an easy task, so
we decided to use WordNet glossary to tag objects, events and affect in images. Our
aim was to get an exhaustive set of relevant high-level semantics for each image in
the database described with tags from a controlled glossary with graph-like structure.
Such tag vocabulary can describe minimum necessary functional relationships between tags and discriminate among different meanings of the same lexical unit (i.e.
word). Furthermore, we wanted all tags to be weighted to indicate their importance in
aggregation of image content.
Representation of picture segments and objects is currently not in our focus as
with some other image annotation tools [4] [5]. However, we plan to scale out the tool
to be able to annotate and store lower level semantics as well. In the future we also
want to add support for other multimedia formats like sounds or video-clips, but
foremost for specific domains of static images such as facial expressions and especially affectively annotated images. Our goal is to optimally reuse the existing code base
and gradually build up the application’s features.
The secondary goal behind the WNtags project was experimentation in information retrieval using the described weighted high-level semantics from a controlled
graph glossary. Instead of using statistical methods and bag-of-words concept in document indexing and retrieval, we would like to see images precisely annotated with
least number of semantically meaningful tags. With formal logical systems based on
Description Logic (DL) [10] these annotation may constitute a knowledge database of
an intelligent expert system. Through the process of reasoning, the annotations could
infer new knowledge about image content, effectively expanding the initial set of
image tags. In this scheme, a search query given by a WNtags user becomes a search
strategy goal in which the system must find the most closely matched images, fetch
them from the database and sort the results by relevance. Procedures of user feedback
and information refinement are also possible.
Finally, WNtags could also be interpreted as a step towards improvement of emotionally annotated databases such as IAPS [6], International Affective Digitized
Sounds (IADS) [11] or Geneva Affective Picture Database (GAPED) [12]. These
databases of multimedia stimuli are important for research of human emotion and
attention [6], but also find their uses in a wide range of research (examples [13][14]).
As mentioned before, their description schemes of multimedia semantics are currently
quite rudimentary and limiting making image retrieval difficult and operator intensive
[7]. Enhancement of multimedia content retrieval in these databases could substantially extend their use and promote further applications in cognitive sciences, psychology, psychiatry and neurology among others areas.
3 Formal image annotation
Optimally the annotation tool should not restrict the users in selection of text labels
for describing images. Most users prefer to provide short and basic-level object labels
(e.g. “child”, “car”, “person”, “airplane”). This is not enough to establish a rich tag
glossary which would enable quality image retrieval. At the same time, there can be a
large variance of terms describing the same object category. Taken together, these
make analysis and retrieval of the object labels difficult, since the system has to be
aware of label synonyms, subsumed labels, member labels, and distinguish between
object identity, events executed by objects, and attributes of objects and events.
We believe that only a knowledge database based on a large upper ontology and
extended with appropriate domain ontologies can encapsulate semantics present in an
image repository [15]. Naturally, domain ontologies would be selected depending on
the repository’s purpose and integrated with the basis upper ontology. Creating such
image repository with large knowledge database and filling it with appropriately annotated images is an immense task. However, a compromise can be made with selection of WordNet as the tagging glossary.
WordNet is a lexical database of English language [8] which may also be interpreted as an informal and lexical ontology [16]. It is readily available, simple to use and –
most importantly – contains over 100,000 concepts, i.e. synsets. This is ideal in terms
of glossary size and encompasses any label a user may enter. However, WordNet’s
linguistic tokens lack the formality of ontological concepts. A formal ontology contains rules specifying complex relations that cannot be captured explicitly with simple
links in graph-like knowledge structure. In spite of these shortcomings, WordNet is
still far superior in image annotation than existing IAPS keywords, which are inconsistent, ambiguous and non-contiguous [7][17].
In WNtags, images are manually annotated with individual senses s1 , s2 , , sn of
WordNet synsets syn1 , syn2 ,
, synm . Each image is tagged with at least three senses by
a group of two or more individuals. Every sense si is weighted in a range wi [0,1]
according to its perceived relevance. Apart from weighted semantic labels, each image img i retained its original semantic description from IAPS iapsi . IAPS keyword
vocabulary V IAPS is unsupervised and each IAPS picture is described with a single
free text keyword. Effectively, iapsi is one keyword from this corpus. IAPS uses
circumplex model of affect to annotate stimuli emotion values [1][6]. In a nutshell,
this model numerically describes the meaning of the emotional impulse in a 3-D space
with the respect to the axes of valence or pleasure (denoted val), the axis of excitation
or arousal (ar) and the axis of dominance over emotion (dom). Values of all axis are
numerical and normalized in interval [1, 9].
Every image img i in WNtags repository retained its original IAPS emotional tuple
value emoi with values vali , ari and domi . If semi is set of all WordNet senses for
ˆ i set of weighted senses semi , and iapsi is one keyword from IAPS vocabimg i , sem
ulary then the cumulative semantic description desci for one image img i stored in
WNtags image repository from becomes a tuple
ˆ i , emoi , iapsi }
desci {sem
(1)
where
semi ( si 1 , si 2 ,
, si n )
ˆ i ( wi 1 si 1 , wi 2 si 2 ,
sem
, wi n si n )
(2)
emoi (vali , ari , domi )
iapsi Vi IAPS
For every sense si WNtags considers WordNet’s built-in semantic relations hypernymy, hyponymy, holonymy and meronymy within a preset node distance d.
WNtags also considers synonyms (coordinate terms) of si . Therefore for every sense
s k in WordNet database and every sense si in desci broadens image description with
all senses of all its neighboring synsets
semi ( si 1 , si 2 ,
, si n )
s s ,s ,
j i1 i 2
, si n
s j , sk d
(3)
As an example, IAPS annotated image (7175.jpg) with weighted labels and interaction with currently implemented WordNet knowledge database is shown in the next
figure:
Fig. 1. An IAPS image in WNtags manually annotated with weighted concepts from a formal
glossary and described in WordNet lexical ontology knowledge database. Each concept (e.g.
“Lamp”) has semantic neighborhood defined with semantic relations and node distance d = 1,
2, 3, …
All indexed images are manually and collaboratively annotated. Personal bias of
users will certainly result in different weights assigned to the image labels. Therefore,
1 k
if there are k weight ratings for si they are averaged out wi wij . Almost all
k j 1
users agreed on sets of 3-5 distinct labels per single image while finding additional
descriptions in many cases proved difficult especially for semantically simple or abstract pictures.
In order to prove the validity of the WNtags concept we have randomly selected
100 emotionally annotated images with different object categories from IAPS database and tagged them using 956 different WordNet synsets. On average a single image is annotated with median of 20 WordNet tags (mean = 20.56436, sd = 2.76917,
min=13, max=28) which were weighted according to their subjective importance in a
particular image.
4 Image retrieval
The described WNtags annotating configuration allows multimodal image retrieval
by three different information dimensions: affect, original free-text keywords and sets
of synonyms from the WordNet lexical ontology. WNtags has modular architecture
and in the current version, to save processing time, semantic similarity is loaded from
the freely available WordNet::Similarity dataset with prepared similarity and relatedness pairs [18].
The search algorithm uses all senses of all collocations (i.e. non-permutable combinations) of query concepts. Some more complex multiword collocations may be
found as specific senses in WordNet. All such detected query senses qsi in query q k
are individually aligned with annotating senses si for all images img l , l 1, 2,..., N img
where N img is the total number of images stored in the repository. Semantic distance
between each pair is calculated together with subjective importance w j of each s j in
image img l . Search goal function is
max
qsi qk , s j imgl
w j sim(qsi , s j )
(4)
An exhaustive search is executed and the retrieved results are sorted from the
highest to the lowest relevance with possible values in range. The final results are
rendered and displayed to the user.
To test the retrieval performance we performed semantic searches. All queries
consisted of one WordNet concept randomly selected within d = 30 node distance
between the nearest image tag. For example (as in Fig. 2.): “aircraft”, “car”, “boat”,
“helicopter”, “road”, “bus”, etc.
Fig. 2. Retrieved results from WNtags repository after querying it with concept “person”.
After N = 40 queries the average precision is 68.93 % and average recall 6.15. As
expected, queries had different recall, i.e. results count. Some queries yielded only
one result, while others produced over 15 IAPS images ranked according to their
semantic similarity with the querying concepts. Precision is calculated for each sequence number in the result set and recall is normalized accordingly. The highest
recall (R = 20) is for query “animal” because of the repository content and the fact
that the most images were already manually tagged with this concept. As can be seen
in the next figure, precision and recall fall as the result size increases, as expected.
The first result has precision 84.21 %, second 81.58 %, third 65.78 % and so on.
Number of currently indexed images is too low for a thorough evaluation of retrieval
performance. However, some indicative conclusions can be drawn from these results.
The aggregated and averaged results for all N = 40 queries are shown in the next figure with query count on x-axis and value range of precision and recall on y-axis. Precision of all queries is indicated with a blue dotted line, while recall is represented
with a solid red line.
Fig. 3. Average retrieval performance after N = 40 queries with unit concepts. Average precision is 68.93 % and average recall 6.15. The highest precision is 84.21 % for the first result in
sequence.
4.1 Discussion
Overall, we found images to be tagged quite exhaustively, which is good for establishing the ground truth in image content, but can be an obstacle for retrieval with
semantically similar concepts leading to unrealistically high precision. As can be seen
in Fig.3, precision oscillates as the number of results increases, which is a consequence of this over-tagging problem. This problem may be easily resolved if only a
randomized subset of tags is used in retrieval.
The second problem is how to define a realistic, useful and at the same time generic measure of semantic relatedness between concepts in ontology. A variety of information-based (corpus dependent) and node distance-based (information independent)
similarity measures have different successes in diverse application domains. Many
such measures exist and finding the most suitable semantic similarity measure is an
ongoing knowledge representation problem.
Another problem is objectivity of tag weight coefficients values. Although potentially a separate source of information valuable for image retrieval, weight values are
a product of individual, and subsequently, potentially biased criterion. Tags with
unequally distributed weights will produce erroneous relatedness with query tag,
especially when coupled with suboptimal semantic distance metrics. However, kappa
statistic may be used to indicate weights that have inadequate inter-annotator agreement. Tags with such weights should be re-annotated or outlier values left out and
weights recalculated. Either way, tag weights have to be consistent across the whole
image repository or they will deteriorate the semantic relatedness function. The best
way to do this is to establish a clear and formal set of rules for calculating the weights
so the influence of personal bias is annulled.
Thirdly, the structure of WordNet’s semantic network may be contrary to intuition
and too complex [19]. Some subsuming concepts such as “artifact, artifact -- (a manmade object taken as a whole)” or “whole, unit -- (an assemblage of parts that is regarded as a single entity…)” are useful for construction of a well-indented taxonomy,
but semantically ambiguous or seemingly redundant to an average person. Intuitively,
it seems that a pruned down and lightweight version of WordNet, with new semantic
relations such as “Relates-To”, “Similar-To” or “Occurs-With”, may be better as a
knowledge taxonomy for image labeling than the complete WordNet, but without
relations that may be lexically ambiguous but could provide fresh and important information content.
WordNet is a huge benefit, but at the same time also a problem. Expressivity of
WNtags image annotations is adequate to support content rich and diverse image
extraction, but qualitatively the lack of formalism and complex functional relationships between concepts proved limiting and led to ambiguities with different concept
senses. For example, a single WordNet concept “snake” has a synonyms “serpent”.
An average WNtags user expects that semantic distance between all synonyms and
another different concept to be the same. However, this is not the case with node
distance metrics based on WordNet senses; semantic neighborhoods in WordNet
semantic graph around synonyms are not identical. This leads to different image retrieval results when querying these concepts, which may be confusing to the user.
For this reason we noticed that it is unnecessary to observe a wide semantic
neighborhood around the search query. Distance of up to 5 or 10 nodes is mostly
enough to pick up the majority of concepts related to a query. Larger distances will
increase recall, but substantially decrease accuracy and precision in semantic similarity. However, an adaptable algorithm may be envisioned to find a good balance between these two opposing parameters. Such algorithm could start with a small semantic neighborhood around the query and then iteratively increase the distance or include different semantic relations until enough related concepts are found or some
other stopping criterion is met. Further research is required to validate this idea.
As mentioned before, it is obvious that only lexically rich knowledge database is
not enough for high quality annotation and retrieval. Ontology with a sufficient number of annotating concepts representing objects, events and their various functional
relations should allow more accurate image retrieval but also retain a rich annotating
glossary which is desirable from the user’s perspective.
5 Conclusion
WNtags is web-based image annotation tool that may be used to manually identify
objects, events and affect that occur in images. Owing to the tool’s friendly user interface, new images can easily be added in the document repository and labeled with
semantic concepts. WNtags performs image retrieval using WordNet’s ontology topology, node distance metrics and collaboratively weighted tags. The results are sorted relative to the query semantic similarity.
The WNtags system was developed as a flexible and modular research tool in areas
of annotation and retrieval of real-life emotionally annotated images, but the problems
outlined in the previous chapter vividly illustrate difficulties with using WordNet as
the knowledge database for image annotation. Only a large ontology deprived of
lexical dependencies is able to formally represent the ground truth meaning of image
content. For this reason we plan to add SUMO to the WNtags knowledge database. In
doing this we would like to retain existing WordNet image labels and map them to
SUMO concepts [20]. The implemented concept relatedness metrics is another point
which requires further work. A lot of false positive retrieval results are due to inaccurate similarity estimates between query and image concepts. We plan to modularly
add different semantic similarity algorithms for information retrieval including information-based metrics.
We would also like to explore modalities of indexing additional multimedia formats apart from pictures, not withstanding specific affective-related domains of static
images such as facial expressions and visual stimuli.
Acknowledgements
This research has been partially supported by the Ministry of Science, Education
and Sports of the Republic of Croatia, grant no. 036-0000000-2029. The WNtags tool
was developed by student team: Anton Grbin (project leader), Aleksandar Dukovski,
Anton Grbin, Jerko Jurin, Dino Milačić, Matija Stepanić, Hrvoje Šimić and Dominik
Trupčević as part of their Programming Project course at the University of Zagreb,
Faculty of Electrical Engineering and Computing.
References
1. Peter, C., Herbon, A.: Emotion representation and physiology assignments in digital systems.
Interacting with Computers, Vol. 18. Elsevier (2006) 139–170
2. Schreiber, A. Th., Dubbeldam, B., Wielemaker, J. Wielinga, B. J.: Ontology-based photo
annotation. IEEE Intelligent Systems, 16(3) (2001) 66–74
3. Rada, R., Mili, H., Bicknell, E., Blettner, M.: Development and application of a metric on
semantic nets. IEEE Transations on Systems, Man, and Cybernetics, 19(1) (1989) 17–30
4. Russell, B. C., Torralba, A., Murphy, K. P., Freeman, W. T.: LabelMe: a database and webbased tool for image annotation. Technical report, MIT (2005)
5. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L.: ImageNet: a large-scale hierarchical image database. Proc. CVPR. (2009) 710–719
6. Lang, P. J., Bradley, M. M., Cuthbert, B. N.: International affective picture system (IAPS):
Affective ratings of pictures and instruction manual (Technical Report A-8). University of
Florida, Gainesville, FL (2008)
7. Horvat, M., Popović, S., Bogunović, N., Ćosić, K.: Tagging multimedia stimuli with ontologies. Proceedings of the 32nd International Convention MIPRO 2009, Croatian Society for
Information and Communication Technology, Electronics and Microelectronics – MIPRO,
Opatija, Croatia (2009) 203–208
8. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press (1998)
9. Pease, A.: Ontology: A Practical Guide. Articulate Software Press, Angwin, CA (2011)
10. Allemang, D., Hendler, J.: Semantic Web for the Working Ontologist: Effective Modeling
in RDFS and OWL. Elsevier (2007)
11. Lang, P. J., Bradley, M. M., Cuthbert, B. N.: International affective digitized sounds
(IADS): Stimuli, instruction manual and affective ratings (Tech. Rep. No. B-2). The Center
for Research in Psychophysiology, University of Florida, U.S.A. (1999)
12. Dan-Glauser, E. S., Scherer, K. R.: The Geneva Affective PicturE Database (GAPED): A
new 730 picture database focusing on valence and normative significance. Behavior Research Methods, in press
13. Tsonos, D., Ikospentaki, K., Kouroupetrolgou, G.: Towards Modeling of Readers’ Emotional State Response for the Automated Annotation of Documents. IEEE World Congress
on Computational Intelligence (WCCI 2008) (2008) 3253–3260
14. Britton, J. C., Taylor, S. F., Sudheimer, K. D., Liberzo, I.: Facial expressions and complex
IAPS pictures: Common and differential networks. NeuroImage, Vol. 31(2006) 906–919
15. Neumann, B., Mӧller, R.: Ontology-based reasoning techniques for multimedia interpretation and retrieval. In Semantic Multimedia and Ontologies: Theory and Applications,
Springer (2008)
16. Niles I., Pease, A.: Towards a standard upper ontology, Proceedings of the international
conference on Formal Ontology in Information Systems, Ogunquit, Maine, USA (2001) 2–9
17. Ćosić, K., Popović, S., Kukolja, D., Horvat, M., Dropuljić, B.: Physiology-driven adaptive
virtual reality stimulation for prevention and treatment of stress related disorders. CyberPsychology, Behavior, and Social Networking, Vol. 13:1 (2010) 73–78
18. Pedersen, T., Patwardhan, S., Michelizzi, J.: WordNet::Similarity - Measuring the Relatedness of Concepts. Proceedings of the Nineteenth National Conference on Artificial Intelligence (AAAI-04), San Jose, CA (Intelligent Systems Demonstration) (2004) 1024–1025
19. Resnik, P.: Semantic Similarity in a Taxonomy: An Information-Based Measure and its
Application to Problems of Ambiguity in Natural Language. Journal Of Artificial Intelligence Research, Vol. 11 (1999) 95–130
20. Niles I., Pease, A.: Linking lexicons and ontologies: Mapping wordnet to the suggested
upper merged ontology. In Proc. of the 2003 Int. Conf. on Information and Knowledge Engineering (2003)
| 2cs.AI
|
Food Recognition using Fusion of Classifiers
based on CNNs
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
arXiv:1709.04864v1 [cs.CV] 14 Sep 2017
Universitat de Barcelona and Computer Vision Center, Spain
{eduardo.aguilar,marc.bolanos,petia.ivanova}@ub.edu
Abstract. With the arrival of convolutional neural networks, the complex problem of food recognition has experienced an important improvement in recent years. The best results have been obtained using methods
based on very deep convolutional ceural cetworks, which show that the
deeper the model,the better the classification accuracy will be obtain.
However, very deep neural networks may suffer from the overfitting problem. In this paper, we propose a combination of multiple classifiers based
on different convolutional models that complement each other and thus,
achieve an improvement in performance. The evaluation of our approach
is done on two public datasets: Food-101 as a dataset with a wide variety
of fine-grained dishes, and Food-11 as a dataset of high-level food categories, where our approach outperforms the independent CNN models.
Keywords: Food Recognition, Fusion Classifiers, CNN
1
Introduction
In the field of computer vision, food recognition has caused a lot of interest
for researchers considering its applicability in solutions that improve people’s
nutrition and hence, their lifestyle [1]. In relation to the healthy diet, traditional
strategies for analyzing food consumption are based on self-reporting and manual
quantification [2]. Hence, the information used to be inaccurate and incomplete
[3]. Having an automatic monitoring system and being able to control the food
consumption is of vital importance, especially for the treatment of individuals
who have eating disorders, want to improve their diet or reduce their weight.
Food recognition is a key element within a food consumption monitoring
system. Originally, it has been approached by using traditional approaches [4,5],
which extracted ad-hoc image features by means of algorithms based mainly on
color, texture and shape. More recently, other approaches focused on using Deep
Learning techniques [5,6,7,8]. In these works, feature extraction algorithms are
not hand-crafted and additionally, the models automatically learn the best way
to discriminate the different classes to be classified. As for the results obtained,
there is a great difference (more than 30%) between the best method based
on hand-crafted features compared to newer methods based on Deep Learning,
where the best results have been obtained with Convolutional Neural Networks
(CNN) architectures that used inception modules [8] or residual networks [7].
2
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
Fig. 1. Example images of Food-101 dataset. Each image represents a dish class.
Food recognition can be considered as a special case of object recognition,
being the most active topic in computer vision lately. The specific part is that
dish classes have a much higher inter-class similarity and intra-class variation
than usual ImageNet objects (cars, animals, rigid objects, etc.) (see Fig.1). If we
analyze the last accuracy increase in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) [9], it has been improving thanks to the depth increase of CNN models [10,11,12,13] and also thanks to the use of CNNs fusion
strategies[11,13]. The main problem of CNNs is the need of large datasets to
avoid overfitting the network as well as the need of high computational power
for training them.
Considering the use of different networks trained on the same data, one can
observe that patterns misclassified by the different models would not necessarily
overlap [14]. This suggests that different classifiers could potentially offer complementary information that can be used to improve the final performance [14].
An option to combine the outputs of different classifiers was proposed in [15],
where the authors used what they call a decision templates scheme instead of
simple aggregation operators such as the product or average. As they showed,
this scheme maintains a good performance using different training set sizes and
is also less sensitive to particular datasets compared to the other schemes.
In this article, we introduce the fusion concept to the CNN framework, with
the purpose of demonstrating that the combination of the classifiers’ output, by
using a decision template scheme, allows to improve the performance on the food
recognition problem. Our contributions are the following: 1) we propose the first
food recognition algorithm that fuses the output of different CNN models, 2) we
show that our CNNs Fusion approach has better performance compared to the
use of CNN models separately, and 3) we demonstrate that our CNNs Fusion
approach keeps a high performance independently of the target (dishes, family
of dishes) and dataset (validating it on 2 public datasets).
The organization of the article is as follows. In section 2, we present the CNNs
Fusion methodology. In section 3, we present the datasets, the experimental setup
and discuss the results. Finally, in section 4, we describe the conclusions.
Food Recognition using Fusion of Classifiers based on CNNs
3
Fig. 2. General scheme of our CNNs Fusion approach.
2
Methodology
In this section, we describe the CNN Fusion methodology (see Fig. 2), which is
composed of two main steps: training K CNN models based on different architectures and fusing the CNN outputs using the decision templates scheme.
2.1
Training of CNN models
The first step in our methodology involves separately training two CNN models.
We chose two different kind of models winners of the ILSVRC in the object
recognition task. Both models won or are based on the winner of the challenges
made in 2014 and 2015 proposing novel architectures: the first based its design in
”inception models” and the second in ”residual networks”. First, each model was
pre-trained on the ILSVRC data. Later, all layers were fine-tuned by a certain
number of epochs, selecting for each one the model that provides the best results
in the validation set and that will be used in the fusion step.
2.2
Decision templates for classifiers fusion
Once we trained the models on the food dataset, we combined the softmax
classifier outputs of each model using the Decision Template (DT) scheme [15].
Let us annotate the output of the last layer of the k-th CNN model as
(ω1,k , . . . , ωC,k ), where c = 1, ..., C is the number of classes and k = 1, ...K
is the index of the CNN model (in our case, K=2). Usually, the softmax function is applied, to obtain
the probability value of model k to classify image x to
ωk,c
a class c: pk,c (x) = PCe eωk,c . Let us consider the k-th decision vector Dk :
c=1
Dk (x) = [pk,1 (x), pk,2 (x), ..., pk,C (x)]
Definition [15]: A Decision Profile, DP for a given image x is defined as:
p1,1 (x) p1,2 (x) ... p1,C (x)
...
DP (x) =
(1)
pK,1 (x) pK,2 (x) ... pK,C (x)
4
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
Definition [15]: Given N training images, a Decision Template is defined
as a set of matrices DT = (DT 1 , . . . , DT C ), where the c-th element is obtained
as the average of the decision profiles (1) on the training images of class c:
PN
j=1 DP (xj ) × Ind(xj , c)
c
,
DT =
PN
j=1 Ind(xj , c)
where Ind(xj , c) is an indicator function with value 1 if the training image
xj has a crisp label c, and 0, otherwise [16].
Finally, the resulting prediction for each image is determined considering the
similarity s(DP (x), DT c (x)) between the decision profile DP (x) of the test image and the decision template of class c, c = 1, . . . , C. Regarding the arguments
of the similarity function s(., .) as fuzzy sets on some universal set with K × C
elements, various fuzzy measures of similarity can be used. We chose different
measures [15], namely 2 measures of similarity, 2 inclusion indices, a consistency
measure and the Euclidean Distance. These measures are formally defined as:
PK PC
c
k=1
i=1 min(DTk,i , DPk,i (x))
,
S1 (DT c , DP (x)) = PK PC
c
k=1
i=1 max(DTk,i , DPk,i (x))
c
− DPk,i (x) : c = 1, . . . , C, k = 1, . . . , K},
S2 (DT c , DP (x)) = 1 − sup{ DTk,i
u
PK PC
c
I1 (DT , DP (x)) =
k=1
c
i=1 min(DTk,i , DPk,i (x))
,
PK PC
c
k=1
i=1 DTk,i
c , DP (x)) : c = 1, . . . , C, k = 1, . . . , K},
I2 (DT c , DP (x)) = inf {max(DTk,i
k,i
u
c
C(DT c , DP (x)) = sup{min(DTk,i
, DPk,i (x)) : c = 1, . . . , C, k = 1, . . . , K},
u
PK PC
− DPk,i (x))2
,
K ×C
c
where DTk,i
is the probability assigned to the class i by the classifier k in the
c is the complement of DT c calculated as 1 − DT c , and DP (x)
DT c , DTk,i
k,i
k,i
k,i
is the probability assigned by the classifier k to the class i in the DP calculated
for the image, x. The final label, L is obtained as the class that maximizes the
similarity, s between DP (x) and DT c : L(x) = argmaxc=1,...,C {s(DT c , DP (x))}.
c
N (DT , DP (x)) = 1 −
3
3.1
k=1
c
i=1 (DTk,i
Experiments
Datasets
The data used to evaluate our approach are two very different images: Food-11
[17] and Food-101 [4], which are chosen in order to verify that the classifiers
Food Recognition using Fusion of Classifiers based on CNNs
5
fusion provides good results regardless of the different properties of the target datasets, such as intra-class variability (the first one is composed of many
dishes of the same general category, while the second one is composed of specific
fine-grained dishes), inter-class similarity, number of images, number of classes,
images acquisition condition, among others.
Food-11 is a dataset for food recognition [17], which contains 16,643 images grouped into 11 general categories of food: pasta, meat, dessert, soup, etc.
(see Fig. 3). The images were collected from existing food datasets (Food-101,
UECFOOD100, UECFOOD256) and social networks (Flickr, Instagram). This
dataset has an unbalanced number of images for each class with an average of
1,513 images per class and a standard deviation of 702. For our experiments, we
used the same data split, images and proportions, provided by the authors [17].
These are divided as 60% for training, 20% for validation and 20% for test. In
total, it consists of 9,866, 3,430 and 3,347 images for each set, respectively.
Fig. 3. Images from the Food-11 dataset. Each image corresponds to a different class.
Food-101 is a standard to evaluate the performance of visual food recognition [4]. This dataset contains 101.000 real-world food images downloaded from
foodspotting.com, which were taken under unconstrained conditions. The authors chose the top 101 most popular classes of food (see Fig. 1) and collected
1,000 images for each class: 75% for training and 25% for testing. With respect
to the classes, these consist of very diverse and fine-grained dishes of various
countries, but also with highly intra-class variation and inter-class similarity in
most occasions. In our experiments, we used the same data splits provided by the
authors. Unlike Food-11, and keeping the procedure followed by other authors
[5,7,8], in this dataset we directly validate and test on the same data split.
3.2
Experimental Setup
Every CNN model was previously trained on the ILSVRC dataset. Following,
we adapted them by changing the output of the models to the number of classes
for each target dataset and fine-tuned the models using the new images. For the
training of the CNN models, we used the Deep Learning framework Keras1 .
The models chosen for Food-101 dataset due to their performance-efficiency
ratio were InceptionV3 [18] and ResNet50 [13]. Both models were trained during
48 epochs with a batch size of 32, and a learning rate of 5 × 10−3 and 1 ×
10−3 , respectively. In addition, we applied a decay of 0.1 during the training
of InceptionV3 and of 0.8 for ResNet50 every 8 epochs. The parameters were
chosen empirically by analyzing the training loss.
1
www.keras.io
6
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
As to the Food-11 dataset, we kept the ResNet50 model, but changed InceptionV3 by GoogLeNet [12], since InceptionV3 did not generalize well over
Food-11. We believe that the reason is the small number of images for each class
not sufficient to avoid over-fitting; the model quickly obtained a good result
in the training set, but a poor performance on the validation set. GoogLeNet
and Resnet50 were trained during 32 epochs with a batch size of 32 and 16,
respectively. The other parameters used for the ResNet50 were the same used
for Food-101. In the case of GoogLeNet, we used a learning rate of 1 × 10−3 and
applied a decay of 0.1 during every 8 epochs, that turned out empirically the
optimal parameters for our problem.
3.3
Data Preprocessing and Metrics
The preprocessing made during the training, validation and testing phases was
the following. During the training of our CNN models, we applied different preprocessing techniques on the images with the aim of increasing the samples and
to prevent the over-fitting of the networks. First, we resized the images keeping
the original aspect ratio as well as satisfying the following criteria: the smallest
side of the resulting images should be greater than or equal to the input size
of the model; and the biggest side should be less than or equal to the maximal
size defined in each model to make random crops on the image. In the case of
InceptionV3, we set to 320 pixels as maximal size, for GoogLeNet and ResNet50
the maximal size was defined as 256 pixels. After resizing the images, inspired
by [8], we enhanced them by means of a series of random distortions such as:
adjusting color balance, contrast, brightness and sharpness. Finally, we made
random crops of the images, with a dimension of 299x299 for InceptionV3 and
of 224x224 for the other models. Then, we applied random horizontal flips with
a probability of 50%, and subtracted the average image value of the ImageNet
dataset. During validation, we applied a similar preprocessing, with the difference that we made a center crop instead of random crops and that we did not
apply random horizontal flips. During test, we followed the same procedure than
in validation (1-Crop evaluation). Furthermore, we also evaluated the CNN using
10-Crops, which are: upper left, upper right, lower left, lower right and center
crop, both in their original setup and also applying an horizontal flip [10]. As for
10-Crops evaluation, the classifier gets a tentative label for each crop, and then
majority voting is used over all predictions. In the cases where two labels are
predicted the same number of times, the final label is assigned comparing their
highest average prediction probability.
We used four metrics to evaluate the performance of our approach, overall
Accuracy (ACC), Precision (P), Recall (R), and F1 score.
3.4
Experimental Results on Food-11
The results obtained during the experimentation on Food-11 dataset are shown
in Table 1 giving the error rate (1 - accuracy) for the best CNN models, compared
to the CNNs Fusion. We report the overall accuracy by processing the test data
Food Recognition using Fusion of Classifiers based on CNNs
7
using two procedures: 1) a center crop (1-Crop), and 2) using 10 different crops
of the image (10-Crops). The experimental results show an error rate of less
than 10 % for all classifiers, achieving a slightly better performance when using
10-Crops. The best accuracy is achieved with our CNNs Fusion approach, which
is about 0.75% better than the best result of the classifiers evaluated separately.
On the other hand, the baseline classification on Food-11 was given by their
authors, who obtained an overall accuracy of 83.5% using GoogLeNet models
fine-tuned in the last six layers without any pre-processing and post-processing
steps. Note that the best results obtained with our approach have been using
the pointwise measures (S2, I2). The particularity of these measures is that they
penalize big differences between corresponding values of DTs and DP being from
the specific class to be assigned as the rest of the class values. From now on, in
this section we only report the results based on the 10-Crops procedure.
Table 1. Overall test set error rate of Food-11 obtained for each model. The distance
measure is shown between parenthesis in the CNNs Fusion models.
Authors
Model
1-Crop 10-Crops N/A
[17]
GoogLeNet
16.5%
us
GoogLeNet
9.89%
9.29%
us
ResNet50
6.57%
6.39%
us
CNNs Fusion (S1 ) 6.36%
5.86%
us
CNNs Fusion (S2 ) 6.12%
5.65%
us
CNNs Fusion (I1 )
6.36%
5.89%
us
CNNs Fusion (I2 )
6.30%
5.65%
6.45%
6.07%
us
CNNs Fusion (C)
us
CNNs Fusion (N)
6.36%
5.92%
-
As shown in Table 2, the CNNs Fusion is able to properly classify not only
the images that were correctly classified by both baselines, but in some occasions
also when one or both fail. This suggests that in some cases both classifiers may
be close to predicting the correct class and combining their outputs can make a
better decision.
Table 2. Percentage of images well-classified and misclassified on Food-11 using our
CNNs Fusion approach, distributed by the results obtained with GoogLeNet (CNN1 )
and ResNet50 (CNN2 ) models independently evaluated.
CNNs evaluated independently
CNNs Fusion Both wrong CNN1 wrong CNN2 wrong Both fine
Well-classified
3.08%
81.77%
54.76%
99.97%
96.92%
18.23%
45.24%
0.03%
Misclassified
Samples misclassified by our model are shown in Fig. 4, where most of them
are produced by mixed items, high inter-class similarity and wrongly labeled
images. We show the ground truth (top) and the predicted class (bottom) for
each sample image.
8
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
Fig. 4. Misclassified Food-11 examples: predicted labels (on the top), and the
groundtruth (on the bottom).
In Table 3, we show the precision, recall and F1 score obtained for each
class separately. By comparing the F1 score, the best performance is achieved
for the class Noodles Pasta and the worst for Dairy products. Specifically, the
class Noddles Pasta only has one image misclassified, which furthermore is a
hard sample, because it contains two classes together (see items mixed in Figure
4). Considering the precision, the worst results are obtained for the class Bread,
which is understandable considering that bread can sometimes be present in
other classes. In the case of recall, the worst results are obtained for Dairy
products, where an error greater than 8% is produced for misclassifying several
images as class Dessert. The cause of this is mainly, because the class Dessert has
a lot of items in their images that could also belong to the class Dairy products
(e.g. frozen yogurt or ice cream) or that are visually similar.
Table 3. Some results obtained on the Food-11 using our CNNs Fusion approach.
Class
#Images Precision Recall
F1
Bread
368
88.95%
91.85%
90.37%
Dairy products
148
89.86%
83.78% 86.71%
Meat
432
94.12%
92.59%
93.35%
Noodles Pasta
147 100.00% 99.32% 99.66%
Rice
96
94.95%
97.92%
96.41%
Vegetable Fruit
231
98.22%
95.67%
96.93%
3.5
Experimental Results on Food-101
The overall accuracy on Food-101 dataset is shown in Table 4 for two classifiers based on CNN models, and also for our CNNs Fusion. The overall accuracy is obtained by means of the evaluation of the prediction using 1-Crop and
10-Crops. The experimental results show better performance (about 1% more)
using 10-Crops instead of 1-Crop. From now on, in this section we only report
the results based on the 10-Crops procedure. In the same way as observed in
Food-11, the best accuracy obtained with our approach was by means of pointwise measures S2, I2, where the latter provides a slightly better performance.
Food Recognition using Fusion of Classifiers based on CNNs
9
Again, the best accuracy is also achieved by the CNNs Fusion, which is about
1.5% higher than the best result of the classifiers evaluated separately. Note that
the best performance on Food-101 (overall accuracy of 90.27%) was obtained using WISeR [7]. In addition, the authors show the performance by another deep
learning-based approaches, in which three CNN models achieved over a 88% (InceptionV3, ResNet200 and WRN [19]). However, WISeR, WRN and ResNet200
models were not considered in our experiments since they need a multi-GPU
server to replicate their results. In addition, those models have 2.5 times more
parameters than the models chosen, which involve a high cost computational
especially during the learning stage. Following the article steps, our best results
replicating the methods were those using InceptionV3 and ResNet50 models used
as a base to evaluate the performance of our CNNs Fusion approach.
Table 4. Overall test set accuracy
Author
Model
[8]
InceptionV3
[7]
ResNet200
[7]
WRN
[7]
WISeR
us
ResNet50
us
InceptionV3
us
CNNs Fusion (S1 )
us
CNNs Fusion (S2 )
us
CNNs Fusion (I1 )
us
CNNs Fusion (I2 )
us
CNNs Fusion (C)
us
CNNs Fusion (N)
of Food-101 obtained for each model.
1-Crop 10-Crops N/A
88.28%
88.38%
88.72%
90.27%
82.31%
83.54%
83.82%
84.98%
85.52%
86.51%
86.07%
86.70%
85.52%
86.51%
85.98%
86.71%
85.24%
86.09%
85.53%
86.50%
-
As shown in Table 5, in this dataset the CNNs Fusion is also able to properly
classify not only the images that were correctly classified for both classifiers,
but also when one or both fail. Therefore, we demonstrate that our proposed
approach maintains its behavior independently of the target dataset.
Table 5. Percentage of images well-classified and misclassified on Food-101 using our
CNNs Fusion approach, distributed by the results obtained with InceptionV3 (CNN1 )
and ResNet50 (CNN2 ) models independently evaluated.
CNNs evaluated independently
CNNs Fusion Both wrong CNN1 wrong CNN2 wrong Both fine
Well-classified
1.95%
73.07%
64.95%
99.97%
Misclassified
98.05%
26.93%
35.05%
0.03%
Some examples of FPs and FNs are shown in Figure 5. Analyzing the misclassified images like Steak, we chose a representive images per class corresponding
to the top three classes with highest error, and in Edamame the unique image misclassified. Regarding the statistical results, in Table 6 are shown the top
3 better and worst classification results on Food-101. We highlight the classes
10
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
Fig. 5. Misclassified examples for the Food-101 classes that obtained the worst (steak)
and best (edamame) classification results by F1 score (groundtruth label - bottom).
with the worst and best results. As for the worst class (Steak), the precision and
recall achieved are 60.32% and 59.60%, respectively. Interestingly, about 26%
error in the precision and 30% error in the recall is produced with only three
classes: Filet mignon, Pork chop and Prime rib. As shown in Fig. 5, these are
fine-grained classes with high inter-class similarities that imply high difficulty
for the classifier, because it is necessary to identify small details that allow to
determine the corresponding class of the images. On the other hand, the best
class (Edamame) was classified achieving 99.60% of precision and 100% of recall. Unlike Steak, Edamame is a simple class to classify, because it has a low
intra-class variation and low inter-class similarities. In other words, the images
in this class have a similar visual appearance and they are quite different from
the images of the other classes.
Table 6. Top 3 better and worst classification results on Food-101.
Class
Precision
Recall
F1
Spaghetti Bolognese
94.47%
95.60%
95.03%
Macarons
97.15%
95.60%
96.37%
Edamame
99.60%
100.00% 99.80%
Steak
60.32%
59.60% 59.96%
Pork Chop
75.71%
63.60%
69.13%
Foie Gras
72.96%
68.00%
70.39%
4
Conclusions
In this paper, we addressed the problem of food recognition and proposed a CNNs
Fusion approach based on the concepts of decision templates and decision profiles
and their similarity that improves the classification performance with respect to
using CNN models separately. Evaluating different similarity measures, we show
that the optimal one is based on the infinimum of the maximum between the
complementary of the decision templates and the decision profile of the test
images. On Food-11, our approach outperforms the baseline accuracy by more
than 10% of accuracy. As for Food-101, we used two CNN architectures providing
the best state of the art results where our CNNs Fusion strategy outperformed
Food Recognition using Fusion of Classifiers based on CNNs
11
them again. As a future work, we plan to evaluate the performance of the CNN
Fusion strategy as a function of the number of CNN models.
Acknowledgement
This work was partially funded by TIN2015-66951-C2, SGR 1219, CERCA,
ICREA Academia’2014, CONICYT Becas Chile, FPU15/01347 and Grant 20141510
(Marató TV3). We acknowledge Nvidia Corp. for the donation of Titan X GPUs.
References
1. Waxman, A., Norum, K.R.: WHO Global Strategy on Diet, Physical Activity and
Health. Food and Nutrition Bulletin. 25, 292-302 (2004).
2. Shim, J.-S., Oh, K., Kim, H.C.: Dietary assessment methods in epidemiologic studies. Epidemiology and Health. e2014009 (2014).
3. Rumpler, W. V, Kramer, M., Rhodes, D.G., Moshfegh, a J., Paul, D.R.: Identifying
sources of reporting error using measured food intake. European Journal of Clinical
Nutrition. 62, 544-552 (2008).
4. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101 - Mining discriminative components with random forests. In: Lecture Notes in Computer Science. pp. 446-461
(2014).
5. Liu, C., Cao, Y., Luo, Y., Chen, G., Vokkarane, V., Ma, Y.: Deepfood: Deep learningbased food image recognition for computer-aided dietary assessment. In: Lecture
Notes in Computer Science. pp. 37-48 (2016).
6. Yanai, K., Kawano, Y.: Food image recognition using deep convolutional network
with pre-training and fine-tuning. In: ICMEW. pp. 1-6 (2015).
7. Martinel, N., Foresti, G.L., Micheloni, C.: Wide-Slice Residual Networks for Food
Recognition. arXiv Prepr. (2016).
8. Hassannejad, H., Matrella, G., Ciampolini, P., De Munari, I., Mordonini, M.,
Cagnoni, S.: Food Image Recognition Using Very Deep Convolutional Networks.
In: Proceedings of the 2nd International Workshop on MADiMa. pp. 41-49 (2016).
9. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large
Scale Visual Recognition Challenge. International Journal of Computer Vision. 115,
211-252 (2015).
10. Krizhevsky, A., Sutskever, I., Geoffrey E., H.: ImageNet Classification with Deep
Convolutional Neural Networks. Adv. Neural Inf. Process. Syst. 25. 1-9 (2012).
11. Zeiler, M.D., Fergus, R.: Visualizing and Understanding Convolutional Networks.
In: Computer VisionECCV. pp. 818-833 (2014).
12. Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan,
D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: CVPR. pp.
1-9 (2015).
13. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition.
In: CVPR. pp. 770-778 (2016).
14. Kittler, J., Hatef, M.: On combining classifiers. IEEE Transactions on Pattern
Analysis and Machine Intelligence. 20, 226-239 (1998).
15. Kuncheva, L.I., Bezdek, J.C., Duin, R.P.W.: Decision templates for multiple classifier fusion: an experimental comparison. Pattern Recognition. 34, 299-314 (2001).
12
Eduardo Aguilar, Marc Bolaños, and Petia Radeva
16. Kuncheva, L.I., Kounchev, R.K. , Zlatev, R.Z.: Aggregation of multiple classification decisions by fuzzy templates. In: EUFIT. 1470-1474 (1995).
17. Singla, A., Yuan, L., Ebrahimi, T.: Food/Non-food Image Classification and Food
Categorization using Pre-Trained GoogLeNet Model. In: Proceedings of the 2nd
International Workshop on MADiMa. pp. 3-11 (2016).
18. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the Inception Architecture for Computer Vision. In: CVPR. pp. 2818-2826 (2016).
19. Zagoruyko, S., Komodakis, N.: Wide Residual Networks. arXiv Prepr. (2016).
| 1cs.CV
|
arXiv:1705.09131v1 [math.GR] 25 May 2017
ON DISCRETE HOMOLOGY OF A FREE PRO-p-GROUP
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Abstract. For a prime p, let F̂p be a finitely generated free pro-p-group of rank ≥ 2. We
show that the second discrete homology group H2 (F̂p , Z/p) is an uncountable Z/p-vector
space. This answers a problem of A.K. Bousfield.
1. Introduction
Let p be a prime. For a profinite group G, there is a natural comparison map
H2disc (G, Z/p) → H2cont (G, Z/p),
which connects discrete and continuous homology groups of G. Here H2disc (G, Z/p) =
H2 (G, Z/p) is the second homology group of G with Z/p-coefficients, where G is viewed
as a discrete group. The continuous homology H2cont (G, Z/p) can be defined as the inverse
limit lim H2 (G/U, Z/p), where U runs over all open normal subgroups of G. The above
←Ð
comparison map H2disc → H2cont is the inverse limit of the coinflation maps H2 (G, Z/p) →
H2 (G/U, Z/p) (see Theorem 2.1 [6]).
The study of the comparison map for different types of pro-p-groups is a fundamental
problem in the theory of profinite groups (see [6] for discussion and references). It is well
known that for a finitely generated free pro-p-group F̂p ,
H2cont (F̂p , Z/p) = 0.
A.K. Bousfield posed the following question in [2], Problem 4.11 (case R = Z/n):
Problem. (Bousfield) Does H2disc (F̂n , Z/n) vanish when F is a finitely generated free group?
Here F̂n is the Z/n-completion of F , which is isomorphic to the product of pro-p-completions
F̂p over prime factors of n (see Prop. 12.3 [2]). That is, the above problem is completely reduced to the case of homology groups H2disc (F̂p , Z/p) for primes p and since
H2cont (F̂p , Z/p) = 0, the problem becomes a question about non-triviality of the kernel of the
comparison map for F̂p .
In [3], A.K. Bousfield proved that, for a finitely generated free pro-p-group F̂p on at least
two generators, the group Hidisc (F̂p , Z/p) is uncountable for i = 2 or i = 3, or both. In
particular, the wedge of two circles S 1 ∨ S 1 is a Z/p-bad space in the sense of Bousfield-Kan.
The group H2disc (F̂p , Z/p) plays a central role in the theory of HZ/p-localizations developed in [2]. It follows immediately from the definition of HZ/p-localization that, for a free
group F , H2disc (F̂p , Z/p) = 0 if and only if F̂p coincides with the HZ/p-localization of F .
In this paper we answer Bousfield’s problem over Z/p. Our main result is the following
1
2
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Main Theorem. For a finitely generated free pro-p-group F̂p of rank ≥ 2, H2disc (F̂p , Z/p)
is uncountable.
There are two cases in the problem of Bousfield, R = Z/n and R = Q. We give the
answer for the case of R = Z/n. The problem over Q is still open: we do not know whether
H2disc (F̂Q , Q) vanishes or not.
The proof is organized as follows.
In Section 2 we consider properties of discrete and continuous homology of profinite
groups. Using a result of Nikolov and Segal [10, Th 1.4], we show that for a finitely
generated profinite group G and a closed normal subgroup H the cokernels of the maps
H2disc (G, Z/p) → H2disc (G/H, Z/p) and H2cont (G, Z/p) → H2cont (G/H, Z/p) coincide (Theorem
2.5):
/ H disc (G/H, Z/p)
2
H2disc (G, Z/p)
ϕ
/ Qdisc
ϕ
/0
≅
H2cont (G, Z/p)
/ H cont (G/H, Z/p)
2
/ Qcont
/ 0.
As a corollary we obtain (Corollary 2.6) that, for a finitely generated free pro-p-group F̂p ,
a continuous epimorphism π ∶ F̂p ↠ G to a pro-p-group induces the exact sequence
(1.1)
ϕ
π∗
H2disc (F̂p , Z/p) Ð→ H2disc (G, Z/p) Ð→ H2cont (G, Z/p) Ð→ 0.
That is, for proving that, for a free group F , H2 (F̂p , Z/p) ≠ 0, it is enough to find a discrete
epimorphism F ↠ G such that, the comparison map of the second homology groups of
the pro-p-completion of G has a nonzero kernel. Observe that, the statements in Section 2
significantly use the theory of profinite groups and there is no direct way to generalize them
for pronilpotent groups. In particular, we don’t see how to prove that, H2 (F̂Z , Z/p) ≠ 0,
where F̂Z is the pronilpotent completion of F .
Section 3 follows the ideas of Bousfield from [3]. Consider the ring of formal power series
Z/p[[x]], and the infinite cyclic group C ∶= ⟨t⟩. We will use the multiplicative notation of the
p-adic integers C ⊗Zp = {tα , α ∈ Zp }. Consider the continuous multiplicative homomorphism
τ ∶ C ⊗ Zp → Z/p[[x]] sending t to 1 − x. The main result of Section 3 is Proposition 3.3,
which claims that the kernel of the multiplication map
(1.2)
Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]] Ð→ Z/p[[x]]
is uncountable.
Our main example is based on the p-lamplighter group Z/p ≀ Z, a finitely generated but
not finitely presented group, which plays a central role in the theory of metabelian groups.
The homological properties of the p-lamplighter group are considered in [9]. The profinite
completion of the p-lamplighter group is considered in [7], it is shown there that it is a semisimilar group generated by finite automaton. We consider the double lamplighter group:
i
i
i
(Z/p)2 ≀ Z = ⟨a, b, c ∣ [b, ba ] = [c, ca ] = [b, ca ] = bp = cp = 1, i ∈ Z⟩,
ON DISCRETE HOMOLOGY OF A FREE PRO-p-GROUP
3
Denote by DL the pro-p-completion of the double lamplighter group. It follows from direct
computations of homology groups, that there is a diagram (in the above notations)
Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]]
/ Z/p[[x]]
⊕
H2disc (DL, Z/p)
H2cont (DL, Z/p),
/
where the left vertical arrow is a split monomorphism and the upper horisontal map is the
multiplication map (see proof of Theorem 4.3). This implies that, for the group DL, the
comparison map H2disc (DL, Z/p) → H2cont (DL, Z/p) has an uncountable kernel. Since the
double lamplighter group is 3-generated, the sequence (1.1) implies that, for a free group F
with at least three generators, H2 (F̂p , Z/p) is uncountable. Finally, we use Lemma 11.2 of
[3] to get the same result for a two generated free group F .
In [2], A.K. Bousfield formulated the following generalization of the above problem for
the class of finitely presented groups (see Problem 4.10 [2], the case R = Z/n). Let G
be a finitely presented group, is it true that HZ/p-localization of G equals to its pro-pcompletion Ĝp ? [The Problem is formulated for HZ/n-localization, but it is reduced to the
case of a prime n = p.] It follows immediately from the definition of HZ/p-localization that,
this problem can be reformulated as follows: is it true that, for a finitely presented group
G, the natural homomorphism H2 (G, Z/p) → H2 (Ĝp , Z/p). It is shown in [2] that this is
true for the class of poly-cyclic groups. The same is true for finitely presented metabelian
groups [8]. Main Theorem of the present paper implies that, for any finitely presented
group P , which maps epimorphically onto the double lamplighter group, the natural map
H2 (P, Z/p) → H2 (P̂p , Z/p) has an uncountable cokernel.
2. Discrete and continuous homology of profinite groups
For a profinite group G and a normal subgroup H, denote by H the closure of H in G
in profinite topology.
Theorem 2.1 ([10, Th 1.4]). Let G be a finitely generated profinite group and H be a closed
normal subgroup of G. Then the subgroup [H, G] is closed in G.
Corollary 2.2. Let G be a finitely generated profinite group and H be a closed normal
subgroup of G. Then the subgroup [H, G] ⋅ H p is closed in G.
Proof. Consider the abelian profinite group H/[H, G]. Then the p-power map H/[H, G] →
H/[H, G] is continuous and its image is equal to ([H, G] ⋅ H p )/[H, G]. Hence ([H, G] ⋅
H p )/[H, G] is a closed subgroup of H/[H, G]. Using that the preimage of a closed set under
continuous function is closed we obtain that [H, G] ⋅ H p is closed.
Lemma 2.3 (mod-p Hopf’s formula). Let G be a (discrete) group and H be its normal
subgroup. Then there is a natural exact sequence
H2 (G, Z/p) Ð→ H2 (G/H, Z/p) Ð→
H ∩ ([G, G]Gp )
Ð→ 0.
[H, G]H p
4
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Proof. It follows from the five term exact sequence
H2 (G, Z/p) Ð→ H2 (G/H, Z/p) Ð→ H1 (H, Z/p)G Ð→ H1 (G, Z/p)
and the equations H1 (H, Z/p)G = H/([H, G]H p ) and H1 (G, Z/p) = G/([G, G]Gp ).
Lemma 2.4 (profinite mod-p Hopf’s formula). Let G be a profinite group and H be its
closed normal subgroup. Then there is a natural exact sequence
H2cont (G, Z/p) Ð→ H2cont (G/H, Z/p) Ð→
H ∩ ([G, G]Gp )
([H, G]H p )
Ð→ 0.
Proof. For the sake of simplicity we set H∗ (−) = H∗discr (−, Z/p) and H∗cont (−) ∶=
H∗cont (−, Z/p). Consider the five term exact sequence (Corollary 7.2.6 of [11])
H2cont (G) Ð→ H2cont (G) Ð→ H0cont (G, H1cont (H)) Ð→ H1cont (G).
Continuous homology and cohomology of profinite groups are Pontryagin dual to each other
(Proposition 6.3.6 of [11]). There are isomorphisms
1
Hcont
(G) = Hom(G/[G, G]Gp , Z/p) = Hom(G/[G, G]Gp , Q/Z),
where Hom denotes the set of continuous homomorphisms (see [12, I.2.3]). It follows that
H1cont (G) = G/[G, G]Gp . Similarly H1cont (H) = H/[H, H]H p . Lemma 6.3.3. of [11] implies that H0cont (G, M ) = M /⟨m − mg ∣ m ∈ M, g ∈ G⟩ for any profinite (Z/p[G])∧ -module
M. Therefore H0cont (G, H1cont (H)) = H/[H, H]H p . The assertion follows.
We denote by ϕ the comparison map
ϕ ∶ H2disc (G, Z/p) → H2cont (G, Z/p).
Theorem 2.5. Let G be a finitely generated profinite group and H a closed normal subgroup
of G. Denote
Qdisc ∶= Coker(H2 (G, Z/p) → H2 (G/H, Z/p))
Qcont ∶= Coker(H2cont (G, Z/p) → H2cont (G/H, Z/p)).
Then the comparison maps ϕ induce an isomorphism Qdisc ≅ Qcont ∶
H2disc (G, Z/p)
/ H disc (G/H, Z/p)
2
ϕ
/ Qdisc
ϕ
/0
≅
H2cont (G, Z/p)
/ H cont (G/H, Z/p)
2
/ Qcont
/0
Proof. This follows from Lemma 2.3, Lemma 2.4 and Corollary 2.2.
Corollary 2.6. Let G be a finitely generated pro-p-group and π ∶ F̂p ↠ G be a continuous epimorphism from the pro-p-completion of a finitely generated free group F. Then the
sequence
ϕ
π∗
H2disc (F̂p , Z/p) Ð→ H2disc (G, Z/p) Ð→ H2cont (G, Z/p) Ð→ 0
is exact.
Proof. This follows from Theorem 2.5 and the fact that H2cont (F̂p , Z/p) = 0.
ON DISCRETE HOMOLOGY OF A FREE PRO-p-GROUP
5
3. Technical results about the ring of power series Z/p[[x]]
In this section we follow to ideas of Bousfield written in Lemmas 10.6, 10.7 of [3]. The
goal of this section is to prove Proposition 3.3
We use the following notation: C = ⟨t⟩ is the infinite cyclic group; C ⊗ Zp is the group of
p-adic integers written multiplicatively as powers of the generator C ⊗ Zp = {tα ∣ α ∈ Zp };
Z/p[[x]] is the ring of power series; Z/p((x)) is the field of formal Laurent series.
i
Lemma 3.1. Let A be a subset of Z/p[[x]]. Denote by Ai the image of A in Z/p[x]/(xp ).
Assume that
i
lim ∣Ai ∣/pp = 0.
i→∞
Then the interior of Z/p[[x]] ∖ A is dense in Z/p[[x]].
s
Proof. Take any power series f and any its neighbourhood of the form f + (xp ). Then for
pi
s
s+i
p
any i the open set f +(xp ) is the disjoint union of smaller open sets ⋃t=1
f +ft +(xp ), where
s
s+i
i+s
s
ft runs over representatives of (xp )/(xp ). Chose i so that ∣As+i ∣/pp ≤ p−p . Then ∣As+i ∣ ≤
i
s+i
pp . Hence the number of elements in Ai+s is lesser the number of open sets f + ft + (xp ).
s+i
It follows that there exists t such that A ∩ (f + ft + (xp )) = ∅. The assertion follows.
Denote by
τ ∶ C ⊗ Zp → Z/p[[x]]
the continuous multiplicative homomorphism sending t to 1 − x. It is well defined because
i
i
(1 − x)p = 1 − xp .
Lemma 3.2. Let K be the subfield of Z/p((x)) generated by the image of τ . Then the degree
of the extension [Z/p((x)) ∶ K] is uncountable.
Proof. Denote the image of the map τ ∶ C ⊗ Zp → Z/p[[x]] by A. Let α = (α1 , . . . , αn ) and
β = (β1 , . . . , βn ), where α1 , . . . , αn , β1 , . . . , βn ∈ Z/p, and k ≥ 1. Denote by Aα,β,k the subset
of Z/p[[x]] consisting of elements that can be written in the form
α1 a1 + ⋅ ⋅ ⋅ + αn an
,
(3.1)
β1 b1 + ⋅ ⋅ ⋅ + βn bn
k
where a1 , . . . , an , b1 , . . . , bn ∈ A and β1 b1 +⋅ ⋅ ⋅ +βn bn ∉ (xp ). Then K ∩Z/p[[x]] = ⋃α,β,k Aα,β,k .
i
Fix some α, β, k. Take i ≥ k and consider the images of A and Aα,β,k in Z/p[x]/(xp ).
i
i
Denote them by Ai and Aiα,β,k . Obviously Ai is the image of the map C/C p → Z/p[x]/(xp )
that sends t to 1−x. Then Ai consists of pi elements. Fix some elements ā1 , . . . , ān , b̄1 , . . . , b̄n ∈
Ai that have preimages a1 , . . . , an , b1 , . . . , bn ∈ A such that the ratio (3.1) is in Aα,β,k . For any
such preimages a1 , . . . , an , b1 , . . . , bn ∈ A the image r̄ of the ratio (3.1) satisfies the equation
r̄ ⋅ (β1 b̄1 + ⋅ ⋅ ⋅ + βn b̄n ) = α1 ā1 + . . . αn ān .
k
Since β1 b̄1 + ⋅ ⋅ ⋅ + βn b̄n ∉ (xp ), the annihilator of β1 b̄1 + ⋅ ⋅ ⋅ + βn b̄n consist of no more than
k
k
pp elements and the equation has no more than pp solutions. Then we have no more than
p2in variants of collections ā1 , . . . , ān , b̄1 , . . . , b̄n ∈ Ai and for any such variant there are no
k
more than pp variants for the image of the ratio. Therefore
(3.2)
k
∣Aiα,β,k ∣ ≤ p2in+p .
6
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Take any sequence of elements v1 , v2 , ⋅ ⋅ ⋅ ∈ Z/p((x)) and prove that ∑∞
m=1 Kvm ≠ Z/p((x)).
Note that x ∈ K because t ↦ 1 − x. Multiplying the elements v1 , v2 , v3 , . . . by powers of x,
we can assume that v1 , v2 , ⋅ ⋅ ⋅ ∈ Z/p[[x]]. Fix some α, β, k as above. Set
Aα,β,k,l = Aα,β,k ⋅ v1 + ⋅ ⋅ ⋅ + Aα,β,k ⋅ vl .
−j
i
Then ∑∞
m=1 Kvm = ⋃α,β,k,l,j Aα,β,k,l ⋅ x . Denote by Aα,β,k,l the image of Aα,β,k,l in
k
Z/p[x]/(xi ). Then (3.2) implies ∣Aiα,β,k,l ∣ ≤ p(2in+p )l . Therefore
i
lim ∣Aiα,β,k,l ∣/pp = 0.
i→∞
By Lemma 3.1 the interior of the complement of Aα,β,k,l is dense in Z/p[[x]]. By the Baire
−j
theorem ∑∞
has empty interior. In particular ∑∞
m=1 Kvm = ⋃α,β,k,l,j Aα,β,k,l ⋅ x
m=1 Kvm ≠
Z/p((x)).
Proposition 3.3. Consider the ring homomorphism Z/p[C ⊗ Zp ] → Z/p[[x]] induced by τ .
Then the kernel of the multiplication map
(3.3)
Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]] Ð→ Z/p[[x]]
is uncountable.
Proof. As in Lemma 3.2 we denote by K the subfield of Z/p((x)) generated by the image
of C ⊗ Zp . Since t ↦ 1 − x, we have x, x−1 ∈ K. Set R ∶= K ∩ Z/p[[x]]. Note that the image of
Z/p[C ⊗ Zp ] lies in R. Consider the multiplication map
µ ∶ Z/p[[x]] ⊗R K → Z/p((x)).
We claim that this is an isomorphism. Construct the map in the inverse direction
κ ∶ Z/p((x)) → Z/p[[x]] ⊗R K
given by
∞
∞
i=−n
i=0
κ( ∑ αi xi ) = ∑ αi+n xi ⊗ x−n .
Since we have ax ⊗ b = a ⊗ xb, κ does not depend on the choice of n, we just have to
chose it big enough. Using this we get that κ is well defined. Obviously µκ = id. Chose
a ⊗ b ∈ Z/p[[x]] ⊗R K. Then b = b1 b−1
2 , where b1 , b2 ∈ R. Since b2 is a power series, we
can chose n such that b2 = xn b3 , where b3 is a power series with nontrivial constant term.
Then b3 is invertible in the ring of power series and b3 , b−1
3 ∈ R because x ∈ K. Hence
−1
−n
−1
−n
a ⊗ b = ab−1
b
⊗
b
=
ab
⊗
x
b
=
ab
b
⊗
x
.
Using
this
presentation,
we see that κµ = id.
1
1 3
3 3
3
Therefore
(3.4)
Z/p[[x]] ⊗R K ≅ Z/p((x)).
Since the image of Z/p[C ⊗ Zp ] lies in R, the tensor product Z/p[[x]] ⊗R Z/p[[x]] is a
quotient of the tensor product Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]] and it is enough to prove that
the kernel of
(3.5)
is uncountable.
Z/p[[x]] ⊗R Z/p[[x]] Ð→ Z/p[[x]]
ON DISCRETE HOMOLOGY OF A FREE PRO-p-GROUP
7
For any ring homomorphism R → S and any R-modules M, N there is an isomorphism
(M ⊗R N ) ⊗R S = (M ⊗R S) ⊗S (N ⊗R S). Using this and the isomorphism (3.4), we obtain
that after application of − ⊗R K to (3.5) we have
(3.6)
Z/p((x)) ⊗K Z/p((x)) Ð→ Z/p((x)).
Assume the contrary that the kernel of the map (3.5) is countable (countable = countable
or finite). It follows that the linear map (3.6) has countably dimensional kernel. Finally,
note that the homomorphism
Λ2K Z/p((x)) → Z/p((x)) ⊗K Z/p((x))
given by a∧b ↦ a⊗b−b⊗a is a monomorphism, its image lies in the kernel and the dimension
of Λ2K Z/p((x)) over K is uncountable because [Z/p((x)) ∶ K] is uncountable (Lemma 3.2).
A contradiction follows.
4. Double lamplighter pro-p-group
Let A be a finitely generated free abelian group written multiplicatively; Z/p[A] be its
group algebra; I be its augmentation ideal and M be a Z/p[A]-module. Then we denote
by M̂ = lim M /M I i its I-adic completion. We embed A into the pro-p-group A ⊗ Zp . We
←Ð
use the following ’multiplicative’ notation aα ∶= a ⊗ α for a ∈ A and α ∈ Zp . Note that for
i
i
i
i
i
any a ∈ A the power ap acts trivially on M /M I p because 1 − ap = (1 − a)p ∈ I p . Then we
can extend the action of A on M̂ to the action of A ⊗ Zp on M̂ in a continuous way.
The proof of the following lemma can be found in [8] but we add it here for completeness.
Lemma 4.1. Let A be a finitely generated free abelian group and M be a finitely generated
Z/p[A]-module. Then
H∗ (A, M ) ≅ H∗ (A, M̂ ) ≅ H∗ (A ⊗ Zp , M̂ ).
Proof. The first isomorphism is proven in [4]. Since Z/p[A] in Noetheran, it follows that
Hn (A, M̂ ) is a finite Z/p-vector space for any n. Prove the second isomorphism. The
action of A ⊗ Zp on M̂ gives an action of A ⊗ Zp on H∗ (A, M̂ ) such that A acts trivially on
H∗ (A, M̂ ). Then we have a homomorphism from A⊗Z/p to a finite group of automorphisms
of Hn (A, M̂ ), whose kernel contains A. Since any subgroup of finite index in A ⊗ Zp is
open (see Theorem 4.2.2 of [11]) and A is dense in A ⊗ Zp , we obtain that the action of
A ⊗ Zp on H∗ (C, M̂ ) is trivial. Note that Zp /Z is a divisible torsion free abelian group,
and hence A ⊗ (Zp /Z) ≅ Q⊕c , where c is the continuous cardinal. Then the second page of
the spectral sequence of the short exact sequence A ↣ A ⊗ Zp ↠ Q⊕c with coefficients in
M̂ is Hn (Q⊕c , Hm (A, M̂ )), where Lm ∶= Hm (A, M̂ ) is a trivial Z/p[Q⊕c ]-module. Then by
universal coefficient theorem we have
0 Ð→ Λn (Q⊕c ) ⊗ Lm Ð→ Hn (Q⊕c , Lm ) Ð→ Tor(Λn−1 (Q⊕c ), Lm ) Ð→ 0.
Since Λn (Q⊕c ) is torsion free and Lm is a Z/p-vector space, we get Λn (Q⊕c ) ⊗ Lm = 0 and
Tor(Λn−1 (Q⊕c ), Lm ) = 0. It follows that Hn (Q⊕c , Lm ) = 0 for n ≥ 1 and H0 (Q⊕c , Lm ) = Lm .
Then the spectral sequence consists only on of one column, and hence H∗ (A ⊗ Zp , M̂ ) =
H∗ (A, M̂ ).
8
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Lemma 4.2. Let A be an abelian group, M be a Z[A]-module and σM ∶ M → M be an
automorphism of the underlying abelian group such that σM (ma) = σM (m)a−1 for any
m ∈ M and a ∈ A. Then there is an isomorphism
(M ⊗ M )A ≅ M ⊗Z[A] M
given by
m ⊗ m′ ↔ m ⊗ σM (m′ ).
Proof. Consider the isomorphism Φ ∶ M ⊗ M → M ⊗ M given by Φ(m ⊗ m′ ) = m ⊗ σ(m′ ).
The group of coinvariants (M ⊗M )A is the quotient of M ⊗M by the subgroup R generated
by elements ma ⊗ m′ a − m ⊗ m′ , where a ∈ A and m, m′ ∈ M. We can write the generators of
R in the following form: ma ⊗ m′ − m ⊗ m′ a−1 . Then Φ(R) is generated by ma ⊗ σM (m′ )
− m ⊗ σM (m′ )a. Using that σM is an automorphism, we can rewrite the generators of R
as follows: ma ⊗ m′ − m ⊗ m′ a. Taking linear combinations of the generators of Φ(R) we
obtain that Φ(R) is generated by elements mλ ⊗ m′ − m ⊗ m′ λ, where λ ∈ Z[A]. Then
(M ⊗ M )/Φ(R) = M ⊗Z[A] M.
The group C = ⟨t⟩ acts on Z/p[[x]] by multiplication on 1 − x. As above, we can extend
the action of C on Z[[x]] to the action of C ⊗ Zp in a continuous way. The group
i
Z/p ≀ C = Z/p[C] ⋊ C = ⟨a, b ∣ [b, ba ] = bp = 1, i ∈ Z⟩
is called the lamplighter group. We consider the ’double version’ of this group, double
lamplighter group:
i
i
i
(Z/p[C] ⊕ Z/p[C]) ⋊ C = ⟨a, b, c ∣ [b, ba ] = [c, ca ] = [b, ca ] = bp = cp = 1, i ∈ Z⟩.
Its pro-p-completion is equal to the semidirect product
DL = (Z/p[[x]] ⊕ Z/p[[x]]) ⋊ (C ⊗ Zp ),
with the described above action of C ⊗ Zp on Z/p[[x]] (see Proposition 4.12 of [8]). We call
the group DL double lamplighter pro-p-group.
Theorem 4.3. The kernel of the comparison homomorphism for the double lamplighter
pro-p-group
ϕ ∶ H2disc (DL, Z/p) Ð→ H2cont (DL, Z/p)
is uncountable.
Proof. For the sake of simplicity we set H2 (−) = H2 (−, Z/p) and H2cont (−) = H2cont (−, Z/p).
Consider the homological spectral sequence E of the short exact sequence Z/p[[x]]2 ↣ DL ↠
2
C ⊗ Zp . Then the zero line of the second page is trivial Ek,0
= Hk (C ⊗ Zp ) = (Λk Zp )⊗ Z/p = 0
for k ≥ 2. Using Lemma 4.1 we obtain Hk (C ⊗ Zp , Z/p[[x]]) = Hk (C, Z/p[C]) = 0 for k ≥ 1,
2
and hence Ek,1
= 0 for k ≥ 1. It follows that
(4.1)
2
H2 (DL) = E0,2
.
For any Z/p-vector space V Künneth formula gives a natural isomorphism
H2 (V ⊕ V ) ≅ (V ⊗ V ) ⊕ H2 (V )2 .
ON DISCRETE HOMOLOGY OF A FREE PRO-p-GROUP
9
Then we have a split monomorphism
(4.2)
2
= H2 (DL).
(Z/p[[x]] ⊗ Z/p[[x]])C⊗Zp ↣ E0,2
It easy to see that the groups DL(i) = ((xi ) ⊕ (xi )) ⋊ (C ⊗ pi Zp ) form a fundamental
system of open normal subgroups. Consider the quotients DL(i) = DL/DL(i) . Then
H2cont (DL) = lim H2 (DL(i) ).
←Ð
2
The short exact sequence Z/p[[x]] ↣ DL ↠ C ⊗ Zp maps onto the short exact sequence
i
(Z/p[x]/(xi ))2 ↣ DL(i) ↠ C/C p . Consider the morphism of corresponding spectral sequences E → (i)E. Using (4.1) we obtain
2
2
Ker(H2 (DL) → H2 (DL(i) )) ⊇ Ker(E2,0
→ (i)E2,0
).
Similarly to (4.2) we have a split monomorphism
2
.
(Z/p[x]/(xi ) ⊗ Z/p[x]/(xi ))C⊗Zp ↣ (i) E2,0
Then we need to prove that the kernel of the map
(Z/p[[x]] ⊗ Z/p[[x]])C⊗Zp Ð→ lim(Z/p[x]/(xi ) ⊗ Z/p[x]/(xi ))C⊗Zp
←Ð
is uncountable.
Consider the antipod σ ∶ Z/p[C] → Z/p[C] i.e. the ring homomorphism given by σ(tn ) =
−n
t . The antipod induces a homomorphism σ ∶ Z/p[x]/(xi ) → Z/p[x]/(xi ) such that σ(1 −
x) = 1 + x + x2 + . . . . It induces the continuous homomorphism σ ∶ Z/p[[x]] → Z/p[[x]] such
that σ(x) = −x − x2 − . . . . Moreover, we consider the antipode σ on Z/p[C ⊗ Zp ]. Note that
the homomorphisms
(4.3)
Z/p[C] → Z/p[C ⊗ Zp ] → Z/p[[x]] → Z/p[x]/(xi )
commute with the antipodes.
By Lemma 4.2 the correspondence a ⊗ b ↔ a ⊗ σ(b) gives isomorphisms
(Z/p[[x]] ⊗ Z/p[[x]])C⊗Zp ≅ Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]],
(Z/p[x]/(xi ) ⊗ Z/p[x]/(xi ))C⊗Zp ≅ Z/p[x]/(xi ) ⊗Z/p[C⊗Zp ] Z/p[x]/(xi ).
Moreover, since Z/p[C ⊗ Zp ] → Z/p[x]/(xi ) is an epimorphism, we obtain
Z/p[x]/(xi ) ⊗Z/p[C⊗Zp ] Z/p[x]/(xi ) ≅ Z/p[x]/(xi ).
Therefore the homomorphism (4.3) is isomorphic to the multiplication homomorphism
Z/p[[x]] ⊗Z/p[C⊗Zp ] Z/p[[x]] Ð→ Z/p[[x]],
whose kernel is uncountable by Proposition 3.3.
5. Proof of Main Theorem
Since the double lamplighter pro-p-group is 3-generated, we have a continuous epimorphism F̂p ↠ DL, where F is the 3-generated free group. Then the statement of the theorem
for three generated free group follows from Proposition 4.3 and Corollary 2.6. Using that
the 3-generated free group is a retract of k-generated free group for k ≥ 3, we obtain the
result for k ≥ 3. The result for two generated free group follows from Lemma 11.2 of [3].
10
SERGEI O. IVANOV AND ROMAN MIKHAILOV
Acknowledgement
The research is supported by the Russian Science Foundation grant N 16-11-10073.
References
[1] N. Bourbaki: Commutative Algebra. Addison-Wesley, Paris (1972).
[2] A. K. Bousfield: Homological localization towers for groups and π-modules, Mem. Amer. Math. Soc,
no. 186, 1977.
[3] A. K. Bousfield: On the p-adic completions of nonnilpotent spaces, Trans. Amer. Math. Soc. 331 (1992),
335–359.
[4] K.S. Brown, E. Dror: The Artin–Rees property and homology, Israel J. Math. 22 (1975) 93–109.
[5] H. Cartan: Algébre d’Eilenberg–MacLane et homotopie, Matematika, 3:6 (1959), 3–46
[6] G. A. Fernandez-Alcober, I. V. Kazatchkov, V. N. Remeslennikov, P. Symonds: Comparison of the
discrete and continuous cohomology groups of a pro-p-group, St. Petersburg Math. J., 19 (2008), 961–973.
[7] R. Grigorchuk, R. Kravchenko: On the lattice of subgroups of the lamplighter group, Int. J. Alg. Comp.
24 (2014), 837–877.
[8] S. O. Ivanov and R. Mikhailov: On a problem of Bousfield for metabelian groups: Advances in Math.
290 (2016), 552—589.
[9] P. H. Kropholler: A note on the cohomology of metabelian groups, Math. Proc. Camb. Math. Soc. 98
(1985), 437–445.
[10] N. Nikolov, D. Segal: On finitely generated profinite groups, I: strong completeness and uniform bounds,
Annals of Math. 165 (2007), 171–238
[11] L. Ribes, P. Zalesskii: Profinite Groups, Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge,
Vol. 40 (2000).
[12] J.-P. Serre: Galois cohomology, Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2002.
Chebyshev Laboratory, St. Petersburg State University, 14th Line, 29b, Saint Petersburg,
199178 Russia
E-mail address: [email protected]
Chebyshev Laboratory, St. Petersburg State University, 14th Line, 29b, Saint Petersburg,
199178 Russia and St. Petersburg Department of Steklov Mathematical Institute
E-mail address: [email protected]
| 4math.GR
|
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
Selective-Candidate Framework with
Similarity Selection Rule for Evolutionary
Optimization
Sheng Xin Zhang, Wing Shing Chan, Member, IEEE, Zi Kang Peng,
Shao Yong Zheng, Senior Member, IEEE and Kit Sang Tang, Senior Member, IEEE
Abstract—This paper proposes to resolve limitations of the
traditional one-reproduction (OR) framework which produces
only one candidate in a single reproduction procedure. A
selective-candidate framework with similarity selection rule
(SCSS) is suggested to make possible, a selective direction of
search. In the SCSS framework, M (M > 1) candidates are
generated from each current solution by independently
conducting the reproduction procedure M times. The winner is
then determined by employing a similarity selection rule. To
maintain balanced exploitation and exploration capabilities, an
efficient similarity selection rule based on the Euclidian distances
between each of the M candidates and the corresponding current
solution is proposed. The SCSS framework can be easily applied
to any evolutionary algorithms or swarm intelligences.
Experiments conducted with 60 benchmark functions show the
superiority of SCSS over OR in three classic, four state-of-the-art
and four up-to-date algorithms.
Index Terms—Selective-candidate, similarity selection,
exploitation and exploration, evolutionary algorithm (EA), swarm
intelligence (SI).
I. INTRODUCTION
C
ONSTRUCTED on a population basis, evolutionary
optimization explores a searching space by iteratively
performing genetic operations (for evolutionary algorithms,
EAs [1, 2]) or social learning processes (for swarm
intelligences, SIs [3]). In order to maintain diversity in the
search, the genetic operations and social learning processes
commonly involve some levels of randomness. For example,
parents are randomly selected in the mutation process of classic
differential evolution (DE) [4, 5]; normal distributions are
utilized in evolution strategy (ES) [6]; and two randomly
distributed numbers are used in the velocity update equation of
particle swarm optimization (PSO) [7]. Due to the randomness,
M (M > 1) independent reproduction performed on the same
current solution would produce M different offspring, which
bring up different building blocks that affects the search
performance differently. However, not only in the classic EAs
[8-11] and SIs [12-14], but many of their invariants (eg.
improved EAs [4, 5, 15-19] and SIs [7, 20-23]), one
reproduction (OR) procedure is usually performed generating
one offspring at a time, and as a result, the candidate for each
current solution is determined in a random manner.
This paper challenges this traditional OR framework and
proposed a selective-candidate framework (SC) to determine
the candidates. SC firstly generates M (M > 1) candidates for
each current solution by independently performing the
reproduction procedures M times. Afterwards, these M
candidates compete, with one of them becoming the final
candidate for each current solution based on a selective rule
(SR). The big challenge here is that SR should be effective and
efficient. On the one hand, it should not involve objective
function evaluation, which may waste function evaluations.
While on the other hand, it is required to provide an excellent
candidate for the next generation. To resolve these issues, a
similarity selection (SS) rule is proposed which measures the
quality of a candidate, based on its Euclidian distance to the
current solution. As shown later, SS is able to secure a balance
in the exploitation and exploration capabilities.
To demonstrate the effectiveness of the proposal, the SCSS
framework is incorporated into two popular EAs (DE and ES)
and one popular SI (PSO). Extensive experiments conducted
using the CEC2014 [24] and CEC2017 [25] test suits show that
SCSS significantly improves the performance of the considered
algorithms, including three classic, four start-of-the-art, and
four up-to-date top algorithms.
The rest of this paper is organized as follows. Section II
briefly reviews the procedures of DE, ES, and PSO. Section III
analyses their drawbacks and describes the proposed
framework. Section IV reports the experimental results.
Finally, Section V concludes this paper.
II. BACKGROUND AND RELATED WORKS
S. X. Zhang, W. S. Chan and K. S. Tang are with the Department of
Electronic Engineering, City University of Hong Kong, Hong Kong
(e-mail:[email protected];
[email protected];
[email protected]).
Z. K. Peng and S. Y. Zheng are with the School of Electronics and
Information Technology, Sun Yat-sen University, Guangzhou, China. (email:
[email protected]; [email protected]).
This work is submitted to the IEEE for possible publication.
A. Review of Evolutionary Algorithms and Swarm Intelligences
In this subsection, we first briefly review and present the
flow of three popular EAs and SIs including DE, ES and PSO.
>
2
1. DE
Differential evolution (DE) as proposed by Storn and Price
[8, 9] is a simple yet powerful EA. At each generation G, three
genetic operations, namely mutation, crossover, and selection
are included.
Initialization: Given a D-dimensional minimization problem,
DE starts with a population P0 = {X1,0, X2,0, …, XNP,0} of NP
individuals which is uniformly sampled from the entire
searching space.
Mutation: Mutation in DE is performed by combining a basic
vector with one or more difference vectors to generate a mutant
vector Vi, G {i = 1, 2, …, NP}. The classic “rand/1” mutation
strategy is formulated as follows.
Vi, G = Xr1, G + F × (Xr2, G - Xr3, G)
(1)
where r1, r2 and r3 are three distinct integers within [1, NP] and
are different from the index i, while F is a mutation factor
between 0 and 1.
Crossover: After mutation, crossover is performed between
the mutant vector Vi, G and the current vector Xi, G to generate a
trial vector Ui, G as follows.
vi , j ,G if rand j (0,1) CR or j jrand
(2)
ui , j ,G
xi , j ,G otherwise
where randj(0,1) is a uniform random number in (0, 1), jrand is a
randomly generated integer from [1, D], and CR is a crossover
factor within [0,1].
Selection: Selection compares the fitness of Ui, G with that of
the corresponding Xi, G and selects the better one to enter into
the next generation.
Ui, G if f (Ui, G) ≤ f (Xi, G)
Xi, G+1 =
(3)
Xi, G otherwise
2. ES
Evolution strategy (ES) first appeared in 1964 at the
Technical University of Berlin (TUB), and was used to solve
hydrodynamic problems [10]. Different versions of ES have
been proposed since this first version. Generally, ES can be
categorized according to the number of parents and offspring
involved in each generation. (1+1)-ES includes only one
parent, which generates one offspring for each generation by
means of Gaussian mutation. (μ + 1)-ES uses μ (μ > 1) parents
to generate one offspring per generation. (μ + λ)-ES utilizes μ
parents to generate λ (λ > μ) offspring and then chooses μ
individuals from the (μ + λ) individuals to enter next
generation, while (μ, λ)-ES chooses μ individuals only from the
λ offspring.
Initialization: Given a D-dimensional minimization problem,
ES starts with an initial population P0 = {X1,0, X2,0,…, Xμ,0} of μ
individuals. Each individual Xi,0 = [xi,1,0, xi,2,0,…, xi, D,0, σi,1,0,
σi,2,0,…, σi, D,0], (i = 1, 2, …, μ) has D variables and D
independent standard deviations. The initial standard deviation
σi,0 is calculated as
i ,0
X i
(4)
D
where ∆Xi is the Euclidian distance between Xi,0 and the fittest
individual in the initial population.
Recombination: At each generation G, recombination is
performed on two randomly selected individuals to produce a
new individual XRi,G {i = 1, 2, …, λ}. Different recombination
strategies are specified as follows:
x p , j ,G , without recombination
(5)
xri , j ,G x p , j ,G or xq , j ,G , discrete recombination
x p , j ,G ( xq , j ,G x p , j ,G ), intermediate recombination
where p and q are the two distinct integers uniformly selected
from the set {1, 2, …, μ}, j = 1, 2, …, D is the dimension to be
recombined and χ is a constant value usually set to 0.5 [1].
Mutation: Following recombination, mutation is performed
to generate λ mutant individuals XMi,G{i = 1, 2, …, λ}as
described by the following:
(6)
i , j ,G i , j ,G exp( N (0,1) Ni (0,1))
xmi , j ,G xri , j ,G N (0, i , j ,G )
(7)
where j = 1, 2, …, D, N(0,1) and Ni(0,1) are two normal
distributions, τ′and τ are constants usually set as unity.
Selection: Select μ fittest individuals from the set of μ + λ
individuals ((μ + λ)-ES), or from the set of λ offspring produced
by mutation ((μ, λ)-ES).
3. PSO
Particle swarm optimization (PSO) as proposed by Kennedy
and Eberhart [12] imitates the swarm behavior of animals, such
as birds flocking and fish schooling. Given a D-dimensional
minimization problem, PSO explores the searching space by
utilizing a swarm of NP particles with each particle associated
with a velocity vector Vi = [vi1, vi2,…, viD] and a position vector
Xi = [xi1, xi2,…, xiD], i = 1, 2, …, NP. During the searching
process, each individual historical best position vector is
recorded in pbesti = [pi1, pi2,…, piD] and the global best position
vector is stored in gbest = [g1, g2,…, gD]. Based on pbesti and
gbest, particles update their velocity and position at each
iteration by using Eq. (8) and (9) respectively:
vij w vij c1 r1 j ( pbestij xij ) c2 r2 j ( gbest j xij ) (8)
xij xij vij
(9)
where w is the inertia weight, c1 and c2 are the acceleration
constants, which are commonly set to 2.0. r1j and r2j are two
uniformly distributed random numbers within (0, 1) for each
dimension j. The updated velocity vij on each dimension is
bounded by a maximum value VMAXj,. If vij exceeds VMAXj, then
it is set as sign(vij) VMAXj.
4. General procedures
From above, the general procedures for EAs and SIs is
summarized as Algorithm 1.
--------------------------------------------------------------------------Algorithm 1. General Procedures of EAs and SIs
--------------------------------------------------------------------------1: Initialize population X = {X1, X2, …, XNP};
2: While the stopping criteria are not met Do
3: Determine the control parameters CP for genetic operations
or social learning ;
4: Produce a new population Y via genetic operations or social
learning on X;
5: Evaluate the fitness of Y;
6: Select solutions as new X from XY to enter next iteration.
>
7: End While
--------------------------------------------------------------------------B. Exploitation and Exploration
Exploitation and exploration are two cornerstones of
evolutionary optimization [26]. In 2013, Črepinšek et al [27]
conducted a comprehensive survey of the research works that
focused on balancing the exploitation and exploration
capabilities (EEC) of EAs. According to [27], the amount of
exploitation and exploration in EAs is not only affected by the
genetic operations, i.e. mutation, crossover, and selection, but
also by the algorithmic parameter setting and the representation
of individuals. Moreover, even when given only one operation,
it is often hard to distinguish between exploitation and
exploration because of the different points of view. As
classified by Črepinšek et al, the existing way of balancing
EEC in EAs can be divided into four categories: diversity
maintenance, diversity control, diversity learning and other
direct methods. Detailed information can be found in [27]. In
the following, the related works that were not included in [27],
will be reviewed. This includes more recent methods that were
proposed after 2013 and related works conducted on a popular
SI algorithm, i.e. PSO.
With respect to the research on DE, in [28], Epitropakis et al
proposed a proximity-based (Pro) mutation operator for DE, in
which the probability of a parent involved in mutation is
inversely proportional to its distance from the mutated
individual. In [29], Gong et al proposed a ranking-based (Rank)
mutation operator, in which better parents are associated with
higher selection probabilities for mutation. In [30], Zheng et al
proposed a collective information based (CIM) mutation
operator to combine the useful information provided by
multiple promising solutions. Pro, Rank and CIM mutation all
focus on exploitation. In [31], Wang et al developed a
multiobjective sorting-based (MS) mutation operator for DE,
which considers the fitness as well as the diversity of the
parents selected for mutation. In [32], Tanabe and Fukunaga
improved the performance of JADE [33] by introducing a
success-history based parameter adaptation (SHA) scheme,
which can generate more diverse control parameters than the
parameter adaptation method of JADE. In [34], Tanabe and
Fukunaga further enhanced the performance of SHADE [32] by
adopting a linear population size reduction (LPSR) scheme.
LPSR maintains a large population to promote diversity at the
earlier evolution stages but a small population for exploitation
at the latter stages. In [35], Wu et al proposed a
multi-population based ensemble of mutation strategies
approach for DE, which includes mutation strategies with
different amounts of EEC in multiple subpopulations to
diversify the search. In [36], Cui et al assigned different
mutation strategies to different sub-populations with different
fitness values. In [37], Tang et al proposed to assign
individual-dependent control parameters and mutation
strategies according to the differences in fitness values. In [38],
Wang et al proposed to use three different mutation strategies
combined with three different pairs of control parameters to
generate multiple solutions that are used for selecting the fittest
ones. In [39], Gong et al took advantage of multiple operators
to generate multiple offspring that are used for selecting
3
solutions with maximum density values by using a cheap
surrogate model. In [40], Yang et al proposed a gene-level
auto-enhanced population diversity (AEPD) mechanism to
improve the diversity of a population. In [41], Du et al
introduced an event-triggered impulsive (ETI) control scheme
into DE. ETI includes stabilizing impulses for exploitation and
destabilizing impulses for exploration, which could be
triggered once the stagnation of DE is detected. In [42], Qiu et
al developed a multiple exponential recombination to improve
the exploration ability of the crossover operation of DE. In [43],
Arabas and Biedrzycki proposed to monitor the population
midpoint in order to enhance the exploitation of EAs.
Considering ES, the simple multi-membered evolution
strategy (SMES) [44] uses a simple diversity mechanism and a
simple feasibility-based comparison mechanism to search the
global optimum more effectively. Meanwhile, a hybrid
panmictic recombination is also included to improve the
exploitation capability. Covariance matrix adaptation ES
(CMA-ES) [18], which is a popular ES variant, adapts the
population to the function landscape by adaptively adjusting
the covariance matrix. Restart CMA-ES with increasing
population size (IPOP-CMA-ES) [45] increases the population
size of CMA-ES and restarts the search when any one of the
preset stopping criterions is met, which allows a more global
search of CMA-ES and enhances the exploration ability.
BI-population CMA-ES (BIPOP-CMA-ES) [46] divides the
population into two interlaced multi-start regimes with the
same function evaluation budget. One regime has an increasing
population size while population size of the other one does not
change rapidly. Particle swarm CMA-ES (PS-CMA-ES) [47]
combines the global exploration advantage of PSO and the
local search ability of CMA-ES to compromise exploration and
exploitation.
Regarding PSO, distance-based locally informed PSO (LIPS)
[20] utilizes the local best experiences, instead of the global
best experience to update the velocities of particles and
consequently exhibits promising performance in multimodal
optimization. Territorial PSO (TPSO) [48] uses a “collision
operator” to prevent premature convergence and the weighted
average of neighbors’ best for efficient exploitation. Diversity
enhanced PSO with neighborhood search (DNSPSO) [49]
introduces an operation which is similar to the crossover
operation of DE to improve the diversity of the swarm. Two
neighborhood search strategies were also introduced in
DNSPSO, which aimed for a balanced EEC. Scattering
learning PSO (SLPSOA) [50] archives high-quality solutions in
an exemplar pool (EP) and subsequently uses them to lead the
swarm to regions with more promising solutions. Besides,
SLPSOA also adopts a local search algorithm to refine
solutions in the newfound regions. Heterogeneous
comprehensive learning PSO (HCLPSO) [51] divides the entire
swarm into two subpopulations including an exploration one
and an exploitation one, with an exploration-enhanced and an
exploitation-enhanced velocity update strategy used in the
corresponding subpopulation, respectively.
>
4
III. PROPOSED FRAMEWORK
A. Motivations
From the review presented in Section II-A, it is found that in
the traditional EAs and SIs, it is common that only one
candidate is generated from a current solution by applying a
reproduction procedure. However, due to the random nature
available in the reproduction procedure, it is not guaranteed that
the candidates will be located in promising searching areas. It is
obvious that, when the reproduction procedure is repeated for
once more, the two candidates from the same current solution
will be different. Thus, a question naturally rises: which one is
more beneficial for the search? This issue has never been
considered in existing EAs and SIs and the canonical
one-reproduction (OR) framework (Algorithm 1) is simply
followed. To alleviate the possible adverse effect from
randomness and improve the performance, in this paper, we
propose a selective-candidate framework with similarity
selection rule (SCSS), in which M candidates (M > 1) are
generated for each of the current solutions by independently
performing the reproduction procedure M times. Afterwards,
only one of them is selected by the selective rule (SR), to
become the final competitor against each current solutions.
B. SC Framework
To demonstrate the potential drawback in the existing OR
framework, the procedures of two representative algorithms,
DE and PSO from EA and SI families respectively are
reviewed.
In the traditional DE with “rand/1” mutation strategy, a
mutant solution V is first generated for each of the current
solutions X, by randomly selecting three vectors for mutation.
Afterwards, the crossover is performed between the pairs of
current solution X and its corresponding mutant solution V to
form a trial solution U. Obviously, the trial solution U inherits a
certain degree of randomness from the mutant solution V,
which makes U unpredictable. For this reason, whether the
newly generated offspring U brings benefit is unknown.
Moreover, due to the random selection of solutions for
mutation, if the genetic procedures (i.e. mutation and
crossover) is repeated once again, the generated U’ is seldom
the same as U. Appropriate determination of the final offspring
from U and U’ may be beneficial for the search. However, the
OR framework indeed misses U’.
For traditional PSO, the velocity and position of each particle
are updated in each iteration. In the velocity update equation,
r1d and r2d introduce randomness in the velocity V in order to
explore the search space. Then, in the position update equation,
each particle utilizes the velocity V to update its position X.
Similar to the U in DE, X in PSO is also random and would vary
for different trials.
To control the amount of randomness in EAs and SIs, this
paper proposes a selective-candidate (SC) framework, which
selectively determines a candidate from M (M > 1) candidates
produced by M reproduction processes. The pseudo-code is
presented in Algorithm 2. It can be seen from Algorithm 2 that
distinct from the traditional OR framework (Algorithm 1), SC
performs the reproduction procedures of OR framework M
times with independent parameter settings (line 4) and
independent reproduction operations (lines 5-7). Thus, for each
current solution Xi, it owns a candidate pool Yim {m = 1, 2, …,
M}. Afterwards, one solution Yi from the corresponding M
candidates is selected for each Xi. (line 10). Besides, the
parameters actually used for each selected solution are
determined (line 11).
It should be pointed out that [38] and [39] also use multiple
candidates, as surveyed in Section II-B. However, they do
indeed belong to the OR framework and the motivation is
essentially different from that of the proposed SC. [38] and [39]
aimed at combining advantages of different operators. To
improve the performance, extra mutation operators have to be
carefully chosen [38, 39]. While SC provides an insight into the
potential drawbacks of a single reproduction process, no extra
operations are required and the baseline can be
“self-improved”. Moreover, experiments conducted in Section
IV also confirm that SC could also improve the “M operators,
M candidates” method.
--------------------------------------------------------------------------Algorithm 2. SC Framework
--------------------------------------------------------------------------1: Initialize population X = {X1, X2, …, XNP};
2: While the stopping criteria are not met Do
3: For m = 1: M
4: Determine the control parameters CPm for genetic
operations or social learning, following the original design
of the baseline;
5: For i = 1: NP Do
6: Produce new solution Yim via genetic operations or social
learning on Xi;
7: End For
8: End For
9: For i = 1: NP Do
10: Determine Yi from Yim {m = 1, 2, …, M};
11: Determine the actually used CPi of Yi from CPim {m = 1,
2, …, M};
12: End For
13: Evaluate the fitness of Y;
14: Select solutions as new X from XY to enter next iteration.
15: End While
--------------------------------------------------------------------------C. SS Rule
Apparently, the core issue of the SC framework is how to
determine the final competitor from M candidates. Due to
different parameter settings combined with different genetic
(for EAs) or different social learning (for SIs) procedures, the
M candidates can be very different. Therefore, the way of
selecting the competitors will directly affect the performance.
An effective rule for selection could bring performance
enhancement while an inappropriate one may even deteriorate
the performance. Moreover, the rule should be efficient,
meaning that the computational load should be light.
In this paper, we propose a similarity selection (SS) rule,
which is based on the Euclidian distance of each of the M
candidates Yim to its corresponding current solution Xi,
calculated as
distim =
D
(Y
j 1
m
i, j
X i , j )2
>
By adjusting SS, the amount of exploitation and exploration
of the SC framework can be directly controlled. For instance,
favoring candidates closest to the current solutions are
exploitative while preferring the ones farthest to the current
solutions could encourage exploration. Moreover, different
from the existing works [28-31,48-51], as reviewed in Section
II-B, that investigated the exploitation and exploration effects
of operations and parameter settings separately, SS focuses on
the candidates, which reveal the combined effects of operations
and parameters in the reproduction process.
However, the appropriate choice of an SS rule for a specific
algorithm depends on the EEC of the given algorithm. It is
assumed that EEC could be represented by a searching radius
(SRAD). A larger SRAD would result in a more explorative
characteristic. The effects of SRAD on the performance of an
algorithm are illustrated in Fig. 1. In Fig. 1 (a), optimizer 1 is
very explorative with a large SRAD. Therefore, the search is
very random and there are little risks suffered from local optima.
However, this large SRAD would make the individuals such as
1 and 2 hard to refine. In Fig. 1(b), optimizer 2 is very
exploitative with a small SRAD. In this case, individuals 1 and
2 focuses more on local searches. It is difficult for them to
move from basin I to basin II, which is important for diversity
enhancement. Different from optimizers 1 and 2, in Fig. 1 (c),
optimizer 3 has a balanced EEC with an appropriate SRAD.
However, one drawback with this optimizer is that the SRAD is
the same for the superior individual 1 and the inferior
individual 2. For the superior individual 1, this SRAD may not
be small enough for an efficient local search while for the
inferior individual 2, the SRAD may not be large enough for it
to jump from basin I to basin II. In Fig. 1(d), optimizer 4 is an
improved version of optimizer 3 based on SC with M = 2. The
possible candidates generated by SC could be close to the
current solutions with a small SRAD (solid line circles in
Fig.1(d)), such as A1, B1,C1 and D1, or away from with a large
SRAD (dotted line circles in Fig. 1(d)), such as A2, B2, C2,
and D2. Indeed, the SRAD size of optimizer 3 ranges between
those of the dotted line circles and the solid line circles of
optimizer 4. Therefore, compared with the SRAD of optimizer
3, the solid line circles of optimizer 4 provide a smaller radius
for local search while the dotted line circles could be large
enough for basin-jumping.
Regarding different cases: 1) for an explorative optimizer
with an SRAD that is too large (Fig.1(a)), the SRAD should be
reduced to concentrate the search; 2) for an exploitative
optimizer with an SRAD that is too small (Fig.1(b)), the SRAD
should be enlarged to encourage exploration to new searching
areas; and 3) for a well-performing optimizer with balanced
EEC, different search tasks should be assigned to solutions with
different potentials. On the one hand, since new best solutions
are likely to be located in the area near the top solutions in the
context of continuous landscape, the superior solutions are
assigned with the exploitation task. They are then compared to
the closest candidates to make them steadily exploit promising
areas. While on the other hand, to prevent the population from
rapid diversity loss, the inferior solutions are assigned with the
farthest candidates.
Two SS schemes are proposed as follows:
5
Fig.1 Illustration of the effects of SRAD on the performance of an algorithm.
Scheme 1: If rank(i) ≤ ceil(NP×GD)
Select the closest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
Else
Select the farthest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
End If
Scheme 2: If randi(0,1) > rank(i)/NP
Select the closest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
Else
Select the farthest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
End If
where rank(i) {1, 2, …,NP} is the fitness ranking of
individual Xi and rank(i)=1 is the best. randi(0,1) is a uniformly
distributed random number within (0,1) for individual Xi {i = 1,
2, …,NP}.
In Scheme 1, the proportion of top individuals preferring the
closest candidates is controlled by a greedy degree parameter
GD in the range [0,1]. Specifically, the superior GD×100%
selects the nearest candidates while the inferior (1 - GD)×100%
portion selects the farthest candidates. The larger the GD value
is, the greedier Scheme 1 becomes.
In Scheme 2, higher ranked individuals are associated with
higher probabilities in using the closest candidates, while lower
ranked ones are likely to utilize the farthest candidates. Scheme
2 is proposed for the reason that it is parameterless. As shown
later in Section IV, it works well for most of the advanced EA
and SI variants.
By adopting SS rule (Scheme 1 for example) in the SC
framework (Algorithm 2), the proposed selective candidate
framework with similarity selection rule (SCSS) is shown in
Algorithm 3. As observed from Algorithm 3, line 8 calculates
the Euclidian distances between each of the candidates Yim and
the current solution Xi. Lines 12-20 determine the final
candidate Yi and the control parameter that is actually used CPi
of Yi.
>
6
--------------------------------------------------------------------------Algorithm 3. SCSS Framework
--------------------------------------------------------------------------1: Initialize population X = {X1, X2, …, XNP};
2: While the stopping criteria are not met Do
3: Determine the fitness ranking rank(i) of each individual i
{i = 1, 2, …, NP};
4: For m = 1: M
5: Determine the control parameters CPm for genetic
operations or social learning, following the original design
of the baseline;
6: For i = 1: NP Do
7:
Produce new solution Yim via genetic operations or social
learning on Xi;
8:
Calculate distim = Euclidian distance (Yim, Xi);
9: End For
10: End For
11: For i = 1: NP Do
12:
If rank(i) ≤ ceil (NP×GD)
13:
index = arg min (distim ) ;
m{1,2,..., M }
14:
15:
16:
17:
Yi = Yiindex;
CPi = CPiindex;
Else
index = arg max (distim ) ;
m{1,2,..., M }
18:
Yi = Yiindex;
19:
CPi = CPiindex;
20:
End If
21: End For
22: Evaluate the fitness of Y;
23: Select solutions as new X from XY to enter next iteration.
24: End While
--------------------------------------------------------------------------D. The SCSS Variants
Based on Algorithm 3, the SCSS variants for the traditional
EAs and SIs can be easily implemented. The work flow of three
SCSS variants, i.e. SCSS-DE, SCSS-ES and SCSS-PSO for the
classic DE, ES, and PSO are given in Algorithms 4, 5 and 6
respectively. The arrows “ ” highlight the differences
between the SCSS variants and the baseline algorithms.
Remark 1: In SCSS framework, the control parameters that
are actually used, CPi of Yi should be determined (lines 15 and
19 in Algorithm 3) for the reason that different reproduction
procedure m may use different CP and the CP may have further
usages. For example, in the well-known JADE [33] and
SHADE [32] algorithms, control parameters F and CR are
generated according to Cauchy and normal distributions,
respectively and after selection, the successful CP are archived
to determine new location parameters of Cauchy and normal
distributions. Thus, in SCSS, the generations of F and CR are
independent in each reproduction procedure m and the
successful CP that are actually used is archived. In Algorithms
4 and 6, this is not shown because the classic DE and PSO use
pre-defined fixed CP, i.e. F and CR in DE and w, c1 and c2 in
PSO.
Remark 2: In PSO, the personal best position of each particle
is regarded as a current solution for the similarity calculation
(line 12 in Algorithm 6).
Remark 3: Different from the one-to-one reproduction
procedures in DE and PSO, λ offspring is generated by using μ
parents in ES. Therefore, we treat the λ new individuals XR
produced by recombination as the current solutions, and their
fitness rankings are calculated to be the average ranking of the
pth and qth individuals used to perform recombination (lines 6
and 7 in Algorithm 5).
--------------------------------------------------------------------------Algorithm 4. SCSS-DE
--------------------------------------------------------------------------1: Set the population size NP, initialize the population P0 =
{X1,0, X2,0, …, XNP,0}, set F and CR, set the generation
counter G = 0;
2: Set GD;
3: While the stopping criteria are not met Do
4: Determine the fitness ranking rank(i) of each individual i
{i = 1, 2,…, NP};
5: For m = 1: M
6: For i = 1: NP Do
-------------------------Mutation -------------------------7:
Generate a mutant vector Vim, G using Eq. (1);
-------------------------Crossover------------------------8:
Generate a trial vector Uim, G using Eq. (2);
9: distim = Euclidian distance (Uim, G, Xi, G);
10: End For
11: End For
12: For i = 1: NP Do
13:
If rank(i) ≤ ceil(NP×GD)
m
14:
index = arg min (disti ) ;
m{1,2,..., M }
15:
16:
17:
Ui, G = Uiindex, G;
Else
index = arg max (distim ) ;
m{1,2,..., M }
18:
Ui, G = Uiindex, G;
19:
End If
20: End For
21: Evaluate the fitness of Ui, G {i = 1, 2, …, NP};
-------------------------Selection--------------------------22: For i = 1: NP Do
23: If f (Ui, G) ≤ f (Xi, G)
24:
Xi, G+1 = Ui, G;
25: Else
26:
Xi, G+1 = Xi, G;
27: End If
28: End For
29: G = G + 1;
30: End While
----------------------------------------------------------------------------------------------------------------------------------------------------Algorithm 5. SCSS-ES
--------------------------------------------------------------------------1: Set the population size μ, initialize the population P0 = {X1,0,
X2,0, …, Xμ,0}, set the generation counter G = 0;
2: Set GD;
3: While the stopping criteria are not met Do
4: Determine the fitness ranking RANK(k) of each individual k
{k = 1, 2,…, μ};
>
7
5: For i = 1: λ Do
----------------------Recombination------------------6: Randomly choose p and q, use the pth and qth individuals
from PG to generate a new individual XRi,G with the
recombination strategy, i.e. Eq. (5);
7: Calculate the fitness rank(i) of individual i{i = 1, 2, …, λ}as
(RANK(p)+ RANK(q))/2;
8: End For
9: For m = 1: M
10: For i = 1: λ Do
-------------------------Mutation--------------------------11:
Use Eq. (6) and (7) to mutate the individual XRi,G
produced
by recombination and generate a mutant
individual XMim,G;
12: distim = Euclidian distance (XMim,G, XRi,G);
13: End For
14: End For
15: For i = 1: λ Do
16:
If rank(i) ≤ ceil(λ ×GD)
17:
index = arg min (distim ) ;
m{1,2,..., M }
18:
19:
20:
XMi,G = XMiindex,G;
Else
index = arg max (distim ) ;
m{1,2,..., M }
21:
XMi,G = XMiindex,G;
22: End If
23: End For
24: Evaluate the fitness of all the new individuals XMi,G {i = 1,
2, …, λ};
-------------------------Selection--------------------------25: Select μ fittest individuals Xi,G {i = 1, 2, …, μ} from the
μ + λ individuals to form a new population PG+1.
26: G = G + 1;
27: End While
----------------------------------------------------------------------------------------------------------------------------------------------------Algorithm 6. SCSS-PSO
--------------------------------------------------------------------------1: Set the swarm size NP, initialize positions X = {X1, X2, …,
XNP}, initialize velocities V = {V1, V2, …, VNP}, record each
particle’s historical best position in pbesti and the global
best position in gbest, set w, c1 and c2, set iteration counter
IT = 0;
2: Set GD;
3: While the stopping criteria are not met Do
4: Determine the fitness ranking rank(i) of each particle i
{i = 1, 2, …, NP};
5: For m = 1: M
6: For i = 1: NP Do
7: For j = 1: D Do
8:
Update vijm using Eq. (8);
9:
Adjust vijm if it exceeds VMAXj;
10:
Update xijm using Eq. (9);
11: End For
12: disti m = Euclidian distance (Xim, pbesti);
13: End For
14: End For
15: For i = 1: NP Do
16:
If rank(i) ≤ ceil(NP×GD)
17:
index = arg min (distim ) ;
Xi = Xiindex;
Else
index = arg max (distim ) ;
m{1,2,..., M }
18:
19:
20:
m{1,2,..., M }
21:
Xi = Xiindex;
22: End If
23: End For
24: For i = 1: NP Do
25: Evaluate the fitness of the new position Xi;
26: If f (Xi) ≤ f (pbesti)
27:
pbesti = Xi;
28: End If
29: If f (Xi) ≤ f (gbest)
30:
gbest = Xi;
31: End If
32: End For
33: IT = IT + 1;
34: End While
--------------------------------------------------------------------------IV. SIMULATION
In this section, the effectiveness of the proposed SCSS
framework and its working mechanism are investigated
through comprehensive experiments conducted using the
CEC2014 [24] and CEC2017 [25] benchmark function sets.
Each function set consists of 30 functions with diverse
mathematic characteristics, such as unimodal, multimodal,
hybrid and composition. Since the CEC function suits are with
bounded constraints, to make the comparison fair, the
constraint handling technique adopted in the SCSS variants is
kept the same as the corresponding baselines.
The solution error value, defined as f(x) - f(x*), is used to
measure the performance of the compared algorithms, where
f(x) is the smallest fitness obtained after 104×D function
evaluations and f(x*) is the fitness of the global optimal x*.
Following [24, 25], solution error values smaller than 10-8 are
considered as zero. For each test function, 51 independent runs
are performed, while the mean and standard deviations of the
solution error values are reported. Besides, in order to draw
statistically sounded conclusions, Wilcoxon signed-rank test
[52] with 5% significance level is applied to compare the
performance. The symbols “-”, “=” and “+” represent that the
baseline algorithms perform significantly worse than, similar to
or better than the corresponding SCSS variants, respectively.
The significant ones are highlighted in bold.
A. Performance Enhancement of Classic EAs and SIs
In this subsection, the proposed SCSS framework is first
integrated with three classic EAs and SIs, i.e. DE and ES from
EA family and PSO from SIs. Performance of the resulting
variants, SCSS-DE, SCSS-ES and SCSS-PSO are compared
with the baseline algorithm, respectively.
Parameters settings for the compared algorithms are
summarized as follows:
>
8
DE and SCSS-DE: NP = 100, F = 0.7, CR = 0.5;
ES and SCSS-ES: μ = 25, λ = 100, intermediate
recombination is used;
PSO and SCSS-PSO: NP = 20, w = 0.9, c1 = 2.0, and c2= 2.0;
In addition, regarding the SS rule, Scheme 1 with GD = 1.0 is
adopted in the three SCSS variants. The reproduction times M
is set to 10 for DE and ES and 2 for PSO respectively.
The experimental results on 30-D and 50-D CEC2014
functions are presented in Tables S1 and S2, respectively, in the
supplementary file, and the comparison results are summarized
in Table I.
From Tables S1, S2 and I, the effectiveness of the proposed
SCSS framework on all the considered algorithms can be
observed. In the total 180 cases, SCSS variants win in 137
(=27+27+15+26+28+14) cases and only lose in 2 cases.
Specifically, in the 30-D cases, SCSS-DE and SCSS-ES both
perform significantly better than their corresponding baseline
in 27 functions and losing only in one function. SCSS-PSO
wins PSO in 15 functions and ties in 15 functions. In the 50-D
case, SCSS-DE, SCSS-ES, and SCSS-PSO win the baselines in
26, 28 and 14 functions and the rests are tie. It should be
remarked that, since the classic algorithms use fixed parameter
settings, these performance improvements are attributed to the
control of the randomness of the reproduction operations by
SCSS, such as the random selection of parents for the mutation
in DE. In summary, the proposed SCSS framework
significantly enhances the performance of these basic
algorithms.
TABLE I COMPARISON RESULTS OF THREE SCSS- BASED
CLASSIC ALGORITHMS WITH THE BASELINES ON CEC2014
TEST FUNCTIONS
-/=/+
DE v.s. SCSS-DE
ES v.s. SCSS-ES
PSO v.s. SCSS-PSO
D = 30
27/2/1
27/2/1
15/15/0
D = 50
26/4/0
28/2/0
14/16/0
B. Performance Enhancement of Advanced EAs and SIs
Thanks to the efforts by EA and SI researchers, the
performance of the classic algorithms had been greatly
improved by many advanced variants. Thus, it is more
important that our proposed method could also enhance these
algorithms. For demonstration, SCSS is incorporated into four
well-known advanced baselines, namely JADE [33], SHADE
[32], CMA-ES [18] and LIPS [20]. Parameter settings for the
compared algorithms are set the same as recommended in their
original literature. Additionally, for the SCSSs, in SCSS-JADE,
SCSS-SHADE, and SCSS-LIPS, Scheme 2 is utilized as the SS
rule while in SCSS-CMA-ES, Scheme 1 with GD = 0 is
adopted. The reproduction times M is set to 2 for SCSS-JADE
and SCSS-SHADE, 4 for SCSS-LIPS and 5 for
SCSS-CMA-ES.
The experimental results on 30-D and 50-D CEC2014
functions are shown in Tables II and S3, respectively, in the
supplementary file and further summarized in Table III.
As observed from Tables II, S3 and III, SCSS also exhibits
significant improvements by incorporating the advanced
algorithms. Out of the total 240 cases, SCSS wins in 134
(=14+14+17+23+16+11+13+26) cases and loses in 17
(=1+0+5+2+1+0+5+3) cases. To be specific, for the advanced
DEs, i.e. JADE and SHADE, SCSS improves their
performance in 55 functions and is inferior in 2 functions. For
CMA-ES, SCSS wins in 17 and 13 cases and loses in 5 cases in
the 30-D and 50-D functions, respectively. For the advanced
PSO algorithm, i.e. LIPS, SCSS-LIPS is superior in more than
20 functions and inferior in far fewer functions in both 30-D
and 50-D cases.
Considering the diverse mathematic properties of the test
functions, it can be concluded that SCSS consistently work well
on various types of functions, including unimodal, multimodal,
hybrid and composition.
TABLE III COMPARISON RESULTS OF FOUR SCSS BASED ADVANCED
ALGORITHMS WITH THE BASELINES ON CEC2014TEST FUNCTIONS
-/=/+
JADE v.s. SCSS-JADE
SHADE v.s. SCSS-SHADE
CMA-ES v.s. SCSS-CMA-ES
LIPS v.s. SCSS-LIPS
D = 30
14/15/1
14/16/0
17/8/5
23/5/2
D = 50
16/13/1
11/19/0
13/12/5
26/1/3
C. Working Mechanism of SCSS
After verifying the effectiveness, this subsection investigates
the factors that would influence its performance. To this end,
three experiments are designed as follows.
Firstly, the performance sensitivity of SCSS to the SS rule is
investigated. Performance of the seven SCSSs, i.e. SCSS-DE,
SCSS-ES,
SCSS-PSO,
SCSS-JADE,
SCSS-SHADE,
SCSS-CMA-ES and SCSS-LIPS with different SS rule (i.e.
Scheme 1 with six GD values, i.e. 0, 0.2, 0.4, 0.6, 0.8, 1.0 and
Scheme 2) are compared with those of the baseline algorithms,
respectively. The M value for all the SCSS variants in this
experiment is set to 2. The comparison results “-/=/+” and P-N
values (defined as the number of “-” minus the number of “+”)
are shown in Table IV.
From Table IV, we have the following observations:
1) For the classic algorithms, including DE, ES, and PSO,
SCSS variants adopting larger GD values perform better than
those with smaller ones. The reason lies in that the classic
algorithms are explorative and lack in exploitation (the case in
Fig.1(a)). Large GD values could encourage exploitation to
remedy the blindness of the search. While small GD values,
such as GD=0, make the algorithms even more explorative and
deteriorate the performance, as can be observed from Table IV.
2) For the advanced algorithms, Scheme 2 is the best choice
for SCSS-SHADE and SCSS-LIPS and the third best choice for
SCSS-JADE. For SCSS-JADE and SCSS-SHADE, the
performance of SCSSs significantly degenerates when GD is
too large (GD = 1.0) or too small (GD = 0). It is because JADE
and SHADE themselves maintain relatively balanced EEC (the
case in Fig.1 (c)). GD = 1.0 would over-emphasize exploitation
and make the algorithms too greedy while a too explorative
setting GD = 0.0 may deteriorate the performance on test
functions which need more exploitation. For SCSS-LIPS, the
performance significantly deteriorates when GD is set to 0.
3) For SCSS-CMA-ES, Scheme 1 with GD = 0 achieves the
best performance, which indicates that the original CMA-ES
(the case in Fig. 1(b)) needs more exploration in order to
improve performance. This is in accord with the statements in
some CMA-ES literature, such as IPOP-CMA-ES [45] and
PS-CMA-ES [47] that CMA-ES could benefit from enhanced
exploration capability.
In conclusion, the choice of a best SS rule depends on the
EEC of the baselines even though Scheme 2 consistently
performs significantly better than or similar to the baselines.
>
9
TABLE II PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES
ON 30-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
JADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
2.04E+03 =
(2.59E+03)
0.00E+00 =
(0.00E+00)
2.08E-05 (1.13E-04)
0.00E+00 =
(0.00E+00)
2.03E+01 (3.12E-02)
8.76E+00 =
(2.72E+00)
3.38E-04 =
(1.71E-03)
0.00E+00 =
(0.00E+00)
2.58E+01 (3.62E+00)
4.49E-03 +
(1.05E-02)
1.66E+03 (2.67E+02)
2.60E-01 (4.06E-02)
2.10E-01 (3.53E-02)
2.24E-01 =
(3.09E-02)
3.11E+00 (4.17E-01)
9.49E+00 =
(3.17E-01)
1.24E+03 (3.35E+02)
2.11E+02 (8.15E+02)
4.52E+00 (6.74E-01)
2.02E+03 =
(2.81E+03)
4.07E+03 (1.89E+04)
1.30E+02 =
(6.92E+01)
3.15E+02 =
(4.02E-13)
2.26E+02 =
(3.11E+00)
2.05E+02 (2.18E+00)
1.00E+02 (3.77E-02)
3.60E+02 =
(5.07E+01)
7.99E+02 =
(2.34E+01)
7.33E+02 (1.60E+01)
1.55E+03 =
(6.33E+02)
14/15/1
SCSSJADE
1.47E+03
(2.14E+03)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.03E+01
(7.09E-02)
7.33E+00
(3.86E+00)
1.93E-04
(1.38E-03)
0.00E+00
(0.00E+00)
2.13E+01
(4.82E+00)
9.39E-03
(1.52E-02)
1.54E+03
(2.28E+02)
2.27E-01
(4.87E-02)
1.85E-01
(3.68E-02)
2.32E-01
(3.71E-02)
2.86E+00
(3.22E-01)
9.34E+00
(4.29E-01)
8.28E+02
(3.47E+02)
4.72E+01
(2.34E+01)
4.01E+00
(8.54E-01)
1.88E+03
(2.44E+03)
2.41E+02
(1.15E+02)
1.10E+02
(6.90E+01)
3.15E+02
(4.02E-13)
2.25E+02
(3.27E+00)
2.03E+02
(6.04E-01)
1.00E+02
(3.56E-02)
3.44E+02
(5.09E+01)
8.01E+02
(1.64E+01)
7.20E+02
(7.10E+01)
1.53E+03
(6.34E+02)
SHADE
1.61E+03 =
(2.04E+03)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.02E+01 (2.78E-02)
6.42E+00 (3.15E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.10E+01 (3.81E+00)
5.31E-03 =
(1.01E-02)
1.48E+03 =
(2.35E+02)
2.10E-01 (2.67E-02)
2.23E-01 (3.61E-02)
2.27E-01 (3.04E-02)
2.97E+00 (3.67E-01)
9.51E+00 =
(3.99E-01)
9.44E+02 (3.12E+02)
3.44E+01 (1.74E+01)
3.95E+00 =
(4.72E-01)
1.09E+01 (4.61E+00)
2.13E+02 =
(1.01E+02)
6.36E+01 =
(4.93E+01)
3.15E+02 =
(4.02E-13)
2.24E+02 =
(1.01E+00)
2.04E+02 (1.04E+00)
1.00E+02 (3.26E-02)
3.16E+02 =
(3.71E+01)
7.95E+02 =
(1.99E+01)
7.25E+02 (1.02E+01)
1.45E+03 (6.13E+02)
14/16/0
SCSSSHADE
1.50E+03
(2.68E+03)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.01E+01
(2.29E-02)
4.12E+00
(3.37E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.92E+01
(3.44E+00)
7.76E-03
(1.17E-02)
1.50E+03
(2.02E+02)
1.68E-01
(2.45E-02)
2.04E-01
(3.18E-02)
2.09E-01
(3.26E-02)
2.59E+00
(3.03E-01)
9.50E+00
(4.24E-01)
5.78E+02
(2.32E+02)
2.05E+01
(1.20E+01)
3.84E+00
(6.58E-01)
8.41E+00
(3.45E+00)
1.90E+02
(1.12E+02)
7.12E+01
(6.10E+01)
3.15E+02
(4.02E-13)
2.24E+02
(1.21E+00)
2.03E+02
(4.63E-01)
1.00E+02
(3.42E-02)
3.21E+02
(4.03E+01)
7.93E+02
(2.17E+01)
7.12E+02
(5.40E+01)
1.19E+03
(3.57E+02)
CMA-ES
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.00E+01 +
(3.27E-05)
4.12E+01 (9.58E+00)
1.64E-03 =
(3.51E-03)
4.08E+02 (8.57E+01)
6.35E+02 (1.23E+02)
4.92E+03 (7.43E+02)
5.10E+03 (8.25E+02)
3.76E-01 (4.02E-01)
2.62E-01 +
(7.72E-02)
3.71E-01 +
(9.68E-02)
3.49E+00 =
(7.56E-01)
1.43E+01 (4.33E-01)
1.56E+03 =
(4.64E+02)
1.35E+02 +
(4.50E+01)
1.01E+01 (2.11E+00)
2.89E+02 (1.01E+02)
1.04E+03 (3.50E+02)
3.07E+02 (2.29E+02)
3.15E+02 +
(3.15E-12)
2.33E+02 (6.83E+00)
2.04E+02 (2.42E+00)
1.31E+02 (1.37E+02)
4.40E+02 (2.10E+02)
4.43E+03 (3.23E+03)
7.88E+02 =
(9.18E+01)
2.30E+03 (5.50E+02)
17/8/5
SCSSCMA-ES
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.13E+01
(5.20E-01)
4.19E+00
(5.18E+00)
1.59E-03
(4.45E-03)
2.31E+02
(2.00E+02)
2.17E+02
(2.74E+02)
3.49E+03
(1.10E+03)
3.58E+03
(1.15E+03)
2.40E-01
(1.01E+00)
4.24E-01
(1.46E-01)
5.66E-01
(2.97E-01)
3.21E+00
(6.63E-01)
1.38E+01
(7.44E-01)
1.71E+03
(3.84E+02)
1.78E+02
(7.13E+01)
6.74E+00
(1.58E+00)
1.49E+02
(5.45E+01)
8.64E+02
(3.05E+02)
1.16E+02
(1.10E+02)
3.15E+02
(2.57E-11)
2.26E+02
(6.96E+00)
2.03E+02
(5.20E-01)
1.26E+02
(1.58E+02)
3.40E+02
(3.93E+01)
1.25E+03
(1.41E+03)
8.00E+02
(1.45E+02)
1.58E+03
(5.95E+02)
LIPS
2.84E+07 (2.65E+07)
2.58E+03 =
(4.30E+03)
3.93E+03 (3.64E+03)
2.74E+02 (1.13E+02)
2.00E+01 +
(8.23E-05)
1.48E+01 (2.70E+00)
1.59E-03 =
(4.86E-03)
5.35E+01 (1.26E+01)
6.29E+01 (1.82E+01)
1.97E+03 (4.14E+02)
2.54E+03 (4.39E+02)
1.78E-01 =
(4.81E-02)
3.06E-01 (6.43E-02)
2.45E-01 +
(3.56E-02)
1.08E+01 (3.87E+00)
1.15E+01 (4.96E-01)
2.89E+05 (3.04E+05)
4.88E+02 =
(7.08E+02)
2.54E+01 (2.49E+01)
1.47E+04 =
(7.71E+03)
1.11E+05 (8.42E+04)
3.27E+02 (1.20E+02)
3.24E+02 (5.26E+00)
2.39E+02 (4.83E+00)
2.16E+02 (3.59E+00)
1.32E+02 (4.40E+01)
6.03E+02 (1.66E+02)
1.78E+03 (3.95E+02)
1.34E+04 (5.19E+04)
3.84E+04 (2.59E+04)
23/5/2
SCSSLIPS
5.42E+06
(6.50E+06)
5.84E+03
(8.14E+03)
2.13E+03
(1.95E+03)
1.40E+02
(6.49E+01)
2.09E+01
(4.90E-02)
7.72E+00
(2.24E+00)
2.37E-03
(4.57E-03)
2.64E+01
(6.79E+00)
3.62E+01
(8.74E+00)
9.61E+02
(2.63E+02)
2.02E+03
(4.10E+02)
7.59E-01
(1.02E+00)
2.75E-01
(5.22E-02)
3.10E-01
(7.15E-02)
3.92E+00
(8.93E-01)
1.06E+01
(4.65E-01)
1.86E+05
(2.99E+05)
4.92E+02
(9.08E+02)
8.85E+00
(2.76E+00)
1.23E+04
(7.41E+03)
4.26E+04
(5.58E+04)
2.28E+02
(1.10E+02)
3.16E+02
(5.73E-01)
2.33E+02
(5.09E+00)
2.11E+02
(1.97E+00)
1.09E+02
(2.68E+01)
4.79E+02
(9.74E+01)
1.12E+03
(1.70E+02)
1.29E+03
(2.46E+02)
1.08E+04
(6.59E+03)
>
10
TABLE IV COMPARISONS RESULTS OF SCSS VARIANTS WITH DIFFERENT SS RULES
AGAINST THE BASELINES ON 30-D CEC2014 TEST FUNCTIONS (M = 2 FOR ALL THE SCSS VARIANTS, BEST ENTRIES ARE
HIGHLIGHTED)
-/=/+ (P-N)
Scheme 1
Scheme 2
DE
ES
PSO
JADE
SHADE
CMA-ES
LIPS
GD = 0
GD = 0.2
GD = 0.4
GD = 0.6
GD = 0.8
GD = 1.0
0/5/25(-25)
0/3/27 (-27)
0/4/26 (-26)
14/9/7 (7)
12/12/6 (6)
13/15/2 (11)
16/5/9 (7)
1/13/16(-15)
25/5/0 (25)
10/15/5(5)
15/11/4 (11)
14/13/3 (11)
6/23/1 (5)
22/4/4 (18)
2/21/7(-5)
26/4/0 (26)
10/18/2(8)
19/8/3 (16)
15/14/1 (14)
0/30/0 (0)
22/5/3 (19)
11/19/0(11)
26/4/0 (26)
14/16/0(14)
15/14/1 (14)
14/16/0 (14)
1/25/4 (-3)
22/5/3 (19)
19/11/0 (19)
26/4/0 (26)
14/16/0(14)
5/11/14 (-9)
5/21/4 (1)
1/19/10 (-9)
21/5/4 (17)
21/8/1(20)
26/4/0 (26)
15/15/0(15)
2/6/22 (-20)
3/9/18 (-15)
1/21/8 (-7)
20/8/2 (18)
5/21/4(1)
26/4/0 (26)
13/17/0(13)
14/15/1 (13)
14/16/0 (14)
2/26/2 (0)
23/4/3 (20)
Fig.2 TD against the rank on 30-D CEC2014 functions F5 and F13. (The similar phenomena can be observed on all the CEC
functions, we only present that of these two functions due to space limitation.)
In the proposed SCSS framework, the selection of the closest
or farthest candidates is conducted based on the fitness ranking
of the current solutions. In this way, SCSS adjusts the level of
exploration/exploitation according to their potentials. In the
second experiment conducted on SCSS-DE and SCSS-SHADE,
SS rule is compared with the randomly selecting (RS) manner.
The total distance TD between the selected candidates and the
current solutions against the rank on 30-D functions F5 and F13
in the median run is shown in Fig. 2.
From Fig. 2, we have the following observations: 1) on the
explorative DE, SS enhances the exploitation on all the ranks,
resulting in smaller TD values than that of RS; 2) on SHADE
with relative balanced EEC, for ranks smaller than NP/2 = 50,
SS yields smaller TD compared to RS, resulting in more
exploitation. While for ranks larger than 50, it is the opposite
case; 3) on SHADE, for RS, TD does not vary much with the
rank while for SS, TD significantly increases with the rank.
Since SRAD can be roughly calculated as TD/Max_Gen, where
Max_Gen is the maximum number of generations and it is the
same for SHADE and SCSS-SHADE, SRAD TD . This
means that SRAD increases with the rank in SS while it is the
same in RS. In other words, SS is a finer strategy; 4) the smaller
TD values of SHADE compared to that of DE reveal that
SHADE is more exploitative than DE. Therefore, unlike the
case in SCSS-DE, enlarge GD in SCSS-SHADE may make the
algorithm too exploitative and deteriorate the performance, as
can be observed from Table IV.
In the third experiment, we compare the performance of
standard SCSSs with those of the two variants:
Variant I: Instead of using SS, RS is adopted in this variant.
Variant II: The opposite version of Scheme 2 is used in this
variant, as follows:
If randi(0,1) > rank(i)/NP
Select the farthest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
Else
Select the closest candidate from Yim
{m = 1, 2, …, M} for individual Xi;
End If
Except these differences, other settings for the variants are
kept the same as the standard cases. As examples, this
experiment was conducted on SCSS-JADE and SCSS-SHADE.
The experimental results on 30-D CEC2014 functions are
reported in Table S4 and summarized in Table V. As seen from
Table V, the standard SCSS algorithms exhibit significantly
better performance than the two compared variants, which
shows the superiority of the proposed SS rule. Besides,
comparing Table S4 with Table II, it is also observed that
Variant II performs significantly worse compared to the
baselines, which means that the opposite version of Scheme 2 is
an inappropriate selective rule. This confirms the illustrations
given in Section III-C and Fig. 1.
TABLE V COMPARISON RESULTS OF SCSS-JADE AND
SCSS-SHADE WITH THE VARIANTS ON 30-D CEC2014 TEST
FUNCTIONS
-/=/+
SCSS-JADE_Variant I v.s. SCSS-JADE
SCSS-JADE_Variant II v.s. SCSS-JADE
SCSS-SHADE_Variant I v.s. SCSS-SHADE
SCSS-SHADE_Variant II v.s. SCSS-SHADE
17/13/0
24/5/1
10/19/1
21/8/1
D. Performance Sensitivity to M
In SCSS, M (M > 1) reproduction procedures should be
performed. Indeed, if M is set to 1, SCSS degenerates to the
traditional OR framework. Apparently, the performance of the
SCSS is influenced by M. The SCSS variants with five different
M values, i.e. M = 2, 3, 4, 5 and 10 are investigated. Except M,
other parameter settings for the compared algorithms are set the
same as those used previously in Sections IV-A and IV-B.
Performance comparisons of the SCSS variants with the
baselines on 30-D CEC2014 functions are summarized in Table
>
11
VI. In addition, to show the dynamic performance variation
with increasing M, the performance of the SCSS variants using
adjacent M settings are also compared with each other, as
shown in Table VII.
From Table VI, it can be observed that all of the M settings
significantly improve the performance of the baselines except
SCSS-JADE and SCSS-SHADE with M = 10, which indicates
that the overall EEC maintained by M = 10 is not suitable for
JADE and SHADE.
In Table VII, for clarity, the algorithms are divided into two
categories. Category 1 includes the SCSS variants which may
perform significantly better with M > 2 than with M = 2, while
category 2 lists the SCSS variants which perform similar to or
even worse with M > 2 than with M = 2.
In category 1, it is observed that the performance of DE and
ES consistently improves when M increases. In this paper, we
only investigate cases up to M=10 because these algorithms
are significantly inferior to the advanced and top algorithms.
Moreover, increasing M will increase the computational
complexity of the algorithm. For CMA-ES and LIPS, SCSS
variants with M = 5 and M = 4 show the best performance,
respectively. It is noticed that in the SCSS-CMA-ES, GD is set
to 0, thus, larger M values would make the algorithm more
explorative.
In category 2, enlarging M does not bring significant
performance improvements. On the contrary, it may even
significantly degrade the performance, eg. M > 4 for JADE and
SHADE, or M > 2 for PSO. The reason is that different from
those in category 1 (eg. DE, ES and LIPS), JADE, SHADE and
PSO include elite individuals in their reproduction processes.
Specifically,
the
top
individuals
used
in
the
“current-to-pbest/1” mutation strategy of JADE and SHADE
and the global best gbest used in the velocity update equation of
PSO. Thus, the algorithm with too large an M value is
potentially too greedy, making the algorithms stuck in local
optima.
TABLE VI PERFORMANCE COMPARISON OF SCSS VARIANTS WITH DIFFERENT
M SETTINGS WITH THE BASELINES (BEST ENTRIES ARE HIGHLIGHTED)
-/=/+ (P-N)
M=2
M=3
M=4
M=5
M = 10
DE
21/8/1 (20)
25/4/1 (24)
23/6/1 (22)
26/3/1 (25)
27/2/1 (26)
ES
26/4/0 (26)
27/2/1 (26)
27/2/1 (26)
27/3/0 (27)
27/3/0 (27)
PSO
13/16/1(12)
12/17/1(11)
10/18/2(8)
10/17/3 (7)
15/15/0 (15)
JADE
14/15/1 (13)
14/15/1 (13)
13/14/3 (10)
8/10/12 (-4)
16/13/1 (15)
SHADE
13/15/2 (11)
14/13/3 (11)
12/14/4 (8)
12/8/10 (2)
14/16/0 (14)
CMA-ES
13/15/2 (11)
15/11/4 (11)
15/11/4 (11)
17/7/6 (11)
17/8/5 (12)
LIPS
23/4/3 (20)
23/4/3 (20)
23/4/3 (20)
20/5/5 (15)
23/5/2(21)
TABLE VII PERFORMANCE COMPARISON BETWEEN SCSS VARIANTS
WITH ADJACENT M SETTINGS
CATEGORY 1
-/=/+ (P-N)
M = 2 V.S. M = 3
M = 3 V.S. M = 4
M = 4 V.S. M = 5
M = 5 V.S. M = 10
DE
17/12/1 (16)
5/24/1 (4)
6/23/1 (5)
12/14/4 (8)
ES
25/5/0 (25)
23/7/0(23)
12/18/0 (12)
28/2/0 (28)
CMA-ES
8/18/4 (4)
7/20/3(4)
1/29/0 (1)
5/18/7 (-2)
LIPS
8/21/1 (7)
2/28/0 (2)
0/29/1 (-1)
3/23/4 (-1)
CATEGORY 2
-/=/+ (P-N)
M = 2 V.S. M = 3
M = 3 V.S. M = 4
M = 4 V.S. M = 5
M = 5 V.S. M = 10
PSO
1/22/7 (-6)
1/28/1 (0)
0/29/1 (-1)
0/28/2 (-2)
JADE
3/25/2 (1)
5/22/3 (2)
4/18/8 (-4)
5/7/18 (-13)
SHADE
4/24/2 (2)
4/23/3 (1)
7/15/8 (-1)
5/13/12 (-7)
E. Application in Top Methods from CEC Competitions
From Sections IV-C and IV-D, advanced SCSS-DEs with
Scheme 2 and M = 2, SCSS-CMA-ES with Scheme 1(GD = 0)
and M = 5 exhibit promising performance. In this subsection,
to demonstrate the flexibility, SCSS is further applied with
these settings to four highly competitive algorithms from the
CEC competitions. Among them, L-SHADE [34] is the winner
of the CEC2014 competition, UMOEA-II [53] and
L-SHADE_EpSin [54] are the joint-winner of the CEC2016
competition and jSO [55] is the best-performing algorithm in
the CEC2017 competition. Parameter settings for the top
algorithms are set the same as the original literature. The
experimental results on 30-D and 50-D functions are presented
in Tables (S5 and S6) and summarized in Table VIII.
From Tables S5, S6 and VIII, it is observed that SCSS also
enhances the performance of these top methods. Out of the total
240 cases, SCSSs wins in 88 (=10+9+8+7+18+10+13+13)
cases and loses in 12 (=2+1+0+2+2+3+0+2) cases. Specifically,
in the 30-D case, SCSS-L-SHADE, SCSS-UMOEA-II,
SCSS-L-SHADE_EpSin, and SCSS-jSO perform significantly
better than the corresponding baseline in 10, 9, 8 and 7 cases
and underperform in 2, 1, 0 and 2 cases, respectively. In the
50-D case, the performance improvements are more significant.
SCSS-L-SHADE, SCSS-UMOEA-II, SCSS-L-SHADE_EpSin
and SCSS-jSO exhibit superior performance in 18, 10, 13 and
13 functions respectively and are inferior in far fewer functions.
These performance enhancements reveal that the proposed
SCSS framework could remedy the drawbacks of the
traditional OR framework, which not only exist in the classic
algorithms, but also in these carefully designed top algorithms.
TABLE VIII COMPARISON RESULTS OF FOUR SCSS BASED
TOP ALGORITHMS WITH THE BASELINES ON CEC2014TEST
FUNCTIONS
-/=/+
L-SHADE v.s. SCSS-L-SHADE
UMOEA-II v.s. SCSS-UMOEA-II
L-SHADE_EpSin v.s. SCSS-L-SHADE_EpSin
jSO v.s. SCSS-jSO
D = 30
10/18/2
9/20/1
8/22/0
7/21/2
D = 50
18/10/2
10/17/3
13/17/0
13/15/2
F. Performance on CEC2017 Test Suit
In this subsection, we further test the SCSS framework on the
newly developed CEC2017 test suit [25]. Parameter settings for
the algorithms are the same as those used in Sections IV-A,
IV-B and IV-E. Tables S7-S12 presents the experimental
results on 30-D and 50-D functions and Table IX summarizes
the comparison results. From Table IX, it is clear that SCSS
also significantly improves the performance of the baselines on
the CEC2017 functions. In the total 660 cases, SCSS wins in
383 cases, ties in 262 cases and loses in 15 cases. This further
confirms the superiority of SCSS over OR on a wide variety of
functions.
TABLE IX COMPARISON RESULTS OF SCSS VARIANTS
WITH THE BASELINES ON CEC2017 TEST SUIT
-/=/+
DE v.s. SCSS-DE
ES v.s. SCSS-ES
PSO v.s. SCSS-PSO
JADE v.s. SCSS-JADE
SHADE v.s. SCSS-SHADE
CMA-ES v.s. SCSS-CMA-ES
LIPS v.s. SCSS-LIPS
L-SHADE v.s. SCSS- L-SHADE
UMOEA-II v.s. SCSS-UMOEA-II
L-SHADE_EpSin v.s. SCSS-L-SHADE_EpSin
jSO v.s. SCSS-jSO
Total
D = 30
D = 50
29/1/0
27/3/0
30/0/0
30/0/0
23/7/0
19/11/0
19/11/0
18/10/2
7/23/0
11/19/0
18/11/1
16/14/0
28/1/1
28/1/1
9/18/3
15/15/0
3/24/3
14/14/2
7/21/2
13/17/0
7/23/0
12/18/0
383/262/15
>
12
G. Scalability Study
To study scalability, the SCSS framework is also tested on
100-D CEC2017 functions. The four top methods are selected
for this experiment and the parameters are set the same as those
used previously.
As shown in Tables S13 and X, SCSS still yields significant
performance improvements in the higher dimensional, which
are much more difficult than the lower ones. In the total of 120
cases, SCSS outperforms in 70 (=20+14+16+20) cases and
underperforms in 6 (=2+2+0+2) cases.
Furthermore, the overall performance of the four
SCSS-based top algorithms and the baselines are compared
according to multiple problem Wilcoxon’s test [56] and
Friedman’s test [56]. Based on multiple problems Wilcoxon’s
test, Table XI shows that the SCSS variants perform
significantly better than the corresponding baselines at α = 0.05.
With respect to the Friedman’s test, Table XII indicates that
SCSS-jSO is the best-performing algorithm, which achieves
the smallest ranking value of 2.76, followed by
SCSS-L-SHADE_EpSin.
TABLE X COMPARISON RESULTS OF FOUR SCSS- BASED TOP
ALGORITHMS WITH THE BASELINES ON 100-D CEC2017
TEST FUNCTIONS
-/=/+
L-SHADE v.s. SCSS- L-SHADE
UMOEA-II v.s. SCSS-UMOEA-II
L-SHADE_EpSin v.s. SCSS-L-SHADE_EpSin
jSO v.s. SCSS-jSO
D = 100
20/8/2
14/14/2
16/14/0
20/8/2
TABLE XI COMPARISON RESULTS OF THE TOP SCSS VARIANTS WITH THE
BASELINES ON 30-D, 50-D AND 100-D CEC2017 BENCHMARK SET
ACCORDING TO MULTI-PROBLEM WILCOXON’S TEST
SCSS-L-SHADE
v.s. L-SHADE
SCSS-UMOEA-II
v.s. UMOEA-II
SCSS-L-SHADE_EpSin
v.s. L-SHADE_EpSin
SCSS-jSO
v.s. jSO
R+
3235.0
R770.0
p-value
0.0E+00
α = 0.05
Yes
3052.5
952.5
1.7E-05
Yes
3077.0
1018.0
3.4E-05
Yes
3710.5
384.5
0.0E+00
Yes
TABLE XII OVERALL PERFORMANCE RANKING OF THE TOP
ALGORITHMS AND THEIR SCSS VARIANTS ON 30-D, 50-D AND
100-D CEC2017 BENCHMARK SET BY FRIEDMAN’S TEST
Algorithm
SCSS-jSO
SCSS-L-SHADE_EpSin
jSO
L-SHADE_EpSin
. SCSS-L-SHADE
SCSS-UMOEA-II
L-SHADE
UMOEA-II
Ranking
2.76
3.60
3.88
4.33
4.57
5.19
5.67
5.96
H. Computational Overhead
To illustrate the computational overhead of the SCSS
framework, the algorithm complexity (AC) of SCSS-L-SHADE
is compared with that of L-SHADE on 30-D, 50-D and 100-D
functions. Table XIII shows the experimental results, where T0,
T1 and T2 are calculated by the instructions in [25] as follows.
Specifically, T0 is the computation time for running the below
test program:
x= 0.55;
for i=1:1000000
x=x + x; x=x/2; x=x*x; x=sqrt(x); x=log(x); x=exp(x);
x=x/(x+2);
end
T1 is the time cost for 200,000 evaluations on Function 18
and T2 is the average time of five independent runs to solve
Function 18 by a specified algorithm with 200,000 evaluations.
AC is defined as (T2 -T1)/T0.
From Table XIII, it can be seen that the algorithm complexity
of SCSS-L-SHADE (AC2) is higher than that of L-SHADE
(AC1). The computational overhead, calculated as
(AC2-AC1)/AC1 in the 30-D, 50-D and 100-D cases is 23%,
33%, and 42% respectively, reflecting that SCSS does not
impose serious computational burdens.
TABLE XIII ALGORITHM COMPLEXITY (AC) COMPARISON BETWEEN
SCSS-L-SHADE AND L-SHADE
T0 [s]
T1 [s]
T2 [s]
AC1
L-SHADE
D = 30
D = 50
0.1114
2.6985
3.4898
5.5600
6.8767
25.6867
30.4031
T0 [s]
T1 [s]
T2 [s]
AC2
SCSS-L-SHADE
D = 30
D = 50
0.1114
2.6985
3.4898
6.2312
8.0110
31.7118
40.5853
D = 100
7.5146
13.4262
53.0664
D = 100
7.5146
15.9492
75.7145
COMPUTATIONAL OVERHEAD
D = 30
D = 50
D = 100
(AC2- AC1)/ AC1
23%
33%
42%
I. More Investigations on SCSS
As pointed out in Section III-B, SC basically focuses on
overcoming the drawback of a single reproduction procedure
and is essentially different from the existing “M operators, M
candidates” method. To investigate whether SC could further
improve this method, in this subsection, we applied SCSS on
the typical CoDE [38] algorithm. CoDE adopts three trial
vector generation strategies to generate three offspring for each
current solution and then select the fittest one to perform the
selection of DE. In the SCSS-CoDE implementation, firstly, for
each trial vector generation strategy, M candidates are
generated by performing M reproduction procedures.
Subsequently, the final candidate is determined from the
corresponding M candidates for each trial vector generation
strategy by using the SS rule. Finally, the three final candidates
for each current solution are evaluated and the fittest one is
selected to perform the selection of DE. Parameter settings for
CoDE and SCSS-CoDE are set the same as those used in the
original literature. The extra settings for SCSS-CoDE are as
follows: M = 2 and Scheme 2 is used as the SS rule. The
experimental results on 30-D and 50-D CEC2017 functions are
shown in Table S14 and the comparison results are summarized
in Table XIV.
As shown in Table XIV, SCSS-CoDE outperforms CoDE in
47 (=23+24) cases and underperforms in 3 (=1+2) cases. These
significant improvements confirm that the proposed SCSS
framework deals with the basic problem that exists in the
>
13
existing evolutionary optimization, including the “M operators,
M candidates” method.
TABLE XIV COMPARISON RESULTS OF SCSS-CODE WITH CODE
ON 30-D AND 50-D CEC2017 TEST FUNCTIONS
-/=/+
CoDE v.s. SCSS-CoDE
D = 30
23/6/1
[7]
[8]
D = 50
24/4/2
[9]
V. CONCLUSION
To address the potential adverse effect of randomness in
one-reproduction scheme, which are commonly used in
evolutionary algorithms, a selective-candidate framework with
similarity selection rule (SCSS) is proposed in this paper. In
SCSS, each current solution owns a pool of M candidates
generated by independently performing the reproduction
procedures M times. The final candidate is then determined
from the pool by a selective rule. In order to provide an
effective and efficient selective rule, a similarity selection rule
based on Euclidian distance measure was designed. It was
subsequently shown that a good balance between exploration
and exploitation capabilities could be obtained. This was
proven through comprehensive experiments conducted on the
CEC2014 and CEC2017 test suits, which demonstrated the
success of the design on several classic, advanced and top
algorithms from EA and SI families.
Future works on the SC framework can be extended in
multiple directions. Firstly, although the proposed SS rule is
already very promising, it is worth exploring other possible
designs in the selective rules. For example, different searching
stages may need different amounts of exploitation and
exploration. Thus, integrating this consideration with SS may
further improve its performance. Secondly, even though the
investigated M and GD values yield promising overall
performance, for a given problem, the algorithm may need
certain settings to achieve the best performance. In this context,
adaptive M and GD may further enhance the performance.
Finally, it is also interesting to investigate the performance of
the proposal on other EAs, including evolution programming
(EP) [10] and estimation of distribution algorithm (EDA) [11,
56], other SIs including artificial bee colony (ABC) [13] and
teaching-learning-based optimization (TLBO) [14] as well as
other kinds of optimization problems, such as multimodal
[58-60], multiobjective [61, 62] and large-scale [63]
optimization.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
T. Bäck, Evolutionary Algorithms in Theory and Practice. London,
U.K.: Oxford Univ. Press, 1996.
K. S. Tang, K. F. Man, S. Kwong, and Q. He, “Genetic algorithms and
their applications,” IEEE Signal Process. Mag., vol. 13, no. 6, pp. 22–37,
Nov. 1996.
E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm Intelligence: From
Natural to Artificial Systems. Oxford, U.K.: Oxford Univ. Press, 1999.
S. Das and P. N. Suganthan, “Differential evolution—A survey of the
state-of-the-art,” IEEE Trans. Evol. Comput., vol. 15, no. 1, pp. 4-31,
Feb. 2011.
S. Das, S. S. Mullick, and P. N. Suganthan, “Recent advances in
differential evolution—An updated survey,” Swarm Evol. Comput., vol.
27, pp. 1–30, Apr. 2016.
N. Hansen, “The CMA evolution strategy: A comparing review,” in
Towards a New Evolutionary Computation: Studies in Fuzziness and
Soft Computing, J. Kacprzyk, Ed. Berlin: Springer-Verlag, 2006, vol.
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
192, pp. 75–102.
Y. del Valle, G. K. Venayagamoorthy, S. Mohagheghi, J.-C. Hernandez,
and R. G. Harley, “Particle swarm optimization: Basic concepts, variants
and applications in power systems,” IEEE Trans. Evol. Comput., vol. 12,
no. 2, pp. 171–195, Apr. 2008.
R. Storn and K. Price, “Differential evolution–A simple and efficient
adaptive scheme for global optimization over continuous spaces,”
Berkeley, CA, Tech. Rep., 1995, tech. Rep. TR-95-012.
R. Storn and K. Price, “Differential evolution–A simple and efficient
heuristic for global optimization over continuous spaces,” J. of Global
Optim., vol. 11, no. 4, pp. 341-359, Dec. 1997.
T. Bäck and H.-P. Schwefel, “An overview of evolutionary algorithms
for parameter optimization,” Evol. Comput., vol. 1, no. 1, pp. 1–23,
1993.
P. Larrañaga and J. A. Lozano, Estimation of Distribution Algorithms: A
New Tool for Evolutionary Computation. Norwell, MA: Kluwer, 2002.
J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc.
IEEE Int. Conf. Neural Netw., vol. 4. 1995, pp. 1942–1948.
D. Karaboga and B. Basturk, “A powerful and efficient algorithm for
numerical function optimization: Artificial bee colony (ABC)
algorithm,” J. Global Optim., vol. 39, no. 3, pp. 459–471, Nov. 2007.
R. V. Rao, V. J. Savsani, and D. P. Vakharia, “Teaching-learning-based
optimization: A novel method for constrained mechanical design
optimization problems,” Comput.-Aided Design, vol. 43, no. 3, pp.
303–315, Mar. 2011.
S.-M. Guo, C.-C. Yang, P.-H. Hsu, and J. S.-H. Tsai, “Improving
differential evolution with a successful-parent-selecting framework,”
IEEE Trans. Evol. Comput., vol. 19, no. 5, pp. 717–730, Oct. 2015.
Y. Wang, Z.-Z. Liu, J. Li, H.-X. Li, and G. G. Yen, “Utilizing cumulative
population distribution information in differential evolution,” Appl. Soft
Comput., vol. 48, pp. 329- 346, Nov. 2016.
S. X. Zhang, S. Y. Zheng, and L. M. Zheng, “An efficient multiple
variants coordination framework for differential evolution,” IEEE
Trans. Cybern., vol. 47, no. 9, pp. 2780-2793, Sep. 2017.
N. Hansen and A. Ostermeier, “Completely derandomized
self-adaptation in evolution strategies,” Evol. Comput., vol. 9, no. 2,
pp.159–195, 2001.
H. -G. Beyer and B. Sendhoff, “Simplify your covariance matrix
adaptation evolution strategy,” IEEE Trans. Evol. Comput., vol. 21, no. 5,
pp. 746–759, Oct. 2017.
B. Y. Qu, P. N. Suganthan, and S. Das, “A distance-based locally
informed particle swarm model for multimodal optimization,” IEEE
Trans. Evol. Comput., vol. 17, no. 3, pp. 387-402, Jun. 2013.
J. J. Liang, A. K. Qin, P. N. Suganthan, and S. Baskar, “Comprehensive
learning particle swarm optimizer for global optimization of multimodal
functions,” IEEE Trans. Evol. Comput., vol. 10, no. 3, pp. 281-295, Jun.
2006.
Z. H. Zhan, J. Zhang, Y. Li, and Y. H. Shi, “Orthogonal learning particle
swarm optimization,” IEEE Trans. Evol. Comput., vol. 15, no. 6, pp.
832– 847, Dec. 2011.
W. Gao, S. Liu, and L. Huang, “A novel artificial bee colony algorithm
based on modified search equation and orthogonal learning,” IEEE
Trans. Cybern., vol. 43, no. 3, pp. 1011–1024, Jun. 2013.
J. J. Liang, B. Y. Qu, and P. N. Suganthan, “Problem definitions and
evaluation criteria for the CEC 2014 special session and competition on
single objective real-parameter numerical optimization,” Zhengzhou
Univ., China, and Nanyang Technol. Univ., Singapore, Dec. 2013.
N. H. Awad, M. Z. Ali, J. J. Liang, B. Y. Qu, and P. N. Suganthan,
“Problem definitions and evaluation criteria for the CEC 2017 special
session and competition on single objective real-parameter numerical
optimization,” Nanyang Technol. Univ., Singapore, Nov. 2016.
A. E. Eiben and A. Schippers, “On evolutionary exploration and
exploitation,” Fundamenta Inform., vol. 35, nos. 1–4 pp. 35–50, 1998.
M. Črepinšek, S.-H. Liu, and M. Mernik, “Exploration and exploitation
in evolutionary algorithms: A survey,” ACM CSUR, vol. 45, no. 3,
Article 35, 2013.
M. G. Epitropakis, D. K. Tasoulis, N. G. Pavlidis, V. P. Plagianakos, and
M. N. Vrahatis, “Enhancing differential evolution utilizing
proximity-based mutation operators,” IEEE Trans. Evol. Comput., vol.
15, no. 1, pp. 99-119, Feb. 2011.
W. Gong and Z. Cai, “Differential evolution with ranking-based
mutation operators,” IEEE Trans. Cybern., vol. 43, no.6, pp. 2066-2081,
Dec. 2013.
L. M. Zheng, S. X. Zhang, K. S. Tang, and S. Y. Zheng, “Differential
evolution powered by collective information,” Inf. Sci., vol. 399, pp.
>
13-29, Aug. 2017.
[31] J. Wang, J. Liao, Y. Zhou, and Y. Cai, “Differential evolution enhanced
with multiobjective sorting-based mutation operators,” IEEE Trans.
Cybern., vol. 44, no.12, pp. 2792-2805, Dec. 2014.
[32] R. Tanabe and A. S. Fukunaga, “Success-history based parameter
adaptation for differential evolution,” in Proc. IEEE Congr. Evol.
Comput., Cancun, Mexico, Jun. 2013, pp. 71-78.
[33] J. Zhang and A. C. Sanderson, “JADE: Adaptive differential evolution
with optional external archive,” IEEE Trans. Evol. Comput., vol. 13, no.
5, pp. 945-958, Oct. 2009.
[34] R. Tanabe and A. S. Fukunaga, “Improving the search performance of
SHADE using linear population size reduction,” in Proc. IEEE Congr.
Evol. Comput., Beijing, China, Jul. 2014, pp. 1658–1665.
[35] G. Wu, R, Mallipeddi, P. N. Suganthan, R. Wang, and H. Chen,
“Differential evolution with multi-population based ensemble of
mutation strategies,” Inf. Sci., vol. 329, pp. 329-345, Feb. 2016.
[36] L. Cui, G. Li, Q. Lin, J. Chen, and N. Lu, “Adaptive differential
evolution algorithm with novel mutation strategies in multiple
sub-populations,” Comput. Oper. Res., vol. 67, pp. 155-173, Mar. 2016.
[37] L. Tang, Y. Dong, and J. Liu, “Differential evolution with an
individual-dependent mechanism,” IEEE Trans. Evol. Comput., vol. 19,
no. 4, pp. 560-574, Aug. 2015.
[38] Y. Wang, Z. Cai, and Q. Zhang, “Differential evolution with composite
trial vector generation strategies and control parameters,” IEEE Trans.
Evol. Comput., vol. 15, no. 1, pp. 55-66, Feb. 2011.
[39] W. Gong, A. Zhou, and Z. Cai, “A multi-operator search strategy based
on cheap surrogate models for evolutionary optimization,” IEEE Trans.
Evol. Comput., vol. 19, no. 5, pp. 746-758, Oct. 2015.
[40] M. Yang, C. Li, Z. Cai, and J. Guan, “Differential evolution with
auto-enhanced population diversity,” IEEE Trans. Cybern., vol. 45,
no.2, pp. 302 –315, Feb. 2015.
[41] W. Du, S. Y. S. Leung, Y. Tang, and A. V. Vasilakos, “Differential
evolution with event-triggered impulsive control,” IEEE Trans. Cybern.,
vol. 47, no. 1, pp. 244-257, Jan. 2017.
[42] X. Qiu, K. C. Tan, and J.-X. Xu, “Multiple exponential recombination
for differential evolution,” IEEE Trans. Cybern., vol. 47, no. 4, pp.
995-1006, Apr. 2017.
[43] J. Arabas and R. Biedrzycki, “Improving evolutionary algorithms in a
continuous domain by monitoring the population midpoint,” IEEE
Trans. Evol. Comput., vol. 21, no. 5, pp. 807–812, Oct. 2017.
[44] E. Mezura-Montes and C. A. C. Coello, “A simple multimembered
evolution strategy to solve constrained optimization problems,” IEEE
Trans. Evol. Comput., vol. 9, no. 1, pp. 1-17, Feb. 2005.
[45] A. Auger and N. Hansen, “A restart CMA evolution strategy with
increasing population size,” in Proc. IEEE CEC, Sep. 2005, pp.
1769–1776.
[46] N. Hansen, “Benchmarking a BI-population CMA-ES on the
BBOB-2009 function testbed,” in Proc. Companion GECCO, 2009, pp.
2389–2396.
[47] C. L. Müller, B. Baumgartner, and I. F. Sbalzarini, “Particle swarm
CMA evolution strategy for the optimization of multi-funnel
landscapes,” in Proc. IEEE Congr. Evol. Comput., 2009, pp. 2685–2692.
[48] B. O. Arani, P. Mirzabeygi, and M. S. Panahi, “An improved PSO
algorithm with a territorial diversity-preserving scheme and enhanced
exploration–exploitation balance,” Swarm Evol. Comput., vol. 11, pp.
1–15, Aug. 2013.
[49] H. Wang, H. Sun, C. Li, S. Rahnamayan, and J. Pan, “Diversity
enhanced particle swarm optimization with neighborhood search,” Inf.
Sci., vol. 223, pp. 119-135, Feb. 2013.
[50] Z. Ren, A. Zhang, C. Wen, and Z. Feng, “A scatter learning particle
swarm optimization algorithm for multimodal problems,” IEEE Trans.
Cybern., vol. 44, no. 7, pp. 1127-1140, Jul. 2014.
[51] N. Lynn and P. N. Suganthan, “Heterogeneous comprehensive learning
particle swarm optimization with enhanced exploration and
exploitation,” Swarm Evol. Comput., vol. 24, pp. 11–24, Oct. 2015.
[52] D. Sheskin, Handbook of Parametric and Nonparametric Statistical
Procedures. London, U.K.: Chapman & Hall, 2003.
[53] S. Elsayed, N. Hamza, and R. Sarker, “Testing united multi-operator
evolutionary algorithms-II on single objective optimization problems,”
in Proc. IEEE Congr. Evol. Comput., Vancouver, BC, Canada, 2016, pp.
2966–2973.
[54] N. H. Awad, M. Z. Ali, P. N. Suganthan, and R. G. Reynolds, “An
ensemble sinusoidal parameter adaptation incorporated with L-SHADE
for solving CEC2014 benchmark problems,” in Proc. IEEE Congr. Evol.
Comput., Vancouver, BC, Canada, 2016, pp. 2958–2965.
14
[55] J. Brest, M. S Maučec, and B. Bošković, “Single objective
real-parameter optimization: algorithm jSO,” in Proc. IEEE Congr.
Evol. Comput., San Sebastian, 2017, pp. 1311–1318.
[56] S. García, A. Fernández, J. Luengo, and F. Herrera, “A study of
statistical techniques and performance measures for genetics-based
machine learning: Accuracy and interpretability,” Soft Comput., vol. 13,
no. 10, pp. 959–977, 2009.
[57] J. Wang, K. Tang, J. A. Lozano, and X. Yao, “Estimation of the
distribution algorithm with a stochastic local search for uncertain
capacitated arc routing problems,” IEEE Trans. Evol. Comput., vol. 20,
no. 1, pp. 96–109, Feb. 2016.
[58] S. Hui and P. N. Suganthan, “Ensemble and arithmetic recombination
based speciation differential evolution for multimodal optimization,”
IEEE Trans. Cybern., vol. 46, no. 1, pp. 64–74, Jan. 2016.
[59] Q. Yang et al, “Adaptive multimodal continuous ant colony
optimization,” IEEE Trans. Evol. Comput., vol. 21, no. 2, pp. 191–205,
Apr. 2017.
[60] X. Li, M. G. Epitropakis, K. Deb, and A. Engelbrecht, “Seeking multiple
solutions: An updated survey on niching methods and their
applications,” IEEE Trans. Evol. Comput., vol. 21, no. 4, pp. 518–538,
Aug. 2017.
[61] K. Deb, S. Agrawal, A. Pratap, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol.
Comput., vol. 6, no. 2, pp. 182–197, Apr. 2002.
[62] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary
algorithm based on decomposition,” IEEE Trans. Evol. Comput., vol. 11,
no. 6, pp. 712–731, Dec. 2007.
[63] M. N. Omidvar, M. Yang, Y. Mei, X. Li, and X. Yao, “DG2: A faster and
more accurate differential grouping for large-scale black-box
optimization,” IEEE Trans. Evol. Comput., in press.
>
15
Supplemental file of “Selective-Candidate
Framework with Similarity Selection Rule for
Evolutionary Optimization”
TABLE CAPTIONS
TABLE S1 PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES ON 30-D
CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S2 PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES ON 50-D
CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S3 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES ON
50-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S4 PERFORMANCE COMPARISONS OF SCSS-JADE AND SCSS-SHADE WITH THEIR CORRESPONDING VARIANTS
AND II ON 30-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
I
TABLE S5 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES ON
CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
30-D
TABLE S6 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES ON
CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
50-D
TABLE S7 PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES ON 30-D
CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S8 PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES ON 50-D
CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S9 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES ON
30-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S10 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES ON
50-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S11 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES ON
CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
30-D
TABLE S12 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES ON
CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
50-D
TABLE S13 PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES ON
CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
TABLE S14 PERFORMANCE COMPARISONS OF SCSS-CODE WITH CODE ON
OVER 51 INDEPENDENT RUNS
100-D
30-D AND 50-D CEC2017 BENCHMARK SET
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S1 PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES
ON 30-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
DE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
8.79E+07 (1.76E+07)
1.67E+03 (4.57E+02)
2.58E+01 (5.46E+00)
1.21E+02 (5.83E+00)
2.09E+01 (5.74E-02)
3.02E+01 (1.09E+00)
5.13E-02 (1.10E-01)
1.19E+02 (7.92E+00)
1.96E+02 (1.10E+01)
3.87E+03 (2.65E+02)
6.63E+03 (2.62E+02)
2.08E+00 (2.08E-01)
5.03E-01 (4.36E-02)
2.91E-01 =
(3.94E-02)
1.92E+01 (8.88E-01)
1.25E+01 (2.36E-01)
2.46E+06 (5.43E+05)
2.97E+04 (1.50E+04)
1.03E+01 (8.07E-01)
4.92E+02 (1.14E+02)
1.96E+05 =
(8.29E+04)
2.10E+02 (6.39E+01)
3.15E+02 (8.41E-05)
2.09E+02 +
(2.65E+00)
2.22E+02 (3.32E+00)
1.00E+02 (4.62E-02)
3.91E+02 (3.43E+01)
9.80E+02 (2.46E+01)
1.11E+04 (2.61E+03)
5.65E+03 (7.77E+02)
27/2/1
SCSSDE
6.42E+07
(1.09E+07)
7.38E-01
(2.87E+00)
1.51E-03
(6.51E-04)
8.75E+01
(1.38E+01)
2.09E+01
(4.86E-02)
2.78E+01
(9.97E-01)
2.93E-07
(1.65E-06)
9.61E+01
(5.72E+00)
1.78E+02
(1.01E+01)
3.26E+03
(2.48E+02)
6.29E+03
(3.00E+02)
1.83E+00
(2.64E-01)
3.91E-01
(4.66E-02)
2.78E-01
(2.97E-02)
1.66E+01
(1.01E+00)
1.24E+01
(2.37E-01)
2.04E+06
(6.43E+05)
2.22E+04
(8.04E+03)
9.35E+00
(3.68E-01)
4.15E+02
(9.72E+01)
2.00E+05
(5.96E+04)
1.84E+02
(6.22E+01)
3.15E+02
(8.55E-10)
2.27E+02
(7.50E-01)
2.17E+02
(2.38E+00)
1.00E+02
(4.67E-02)
3.57E+02
(5.34E+01)
8.73E+02
(5.46E+01)
8.03E+03
(2.86E+03)
3.75E+03
(6.59E+02)
ES
1.28E+08 (3.16E+07)
2.36E+09 (5.00E+08)
5.39E+04 (7.87E+03)
3.63E+02 (4.45E+01)
2.09E+01 +
(6.05E-02)
2.30E+01 (1.65E+00)
2.31E+01 (3.79E+00)
1.75E+02 (1.64E+01)
2.25E+02 (1.58E+01)
5.00E+03 (4.42E+02)
7.00E+03 (2.88E+02)
2.41E+00 =
(2.93E-01)
8.01E-01 (1.45E-01)
3.94E+00 (2.41E+00)
7.97E+01 (3.81E+01)
1.28E+01 (1.96E-01)
4.11E+06 (1.48E+06)
2.59E+07 (7.24E+06)
3.04E+01 (4.91E+00)
1.44E+04 (6.16E+03)
1.04E+06 (3.86E+05)
6.05E+02 (1.56E+02)
3.39E+02 (4.82E+00)
2.61E+02 (2.40E+00)
2.24E+02 (3.72E+00)
1.01E+02 (1.05E-01)
6.77E+02 =
(1.67E+02)
1.40E+03 (5.09E+01)
8.05E+05 (3.60E+05)
5.64E+04 (1.75E+04)
27/2/1
SCSSES
4.04E+07
(1.19E+07)
6.38E+08
(1.18E+08)
1.73E+04
(3.61E+03)
1.98E+02
(1.84E+01)
2.10E+01
(3.16E-02)
1.44E+01
(1.30E+00)
6.29E+00
(8.83E-01)
1.14E+02
(1.20E+01)
1.91E+02
(9.97E+00)
2.91E+03
(3.74E+02)
6.62E+03
(2.70E+02)
2.40E+00
(2.49E-01)
5.56E-01
(8.57E-02)
4.82E-01
(1.57E-01)
2.56E+01
(1.70E+00)
1.21E+01
(2.84E-01)
1.32E+06
(5.34E+05)
5.81E+06
(1.96E+06)
1.72E+01
(1.70E+00)
2.69E+03
(1.00E+03)
3.85E+05
(1.93E+05)
2.20E+02
(8.85E+01)
3.23E+02
(1.35E+00)
2.44E+02
(3.75E+00)
2.12E+02
(1.65E+00)
1.03E+02
(1.43E+01)
6.44E+02
(2.69E+01)
1.20E+03
(3.83E+01)
1.35E+05
(4.13E+04)
1.72E+04
(4.02E+03)
PSO
6.48E+08 (5.02E+08)
1.16E+10 =
(1.48E+10)
3.31E+06 (6.23E+06)
5.46E+02 (4.61E+02)
2.12E+01 (2.10E-01)
3.21E+01 =
(4.42E+00)
2.26E+02 =
(1.52E+02)
2.45E+02 (7.72E+01)
3.53E+02 (1.21E+02)
6.18E+03 (1.10E+03)
6.39E+03 (1.45E+03)
5.28E+00 (2.35E+00)
4.24E+00 =
(1.91E+00)
5.89E+01 =
(5.14E+01)
5.52E+05 (1.19E+06)
1.40E+01 (5.02E-01)
3.75E+07 (3.49E+07)
1.05E+09 =
(6.53E+08)
1.70E+02 =
(1.01E+02)
8.35E+06 (1.68E+07)
1.86E+07 (7.19E+06)
1.64E+03 =
(4.52E+02)
4.44E+02 =
(1.21E+02)
2.82E+02 =
(4.63E+01)
2.25E+02 =
(2.30E+01)
1.20E+02 =
(3.54E+01)
7.31E+02 =
(3.61E+02)
3.30E+03 =
(7.15E+02)
1.11E+08 (5.53E+07)
4.09E+05 =
(4.43E+05)
15/15/0
SCSSPSO
3.58E+08
(3.79E+08)
9.58E+09
(1.20E+10)
2.62E+05
(1.19E+05)
4.20E+02
(7.15E+02)
2.07E+01
(1.92E-01)
3.08E+01
(2.85E+00)
2.35E+02
(1.26E+02)
1.56E+02
(3.57E+01)
2.21E+02
(5.35E+01)
5.54E+03
(1.08E+03)
5.32E+03
(8.75E+02)
1.23E+00
(5.92E-01)
3.64E+00
(2.04E+00)
6.03E+01
(3.69E+01)
8.81E+04
(2.69E+05)
1.25E+01
(5.58E-01)
2.15E+07
(2.51E+07)
1.07E+09
(5.72E+08)
1.71E+02
(8.87E+01)
2.84E+04
(3.60E+04)
1.57E+07
(7.86E+06)
1.48E+03
(3.92E+02)
4.41E+02
(1.59E+02)
2.71E+02
(3.27E+01)
2.24E+02
(1.60E+01)
1.14E+02
(1.97E+01)
7.16E+02
(3.87E+02)
3.08E+03
(5.58E+02)
8.28E+07
(3.87E+07)
4.75E+05
(5.59E+05)
16
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S2
PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES
ON 50-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
DE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
3.54E+08 (5.04E+07)
5.91E+07 (2.43E+07)
6.57E+04 (5.94E+03)
1.27E+02 (9.76E+00)
2.11E+01 =
(3.70E-02)
5.85E+01 (2.01E+00)
7.40E-01 (4.01E-02)
2.84E+02 (1.13E+01)
4.16E+02 (1.48E+01)
9.29E+03 (3.40E+02)
1.30E+04 (4.19E+02)
3.07E+00 =
(3.18E-01)
6.64E-01 (6.12E-02)
3.81E-01 (5.99E-02)
4.60E+01 (2.11E+00)
2.23E+01 =
(2.15E-01)
1.62E+07 (3.73E+06)
6.04E+04 =
(2.53E+04)
2.64E+01 (1.97E+00)
2.13E+04 (4.98E+03)
7.34E+06 (2.02E+06)
1.23E+03 (1.31E+02)
3.44E+02 (4.57E-04)
2.70E+02 (2.49E+00)
2.70E+02 (7.07E+00)
1.01E+02 (5.29E-02)
1.53E+03 (6.89E+01)
1.50E+03 (4.07E+01)
1.19E+05 (2.35E+04)
3.08E+04 (7.44E+03)
26/4/0
SCSSDE
2.52E+08
(3.47E+07)
6.16E+05
(4.50E+05)
4.18E+04
(4.19E+03)
9.80E+01
(4.10E-01)
2.11E+01
(3.34E-02)
5.58E+01
(2.18E+00)
3.12E-03
(1.13E-03)
2.50E+02
(1.01E+01)
3.87E+02
(1.27E+01)
8.63E+03
(2.58E+02)
1.28E+04
(3.12E+02)
3.03E+00
(2.63E-01)
5.77E-01
(5.15E-02)
3.16E-01
(4.46E-02)
3.65E+01
(1.21E+00)
2.22E+01
(2.14E-01)
1.29E+07
(2.76E+06)
6.00E+04
(2.50E+04)
2.28E+01
(3.08E+00)
1.66E+04
(4.97E+03)
5.94E+06
(1.44E+06)
1.10E+03
(1.65E+02)
3.44E+02
(4.76E-08)
2.59E+02
(1.11E+00)
2.56E+02
(6.63E+00)
1.01E+02
(5.29E-02)
1.12E+03
(2.28E+02)
1.34E+03
(1.02E+02)
4.85E+04
(1.55E+04)
1.22E+04
(3.53E+03)
ES
3.05E+08 (6.47E+07)
7.34E+09 (8.74E+08)
1.11E+05 (1.29E+04)
9.20E+02 (1.29E+02)
2.11E+01 =
(3.41E-02)
4.58E+01 (3.21E+00)
7.30E+01 (1.12E+01)
3.55E+02 (2.62E+01)
4.55E+02 (2.06E+01)
1.04E+04 (6.40E+02)
1.33E+04 (4.61E+02)
3.36E+00 =
(1.65E-01)
1.00E+00 (1.20E-01)
1.86E+01 (2.93E+00)
1.00E+03 (6.48E+02)
2.25E+01 (2.04E-01)
1.86E+07 (5.46E+06)
1.21E+08 (3.30E+07)
1.08E+02 (8.68E+00)
4.61E+04 (1.35E+04)
8.76E+06 (3.58E+06)
1.96E+03 (2.41E+02)
4.13E+02 (1.17E+01)
3.32E+02 (4.01E+00)
2.62E+02 (8.52E+00)
1.01E+02 (1.02E-01)
1.47E+03 (7.28E+01)
2.88E+03 (1.73E+02)
7.51E+06 (2.39E+06)
2.14E+05 (5.44E+04)
28/2/0
SCSSES
1.07E+08
(2.76E+07)
1.91E+09
(2.33E+08)
4.58E+04
(6.60E+03)
3.48E+02
(4.18E+01)
2.11E+01
(3.18E-02)
3.01E+01
(1.97E+00)
2.03E+01
(2.29E+00)
2.57E+02
(1.80E+01)
3.93E+02
(1.54E+01)
7.19E+03
(5.78E+02)
1.30E+04
(3.42E+02)
3.35E+00
(2.80E-01)
7.09E-01
(7.69E-02)
7.54E-01
(2.71E-01)
6.10E+01
(6.80E+00)
2.19E+01
(3.30E-01)
6.89E+06
(2.11E+06)
3.35E+07
(7.82E+06)
7.45E+01
(1.37E+01)
8.44E+03
(2.92E+03)
3.37E+06
(1.21E+06)
1.43E+03
(2.43E+02)
3.68E+02
(3.36E+00)
3.04E+02
(2.60E+00)
2.33E+02
(3.64E+00)
1.39E+02
(5.25E+01)
1.11E+03
(5.19E+01)
2.22E+03
(1.03E+02)
1.43E+06
(4.17E+05)
5.93E+04
(1.09E+04)
PSO
1.76E+09 =
(1.23E+09)
3.32E+10 =
(2.20E+10)
2.53E+06 (4.59E+06)
4.68E+03 =
(7.39E+03)
2.14E+01 (1.47E-01)
6.53E+01 =
(6.74E+00)
5.15E+02 =
(3.07E+02)
4.39E+02 (1.37E+02)
6.84E+02 (1.90E+02)
1.04E+04 (2.72E+03)
1.12E+04 =
(2.15E+03)
6.22E+00 (2.20E+00)
3.70E+00 (1.59E+00)
9.33E+01 =
(6.57E+01)
1.24E+06 =
(2.92E+06)
2.39E+01 (4.33E-01)
1.22E+08 (1.04E+08)
2.12E+09 =
(1.75E+09)
1.10E+03 =
(5.99E+02)
1.47E+07 (2.67E+07)
5.25E+07 (4.98E+07)
4.10E+03 (2.76E+03)
6.60E+02 =
(3.13E+02)
4.89E+02 (1.19E+02)
2.67E+02 (3.49E+01)
1.31E+02 =
(6.30E+01)
2.31E+03 =
(1.61E+02)
5.79E+03 =
(1.52E+03)
3.44E+08 =
(6.58E+07)
7.39E+06 =
(6.79E+06)
14/16/0
SCSSPSO
1.33E+09
(1.17E+09)
3.27E+10
(2.35E+10)
3.09E+05
(2.22E+05)
3.45E+03
(3.92E+03)
2.09E+01
(1.40E-01)
6.41E+01
(5.36E+00)
4.27E+02
(2.85E+02)
3.30E+02
(6.54E+01)
4.63E+02
(8.24E+01)
8.58E+03
(2.20E+03)
1.05E+04
(1.89E+03)
1.49E+00
(4.16E-01)
2.72E+00
(1.97E+00)
7.09E+01
(5.66E+01)
1.19E+06
(2.51E+06)
2.09E+01
(8.38E-01)
8.22E+07
(8.91E+07)
1.88E+09
(2.65E+09)
1.04E+03
(5.94E+02)
1.11E+05
(1.55E+05)
3.14E+07
(3.59E+07)
3.12E+03
(1.46E+03)
6.55E+02
(3.68E+02)
4.29E+02
(9.35E+01)
2.46E+02
(2.31E+01)
1.32E+02
(4.67E+01)
2.26E+03
(1.50E+02)
6.02E+03
(1.59E+03)
3.33E+08
(9.99E+07)
7.00E+06
(7.99E+06)
17
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S3
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES
ON 50-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
JADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
18
1.88E+04 =
(1.26E+04)
0.00E+00 =
(0.00E+00)
3.06E+03 (2.03E+03)
1.37E+01 =
(3.36E+01)
2.04E+01 (3.27E-02)
1.59E+01 =
(6.47E+00)
4.15E-03 =
(5.75E-03)
0.00E+00 =
(0.00E+00)
5.43E+01 (7.72E+00)
1.05E-02 =
(9.47E-03)
3.82E+03 (2.72E+02)
2.61E-01 (3.01E-02)
3.13E-01 (4.70E-02)
3.00E-01 =
(2.93E-02)
7.27E+00 (8.65E-01)
1.77E+01 =
(5.34E-01)
2.29E+03 =
(6.74E+02)
1.64E+02 =
(4.16E+01)
1.48E+01 (5.97E+00)
8.19E+03 (6.72E+03)
1.29E+03 (4.85E+02)
4.78E+02 (1.66E+02)
3.44E+02 +
(4.55E-13)
2.74E+02 =
(2.05E+00)
2.23E+02 (3.19E+00)
1.04E+02 (1.95E+01)
4.65E+02 (5.76E+01)
1.15E+03 (3.72E+01)
8.81E+02 =
(5.80E+01)
9.78E+03 (7.82E+02)
16/13/1
SCSSJADE
1.97E+04
(1.52E+04)
0.00E+00
(0.00E+00)
2.01E+03
(2.98E+03)
2.32E+01
(4.20E+01)
2.02E+01
(2.06E-01)
1.67E+01
(6.84E+00)
2.42E-03
(4.81E-03)
0.00E+00
(0.00E+00)
3.86E+01
(8.83E+00)
1.25E-02
(1.56E-02)
3.53E+03
(2.87E+02)
2.14E-01
(7.30E-02)
2.75E-01
(3.91E-02)
3.18E-01
(9.22E-02)
5.94E+00
(6.97E-01)
1.80E+01
(1.05E+00)
2.53E+03
(7.80E+02)
1.66E+02
(4.06E+01)
1.06E+01
(5.22E+00)
1.99E+03
(4.70E+03)
2.36E+04
(1.61E+05)
3.76E+02
(1.61E+02)
3.44E+02
(5.16E-13)
2.75E+02
(1.89E+00)
2.11E+02
(6.51E+00)
1.00E+02
(8.92E-02)
4.35E+02
(5.42E+01)
1.12E+03
(3.47E+01)
8.94E+02
(9.69E+01)
9.26E+03
(8.07E+02)
SHADE
2.24E+04 =
(1.14E+04)
0.00E+00 =
(0.00E+00)
3.13E-06 (1.39E-05)
2.81E+01 (4.30E+01)
2.02E+01 (2.34E-02)
6.87E+00 =
(5.99E+00)
1.59E-03 =
(3.91E-03)
0.00E+00 =
(0.00E+00)
4.03E+01 =
(5.05E+00)
5.14E-03 =
(8.35E-03)
3.65E+03 =
(3.25E+02)
2.07E-01 (2.79E-02)
3.20E-01 =
(3.32E-02)
2.86E-01 =
(6.25E-02)
6.35E+00 (7.66E-01)
1.79E+01 =
(4.14E-01)
2.74E+03 =
(8.65E+02)
1.47E+02 =
(4.44E+01)
1.63E+01 (7.08E+00)
1.92E+02 (6.69E+01)
1.40E+03 (4.92E+02)
3.76E+02 =
(1.18E+02)
3.44E+02 =
(4.31E-13)
2.73E+02 =
(1.93E+00)
2.18E+02 (5.01E+00)
1.02E+02 (1.40E+01)
3.91E+02 =
(4.89E+01)
1.13E+03 =
(4.00E+01)
9.01E+02 =
(6.55E+01)
9.35E+03 (6.62E+02)
11/19/0
SCSSSHADE
2.66E+04
(1.09E+04)
0.00E+00
(0.00E+00)
1.02E-07
(3.42E-07)
3.08E+01
(4.60E+01)
2.02E+01
(2.30E-02)
5.35E+00
(4.96E+00)
1.69E-03
(4.22E-03)
0.00E+00
(0.00E+00)
3.95E+01
(5.80E+00)
9.06E-03
(1.30E-02)
3.55E+03
(3.46E+02)
1.71E-01
(2.59E-02)
3.12E-01
(4.02E-02)
2.69E-01
(4.02E-02)
5.66E+00
(5.90E-01)
1.79E+01
(3.62E-01)
2.74E+03
(8.27E+02)
1.39E+02
(4.31E+01)
1.28E+01
(4.48E+00)
1.10E+02
(4.37E+01)
1.01E+03
(3.33E+02)
3.38E+02
(1.09E+02)
3.44E+02
(4.50E-13)
2.72E+02
(1.89E+00)
2.11E+02
(6.05E+00)
1.00E+02
(5.89E-02)
3.79E+02
(4.65E+01)
1.11E+03
(3.05E+01)
9.02E+02
(6.54E+01)
8.87E+03
(6.64E+02)
CMA-ES
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
3.28E+01 =
(4.68E+01)
2.00E+01 +
(1.77E-06)
7.68E+01 (1.08E+01)
5.32E-04 =
(2.22E-03)
7.39E+02 (1.09E+02)
1.13E+03 (2.41E+02)
8.43E+03 (9.42E+02)
8.23E+03 (9.32E+02)
2.71E-01 (2.55E-01)
3.48E-01 +
(7.71E-02)
4.43E-01 +
(2.50E-01)
6.41E+00 =
(1.25E+00)
2.38E+01 =
(5.19E-01)
2.69E+03 =
(6.15E+02)
2.30E+02 +
(4.57E+01)
1.84E+01 (2.57E+00)
4.44E+02 (1.22E+02)
1.70E+03 =
(4.32E+02)
4.19E+02 (2.61E+02)
3.44E+02 =
(2.32E-05)
3.67E+02 (5.44E+02)
2.05E+02 (9.61E-01)
1.17E+02 +
(5.81E+01)
5.33E+02 (1.06E+02)
7.61E+03 (5.87E+03)
8.86E+02 =
(6.70E+01)
9.31E+03 =
(7.96E+02)
13/12/5
SCSSCMA-ES
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.35E+01
(3.42E+01)
2.14E+01
(3.67E-01)
1.74E+01
(1.85E+01)
6.77E-04
(2.42E-03)
6.12E+02
(2.31E+02)
5.88E+02
(4.78E+02)
7.21E+03
(1.17E+03)
7.25E+03
(1.10E+03)
7.63E-02
(4.56E-01)
8.08E-01
(1.59E-01)
1.26E+00
(4.03E-01)
6.02E+00
(1.20E+00)
2.40E+01
(6.18E-01)
2.60E+03
(5.98E+02)
2.67E+02
(7.08E+01)
1.46E+01
(2.30E+00)
2.71E+02
(8.53E+01)
1.62E+03
(3.71E+02)
3.20E+02
(2.11E+02)
3.44E+02
(2.38E-05)
2.76E+02
(2.43E+00)
2.05E+02
(2.18E-01)
1.09E+02
(4.04E+01)
4.57E+02
(7.00E+01)
4.39E+03
(2.98E+03)
8.94E+02
(8.74E+01)
9.45E+03
(1.09E+03)
LIPS
1.29E+08 (7.81E+07)
7.57E+02 +
(1.40E+03)
1.67E+04 (6.05E+03)
7.09E+02 (3.77E+02)
2.00E+01 +
(1.49E-05)
3.71E+01 (4.26E+00)
5.88E-03 (1.93E-02)
1.44E+02 (1.89E+01)
1.81E+02 (2.84E+01)
4.33E+03 (5.04E+02)
5.15E+03 (4.95E+02)
2.63E-01 (7.48E-02)
4.31E-01 =
(5.93E-02)
2.71E-01 +
(3.14E-02)
7.62E+01 (4.32E+01)
2.05E+01 (6.41E-01)
4.00E+06 (5.97E+06)
3.26E+02 (1.64E+02)
5.78E+01 (2.86E+01)
3.02E+04 (1.09E+04)
5.78E+05 (4.16E+05)
8.43E+02 (2.08E+02)
3.77E+02 (1.34E+01)
2.95E+02 (6.01E+00)
2.40E+02 (8.81E+00)
1.66E+02 (4.65E+01)
1.39E+03 (1.29E+02)
4.52E+03 (7.42E+02)
8.33E+06 (4.37E+07)
2.84E+05 (1.17E+05)
26/1/3
SCSSLIPS
8.45E+06
(1.32E+07)
1.72E+03
(2.71E+03)
1.14E+04
(5.51E+03)
2.08E+02
(5.28E+01)
2.11E+01
(3.62E-02)
2.33E+01
(3.96E+00)
7.25E-04
(2.57E-03)
6.73E+01
(1.23E+01)
1.08E+02
(2.14E+01)
2.52E+03
(4.62E+02)
4.20E+03
(6.68E+02)
6.84E-01
(1.12E+00)
4.12E-01
(5.70E-02)
3.48E-01
(1.19E-01)
1.20E+01
(2.95E+00)
1.94E+01
(6.37E-01)
7.38E+05
(1.42E+06)
2.53E+02
(7.76E+01)
4.25E+01
(2.26E+01)
1.91E+04
(7.19E+03)
1.71E+05
(1.07E+05)
5.69E+02
(1.88E+02)
3.50E+02
(1.83E+00)
2.80E+02
(3.16E+00)
2.25E+02
(4.59E+00)
1.36E+02
(4.82E+01)
9.91E+02
(8.80E+01)
2.55E+03
(3.27E+02)
2.09E+03
(5.43E+02)
6.41E+04
(2.21E+04)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S4
19
PERFORMANCE COMPARISONS OF SCSS-JADE AND SCSS-SHADE WITH THEIR CORRESPONDING VARIANTS I
AND II ON 30-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
SCSSSCSSSCSSSCSSSCSSJADE
JADE
SHADE
SHADE
JADE
Variant I
Variant II
Variant I
Variant II
2.27E+03 = 1.81E+05 1.47E+03
1.62E+03 = 2.96E+03 F1
(2.68E+03) (1.28E+06) (2.14E+03) (2.22E+03) (2.97E+03)
0.00E+00 = 0.00E+00 =
0.00E+00
0.00E+00 = 0.00E+00 =
F2
(0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00)
5.01E-04 2.90E+00 0.00E+00
0.00E+00 = 0.00E+00 =
F3
(2.65E-03)
(3.03E+00) (0.00E+00) (0.00E+00) (0.00E+00)
0.00E+00 = 0.00E+00 =
0.00E+00
0.00E+00 = 0.00E+00 =
F4
(0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00)
2.03E+01 2.03E+01 2.03E+01
2.02E+01 2.02E+01 F5
(3.41E-02)
(2.70E-02)
(7.09E-02)
(3.08E-02)
(2.61E-02)
1.12E+01 1.24E+01 7.33E+00
5.98E+00 7.78E+00 F6
(1.48E+00) (1.20E+00) (3.86E+00) (3.32E+00) (2.87E+00)
0.00E+00 = 0.00E+00 =
1.93E-04
0.00E+00 = 0.00E+00 =
F7
(0.00E+00) (0.00E+00)
(1.38E-03)
(0.00E+00) (0.00E+00)
0.00E+00 = 0.00E+00 =
0.00E+00
0.00E+00 = 0.00E+00 =
F8
(0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00) (0.00E+00)
2.65E+01 3.91E+01 2.13E+01
2.07E+01 2.76E+01 F9
(4.01E+00) (5.50E+00) (4.82E+00) (3.73E+00) (3.80E+00)
7.76E-03 =
2.45E-03 +
9.39E-03
3.67E-03 +
3.27E-03 +
F10
(1.50E-02)
(6.77E-03)
(1.52E-02)
(9.03E-03)
(7.65E-03)
1.70E+03 2.24E+03 1.54E+03
1.56E+03 = 1.95E+03 F11
(2.31E+02) (1.84E+02) (2.28E+02) (2.12E+02) (2.02E+02)
2.66E-01 3.76E-01 2.27E-01
2.04E-01 3.28E-01 F12
(3.70E-02)
(3.71E-02)
(4.87E-02)
(3.03E-02)
(2.69E-02)
2.26E-01 2.59E-01 1.85E-01
2.24E-01 2.47E-01 F13
(3.26E-02)
(3.58E-02)
(3.68E-02)
(3.53E-02)
(3.04E-02)
2.37E-01 =
2.46E-01 2.32E-01
2.18E-01 =
2.41E-01 F14
(3.40E-02)
(3.02E-02)
(3.71E-02)
(2.65E-02)
(2.56E-02)
3.20E+00 4.30E+00 2.86E+00
2.96E+00 3.76E+00 F15
(4.24E-01)
(4.90E-01)
(3.22E-01)
(3.53E-01)
(4.39E-01)
9.50E+00 9.91E+00 9.34E+00
9.48E+00 = 9.70E+00 F16
(2.97E-01)
(2.48E-01)
(4.29E-01)
(3.26E-01)
(3.76E-01)
6.28E+04 2.85E+05 8.28E+02
1.01E+03 1.28E+03 F17
(2.10E+05) (4.29E+05) (3.47E+02) (4.17E+02) (3.31E+02)
5.84E+02 2.85E+03 4.72E+01
3.55E+01 7.89E+01 F18
(2.39E+03) (3.60E+03) (2.34E+01) (2.02E+01) (2.74E+01)
4.56E+00 4.86E+00 4.01E+00
3.92E+00 = 4.33E+00 F19
(7.11E-01)
(7.86E-01)
(8.54E-01)
(4.88E-01)
(4.82E-01)
4.04E+03 3.53E+03 1.88E+03
1.09E+01 = 2.50E+01 F20
(2.18E+03) (2.22E+03) (2.44E+03) (6.71E+00) (1.35E+01)
1.87E+04 7.95E+04 2.41E+02
2.32E+02 4.46E+02 F21
(4.69E+04) (8.65E+04) (1.15E+02) (1.02E+02) (1.96E+02)
1.26E+02 = 1.64E+02 1.10E+02
6.76E+01 = 9.88E+01 F22
(6.15E+01) (7.95E+01) (6.90E+01) (5.10E+01) (5.80E+01)
3.15E+02 = 3.15E+02 =
3.15E+02
3.15E+02 = 3.15E+02 =
F23
(4.02E-13)
(2.48E-11)
(4.02E-13)
(4.02E-13)
(4.02E-13)
2.25E+02 = 2.26E+02 2.25E+02
2.24E+02 = 2.24E+02 =
F24
(1.98E+00) (3.01E+00) (3.27E+00) (1.11E+00) (1.03E+00)
2.05E+02 2.05E+02 2.03E+02
2.04E+02 2.05E+02 F25
(1.67E+00) (2.05E+00)
(6.04E-01)
(9.84E-01)
(1.99E+00)
1.00E+02 1.00E+02 1.00E+02
1.00E+02 = 1.00E+02 F26
(4.30E-02)
(3.71E-02)
(3.56E-02)
(3.54E-02)
(3.07E-02)
3.47E+02 = 3.61E+02 3.44E+02
3.14E+02 = 3.19E+02 F27
(5.14E+01) (5.23E+01) (5.09E+01) (3.36E+01) (4.00E+01)
8.02E+02 = 8.15E+02 8.01E+02
7.90E+02 = 7.96E+02 =
F28
(2.01E+01) (1.91E+01) (1.64E+01) (1.68E+01) (1.88E+01)
8.35E+02 1.28E+03 7.20E+02
7.22E+02 = 7.34E+02 F29
(2.95E+02) (4.43E+02) (7.10E+01) (2.06E+01) (1.92E+01)
1.73E+03 = 1.97E+03 1.53E+03
1.33E+03 = 1.54E+03 F30
(6.61E+02) (6.55E+02) (6.34E+02) (3.97E+02) (5.46E+02)
-/=/+
17/13/0
24/5/1
10/19/1
21/8/1
Note: The smallest mean values achieved on each function are underlined.
SCSSSHADE
1.50E+03
(2.68E+03)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.01E+01
(2.29E-02)
4.12E+00
(3.37E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.92E+01
(3.44E+00)
7.76E-03
(1.17E-02)
1.50E+03
(2.02E+02)
1.68E-01
(2.45E-02)
2.04E-01
(3.18E-02)
2.09E-01
(3.26E-02)
2.59E+00
(3.03E-01)
9.50E+00
(4.24E-01)
5.78E+02
(2.32E+02)
2.05E+01
(1.20E+01)
3.84E+00
(6.58E-01)
8.41E+00
(3.45E+00)
1.90E+02
(1.12E+02)
7.12E+01
(6.10E+01)
3.15E+02
(4.02E-13)
2.24E+02
(1.21E+00)
2.03E+02
(4.63E-01)
1.00E+02
(3.42E-02)
3.21E+02
(4.03E+01)
7.93E+02
(2.17E+01)
7.12E+02
(5.40E+01)
1.19E+03
(3.57E+02)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S5
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES
ON 30-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
L-SHADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
20
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.01E+01 (3.46E-02)
9.01E-03 =
(6.43E-02)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
7.22E+00 =
(1.33E+00)
5.72E-03 =
(1.11E-02)
1.23E+03 =
(1.92E+02)
1.73E-01 =
(2.13E-02)
1.05E-01 =
(1.35E-02)
2.38E-01 (2.69E-02)
2.28E+00 (2.93E-01)
8.51E+00 +
(3.61E-01)
2.09E+02 (1.13E+02)
6.89E+00 (3.23E+00)
3.75E+00 (5.74E-01)
2.84E+00 =
(1.04E+00)
9.08E+01 (7.29E+01)
2.45E+01 (3.35E+00)
3.15E+02 =
(4.02E-13)
2.24E+02 (1.46E+00)
2.03E+02 (5.33E-02)
1.00E+02 =
(1.47E-02)
3.00E+02 +
(1.25E-13)
8.35E+02 =
(1.83E+01)
7.16E+02 =
(2.52E+00)
1.40E+03 =
(6.66E+02)
10/18/2
SCSSL-SHADE
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.01E+01
(5.37E-02)
9.01E-03
(6.43E-02)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
7.38E+00
(1.63E+00)
7.35E-03
(1.37E-02)
1.24E+03
(1.85E+02)
1.65E-01
(3.01E-02)
1.08E-01
(1.56E-02)
1.90E-01
(2.41E-02)
2.16E+00
(2.47E-01)
8.65E+00
(4.00E-01)
8.89E+01
(4.59E+01)
3.01E+00
(1.50E+00)
3.08E+00
(6.64E-01)
2.59E+00
(1.07E+00)
3.33E+01
(5.40E+01)
2.31E+01
(2.00E+00)
3.15E+02
(3.18E-13)
2.22E+02
(3.44E+00)
2.03E+02
(4.10E-02)
1.00E+02
(1.38E-02)
3.00E+02
(2.16E-13)
8.33E+02
(1.96E+01)
7.15E+02
(1.55E+00)
1.37E+03
(6.31E+02)
UMOEA-II
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.00E+01 =
(1.03E-03)
1.99E-01 =
(1.35E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
8.97E+00 =
(1.79E+00)
1.63E-03 =
(5.65E-03)
1.41E+03 =
(3.01E+02)
1.01E-01 =
(5.51E-02)
1.14E-01 =
(1.81E-02)
2.29E-01 (2.52E-02)
2.44E+00 =
(4.60E-01)
9.15E+00 +
(5.25E-01)
1.29E+02 (7.85E+01)
4.85E+00 (1.76E+00)
2.69E+00 (6.23E-01)
3.57E+00 =
(1.41E+00)
7.84E+01 (7.25E+01)
3.43E+01 (2.47E+01)
3.15E+02 =
(4.02E-13)
2.24E+02 (1.95E+00)
2.03E+02 (3.95E-02)
1.00E+02 =
(1.92E-02)
3.02E+02 =
(1.40E+01)
8.39E+02 =
(1.42E+01)
7.17E+02 (3.10E+00)
9.28E+02 =
(3.55E+02)
9/20/1
SCSSUMOEA-II
L-SHADE_
EpSin
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.00E+01
(4.78E-05)
4.24E-06
(1.86E-05)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
9.03E+00
(2.07E+00)
4.08E-03
(8.35E-03)
1.43E+03
(3.18E+02)
1.08E-01
(6.90E-02)
1.09E-01
(2.15E-02)
2.10E-01
(3.27E-02)
2.29E+00
(5.34E-01)
9.57E+00
(6.20E-01)
7.77E+01
(4.25E+01)
3.89E+00
(1.47E+00)
2.23E+00
(6.65E-01)
3.72E+00
(1.34E+00)
2.43E+01
(4.11E+01)
2.54E+01
(4.05E+00)
3.15E+02
(4.02E-13)
2.22E+02
(4.63E+00)
2.03E+02
(4.46E-02)
1.00E+02
(1.98E-02)
3.02E+02
(1.40E+01)
8.35E+02
(1.53E+01)
7.16E+02
(2.28E+00)
9.35E+02
(4.83E+02)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.01E+01 (2.98E-02)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
1.31E+01 (1.94E+00)
4.49E-03 =
(9.60E-03)
1.14E+03 =
(2.09E+02)
1.54E-01 =
(2.30E-02)
1.34E-01 (1.64E-02)
1.93E-01 =
(2.91E-02)
2.37E+00 (2.41E-01)
8.30E+00 =
(4.58E-01)
1.94E+02 (8.71E+01)
6.02E+00 =
(2.44E+00)
2.63E+00 =
(8.21E-01)
2.34E+00 =
(1.06E+00)
9.09E+01 =
(7.94E+01)
5.17E+01 (5.09E+01)
3.15E+02 =
(4.02E-13)
2.11E+02 =
(1.10E+01)
2.03E+02 =
(3.95E-02)
1.00E+02 (1.25E-02)
3.00E+02 (1.85E-13)
8.37E+02 =
(1.56E+01)
7.22E+02 =
(1.17E+01)
1.46E+03 =
(6.33E+02)
8/22/0
SCSSL-SHADE_
EpSin
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.01E+01
(4.75E-02)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.24E+01
(2.15E+00)
4.90E-03
(1.07E-02)
1.16E+03
(2.03E+02)
1.46E-01
(2.77E-02)
1.24E-01
(1.61E-02)
1.93E-01
(2.44E-02)
2.24E+00
(2.91E-01)
8.26E+00
(3.76E-01)
1.42E+02
(8.41E+01)
5.68E+00
(2.09E+00)
2.78E+00
(6.45E-01)
2.67E+00
(1.18E+00)
9.96E+01
(8.91E+01)
3.76E+01
(3.85E+01)
3.15E+02
(4.16E-13)
2.11E+02
(1.10E+01)
2.03E+02
(3.24E-02)
1.00E+02
(1.64E-02)
3.00E+02
(9.09E-14)
8.37E+02
(1.81E+01)
7.20E+02
(6.36E+00)
1.51E+03
(6.72E+02)
jSO
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.09E+01 =
(8.04E-02)
8.61E-06 =
(3.52E-05)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
8.76E+00 (1.97E+00)
1.43E+00 =
(1.02E+00)
1.20E+03 =
(2.73E+02)
4.17E-01 +
(4.93E-01)
1.37E-01 +
(2.24E-02)
2.26E-01 =
(4.08E-02)
2.37E+00 (2.73E-01)
8.58E+00 =
(7.71E-01)
6.38E+01 =
(2.31E+01)
2.14E+00 =
(1.23E+00)
2.04E+00 =
(7.16E-01)
2.04E+00 =
(8.67E-01)
2.86E+01 =
(4.42E+01)
2.91E+01 (2.45E+01)
3.15E+02 =
(4.16E-13)
2.09E+02 (1.08E+01)
2.03E+02 =
(2.75E-02)
1.00E+02 =
(2.13E-02)
3.00E+02 =
(2.30E-13)
8.25E+02 (2.15E+01)
7.16E+02 (2.07E+00)
6.20E+02 (1.67E+02)
7/21/2
SCSSjSO
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
2.09E+01
(4.80E-02)
1.02E-02
(7.27E-02)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
7.57E+00
(1.62E+00)
1.64E+00
(9.94E-01)
1.26E+03
(2.45E+02)
9.00E-01
(7.61E-01)
1.52E-01
(3.04E-02)
2.30E-01
(3.63E-02)
2.13E+00
(3.37E-01)
8.60E+00
(7.27E-01)
6.22E+01
(2.13E+01)
2.19E+00
(1.17E+00)
1.86E+00
(6.30E-01)
1.97E+00
(8.07E-01)
1.18E+01
(8.29E+00)
2.31E+01
(3.73E+00)
3.15E+02
(4.02E-13)
2.02E+02
(5.83E+00)
2.03E+02
(2.60E-02)
1.00E+02
(2.44E-02)
3.00E+02
(1.23E-05)
8.16E+02
(1.94E+01)
7.15E+02
(1.17E+00)
5.70E+02
(1.73E+02)
Note: In our experiments, the structural bias introduced in the original UMOEA-II and L-SHADE_EpSin algorithms were removed according to the
suggestions in [1]. In detail, in UMOEA-II and SCSS-UMOEA-II, the mutation strategy Vi, G = Fi × Xr1, G + (Xr2, G - Xr3, G) was modified as Vi, G = Xr1, G +
(Xr2, G - Xr3, G) by setting Fi = 1. In L-SHADE_EpSin and SCSS-L-SHADE_EpSin, the local search procedures were skipped.
[1] A. P. Piotrowski and J. J. Napiorkowski, “Some metaheuristics should be simplified,” Inf. Sci., vol. 427, pp. 32-62, Feb. 2018.
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S6
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES
ON 50-D CEC2014 BENCHMARK SET OVER 51 INDEPENDENT RUNS
L-SHADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
21
9.71E+02 (1.66E+03)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
8.23E+01 =
(3.38E+01)
2.03E+01 (3.08E-02)
9.14E-02 (2.74E-01)
0.00E+00 =
(0.00E+00)
3.64E-08 (3.14E-08)
1.26E+01 (2.44E+00)
1.72E-01 (5.24E-02)
3.42E+03 (3.46E+02)
2.44E-01 (3.53E-02)
1.60E-01 (1.74E-02)
3.23E-01 (4.96E-02)
5.30E+00 (5.66E-01)
1.69E+01 +
(4.35E-01)
1.63E+03 (3.52E+02)
1.05E+02 (1.38E+01)
8.11E+00 +
(1.87E+00)
1.45E+01 (3.75E+00)
5.59E+02 (1.62E+02)
1.03E+02 =
(7.30E+01)
3.44E+02 =
(3.20E-13)
2.75E+02 (4.98E-01)
2.05E+02 (3.48E-01)
1.00E+02 =
(1.98E-02)
3.42E+02 (2.68E+01)
1.13E+03 =
(3.69E+01)
8.04E+02 =
(3.34E+01)
8.59E+03 =
(4.15E+02)
18/10/2
SCSSL-SHADE
1.04E+02
(5.89E+02)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
7.62E+01
(4.00E+01)
2.02E+01
(8.40E-02)
5.69E-02
(2.45E-01)
0.00E+00
(0.00E+00)
2.37E-08
(4.23E-08)
1.18E+01
(2.02E+00)
1.38E-01
(5.18E-02)
3.28E+03
(3.38E+02)
2.11E-01
(3.26E-02)
1.50E-01
(2.08E-02)
2.49E-01
(9.34E-02)
4.99E+00
(4.75E-01)
1.71E+01
(4.88E-01)
5.59E+02
(2.32E+02)
2.30E+01
(6.42E+00)
9.64E+00
(1.45E+00)
7.96E+00
(1.96E+00)
3.42E+02
(1.11E+02)
9.95E+01
(7.03E+01)
3.44E+02
(3.46E-13)
2.74E+02
(1.13E+00)
2.05E+02
(2.33E-01)
1.00E+02
(1.66E-02)
3.35E+02
(2.17E+01)
1.12E+03
(3.09E+01)
8.02E+02
(3.22E+01)
8.53E+03
(3.14E+02)
UMOEA-II
1.17E-03 (9.11E-04)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
2.69E+01 +
(4.42E+01)
2.00E+01 =
(6.24E-04)
3.49E-01 (4.91E-01)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
1.60E+01 (4.61E+00)
1.30E+00 +
(1.19E+00)
3.94E+03 =
(7.60E+02)
1.63E-01 =
(1.06E-01)
1.63E-01 =
(2.40E-02)
3.01E-01 (2.29E-02)
5.39E+00 =
(1.04E+00)
1.84E+01 +
(7.63E-01)
1.11E+03 (3.60E+02)
5.70E+01 (2.14E+01)
8.17E+00 =
(2.20E+00)
1.34E+01 (3.52E+00)
4.38E+02 (1.27E+02)
1.81E+02 =
(8.35E+01)
3.44E+02 =
(4.67E-13)
2.75E+02 (8.57E-01)
2.05E+02 =
(2.98E-01)
1.00E+02 =
(2.50E-02)
3.34E+02 (3.31E+01)
1.12E+03 =
(2.83E+01)
8.05E+02 =
(4.27E+01)
8.62E+03 =
(4.71E+02)
10/17/3
SCSSUMOEA-II
L-SHADE_
EpSin
5.83E-04
(3.83E-04)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
5.00E+01
(4.95E+01)
2.00E+01
(4.88E-06)
8.13E-02
(3.21E-01)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.39E+01
(3.94E+00)
3.53E+00
(2.29E+00)
3.93E+03
(6.03E+02)
1.68E-01
(1.06E-01)
1.60E-01
(2.33E-02)
2.63E-01
(2.99E-02)
5.13E+00
(1.06E+00)
1.86E+01
(6.65E-01)
3.94E+02
(1.81E+02)
1.56E+01
(4.28E+00)
7.66E+00
(2.39E+00)
9.33E+00
(3.05E+00)
3.49E+02
(1.32E+02)
1.93E+02
(1.19E+02)
3.44E+02
(4.73E-13)
2.75E+02
(7.27E-01)
2.05E+02
(3.00E-01)
1.00E+02
(2.05E-02)
3.23E+02
(2.59E+01)
1.11E+03
(2.69E+01)
7.95E+02
(3.95E+01)
8.64E+03
(5.04E+02)
1.33E-02 (7.34E-02)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
5.65E+01 (4.83E+01)
2.03E+01 (3.24E-02)
2.04E-04 (2.15E-04)
0.00E+00 =
(0.00E+00)
3.53E-09 =
(2.52E-08)
3.03E+01 (5.20E+00)
4.17E-02 =
(2.19E-02)
3.09E+03 =
(3.06E+02)
2.16E-01 (2.70E-02)
2.06E-01 (2.08E-02)
1.89E-01 (2.33E-02)
5.68E+00 (4.74E-01)
1.67E+01 (3.44E-01)
3.60E+02 =
(1.60E+02)
1.89E+01 =
(6.40E+00)
9.99E+00 (8.84E-01)
6.04E+00 =
(2.23E+00)
3.25E+02 =
(9.65E+01)
9.35E+01 (6.13E+01)
3.44E+02 =
(2.93E-13)
2.68E+02 =
(1.23E+00)
2.05E+02 =
(1.39E-01)
1.00E+02 (4.98E-02)
3.17E+02 =
(2.28E+01)
1.14E+03 =
(3.72E+01)
8.05E+02 =
(2.77E+01)
8.50E+03 =
(3.71E+02)
13/17/0
SCSSL-SHADE_
EpSin
5.13E-05
(3.62E-04)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
4.12E+01
(4.81E+01)
2.02E+01
(7.18E-02)
2.14E-05
(4.97E-05)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
1.90E+01
(5.72E+00)
3.73E-02
(1.78E-02)
3.00E+03
(3.23E+02)
1.99E-01
(2.81E-02)
1.90E-01
(2.35E-02)
1.84E-01
(3.13E-02)
5.04E+00
(5.05E-01)
1.65E+01
(4.28E-01)
3.51E+02
(1.39E+02)
1.83E+01
(6.76E+00)
9.76E+00
(8.22E-01)
5.93E+00
(1.86E+00)
3.08E+02
(1.05E+02)
6.34E+01
(5.00E+01)
3.44E+02
(3.18E-13)
2.68E+02
(1.50E+00)
2.05E+02
(9.28E-02)
1.00E+02
(3.46E-02)
3.25E+02
(2.34E+01)
1.14E+03
(3.83E+01)
8.13E+02
(4.03E+01)
8.60E+03
(4.33E+02)
jSO
1.49E+01 (3.06E+01)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
5.02E+01 =
(4.93E+01)
2.11E+01 =
(5.59E-02)
3.80E-03 (5.50E-03)
0.00E+00 =
(0.00E+00)
0.00E+00 +
(0.00E+00)
1.59E+01 (3.69E+00)
9.92E+00 =
(3.90E+00)
3.22E+03 =
(3.37E+02)
3.69E-01 =
(4.10E-01)
1.92E-01 =
(2.83E-02)
2.91E-01 (4.34E-02)
5.18E+00 (4.85E-01)
1.70E+01 +
(9.41E-01)
3.51E+02 (1.70E+02)
1.08E+01 (3.24E+00)
9.25E+00 (8.19E-01)
5.67E+00 =
(1.95E+00)
3.03E+02 (9.88E+01)
1.51E+02 (1.00E+02)
3.44E+02 =
(3.03E-13)
2.72E+02 (1.80E+00)
2.05E+02 (1.82E-01)
1.00E+02 =
(2.37E-02)
3.10E+02 (1.85E+01)
1.09E+03 =
(2.81E+01)
8.04E+02 =
(4.11E+01)
8.38E+03 =
(3.90E+02)
13/15/2
SCSSjSO
1.59E+00
(2.80E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
5.79E+01
(4.86E+01)
2.11E+01
(5.17E-02)
3.66E-02
(1.44E-01)
0.00E+00
(0.00E+00)
1.82E-09
(6.41E-09)
1.13E+01
(2.93E+00)
8.40E+00
(3.24E+00)
3.26E+03
(3.75E+02)
7.48E-01
(7.45E-01)
2.01E-01
(4.22E-02)
2.73E-01
(4.15E-02)
4.68E+00
(6.92E-01)
1.73E+01
(7.30E-01)
1.76E+02
(1.11E+02)
7.21E+00
(2.16E+00)
8.56E+00
(7.29E-01)
5.17E+00
(1.71E+00)
2.36E+02
(8.45E+01)
1.03E+02
(8.34E+01)
3.44E+02
(3.46E-13)
2.70E+02
(2.18E+00)
2.05E+02
(1.35E-01)
1.00E+02
(3.87E-02)
3.10E+02
(1.84E+01)
1.08E+03
(3.04E+01)
8.03E+02
(4.48E+01)
8.30E+03
(3.38E+02)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S7
PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES
ON 30-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
DE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
9.02E+05 (1.64E+05)
1.16E+03 (2.02E+02)
5.47E+04 (6.63E+03)
8.80E+01 (3.19E-01)
1.92E+02 (8.66E+00)
3.08E-02 (6.00E-03)
2.42E+02 (1.22E+01)
1.94E+02 (9.61E+00)
1.61E+01 (6.51E+00)
6.81E+03 (2.63E+02)
1.65E+02 (1.31E+01)
2.37E+07 (5.42E+06)
1.37E+05 (5.32E+04)
3.76E+02 =
(1.11E+02)
1.61E+03 (3.81E+02)
1.21E+03 (1.39E+02)
2.42E+02 (3.93E+01)
1.07E+06 (4.46E+05)
2.73E+03 (1.11E+03)
2.69E+02 (4.75E+01)
3.87E+02 (1.14E+01)
1.00E+02 (2.22E-04)
5.33E+02 (9.35E+00)
6.02E+02 (1.07E+01)
3.87E+02 (2.96E-02)
2.87E+03 (9.63E+01)
5.28E+02 (6.09E+00)
4.71E+02 (2.26E+01)
1.07E+03 (1.11E+02)
1.32E+05 (3.38E+04)
29/1/0
SCSSDE
3.07E+01
(2.90E+01)
6.81E+02
(8.83E+01)
4.64E+04
(7.57E+03)
8.66E+01
(1.11E-01)
1.72E+02
(9.67E+00)
3.68E-04
(1.29E-04)
2.12E+02
(1.02E+01)
1.70E+02
(1.34E+01)
2.87E-08
(3.41E-08)
6.53E+03
(3.29E+02)
1.40E+02
(1.47E+01)
1.54E+07
(3.15E+06)
1.11E+05
(2.81E+04)
3.75E+02
(1.23E+02)
1.00E+03
(2.20E+02)
1.03E+03
(1.24E+02)
1.96E+02
(2.59E+01)
7.54E+05
(2.91E+05)
1.07E+03
(4.87E+02)
2.13E+02
(5.05E+01)
3.69E+02
(1.03E+01)
1.00E+02
(2.30E+00)
5.16E+02
(1.04E+01)
5.93E+02
(1.06E+01)
3.87E+02
(5.27E-02)
2.54E+03
(3.81E+02)
5.20E+02
(4.04E+00)
4.34E+02
(1.56E+01)
9.56E+02
(9.60E+01)
6.02E+04
(1.66E+04)
ES
2.22E+09 (3.28E+08)
1.40E+03 (3.33E+02)
5.76E+04 (8.64E+03)
3.42E+02 (5.43E+01)
2.29E+02 (1.27E+01)
1.75E+01 (2.23E+00)
3.42E+02 (2.96E+01)
2.29E+02 (1.32E+01)
9.68E+02 (1.64E+02)
7.11E+03 (3.29E+02)
7.14E+02 (1.14E+02)
2.06E+08 (5.45E+07)
5.58E+07 (1.52E+07)
1.02E+05 (4.91E+04)
6.29E+06 (3.50E+06)
1.76E+03 (1.99E+02)
5.66E+02 (1.19E+02)
3.37E+06 (1.44E+06)
9.81E+06 (4.22E+06)
5.69E+02 (1.13E+02)
4.18E+02 (1.49E+01)
4.45E+02 (4.39E+01)
5.91E+02 (1.44E+01)
6.54E+02 (1.58E+01)
5.28E+02 (2.85E+01)
3.42E+03 (1.40E+02)
6.08E+02 (1.24E+01)
6.24E+02 (2.78E+01)
1.35E+03 (1.56E+02)
1.35E+07 (4.56E+06)
30/0/0
SCSSES
5.70E+08
(9.32E+07)
3.95E+02
(6.91E+01)
1.95E+04
(3.37E+03)
1.79E+02
(1.39E+01)
1.91E+02
(9.79E+00)
8.57E+00
(8.43E-01)
2.79E+02
(1.22E+01)
1.92E+02
(1.05E+01)
2.48E+02
(5.18E+01)
6.72E+03
(3.18E+02)
3.06E+02
(4.13E+01)
5.81E+07
(1.50E+07)
1.47E+07
(4.98E+06)
3.43E+04
(1.59E+04)
1.61E+06
(8.38E+05)
1.11E+03
(2.30E+02)
2.52E+02
(5.86E+01)
1.12E+06
(5.61E+05)
2.39E+06
(1.13E+06)
2.83E+02
(6.57E+01)
3.84E+02
(1.25E+01)
3.54E+02
(8.91E+02)
5.48E+02
(1.27E+01)
6.20E+02
(1.16E+01)
4.34E+02
(1.15E+01)
2.93E+03
(1.39E+02)
5.58E+02
(6.78E+00)
5.11E+02
(1.80E+01)
8.94E+02
(8.78E+01)
3.81E+06
(1.05E+06)
PSO
1.34E+10 =
(1.24E+10)
4.27E+04 =
(2.74E+04)
2.19E+08 (1.26E+09)
4.19E+03 =
(4.34E+03)
2.97E+02 (1.04E+02)
5.84E+01 (1.84E+01)
5.50E+02 (1.60E+02)
3.24E+02 (1.09E+02)
1.08E+04 (5.86E+03)
6.78E+03 (1.49E+03)
9.51E+04 (4.75E+05)
2.11E+09 =
(1.92E+09)
2.80E+09 (3.45E+09)
1.48E+07 (1.08E+07)
3.27E+08 (5.32E+08)
2.66E+03 =
(8.21E+02)
1.33E+03 (5.57E+02)
1.48E+08 (1.40E+08)
2.86E+08 (4.04E+08)
9.93E+02 (4.63E+02)
5.52E+02 (9.18E+01)
6.75E+03 (1.61E+03)
9.15E+02 (1.24E+02)
1.27E+03 (8.25E+01)
8.38E+02 (5.39E+02)
8.03E+03 =
(1.85E+03)
1.52E+03 (2.04E+02)
2.22E+03 =
(1.78E+03)
2.30E+03 (5.94E+02)
5.64E+08 (6.82E+08)
23/7/0
SCSSPSO
1.08E+10
(9.30E+09)
3.87E+04
(2.65E+04)
1.11E+05
(8.98E+04)
3.21E+03
(4.13E+03)
2.20E+02
(5.09E+01)
4.51E+01
(1.29E+01)
4.37E+02
(1.43E+02)
2.05E+02
(4.92E+01)
4.40E+03
(1.96E+03)
5.01E+03
(1.02E+03)
3.57E+03
(5.36E+03)
2.42E+09
(2.53E+09)
1.89E+09
(2.76E+09)
1.01E+07
(9.21E+06)
4.54E+04
(6.83E+04)
2.50E+03
(7.43E+02)
1.05E+03
(4.64E+02)
4.66E+07
(8.29E+07)
6.32E+07
(8.57E+07)
7.25E+02
(2.56E+02)
4.64E+02
(5.95E+01)
6.00E+03
(1.29E+03)
8.69E+02
(1.24E+02)
1.21E+03
(9.28E+01)
6.88E+02
(3.60E+02)
7.76E+03
(.39E+03)
1.42E+03
(2.62E+02)
1.59E+03
(1.28E+03)
1.98E+03
(5.06E+02)
4.28E+08
(6.88E+08)
22
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S8
PERFORMANCE COMPARISONS OF THREE SCSS-BASED CLASSIC ALGORITHMS WITH THE BASELINES
ON 50-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
DE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
1.17E+08 (6.83E+07)
6.09E+03 (8.14E+02)
1.35E+05 =
(1.41E+04)
3.18E+02 (1.94E+01)
4.15E+02 (1.46E+01)
2.48E+00 (2.34E-01)
5.28E+02 (1.95E+01)
4.09E+02 (1.68E+01)
1.13E+03 (2.00E+02)
1.29E+04 (3.28E+02)
2.87E+02 (2.05E+01)
3.44E+08 (5.36E+07)
7.29E+05 (3.28E+05)
3.34E+05 =
(1.08E+05)
5.47E+04 =
(2.13E+04)
3.00E+03 (2.29E+02)
1.75E+03 (1.35E+02)
7.37E+06 (2.04E+06)
4.60E+04 (1.78E+04)
1.43E+03 (1.24E+02)
6.10E+02 (1.50E+01)
1.27E+04 (2.54E+03)
8.31E+02 (1.30E+01)
8.79E+02 (1.78E+01)
5.78E+02 (9.35E+00)
5.16E+03 (1.52E+02)
5.80E+02 (1.47E+01)
4.98E+02 (1.04E+01)
2.15E+03 (1.69E+02)
2.94E+07 (5.52E+06)
27/3/0
SCSSDE
1.00E+06
(9.08E+05)
4.33E+03
(4.76E+02)
1.30E+05
(1.46E+04)
1.85E+02
(3.52E+01)
3.78E+02
(1.17E+01)
9.22E-02
(2.22E-02)
4.68E+02
(1.20E+01)
3.78E+02
(1.66E+01)
2.75E+01
(1.40E+01)
1.28E+04
(2.92E+02)
2.43E+02
(1.64E+01)
2.37E+08
(3.76E+07)
5.70E+05
(1.72E+05)
3.31E+05
(8.21E+04)
5.38E+04
(1.52E+04)
2.90E+03
(2.23E+02)
1.53E+03
(1.65E+02)
5.52E+06
(1.87E+06)
3.21E+04
(8.48E+03)
1.27E+03
(1.21E+02)
5.82E+02
(1.37E+01)
1.04E+04
(3.87E+03)
7.98E+02
(1.16E+01)
8.60E+02
(1.14E+01)
4.81E+02
(7.85E-01)
4.82E+03
(1.57E+02)
5.63E+02
(2.19E+01)
4.62E+02
(2.70E-01)
1.94E+03
(1.75E+02)
1.84E+07
(2.69E+06)
ES
7.17E+09 (9.69E+08)
4.19E+03 (6.63E+02)
1.46E+05 (1.89E+04)
9.60E+02 (1.40E+02)
4.49E+02 (1.77E+01)
2.36E+01 (2.00E+00)
6.83E+02 (3.85E+01)
4.48E+02 (1.66E+01)
3.69E+03 (6.61E+02)
1.34E+04 (2.78E+02)
2.45E+03 (4.30E+02)
1.54E+09 (2.98E+08)
3.07E+08 (8.24E+07)
1.10E+06 (4.98E+05)
6.43E+07 (2.12E+07)
3.65E+03 (2.64E+02)
2.42E+03 (2.00E+02)
1.41E+07 (4.12E+06)
3.16E+07 (1.10E+07)
1.80E+03 (1.79E+02)
6.41E+02 (1.89E+01)
1.30E+04 (2.91E+03)
9.15E+02 (1.95E+01)
9.66E+02 (1.88E+01)
1.12E+03 (8.82E+01)
5.92E+03 (2.32E+02)
1.06E+03 (6.83E+01)
1.10E+03 (1.02E+02)
2.74E+03 (2.77E+02)
1.79E+08 (3.52E+07)
30/0/0
SCSSES
1.87E+09
(2.89E+08)
1.45E+03
(2.19E+02)
6.90E+04
(9.99E+03)
4.14E+02
(3.91E+01)
3.91E+02
(1.67E+01)
1.21E+01
(1.26E+00)
5.45E+02
(1.91E+01)
3.92E+02
(1.76E+01)
9.68E+02
(1.64E+02)
1.29E+04
(2.95E+02)
7.74E+02
(8.54E+01)
4.40E+08
(9.39E+07)
7.84E+07
(1.81E+07)
3.50E+05
(1.52E+05)
1.60E+07
(5.57E+06)
2.73E+03
(4.57E+02)
1.83E+03
(2.86E+02)
4.95E+06
(2.03E+06)
8.85E+06
(2.57E+06)
1.43E+03
(1.86E+02)
5.88E+02
(1.66E+01)
1.24E+04
(3.04E+03)
8.37E+02
(1.82E+01)
8.99E+02
(1.61E+01)
7.34E+02
(3.28E+01)
5.12E+03
(2.13E+02)
8.04E+02
(4.84E+01)
6.78E+02
(3.73E+01)
1.77E+03
(2.17E+02)
6.11E+07
(1.10E+07)
PSO
1.99E+10 =
(2.15E+10)
3.79E+04 =
(3.16E+04)
7.63E+09 (2.14E+10)
3.28E+03 =
(3.45E+03)
6.31E+02 (1.73E+02)
6.59E+01 =
(1.82E+01)
1.23E+03 =
(4.16E+02)
5.95E+02 (1.66E+02)
4.03E+04 (1.94E+04)
1.23E+04 (2.35E+03)
1.07E+05 =
(3.46E+05)
2.42E+10 (1.76E+10)
1.28E+10 =
(1.23E+10)
5.92E+07 =
(6.40E+07)
5.83E+08 (1.32E+09)
4.31E+03 (1.29E+03)
1.15E+04 (2.69E+04)
1.01E+08 (1.16E+08)
1.79E+09 (1.77E+09)
2.32E+03 (7.55E+02)
9.06E+02 (1.71E+02)
1.26E+04 (1.74E+03)
1.90E+03 (2.25E+02)
1.77E+03 (9.28E+01)
2.35E+03 (2.33E+03)
1.51E+04 =
(1.50E+03)
2.32E+03 (4.04E+02)
3.93E+03 =
(2.10E+03)
4.82E+03 (3.13E+03)
1.56E+09 =
(1.69E+09)
19/11/0
SCSSPSO
1.85E+10
(1.92E+10)
3.85E+04
(3.46E+04)
3.71E+05
(2.76E+05)
3.16E+03
(3.65E+03)
4.62E+02
(6.97E+01)
6.58E+01
(9.09E+00)
1.22E+03
(3.62E+02)
4.79E+02
(6.00E+01)
1.80E+04
(5.25E+03)
1.05E+04
(2.00E+03)
2.90E+04
(2.09E+04)
1.45E+10
(1.18E+10)
1.02E+10
(9.55E+09)
7.28E+07
(1.04E+08)
3.71E+07
(1.78E+08)
3.39E+03
(8.56E+02)
4.79E+03
(5.70E+03)
4.88E+07
(7.46E+07)
1.12E+09
(1.62E+09)
1.74E+03
(4.95E+02)
7.82E+02
(1.14E+02)
1.09E+04
(1.64E+03)
1.80E+03
(2.11E+02)
1.71E+03
(1.29E+02)
1.52E+03
(1.59E+03)
1.44E+04
(2.11E+03)
2.11E+03
(3.63E+02)
3.96E+03
(2.79E+03)
3.98E+03
(4.28E+03)
1.20E+09
(1.37E+09)
23
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S9
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES
ON 30-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
JADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
24
0.00E+00 =
(0.00E+00)
1.58E-05 =
(8.56E-06)
1.18E+04 (1.92E+04)
5.18E+01 =
(2.08E+01)
2.83E+01 (4.01E+00)
0.00E+00 =
(0.00E+00)
5.61E+01 (3.87E+00)
2.84E+01 (5.00E+00)
2.13E-02 =
(9.01E-02)
1.88E+03 (2.70E+02)
3.37E+01 (2.26E+01)
1.48E+03 =
(8.87E+02)
4.36E+01 =
(2.16E+01)
9.70E+03 (1.12E+04)
1.94E+03 (3.78E+03)
3.92E+02 (1.27E+02)
8.33E+01 (2.86E+01)
5.06E+04 (7.16E+04)
1.88E+03 (4.75E+03)
9.72E+01 (5.22E+01)
2.28E+02 (4.78E+00)
1.00E+02 =
(2.56E-05)
3.75E+02 (6.33E+00)
4.40E+02 (4.90E+00)
3.87E+02 (1.86E-01)
1.19E+03 (1.51E+02)
5.01E+02 =
(7.16E+00)
3.41E+02 =
(5.64E+01)
4.85E+02 (2.28E+01)
2.79E+03 =
(2.00E+03)
19/11/0
SCSSJADE
0.00E+00
(0.00E+00)
1.70E-05
(9.99E-06)
7.74E+02
(5.53E+03)
5.14E+01
(2.06E+01)
2.17E+01
(4.50E+00)
0.00E+00
(0.00E+00)
5.19E+01
(4.41E+00)
2.39E+01
(4.09E+00)
7.02E-03
(2.43E-02)
1.79E+03
(2.39E+02)
2.28E+01
(2.00E+01)
1.30E+03
(7.31E+02)
3.92E+01
(1.61E+01)
2.05E+03
(7.03E+03)
1.14E+02
(6.60E+02)
3.27E+02
(1.28E+02)
7.21E+01
(2.09E+01)
7.69E+03
(3.87E+04)
1.20E+01
(6.37E+00)
7.83E+01
(4.58E+01)
2.22E+02
(4.93E+00)
1.39E+02
(2.76E+02)
3.71E+02
(6.99E+00)
4.36E+02
(5.27E+00)
3.87E+02
(1.72E-01)
1.16E+03
(8.12E+01)
5.03E+02
(7.65E+00)
3.34E+02
(5.44E+01)
4.74E+02
(1.52E+01)
2.13E+03
(1.42E+02)
SHADE
0.00E+00 =
(0.00E+00)
1.77E-05 =
(1.03E-05)
0.00E+00 =
(0.00E+00)
5.47E+01 =
(1.62E+01)
1.99E+01 =
(3.24E+00)
0.00E+00 =
(0.00E+00)
5.09E+01 (3.87E+00)
2.16E+01 =
(3.42E+00)
0.00E+00 =
(0.00E+00)
1.73E+03 =
(2.71E+02)
2.10E+01 =
(2.53E+01)
2.03E+03 (2.68E+03)
3.84E+01 (1.76E+01)
2.73E+01 =
(5.83E+00)
1.32E+01 =
(9.70E+00)
2.91E+02 (1.16E+02)
4.83E+01 =
(1.29E+01)
7.32E+01 (4.20E+01)
7.83E+00 =
(3.06E+00)
6.23E+01 =
(3.64E+01)
2.21E+02 =
(3.13E+00)
1.00E+02 =
(1.00E-13)
3.68E+02 =
(4.87E+00)
4.38E+02 (3.82E+00)
3.87E+02 (1.38E-01)
1.12E+03 =
(6.24E+01)
5.02E+02 =
(5.62E+00)
3.34E+02 =
(5.47E+01)
4.63E+02 =
(2.62E+01)
2.10E+03 =
(1.27E+02)
7/23/0
SCSSSHADE
0.00E+00
(0.00E+00)
1.39E-05
(8.49E-06)
0.00E+00
(0.00E+00)
5.29E+01
(1.76E+01)
1.97E+01
(3.18E+00)
0.00E+00
(0.00E+00)
4.92E+01
(2.84E+00)
2.07E+01
(3.64E+00)
0.00E+00
(0.00E+00)
1.72E+03
(2.46E+02)
2.13E+01
(2.47E+01)
1.20E+03
(5.83E+02)
2.68E+01
(1.20E+01)
2.61E+01
(4.08E+00)
1.05E+01
(5.76E+00)
2.43E+02
(1.35E+02)
5.10E+01
(9.63E+00)
3.43E+01
(1.53E+01)
7.40E+00
(2.40E+00)
5.43E+01
(3.33E+01)
2.20E+02
(3.86E+00)
1.00E+02
(1.00E-13)
3.66E+02
(5.71E+00)
4.36E+02
(3.77E+00)
3.87E+02
(1.33E-01)
1.09E+03
(6.26E+01)
5.02E+02
(4.92E+00)
3.30E+02
(4.90E+01)
4.65E+02
(1.66E+01)
2.08E+03
(1.39E+02)
CMA-ES
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
3.99E+01 +
(2.74E+01)
6.58E+02 (2.22E+02)
9.91E+01 (1.56E+01)
3.66E+03 (1.11E+03)
5.79E+02 (1.43E+02)
1.37E+04 (3.23E+03)
4.93E+03 (5.98E+02)
1.67E+02 (5.67E+01)
1.51E+03 =
(3.69E+02)
1.57E+03 =
(7.42E+02)
1.85E+02 =
(5.74E+01)
3.09E+02 =
(1.32E+02)
5.92E+02 (2.96E+02)
2.80E+02 (2.03E+02)
2.07E+02 =
(8.94E+01)
2.04E+02 (8.72E+01)
1.38E+03 (3.73E+02)
4.92E+02 (2.67E+02)
5.70E+03 (1.03E+03)
1.99E+03 (8.26E+02)
4.74E+02 =
(9.73E+01)
3.87E+02 (2.74E+00)
1.20E+03 (4.75E+02)
8.04E+02 (1.74E+03)
3.51E+02 =
(6.13E+01)
7.88E+02 (1.84E+02)
2.22E+03 =
(2.09E+02)
18/11/1
SCSSCMA-ES
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
4.30E+01
(2.55E+01)
1.34E+02
(2.26E+02)
3.99E+01
(4.70E+01)
2.71E+02
(8.12E+02)
1.60E+02
(2.05E+02)
5.85E+03
(7.13E+03)
4.05E+03
(1.01E+03)
1.20E+02
(3.97E+01)
1.55E+03
(3.41E+02)
1.35E+03
(7.07E+02)
1.66E+02
(5.33E+01)
2.83E+02
(1.36E+02)
3.36E+02
(2.36E+02)
1.45E+02
(9.83E+01)
1.98E+02
(7.43E+01)
1.73E+02
(6.95E+01)
2.05E+02
(1.65E+02)
3.03E+02
(1.56E+02)
3.05E+03
(2.50E+03)
6.46E+02
(6.87E+02)
4.57E+02
(1.09E+01)
3.87E+02
(2.71E-02)
1.20E+03
(3.22E+02)
4.86E+02
(1.08E+01)
3.42E+02
(5.34E+01)
6.36E+02
(1.25E+02)
2.19E+03
(2.20E+02)
LIPS
8.03E+02 +
(1.45E+03)
2.33E+01 (9.02E+01)
1.60E+04 (7.66E+03)
1.64E+02 (9.39E+01)
6.43E+01 (1.35E+01)
8.27E+00 (5.05E+00)
9.77E+01 (2.10E+01)
6.23E+01 (1.31E+01)
6.01E+02 (4.21E+02)
2.80E+03 (4.44E+02)
1.99E+02 (1.41E+02)
1.85E+06 (6.00E+06)
5.74E+03 (5.63E+03)
1.40E+04 (1.13E+04)
2.35E+03 (3.05E+03)
7.30E+02 (2.21E+02)
2.89E+02 (1.19E+02)
1.71E+05 (1.53E+05)
1.55E+03 =
(1.99E+03)
3.21E+02 (1.02E+02)
2.65E+02 (1.55E+01)
1.58E+02 (4.06E+02)
4.45E+02 (3.32E+01)
5.00E+02 (2.83E+01)
4.29E+02 (2.71E+01)
1.47E+03 (8.10E+02)
6.12E+02 (2.52E+01)
5.00E+02 (9.70E+01)
9.73E+02 (1.78E+02)
1.19E+05 (1.81E+05)
28/1/1
SCSSLIPS
2.73E+03
(4.18E+03)
1.87E-03
(1.95E-04)
7.74E+03
(3.55E+03)
1.11E+02
(4.93E+01)
3.43E+01
(9.30E+00)
4.58E-01
(5.87E-01)
7.32E+01
(1.09E+01)
3.58E+01
(8.35E+00)
2.07E+01
(2.53E+01)
2.15E+03
(3.40E+02)
8.58E+01
(4.31E+01)
1.78E+05
(2.11E+05)
2.78E+03
(4.82E+03)
8.81E+03
(2.02E+04)
1.40E+03
(2.16E+03)
4.78E+02
(1.61E+02)
1.52E+02
(6.88E+01)
1.16E+05
(6.72E+04)
1.61E+03
(3.30E+03)
1.83E+02
(7.84E+01)
2.39E+02
(9.85E+00)
1.00E+02
(2.11E-13)
3.91E+02
(1.08E+01)
4.49E+02
(1.02E+01)
3.99E+02
(1.32E+01)
1.14E+03
(5.73E+02)
5.56E+02
(1.69E+01)
3.90E+02
(7.31E+01)
7.05E+02
(7.69E+01)
1.20E+04
(5.61E+03)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S10
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED ADVANCED ALGORITHMS WITH THE BASELINES
ON 50-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
JADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
25
0.00E+00 =
(0.00E+00)
4.21E-05 +
(1.21E-05)
1.42E+04 (3.38E+04)
5.46E+01 =
(5.18E+01)
5.18E+01 (9.01E+00)
0.00E+00 +
(0.00E+00)
9.89E+01 (8.16E+00)
5.43E+01 (8.64E+00)
1.44E+00 =
(1.52E+00)
3.70E+03 (3.77E+02)
1.57E+02 (5.18E+01)
7.02E+03 =
(6.81E+03)
2.52E+02 =
(1.52E+02)
6.91E+04 (1.19E+05)
1.13E+03 (2.51E+03)
9.06E+02 (1.65E+02)
6.40E+02 (1.59E+02)
1.82E+05 (4.33E+05)
9.41E+02 (2.46E+03)
4.74E+02 (1.35E+02)
2.54E+02 (1.03E+01)
3.68E+03 (1.67E+03)
4.79E+02 (1.09E+01)
5.40E+02 (8.46E+00)
5.23E+02 =
(3.28E+01)
1.63E+03 (1.22E+02)
5.58E+02 =
(2.58E+01)
4.91E+02 =
(2.25E+01)
4.60E+02 =
(6.92E+01)
6.64E+05 =
(9.01E+04)
18/10/2
SCSSJADE
0.00E+00
(0.00E+00)
4.93E-05
(1.63E-05)
0.00E+00
(0.00E+00)
5.37E+01
(5.01E+01)
3.98E+01
(9.33E+00)
5.77E-07
(2.18E-06)
8.94E+01
(8.04E+00)
4.17E+01
(8.53E+00)
1.46E+00
(1.26E+00)
3.49E+03
(3.97E+02)
1.32E+02
(3.61E+01)
6.57E+03
(3.92E+03)
2.10E+02
(1.23E+02)
5.09E+03
(2.12E+04)
1.92E+02
(9.30E+01)
7.24E+02
(1.67E+02)
5.52E+02
(1.55E+02)
1.59E+02
(1.54E+02)
1.19E+02
(4.55E+01)
3.97E+02
(1.28E+02)
2.41E+02
(8.60E+00)
3.41E+03
(1.45E+03)
4.65E+02
(1.01E+01)
5.29E+02
(6.59E+00)
5.20E+02
(3.62E+01)
1.50E+03
(1.34E+02)
5.55E+02
(2.94E+01)
4.94E+02
(2.11E+01)
4.72E+02
(7.48E+01)
6.56E+05
(8.03E+04)
SHADE
0.00E+00 =
(0.00E+00)
5.08E-05 =
(1.48E-05)
0.00E+00 =
(0.00E+00)
6.40E+01 =
(5.03E+01)
4.35E+01 (5.40E+00)
1.59E-06 =
(2.26E-06)
8.91E+01 (5.48E+00)
4.21E+01 =
(6.54E+00)
3.87E-01 =
(3.94E-01)
3.48E+03 =
(3.77E+02)
8.67E+01 (2.71E+01)
5.66E+03 =
(3.09E+03)
2.94E+02 (1.94E+02)
1.82E+02 (4.59E+01)
2.52E+02 (1.05E+02)
7.26E+02 =
(1.83E+02)
4.78E+02 =
(1.37E+02)
1.38E+02 (8.50E+01)
1.14E+02 (4.32E+01)
3.46E+02 =
(1.19E+02)
2.44E+02 =
(6.19E+00)
3.50E+03 =
(1.50E+03)
4.66E+02 (8.46E+00)
5.35E+02 (8.93E+00)
5.15E+02 =
(3.61E+01)
1.45E+03 (9.07E+01)
5.37E+02 =
(1.88E+01)
4.82E+02 =
(2.44E+01)
4.38E+02 =
(5.83E+01)
6.57E+05 =
(7.82E+04)
11/19/0
SCSSSHADE
0.00E+00
(0.00E+00)
5.41E-05
(1.87E-05)
0.00E+00
(0.00E+00)
5.50E+01
(4.53E+01)
3.89E+01
(6.36E+00)
1.67E-06
(1.87E-06)
8.60E+01
(5.82E+00)
4.10E+01
(7.27E+00)
3.55E-01
(4.33E-01)
3.43E+03
(3.50E+02)
6.88E+01
(1.66E+01)
6.95E+03
(4.86E+03)
1.33E+02
(5.36E+01)
8.43E+01
(2.75E+01)
1.28E+02
(5.77E+01)
7.44E+02
(1.31E+02)
4.90E+02
(1.25E+02)
1.10E+02
(7.29E+01)
7.53E+01
(3.39E+01)
3.27E+02
(9.96E+01)
2.42E+02
(7.15E+00)
3.27E+03
(1.57E+03)
4.60E+02
(8.48E+00)
5.30E+02
(6.90E+00)
5.08E+02
(3.75E+01)
1.41E+03
(9.53E+01)
5.31E+02
(1.33E+01)
4.85E+02
(2.38E+01)
4.46E+02
(5.42E+01)
6.54E+05
(6.50E+04)
CMA-ES
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
4.34E+01 =
(4.79E+01)
1.03E+03 (1.78E+02)
9.54E+01 (1.04E+01)
6.42E+03 (1.55E+03)
1.09E+03 (2.12E+02)
3.08E+04 =
(5.49E+03)
8.04E+03 (9.92E+02)
2.88E+02 (6.63E+01)
2.66E+03 =
(6.49E+02)
2.55E+03 =
(7.76E+02)
3.16E+02 =
(7.64E+01)
4.88E+02 =
(1.68E+02)
9.06E+02 (3.97E+02)
9.86E+02 (2.57E+02)
3.60E+02 =
(1.23E+02)
2.71E+02 =
(1.30E+02)
2.37E+03 (5.04E+02)
7.97E+02 (4.85E+02)
9.11E+03 (1.09E+03)
3.18E+03 (6.79E+02)
7.00E+02 (2.49E+02)
5.02E+02 =
(3.32E+01)
1.90E+03 (5.02E+02)
7.55E+02 (1.17E+03)
4.70E+02 =
(2.01E+01)
1.04E+03 (2.96E+02)
7.86E+05 =
(1.45E+05)
16/14/0
SCSSCMA-ES
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
3.61E+01
(4.31E+01)
6.32E+02
(4.78E+02)
7.49E+01
(3.66E+01)
1.65E+03
(2.74E+03)
5.94E+02
(4.60E+02)
2.64E+04
(1.16E+04)
7.19E+03
(1.22E+03)
2.08E+02
(5.01E+01)
2.64E+03
(6.45E+02)
2.28E+03
(7.63E+02)
2.97E+02
(9.08E+01)
4.84E+02
(1.20E+02)
5.49E+02
(3.04E+02)
5.71E+02
(2.25E+02)
3.31E+02
(1.07E+02)
2.43E+02
(7.61E+01)
8.23E+02
(8.32E+02)
4.13E+02
(3.21E+02)
7.94E+03
(1.30E+03)
1.20E+03
(1.18E+03)
5.72E+02
(2.19E+01)
4.94E+02
(2.97E+01)
1.76E+03
(5.10E+02)
4.76E+02
(1.37E+01)
4.64E+02
(1.60E+01)
6.93E+02
(1.73E+02)
7.87E+05
(1.72E+05)
LIPS
1.17E+03 +
(2.02E+03)
7.62E+02 (7.84E+02)
9.27E+04 (2.23E+04)
6.66E+02 (3.39E+02)
1.68E+02 (2.62E+01)
2.41E+01 (5.43E+00)
3.74E+02 (6.09E+01)
1.74E+02 (3.49E+01)
4.44E+03 (1.45E+03)
5.14E+03 (6.66E+02)
2.35E+03 (2.45E+03)
1.35E+07 (4.17E+07)
6.58E+03 (3.64E+03)
1.32E+05 (3.30E+05)
1.97E+03 (1.89E+03)
1.44E+03 (3.37E+02)
1.16E+03 (2.11E+02)
1.21E+06 (2.22E+06)
3.34E+03 =
(4.99E+03)
6.79E+02 (1.67E+02)
3.60E+02 (3.55E+01)
4.55E+03 (2.41E+03)
7.13E+02 (6.14E+01)
7.71E+02 (7.71E+01)
9.66E+02 (2.15E+02)
3.87E+03 (6.48E+02)
1.19E+03 (9.61E+01)
1.49E+03 (4.96E+02)
2.02E+03 (3.35E+02)
3.31E+07 (1.45E+07)
28/1/1
SCSSLIPS
2.89E+03
(4.25E+03)
3.25E-03
(4.46E-04)
6.53E+04
(1.57E+04)
2.52E+02
(7.79E+01)
1.00E+02
(2.00E+01)
4.92E+00
(2.13E+00)
1.74E+02
(2.69E+01)
1.02E+02
(1.71E+01)
8.85E+02
(5.90E+02)
4.24E+03
(6.02E+02)
2.58E+02
(8.87E+01)
1.84E+06
(1.55E+06)
1.16E+03
(7.74E+02)
2.61E+04
(2.66E+04)
8.09E+02
(6.53E+02)
9.12E+02
(2.46E+02)
7.70E+02
(1.70E+02)
3.56E+05
(2.38E+05)
3.26E+03
(5.11E+03)
4.60E+02
(1.57E+02)
3.01E+02
(1.72E+01)
3.92E+03
(1.87E+03)
5.59E+02
(2.46E+01)
6.05E+02
(1.99E+01)
6.35E+02
(4.87E+01)
2.19E+03
(6.09E+02)
8.66E+02
(6.62E+01)
6.25E+02
(5.57E+01)
1.12E+03
(1.80E+02)
4.90E+06
(1.58E+06)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S11
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES
ON 30-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
L-SHADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
26
0.00E+00 =
(0.00E+00)
4.06E-09 (8.59E-09)
0.00E+00 =
(0.00E+00)
5.86E+01 =
(3.75E-14)
7.02E+00 =
(1.52E+00)
3.38E-09 =
(1.98E-08)
3.79E+01 +
(1.18E+00)
7.11E+00 =
(1.58E+00)
0.00E+00 =
(0.00E+00)
1.41E+03 =
(2.31E+02)
3.73E+01 (2.91E+01)
1.04E+03 (3.37E+02)
1.92E+01 (4.61E+00)
2.19E+01 +
(1.22E+00)
3.54E+00 (1.56E+00)
4.00E+01 =
(2.74E+01)
3.29E+01 =
(6.27E+00)
2.23E+01 (1.28E+00)
5.96E+00 =
(1.87E+00)
3.01E+01 =
(5.93E+00)
2.08E+02 =
(1.65E+00)
1.00E+02 =
(9.20E-14)
3.54E+02 =
(3.16E+00)
4.28E+02 =
(1.58E+00)
3.87E+02 (1.97E-02)
9.85E+02 (3.55E+01)
5.07E+02 =
(4.03E+00)
3.39E+02 =
(5.61E+01)
4.36E+02 +
(7.53E+00)
1.99E+03 (5.56E+01)
9/18/3
SCSSL-SHADE
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
5.86E+01
(3.27E-14)
7.61E+00
(1.58E+00)
1.14E-08
(3.73E-08)
3.91E+01
(2.03E+00)
8.09E+00
(2.13E+00)
0.00E+00
(0.00E+00)
1.44E+03
(2.33E+02)
3.36E+01
(2.90E+01)
6.95E+02
(3.16E+02)
1.73E+01
(4.88E+00)
2.22E+01
(3.11E+00)
2.80E+00
(1.34E+00)
3.43E+01
(1.48E+01)
3.44E+01
(5.90E+00)
2.04E+01
(2.79E+00)
5.90E+00
(2.05E+00)
2.99E+01
(4.37E+00)
2.08E+02
(1.53E+00)
1.00E+02
(1.00E-13)
3.54E+02
(2.98E+00)
4.28E+02
(1.87E+00)
3.87E+02
(1.26E-02)
9.65E+02
(3.66E+01)
5.06E+02
(5.63E+00)
3.27E+02
(4.88E+01)
4.42E+02
(1.15E+01)
1.97E+03
(4.32E+01)
UMOEA-II
0.00E+00 =
(0.00E+00)
4.14E-08 =
(5.51E-08)
0.00E+00 =
(0.00E+00)
5.86E+01 =
(4.90E-14)
8.29E+00 =
(2.19E+00)
1.81E-08 =
(8.05E-08)
4.04E+01 =
(2.73E+00)
8.45E+00 =
(1.86E+00)
0.00E+00 =
(0.00E+00)
1.69E+03 =
(3.17E+02)
1.34E+01 =
(2.02E+01)
8.28E+02 (3.18E+02)
1.53E+01 =
(6.24E+00)
2.22E+01 =
(3.42E+00)
3.30E+00 (1.70E+00)
9.31E+01 =
(9.08E+01)
4.07E+01 +
(8.68E+00)
2.15E+01 =
(6.94E-01)
6.38E+00 =
(1.91E+00)
4.27E+01 =
(9.05E+00)
2.09E+02 =
(2.11E+00)
1.00E+02 =
(1.39E-13)
3.54E+02 =
(4.25E+00)
4.28E+02 +
(2.39E+00)
3.87E+02 (2.43E-02)
9.51E+02 =
(3.60E+01)
5.03E+02 =
(4.75E+00)
3.20E+02 =
(4.37E+01)
4.38E+02 +
(1.62E+01)
1.97E+03 =
(3.05E+01)
3/24/3
SCSSUMOEA-II
L-SHADE_
EpSin
0.00E+00
(0.00E+00)
3.23E-08
(5.00E-08)
0.00E+00
(0.00E+00)
5.87E+01
(7.78E-01)
8.54E+00
(2.06E+00)
6.71E-09
(2.74E-08)
4.06E+01
(2.68E+00)
8.54E+00
(2.36E+00)
0.00E+00
(0.00E+00)
1.63E+03
(3.04E+02)
1.53E+01
(2.34E+01)
2.84E+02
(1.85E+02)
1.61E+01
(5.99E+00)
2.22E+01
(4.58E+00)
2.83E+00
(2.22E+00)
7.11E+01
(8.16E+01)
4.46E+01
(1.00E+01)
2.13E+01
(7.26E-01)
7.13E+00
(2.35E+00)
3.97E+01
(7.88E+00)
2.10E+02
(2.43E+00)
1.00E+02
(1.87E-13)
3.54E+02
(3.85E+00)
4.29E+02
(2.35E+00)
3.87E+02
(1.71E-02)
9.52E+02
(4.31E+01)
5.01E+02
(6.09E+00)
3.26E+02
(4.74E+01)
4.45E+02
(1.19E+01)
1.98E+03
(3.66E+01)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
0.00E+00 =
(0.00E+00)
5.86E+01 =
(2.88E-14)
1.22E+01 (1.60E+00)
8.05E-09 =
(3.25E-08)
4.35E+01 (2.48E+00)
1.35E+01 (1.50E+00)
0.00E+00 =
(0.00E+00)
1.35E+03 =
(1.90E+02)
1.58E+01 =
(2.30E+01)
4.03E+02 =
(2.22E+02)
1.42E+01 =
(6.02E+00)
2.13E+01 =
(4.65E+00)
2.41E+00 =
(1.44E+00)
5.09E+01 (4.44E+01)
2.83E+01 =
(6.47E+00)
2.13E+01 =
(9.45E-01)
5.24E+00 =
(1.63E+00)
2.83E+01 =
(7.68E+00)
2.12E+02 (2.62E+00)
1.00E+02 =
(1.00E-13)
3.55E+02 =
(2.86E+00)
4.29E+02 (2.73E+00)
3.87E+02 =
(5.91E-03)
9.55E+02 (3.92E+01)
5.05E+02 =
(4.52E+00)
3.06E+02 +
(2.63E+01)
4.29E+02 +
(6.34E+00)
1.99E+03 =
(7.24E+01)
7/21/2
SCSSL-SHADE_
EpSin
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
0.00E+00
(0.00E+00)
5.86E+01
(2.93E-14)
1.06E+01
(2.43E+00)
0.00E+00
(0.00E+00)
4.19E+01
(2.75E+00)
1.26E+01
(2.46E+00)
0.00E+00
(0.00E+00)
1.28E+03
(2.38E+02)
1.97E+01
(2.55E+01)
3.77E+02
(2.15E+02)
1.54E+01
(5.86E+00)
2.26E+01
(1.20E+00)
2.58E+00
(1.61E+00)
3.12E+01
(3.38E+01)
2.91E+01
(5.86E+00)
2.13E+01
(9.30E-01)
5.10E+00
(1.87E+00)
2.60E+01
(5.45E+00)
2.10E+02
(2.50E+00)
1.00E+02
(1.00E-13)
3.55E+02
(3.71E+00)
4.27E+02
(2.07E+00)
3.87E+02
(5.70E-03)
9.35E+02
(4.45E+01)
5.05E+02
(4.34E+00)
3.24E+02
(4.66E+01)
4.35E+02
(8.65E+00)
1.99E+03
(5.68E+01)
jSO
0.00E+00 =
(0.00E+00)
6.65E-08 =
(9.56E-08)
0.00E+00 =
(0.00E+00)
5.86E+01 =
(2.13E-14)
8.32E+00 (1.74E+00)
9.39E-09 =
(3.29E-08)
3.84E+01 (1.83E+00)
8.81E+00 (2.17E+00)
0.00E+00 =
(0.00E+00)
1.49E+03 =
(2.66E+02)
9.87E+00 =
(1.89E+01)
1.66E+02 (8.86E+01)
1.60E+01 =
(5.76E+00)
2.20E+01 =
(1.08E+00)
1.26E+00 =
(8.34E-01)
6.50E+01 =
(6.92E+01)
3.45E+01 (7.04E+00)
2.08E+01 =
(3.79E-01)
4.53E+00 =
(1.90E+00)
3.01E+01 =
(8.53E+00)
2.09E+02 (1.93E+00)
1.00E+02 =
(9.20E-14)
3.51E+02 (3.46E+00)
4.26E+02 =
(2.38E+00)
3.87E+02 =
(5.99E-03)
9.30E+02 =
(3.65E+01)
4.97E+02 =
(6.63E+00)
3.13E+02 =
(3.54E+01)
4.32E+02 =
(1.58E+01)
1.97E+03 =
(1.68E+01)
7/23/0
SCSSjSO
0.00E+00
(0.00E+00)
9.39E-08
(9.54E-08)
0.00E+00
(0.00E+00)
5.86E+01
(2.41E-14)
7.49E+00
(1.80E+00)
1.74E-08
(4.45E-08)
3.75E+01
(1.33E+00)
7.57E+00
(2.04E+00)
0.00E+00
(0.00E+00)
1.54E+03
(2.18E+02)
6.46E+00
(1.39E+01)
8.34E+01
(7.27E+01)
1.63E+01
(4.50E+00)
2.14E+01
(3.19E+00)
1.03E+00
(8.73E-01)
5.02E+01
(6.73E+01)
3.17E+01
(7.19E+00)
1.95E+01
(4.82E+00)
4.06E+00
(1.43E+00)
2.75E+01
(7.25E+00)
2.08E+02
(2.04E+00)
1.00E+02
(1.00E-13)
3.50E+02
(3.15E+00)
4.26E+02
(3.06E+00)
3.87E+02
(6.30E-03)
9.25E+02
(4.04E+01)
4.95E+02
(7.76E+00)
3.02E+02
(1.60E+01)
4.27E+02
(2.42E+01)
1.97E+03
(1.11E+01)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S12
L-SHADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
0.00E+00 =
(0.00E+00)
5.01E-06 (3.12E-06)
0.00E+00 =
(0.00E+00)
7.23E+01 =
(4.94E+01)
1.19E+01 =
(2.46E+00)
7.12E-08 (2.58E-07)
6.50E+01 =
(2.23E+00)
1.21E+01 =
(2.39E+00)
0.00E+00 =
(0.00E+00)
3.32E+03 (2.81E+02)
4.80E+01 (6.64E+00)
2.07E+03 =
(5.21E+02)
6.52E+01 (2.98E+01)
3.06E+01 (3.73E+00)
4.53E+01 (1.40E+01)
3.76E+02 =
(1.36E+02)
2.32E+02 =
(6.72E+01)
5.06E+01 (1.72E+01)
3.50E+01 (1.39E+01)
1.56E+02 =
(4.95E+01)
2.16E+02 (2.26E+00)
2.84E+03 =
(1.53E+03)
4.33E+02 (4.04E+00)
5.12E+02 (3.01E+00)
4.82E+02 (4.55E+00)
1.21E+03 (4.31E+01)
5.43E+02 =
(2.15E+01)
4.64E+02 (1.51E+01)
3.53E+02 =
(1.08E+01)
6.68E+05 =
(8.12E+04)
15/15/0
27
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES
ON 50-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
SCSSL-SHADE
0.00E+00
(0.00E+00)
1.66E-06
(9.79E-07)
0.00E+00
(0.00E+00)
7.34E+01
(5.05E+01)
1.20E+01
(1.99E+00)
2.22E-08
(6.76E-08)
6.46E+01
(2.12E+00)
1.17E+01
(2.56E+00)
0.00E+00
(0.00E+00)
3.12E+03
(3.27E+02)
3.37E+01
(4.65E+00)
2.10E+03
(4.81E+02)
5.09E+01
(2.89E+01)
2.48E+01
(2.30E+00)
2.77E+01
(3.82E+00)
3.49E+02
(1.17E+02)
2.04E+02
(9.33E+01)
2.80E+01
(3.87E+00)
1.71E+01
(3.01E+00)
1.72E+02
(6.37E+01)
2.14E+02
(2.74E+00)
3.33E+03
(8.42E+02)
4.30E+02
(4.60E+00)
5.11E+02
(2.81E+00)
4.81E+02
(3.57E+00)
1.17E+03
(3.93E+01)
5.38E+02
(1.56E+01)
4.60E+02
(5.68E+00)
3.57E+02
(1.44E+01)
6.51E+05
(8.03E+04)
UMOEA-II
0.00E+00 =
(0.00E+00)
1.37E-05 (6.95E-06)
3.00E-10 +
(2.14E-09)
7.22E+01 =
(4.97E+01)
1.61E+01 (4.55E+00)
1.66E-04 (5.76E-04)
7.04E+01 =
(5.17E+00)
1.58E+01 =
(4.09E+00)
0.00E+00 =
(0.00E+00)
3.75E+03 =
(5.99E+02)
4.42E+01 (9.48E+00)
2.17E+03 =
(5.36E+02)
4.69E+01 (1.73E+01)
2.85E+01 (3.30E+00)
3.45E+01 (6.42E+00)
4.58E+02 =
(1.68E+02)
3.14E+02 =
(1.18E+02)
3.26E+01 (7.70E+00)
2.08E+01 (3.32E+00)
2.60E+02 =
(1.20E+02)
2.20E+02 (5.20E+00)
2.82E+03 =
(2.11E+03)
4.42E+02 (8.43E+00)
5.12E+02 =
(4.82E+00)
4.82E+02 (6.18E+00)
1.21E+03 =
(6.22E+01)
5.36E+02 (1.67E+01)
4.73E+02 (2.25E+01)
3.62E+02 +
(1.91E+01)
6.68E+05 =
(1.02E+05)
14/14/2
SCSSUMOEA-II
L-SHADE_
EpSin
0.00E+00
(0.00E+00)
6.55E-06
(4.16E-06)
1.54E-08
(2.31E-08)
8.27E+01
(5.36E+01)
1.43E+01
(3.11E+00)
1.16E-07
(2.28E-07)
6.85E+01
(5.14E+00)
1.43E+01
(4.17E+00)
0.00E+00
(0.00E+00)
3.64E+03
(5.22E+02)
3.16E+01
(4.51E+00)
2.01E+03
(4.99E+02)
3.56E+01
(1.57E+01)
2.70E+01
(2.35E+00)
2.69E+01
(3.14E+00)
4.07E+02
(1.69E+02)
3.01E+02
(1.07E+02)
2.60E+01
(2.90E+00)
1.70E+01
(3.00E+00)
2.80E+02
(1.16E+02)
2.18E+02
(4.64E+00)
2.78E+03
(2.16E+03)
4.37E+02
(7.54E+00)
5.11E+02
(3.86E+00)
4.81E+02
(2.33E+00)
1.19E+03
(5.77E+01)
5.31E+02
(1.78E+01)
4.64E+02
(1.55E+01)
3.84E+02
(1.93E+01)
6.38E+05
(5.48E+04)
0.00E+00 =
(0.00E+00)
2.23E-07 (1.36E-07)
0.00E+00 =
(0.00E+00)
5.04E+01 =
(4.38E+01)
2.90E+01 (6.65E+00)
2.57E-07 (3.41E-07)
7.98E+01 (7.02E+00)
3.07E+01 (3.99E+00)
0.00E+00 =
(0.00E+00)
3.07E+03 (2.91E+02)
2.75E+01 =
(2.01E+00)
1.38E+03 =
(3.79E+02)
3.76E+01 =
(2.60E+01)
2.71E+01 =
(2.68E+00)
2.51E+01 =
(3.17E+00)
3.31E+02 (1.25E+02)
2.40E+02 (6.48E+01)
2.53E+01 =
(2.70E+00)
1.62E+01 =
(3.11E+00)
1.35E+02 (5.03E+01)
2.30E+02 (6.27E+00)
1.54E+03 =
(1.62E+03)
4.43E+02 (6.60E+00)
5.13E+02 (5.58E+00)
4.80E+02 =
(1.44E-02)
1.27E+03 (7.63E+01)
5.33E+02 =
(1.56E+01)
4.60E+02 =
(6.84E+00)
3.49E+02 =
(9.11E+00)
6.50E+05 =
(6.32E+04)
13/17/0
SCSSL-SHADE_
EpSin
0.00E+00
(0.00E+00)
9.62E-08
(6.14E-08)
0.00E+00
(0.00E+00)
4.51E+01
(3.97E+01)
1.94E+01
(6.64E+00)
4.20E-08
(6.98E-08)
7.15E+01
(5.69E+00)
1.96E+01
(6.59E+00)
0.00E+00
(0.00E+00)
2.89E+03
(2.90E+02)
2.71E+01
(2.06E+00)
1.36E+03
(3.67E+02)
4.29E+01
(2.23E+01)
2.67E+01
(2.57E+00)
2.39E+01
(2.44E+00)
2.68E+02
(1.16E+02)
2.04E+02
(8.12E+01)
2.46E+01
(2.15E+00)
1.56E+01
(2.97E+00)
1.07E+02
(2.47E+01)
2.20E+02
(6.07E+00)
2.10E+03
(1.46E+03)
4.35E+02
(7.00E+00)
5.08E+02
(4.57E+00)
4.81E+02
(3.52E+00)
1.18E+03
(1.08E+02)
5.28E+02
(1.16E+01)
4.60E+02
(6.84E+00)
3.49E+02
(1.14E+01)
6.72E+05
(8.23E+04)
jSO
0.00E+00 =
(0.00E+00)
1.38E-05 =
(8.23E-06)
0.00E+00 =
(0.00E+00)
5.85E+01 =
(4.56E+01)
1.56E+01 (2.65E+00)
4.10E-07 =
(5.52E-07)
6.66E+01 (3.10E+00)
1.69E+01 (3.43E+00)
0.00E+00 =
(0.00E+00)
3.21E+03 (3.78E+02)
2.66E+01 (3.13E+00)
1.61E+03 (4.42E+02)
3.17E+01 =
(2.01E+01)
2.50E+01 =
(2.34E+00)
2.37E+01 (2.77E+00)
4.77E+02 =
(1.36E+02)
2.93E+02 =
(1.10E+02)
2.46E+01 (2.42E+00)
1.42E+01 (2.73E+00)
1.17E+02 =
(6.45E+01)
2.17E+02 (2.73E+00)
1.07E+03 =
(1.61E+03)
4.30E+02 (6.16E+00)
5.08E+02 =
(4.54E+00)
4.81E+02 (2.32E+00)
1.13E+03 =
(4.90E+01)
5.14E+02 =
(1.01E+01)
4.59E+02 =
(3.03E-13)
3.65E+02 =
(1.52E+01)
6.08E+05 =
(3.03E+04)
12/18/0
SCSSjSO
0.00E+00
(0.00E+00)
1.48E-05
(8.26E-06)
0.00E+00
(0.00E+00)
4.87E+01
(4.11E+01)
1.26E+01
(2.70E+00)
2.85E-07
(5.12E-07)
6.33E+01
(2.66E+00)
1.20E+01
(2.67E+00)
0.00E+00
(0.00E+00)
3.05E+03
(3.63E+02)
2.50E+01
(4.12E+00)
1.29E+03
(3.66E+02)
2.60E+01
(2.09E+01)
2.51E+01
(2.46E+00)
2.12E+01
(1.81E+00)
4.45E+02
(1.55E+02)
2.61E+02
(1.04E+02)
2.24E+01
(1.14E+00)
1.17E+01
(2.65E+00)
1.14E+02
(6.57E+01)
2.14E+02
(3.27E+00)
1.63E+03
(1.79E+03)
4.26E+02
(6.54E+00)
5.07E+02
(3.77E+00)
4.81E+02
(3.15E+00)
1.12E+03
(5.07E+01)
5.10E+02
(1.37E+01)
4.59E+02
(3.32E-13)
3.65E+02
(1.40E+01)
6.04E+05
(2.57E+04)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
TABLE S13
L-SHADE
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
-/=/+
0.00E+00 =
(0.00E+00)
3.16E-04 +
(5.07E-05)
5.47E-06 +
(6.19E-06)
2.01E+02 (7.69E+00)
3.78E+01 (7.64E+00)
1.37E-03 (8.75E-04)
1.51E+02 (4.80E+00)
3.92E+01 (5.48E+00)
1.56E-01 (2.22E-01)
1.14E+04 (6.11E+02)
3.86E+02 (9.53E+01)
2.37E+04 =
(1.05E+04)
1.36E+03 (8.06E+02)
2.55E+02 (3.25E+01)
2.50E+02 =
(4.87E+01)
1.79E+03 (2.58E+02)
1.20E+03 (2.21E+02)
2.15E+02 =
(4.60E+01)
1.77E+02 (2.31E+01)
1.57E+03 (2.42E+02)
2.69E+02 (5.81E+00)
1.19E+04 (5.24E+02)
5.68E+02 =
(7.98E+00)
9.19E+02 (8.98E+00)
7.46E+02 =
(3.47E+01)
3.41E+03 (1.02E+02)
6.58E+02 (1.38E+01)
5.28E+02 =
(2.19E+01)
1.53E+03 =
(1.92E+02)
2.43E+03 (1.45E+02)
20/8/2
28
PERFORMANCE COMPARISONS OF FOUR SCSS-BASED TOP ALGORITHMS WITH THE BASELINES
ON 100-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
SCSSL-SHADE
0.00E+00
(0.00E+00)
3.41E-04
(5.81E-05)
1.07E-03
(1.73E-03)
2.00E+02
(8.00E+00)
2.69E+01
(6.48E+00)
5.37E-04
(4.36E-04)
1.38E+02
(4.48E+00)
2.75E+01
(5.11E+00)
1.42E-02
(6.64E-02)
1.05E+04
(4.67E+02)
1.54E+02
(5.30E+01)
2.25E+04
(8.53E+03)
2.45E+02
(7.34E+01)
1.01E+02
(2.01E+01)
2.59E+02
(4.34E+01)
1.55E+03
(2.39E+02)
1.04E+03
(2.00E+02)
2.11E+02
(5.33E+01)
1.63E+02
(2.46E+01)
1.50E+03
(1.79E+02)
2.59E+02
(4.38E+00)
1.12E+04
(6.26E+02)
5.67E+02
(7.15E+00)
9.12E+02
(8.61E+00)
7.44E+02
(3.50E+01)
3.31E+03
(9.92E+01)
6.47E+02
(1.57E+01)
5.34E+02
(2.30E+01)
1.48E+03
(1.83E+02)
2.34E+03
(1.32E+02)
UMOEA-II
0.00E+00 =
(0.00E+00)
9.66E-05 =
(1.75E-05)
2.84E-06 +
(3.01E-06)
1.87E+02 =
(4.03E+01)
3.53E+01 (7.62E+00)
8.12E-03 (5.54E-03)
1.41E+02 (9.72E+00)
3.60E+01 (7.09E+00)
5.35E-01 (5.13E-01)
1.19E+04 =
(1.25E+03)
4.27E+02 (1.03E+02)
4.52E+03 =
(8.56E+02)
3.60E+02 (1.47E+02)
2.35E+02 (3.25E+01)
2.67E+02 (5.38E+01)
1.67E+03 =
(4.55E+02)
1.36E+03 =
(3.13E+02)
2.35E+02 =
(6.29E+01)
1.76E+02 (2.65E+01)
1.93E+03 =
(3.61E+02)
2.56E+02 =
(6.84E+00)
1.27E+04 =
(1.81E+03)
5.70E+02 =
(9.40E+00)
9.22E+02 (8.89E+00)
7.49E+02 (2.76E+01)
3.42E+03 (9.37E+01)
6.41E+02 (1.79E+01)
5.18E+02 +
(3.80E+01)
1.40E+03 =
(2.46E+02)
2.36E+03 =
(1.26E+02)
14/14/2
SCSSUMOEA-II
L-SHADE_
EpSin
0.00E+00
(0.00E+00)
9.31E-05
(1.35E-05)
6.60E-06
(4.57E-06)
1.93E+02
(3.12E+01)
2.79E+01
(7.14E+00)
2.61E-03
(2.27E-03)
1.36E+02
(9.40E+00)
2.78E+01
(7.23E+00)
9.17E-02
(1.35E-01)
1.13E+04
(1.59E+03)
1.58E+02
(4.12E+01)
4.86E+03
(1.42E+03)
1.64E+02
(4.77E+01)
7.25E+01
(1.56E+01)
2.21E+02
(4.82E+01)
1.64E+03
(4.27E+02)
1.28E+03
(2.62E+02)
2.16E+02
(4.72E+01)
1.52E+02
(2.50E+01)
1.89E+03
(3.11E+02)
2.55E+02
(6.49E+00)
1.25E+04
(1.61E+03)
5.70E+02
(1.34E+01)
9.16E+02
(1.16E+01)
7.29E+02
(3.77E+01)
3.32E+03
(9.49E+01)
6.32E+02
(1.61E+01)
5.28E+02
(3.07E+01)
1.48E+03
(2.33E+02)
2.36E+03
(1.53E+02)
0.00E+00 =
(0.00E+00)
1.58E-04 (4.22E-05)
5.35E-09 (1.11E-08)
2.04E+02 =
(9.79E+00)
6.06E+01 (7.15E+00)
3.51E-05 (1.38E-05)
1.67E+02 (9.13E+00)
5.73E+01 (9.38E+00)
0.00E+00 =
(0.00E+00)
1.05E+04 (5.15E+02)
4.16E+01 =
(2.39E+01)
5.28E+03 (1.39E+03)
7.92E+01 =
(2.87E+01)
5.13E+01 =
(8.93E+00)
7.28E+01 =
(3.14E+01)
1.55E+03 (2.51E+02)
1.16E+03 (1.72E+02)
7.92E+01 =
(2.19E+01)
5.22E+01 =
(6.65E+00)
1.44E+03 (1.96E+02)
2.83E+02 (1.41E+01)
1.08E+04 (5.90E+02)
5.98E+02 (7.21E+00)
9.37E+02 (2.15E+01)
6.93E+02 =
(4.53E+01)
3.24E+03 (2.51E+02)
5.92E+02 =
(1.37E+01)
5.15E+02 =
(1.95E+01)
1.23E+03 =
(1.62E+02)
2.34E+03 =
(1.35E+02)
16/14/0
SCSSL-SHADE_
EpSin
0.00E+00
(0.00E+00)
1.38E-04
(4.25E-05)
2.20E-10
(1.57E-09)
2.05E+02
(1.11E+01)
4.15E+01
(6.26E+00)
9.41E-06
(5.14E-06)
1.45E+02
(5.70E+00)
3.87E+01
(6.26E+00)
0.00E+00
(0.00E+00)
9.57E+03
(4.63E+02)
4.26E+01
(2.91E+01)
4.62E+03
(7.33E+02)
8.36E+01
(3.44E+01)
4.86E+01
(6.46E+00)
7.73E+01
(2.83E+01)
1.31E+03
(2.61E+02)
9.23E+02
(1.76E+02)
7.59E+01
(1.83E+01)
5.09E+01
(5.78E+00)
1.23E+03
(1.89E+02)
2.64E+02
(5.61E+00)
9.54E+03
(5.05E+02)
5.92E+02
(6.32E+00)
9.08E+02
(8.10E+00)
6.89E+02
(4.55E+01)
3.06E+03
(9.06E+01)
5.90E+02
(1.81E+01)
5.22E+02
(2.30E+01)
1.21E+03
(1.42E+02)
2.37E+03
(1.92E+02)
jSO
0.00E+00 =
(0.00E+00)
3.10E-04 +
(5.45E-05)
2.71E-06 +
(2.72E-06)
1.94E+02 =
(2.35E+01)
4.29E+01 (7.17E+00)
1.61E-04 (4.30E-04)
1.41E+02 (6.94E+00)
4.31E+01 (5.58E+00)
4.60E-02 (1.11E-01)
9.71E+03 (6.59E+02)
1.06E+02 (3.82E+01)
2.05E+04 (1.06E+04)
1.60E+02 (4.19E+01)
6.28E+01 (1.18E+01)
1.64E+02 (4.20E+01)
1.84E+03 =
(3.15E+02)
1.26E+03 (2.63E+02)
1.76E+02 (4.05E+01)
1.07E+02 (2.14E+01)
1.38E+03 =
(2.44E+02)
2.64E+02 (6.56E+00)
1.07E+04 (6.27E+02)
5.69E+02 =
(1.37E+01)
9.01E+02 (1.04E+01)
7.18E+02 =
(3.87E+01)
3.20E+03 (8.46E+01)
5.86E+02 (2.05E+01)
5.29E+02 =
(2.78E+01)
1.33E+03 (2.02E+02)
2.31E+03 =
(1.23E+02)
20/8/2
SCSSjSO
0.00E+00
(0.00E+00)
3.66E-04
(6.77E-05)
1.52E-04
(1.69E-04)
1.96E+02
(1.09E+01)
2.84E+01
(5.43E+00)
1.68E-05
(1.18E-05)
1.27E+02
(4.53E+00)
2.99E+01
(5.62E+00)
0.00E+00
(0.00E+00)
9.23E+03
(6.08E+02)
7.21E+01
(3.10E+01)
1.41E+04
(8.02E+03)
1.12E+02
(2.79E+01)
3.95E+01
(4.08E+00)
9.73E+01
(3.56E+01)
1.74E+03
(2.99E+02)
1.13E+03
(2.20E+02)
1.11E+02
(3.07E+01)
5.22E+01
(5.72E+00)
1.29E+03
(2.12E+02)
2.49E+02
(5.18E+00)
1.01E+04
(6.70E+02)
5.67E+02
(1.14E+01)
8.96E+02
(7.84E+00)
7.13E+02
(4.26E+01)
3.12E+03
(9.03E+01)
5.77E+02
(2.28E+01)
5.25E+02
(2.86E+01)
1.25E+03
(1.82E+02)
2.27E+03
(1.06E+02)
> arXiv preprint, arXiv:1712.06338 Source code available at https://shengxinzhang.github.io/
29
TABLE S14 PERFORMANCE COMPARISONS OF SCSS-CODE WITH CODE
ON 30-D AND 50-D CEC2017 BENCHMARK SET OVER 51 INDEPENDENT RUNS
D = 30
CoDE
SCSSCoDE
1.19E-06
(1.15E-06)
4.69E-04
(2.04E-04)
3.76E-06
(5.86E-06)
5.80E+01
(5.64E+00)
1.07E+02
(8.96E+00)
9.26E-06
(4.18E-06)
1.59E+02
(1.19E+01)
1.08E+02
(8.81E+00)
4.95E+01
(2.87E+01)
4.20E+03
(3.00E+02)
4.76E+01
(2.09E+01)
1.41E+04
(1.52E+04)
7.89E+01
(1.48E+01)
2.94E+01
(1.43E+01)
2.32E+01
(7.48E+00)
4.52E+02
(1.48E+02)
6.76E+01
(1.47E+01)
3.29E+01
(8.95E+00)
2.03E+01
(4.11E+00)
6.08E+01
(5.91E+01)
3.12E+02
(8.96E+00)
1.00E+02
(6.48E-11)
4.45E+02
(1.17E+01)
5.32E+02
(1.48E+01)
3.87E+02
(3.93E-02)
1.99E+03
(2.67E+02)
4.93E+02
(1.25E+01)
3.06E+02
(2.54E+01)
5.92E+02
(5.77E+01)
2.02E+03
(6.14E+01)
D = 50
CoDE
SCSSCoDE
6.99E+02
(1.22E+03)
1.98E-02
(2.14E-02)
9.35E-03
(3.16E-02)
7.52E+01
(5.06E+01)
2.93E+02
(1.65E+01)
4.92E-03
(1.27E-03)
3.91E+02
(1.74E+01)
2.91E+02
(1.36E+01)
1.41E+03
(4.20E+02)
9.32E+03
(4.30E+02)
1.22E+02
(1.49E+01)
1.11E+05
(6.15E+04)
6.31E+02
(6.44E+02)
8.88E+01
(2.31E+01)
9.95E+01
(1.88E+01)
1.19E+03
(2.57E+02)
8.15E+02
(1.56E+02)
2.01E+03
(2.27E+03)
5.27E+01
(9.20E+00)
5.20E+02
(1.55E+02)
4.93E+02
(1.57E+01)
9.46E+03
(1.95E+03)
7.11E+02
(1.81E+01)
8.13E+02
(2.05E+01)
4.81E+02
(2.80E+00)
3.98E+03
(1.75E+02)
5.16E+02
(1.32E+01)
4.60E+02
(6.84E+00)
8.34E+02
(1.21E+02)
5.86E+05
(2.02E+04)
3.94E-05 2.22E+03 (2.93E-05)
(2.62E+03)
4.44E-04 =
1.67E-01 F2
(1.76E-04)
(2.20E-01)
6.33E-06 =
7.93E-03 =
F3
(1.59E-05)
(2.18E-02)
5.89E+01 7.67E+01 =
F4
(1.32E+00)
(5.50E+01)
1.17E+02 3.10E+02 F5
(8.30E+00)
(1.60E+01)
6.72E-06 +
3.41E-03 +
F6
(4.14E-06)
(8.91E-04)
1.70E+02 4.05E+02 F7
(1.24E+01)
(1.70E+01)
1.23E+02 3.10E+02 F8
(8.54E+00)
(1.37E+01)
1.07E+02 1.98E+03 F9
(6.69E+01)
(5.34E+02)
4.59E+03 9.75E+03 F10
(2.80E+02)
(3.75E+02)
6.88E+01 1.47E+02 F11
(2.04E+01)
(1.74E+01)
1.18E+04 =
9.97E+04 =
F12
(8.03E+03)
(3.86E+04)
8.63E+01 6.74E+02 =
F13
(1.16E+01)
(7.32E+02)
4.27E+01 1.05E+02 F14
(1.35E+01)
(2.07E+01)
3.20E+01 1.14E+02 F15
(7.91E+00)
(1.56E+01)
4.81E+02 =
1.38E+03 F16
(1.75E+02)
(2.61E+02)
8.38E+01 8.77E+02 F17
(2.79E+01)
(1.72E+02)
3.63E+01 =
1.29E+03 +
F18
(1.23E+01)
(8.91E+02)
2.34E+01 6.10E+01 F19
(2.94E+00)
(9.09E+00)
7.52E+01 6.15E+02 F20
(6.08E+01)
(1.48E+02)
3.23E+02 5.14E+02 F21
(7.97E+00)
(1.49E+01)
1.00E+02 1.03E+04 F22
(1.87E-10)
(3.63E+02)
4.60E+02 7.33E+02 F23
(8.35E+00)
(1.70E+01)
5.52E+02 8.39E+02 F24
(1.33E+01)
(2.30E+01)
3.87E+02 4.81E+02 F25
(3.05E-02)
(2.76E+00)
2.10E+03 4.18E+03 F26
(2.77E+02)
(1.40E+02)
5.01E+02 5.35E+02 F27
(1.00E+01)
(1.24E+01)
3.10E+02 4.60E+02 F28
(3.17E+01)
(6.84E+00)
6.37E+02 9.55E+02 F29
(6.78E+01)
(1.51E+02)
2.02E+03 =
5.83E+05 F30
(8.17E+01)
(1.11E+04)
-/=/+
23/6/1
24/4/2
Note: In our CoDE and SCSS-CoDE implementations, following the classic design of DE, for each mutation strategy involved, the indices are mutually
different (restrained condition), this differs from the source code provided by Dr. Y. Wang, in which for the “current-to-rand/1” strategy, the indices could be
the same. We empirically found that when both with the restrained condition, SCSS-CoDE significantly outperforms CoDE. But when the restrained
condition is removed in both, SCSS-CoDE performs similarly to CoDE with the investigated M value and SS rule.
F1
| 9cs.NE
|
1
Geometric Properties of Isostables and Basins of
Attraction of Monotone Systems
arXiv:1705.02853v1 [cs.SY] 8 May 2017
Aivar Sootla and Alexandre Mauroy
Abstract—In this paper, we study geometric properties of
basins of attraction of monotone systems. Our results are based
on a combination of monotone systems theory and spectral
operator theory. We exploit the framework of the Koopman
operator, which provides a linear infinite-dimensional description
of nonlinear dynamical systems and spectral operator-theoretic
notions such as eigenvalues and eigenfunctions. The sublevel sets
of the dominant eigenfunction form a family of nested forwardinvariant sets and the basin of attraction is the largest of these
sets. The boundaries of these sets, called isostables, allow studying
temporal properties of the system. Our first observation is that
the dominant eigenfunction is increasing in every variable in the
case of monotone systems. This is a strong geometric property
which simplifies the computation of isostables. We also show
how variations in basins of attraction can be bounded under
parametric uncertainty in the vector field of monotone systems.
Finally, we study the properties of the parameter set for which
a monotone system is multistable. Our results are illustrated on
several systems of two to four dimensions.
Index Terms—Monotone Systems, Koopman Operator, Computation of Isostables, Computation of Basins of Attraction,
Genetic Toggle Switch
I. I NTRODUCTION
In many applications, such as economics and biology, the
states of linear dynamical systems take only nonnegative
values. These systems are called positive and have received
considerable attention in the context of systems theory [2],
[3], model reduction [4], [5], distributed control [6], [7],
etc. One of the main tools to study such systems is the
Perron-Frobenius theorem (see e.g. [8]), which describes some
spectral properties of the drift matrix in a linear positive
system. In the nonlinear setting, positive systems have a couple
of generalizations, one of which is the class of cooperative
monotone systems (see e.g. [9]). Similarly to the linear case,
cooperative monotone systems generate trajectories (or flows)
which are increasing functions in every argument with respect
to the initial state and for every time. With a slight abuse of
terminology, we will refer to cooperative monotone systems
simply as monotone. Over the years there were a number of
Aivar Sootla is with the Department of Engineering Science,
University of Oxford, Parks Road, Oxford, OX1 3PJ, UK
[email protected].
Alexandre Mauroy is with Namur Center for Complex Systems (naXys)
and Department of Mathematics, University of Namur, B-5000, Belgium
[email protected]
The authors would like to thank Prof Zhao for explaining some aspects
of the results in [1]. Most of this work was performed, while Aivar Sootla
and Alexandre Mauroy were with Montefiore Institute, Liège University and
were supported by F.R.S.– FNRS postdoctoral grant and a BELSPO Return
Grant, respectively. Currently, Aivar Sootla is supported by the EPSRC Grant
EP/M002454/1.
developments in monotone systems theory [10], [11] as well
as applications such as finance [12], energy networks [13],
ventilation systems [14], biology [15]–[17], etc.
The Koopman operator (see e.g. [18]) provides a framework
that allows to define and study spectral properties of dynamical
systems in the basins of attraction of their hyperbolic attractors. The operator has eigenvalues and eigenfunctions (i.e.,
infinite dimensional eigenvectors), which are directly related to
the geometric properties of the system. For instance, the level
sets of the dominant eigenfunction (that is, the eigenfunction
corresponding to the eigenvalue with the maximal real part)
are called isostables [19] and contain the initial conditions
of trajectories that converge synchronously toward the fixed
point. In addition, the interior of the sublevel set at infinity
is the basin of attraction of the fixed point. Hence isostables
serve as a convenient refinement of the basins of attraction and
add further details to the geometric description of the system.
In [9], it was mentioned that the flow of a monotone system
can be seen as a positive operator. Hence the authors argued
that an operator version of the Perron-Frobenius theorem,
which is called the Krein-Rutman theorem, can be applied.
However, the investigation into spectral properties of these
operators lacked, probably since spectral theory of such operators was not well-developed. This gap can be filled by
the Koopman operator framework, which may pave the way
to formulate a version of the Perron-Frobenius theorem for
monotone systems admitting a stable hyperbolic fixed point.
In this paper, we first provide a spectral characterization
of a so-called maximal Lyapunov function, which is used to
compute basins of attraction in [20]. To do so we use the
eigenfunctions of the Koopman operator. In the case of monotone systems, we show that the maximal Lyapunov functions
can be constructed with only one eigenfunction under some
mild assumptions. We proceed by studying the properties of
the isostables, which we connect to the properties of basins of
attraction. Basins of attraction have been extensively studied
in the case of monotone systems [1], [21], [22]. In [23],
we showed that the isostables of monotone systems have
properties similar to the boundaries of basins of attraction.
In this paper, we expand these arguments using properties
of general increasing functions and order-convexity. Orderconvexity is a strong geometric property that is well-suited
to describe the behavior of monotone systems.
We proceed by studying systems with two asymptotically
stable fixed points (i.e. bistable systems). We consider a class
of bistable (not necessarily monotone) systems, whose vector
fields can be bounded from below and above by the vector
fields of two bistable monotone systems. For this class of
2
systems, basins of attraction can be estimated using the basins
of these monotone bounding systems. We note that the idea
of bounding a system with monotone ones is not novel and
appears in many works (see e.g., [11], [24]). This approach
is then extended to estimate basins of attraction of monotone
systems under parametric uncertainty. A preliminary study on
estimating basins of attraction under parameter uncertainty
was performed in [23], and we generalize it in this paper by
providing easy to verify assumptions. Furthermore, we study
the properties of the parameter set for which a monotone
system is (at least) bistable. We illustrate our theoretical
findings with several numerical examples.
Our theoretical results are complemented by a discussion
on methods for computing inner and outer estimates on basins
of attraction. We cover methods based on linear algebra [25]
and sum-of-squares programming [26], [27], and their relation
to the Koopman operator framework. However, in the case of
monotone systems, we propose to exploit a different technique.
According to properties of monotone systems, we can build
inner and outer approximations of the basin of attraction by
computing flows starting from a finite number of points. This
allows us to derive data-sampled algorithms. We discuss two
conceptually similar algorithms exploiting this idea [23], [28].
We also exploit some results and techniques from [29],
which considers the problem of switching between the stable
fixed points of a bistable monotone system by using a pulse
control signal. In that work, the authors introduced the concept
of switching set, which is reminiscent of basins of attraction,
but in the space of control parameters. In [30], the concept of
eigenfunctions was extended to the pulse control problem.
The rest of the paper is organized as follows. In Section II,
we introduce the main properties of the Koopman operator and
monotone systems. In Section III we present the properties of
order-convex sets, which provide basic but strong topological
tools for monotone systems. In Subsection IV-A we discuss
spectral properties of monotone systems and build (maximal)
Lyapunov functions using eigenfunctions of the Koopman
operator. We study geometric properties and investigate the
behavior of basins of attraction of monotone systems under
parameter variations in Subsections IV-B and IV-C, respectively. We discuss methods to compute isostables and basins
of attraction in Section V. We provide numerical examples in
Section VI and conclude in Section VII.
depend on parameters, we will simply omit the notation p (e.g.
f (x), J(x)). If x∗ is a stable fixed point of f (x), we assume
that the eigenvectors of J(x∗ ) are linearly independent (i.e.,
J(x∗ ) is diagonalizable). We denote the eigenvalues of J(x∗ )
by λi for i =P
1, . . . , k and assume that they have multiplicities
k
µi such that i=1 µi = n.
Koopman Operator. We limit our study of the Koopman
operator to a basin of attraction B(x∗ ) of an attractive fixed
point x∗ .
Definition 1: The basin of attraction B(x∗ ) of an attractive
fixed point x∗ for the system ẋ = f (x) is the set of initial
∗
conditionsn x such that the flow φf (t, x)
o converges to x , i.e.
∗
∗
n
B(x ) = x ∈ R lim φf (t, x) = x .
t→∞
∗
When the fixed point x can be understood from the
context, we write simply B. We will assume that x∗ is a
stable hyperbolic fixed point, that is, the eigenvalues λj of
the Jacobian matrix J(x∗ ) are such that <(λj ) < 0 for all j.
Definition 2: The Koopman semigroup of operators associated with ẋ = f (x) is defined as
U t g(x) = g(φf (t, x)),
where the functions g : Rn → C are called observables, x ∈
B(x∗ ) and φf (t, x) is a solution to ẋ = f (x).
The Koopman semigroup is linear [31], so that it is natural
to study its spectral properties. Consider the system with f ∈
C 2 on the basin of attraction B(x∗ ) of a stable hyperbolic fixed
point. We define the Koopman eigenfunctions as a nontrivial
function sj satisfying U t s(x) = s(φf (t, x)) = s(x) eλt , where
eλt belongs to the point spectrum of U t and we refer to such
λ ∈ C as Koopman eigenvalues. In particular, the eigenvalues
λj of the Jacobian matrix J(x∗ ) are Koopman eigenvalues
under the assumptions above. Furthermore, it can be shown
that there exist n eigenfunctions sj ∈ C 1 (B) associated with
eigenvalues λj (see e.g. [25]), and:
(f (x))T ∇sj (x) = λj sj (x),
Throughout the paper we consider parameter-dependent
systems of the form
ẋ = f (x, p),
x(0) = x0 ,
(1)
with f : D × P → Rn , D ⊂ Rn , and P ⊂ Rm for some
integers n and m. We define the flow map φf : R × D × P →
Rn , where φf (t, x0 , p) is a solution to the system (1) with
the initial condition x0 and the parameter p. We assume
that f (x, p) is continuous in (x, p) on D × P and twice
continuously differentiable in x for every fixed p, unless it is
stated otherwise. We denote the Jacobian matrix of f (x, p) as
J(x, p) for every p. When we consider systems which do not
(3)
If f is analytic and if the eigenvalues λj are simple (i.e.,
µj = 1 for all j), then the flow of the system can be expressed
(at least locally) through the following expansion (cf. [19]):
φf (t, x) = x∗ +
n
X
sj (x)vj eλj t +
(4)
j=1
X
II. P RELIMINARIES
(2)
vk1 ,...,kn sk11 (x) · · · sknn (x)e(k1 λ1 +...kn λl )t ,
k1 ,...,kn ∈N0
k1 +···+kn >1
where N0 is the set of nonnegative integers, vj are the
right eigenvectors of J(x∗ ) corresponding to λj , the vectors
vk1 ,...,kn are the so-called Koopman modes (see [25], [31]
for more details). We also note that it is implicitly assumed
in (4) that viT ∇sj (x∗ ) = 0 for all i 6= j and viT ∇si (x∗ ) = 1.
In the case of a linear system ẋ = Ax where the matrix
A has the left eigenvectors wi , the eigenfunctions si (x) are
equal to wiT x and the expansion (4) has only the finite sum
(i.e., vk1 ,...,kn = 0). A similar (but lengthy) expansion can be
obtained if the eigenvalues λj are not simple and have linearly
dependent eigenvectors (see e.g., [32]).
3
Let λj be such that 0 > <(λ1 ) > <(λj ), j 6= 1, then the
eigenfunction s1 , which we call dominant, can be computed
using the so-called Laplace average [19]:
Zt
1
∗
gλ (x) = lim
(g ◦ φf (s, x))e−λs ds.
(5)
t→∞ t
0
For all g ∈ C that satisfy g(x∗ ) = 0 and v1T ∇g(x∗ ) 6= 0,
the Laplace average gλ∗1 is equal to s1 up to a multiplication
with a scalar. Note that we do not require the knowledge of
B(x∗ ) in order to compute s1 , since the limit in (5) does not
converge to a finite value for x 6∈ B(x∗ ). The eigenfunctions
sj with j ≥ 2 (non-dominant eigenfunctions) are generally
harder to compute and are not considered in the present study.
The eigenfunction s1 captures the asymptotic behavior of
the system. In order to support this statement we consider the
following definition.
Definition 3: Let s1 be a C 1 (B) eigenfunction corresponding to λ1 such that 0 > <(λ1 ) > <(λj ) for j ≥ 2. The isostables ∂Bα for α ≥ 0 are boundaries of the sublevel sets Bα =
{x ∈ B||s1 (x)| ≤ α}, that is, ∂Bα = {x ∈ B||s1 (x)| = α}.
In the case of a linear system ẋ = Ax, the isostables are the
level sets of |s1 (x)| = |w1T x|, where w1 is the left eigenvector
of A associated with the dominant eigenvalue λ1 . If λ1 is real,
then s1 is real and we
n
o will use the
n following notationo∂+ Bα =
x ∈ B s1 (x) = α , ∂− Bα = x ∈ B s1 (x) = −α for α ≥
0. Furthermore, when λ1 is real and simple, it follows from (4)
that the trajectories starting from the isostable ∂Bα share the
same asymptotic evolution φf (t, x) → x∗ + v1 αeλ1 t with
t → ∞. This implies that the isostables contain the initial
conditions of trajectories that converge synchronously toward
the fixed point. In particular, trajectories starting from the same
isostable ∂Bα1 reach other isostables ∂Bα2 (with α2 < α1 )
after a time T = ln (α1 /α2 ) /|<(λ1 )|. If λ1 is not simple, then
the isostables are not unique. However, we will choose specific
isostables in the case of monotone systems (see Section IV).
It can be shown that Bα = B as α → ∞, so that we
will also use the notations B∞ and ∂B∞ to denote the basin
of attraction and its boundary, respectively. More information
about the isostables and their general definition using the flows
of the system can be found in [19].
Using the same tools as in [25], it can be shown (provided
that the eigenvectors of J(x∗ ) are linearly independent) that
n
1/p
P
p
the function Wβ (x) =
βi |si (x)|
, with p ≥ 1
1
i=1
and βi > 0, is a Lyapunov function, that is Wβ ∈ C 1 (B),
Wβ (x∗ ) = 0, Wβ (x) > 0, and Ẇβ (x) < 0 for all x ∈ B\x∗ .
The properties Wβ (x∗ ) = 0 and Wβ (x) > 0 for x 6= x∗ stem
from the fact that the zero level sets of si ’s intersect only in
the fixed point (see [25]). On the other hand, one can show
that Ẇβ ≤ −<{λ1 }kxkp by direct computation and definition
of si ’s. In [25], it was also discussed that Wβ can be used
to estimate a basin of attraction of the system if all si can be
computed, since the function becomes infinite on the boundary
∂B of the basin of attraction. This property is reminiscent of
the definition of the maximal Lyapunov function [20].
Definition 4: A function Vm : Rn → R ∪ {+∞} is
called a maximal Lyapunov function for the system ẋ = f (x)
admitting an asymptotically stable fixed point x∗ with a basin
of attraction B, if
1) Vm (x∗ ) = 0, Vm (x) > 0 for all x ∈ B\x∗ ;
2) Vm (x) < ∞ if and only if x ∈ B;
3) Vm (x) → ∞ as x → ∂B and/or kxk → +∞;
4) V̇m is well defined for all x ∈ B, and V̇m (x) is negative
definite for all x ∈ B\x∗ .
In particular, it is straightforward to show that
(
Wβ (x) x ∈ B
V (x) =
∞
otherwise
is a maximal Lyapunov function and V ∈ C 1 (B) provided that
the fixed point x∗ is stable and hyperbolic, and f ∈ C 2 (B).
Partial Orders and Monotone Systems. We define a partial
order as follows: x y if and only if x − y ∈ Rn≥0 . In other
words, x y means that x is larger or equal to y entrywise.
We write x 6 y if the relation x y does not hold. We
will also write x y if x y and x 6= y, and x y if
x − y ∈ Rn>0 . Note that cones K more general than Rn>0 can
also be used to define partial orders [33], but unless stated
otherwise we will consider K = Rn≥0 . Systems whose flows
preserve a partial order relation are called monotone systems.
Definition 5: The system is monotone on D × P if
φf (t, x, p) φf (t, y, q) for all t ≥ 0 and for all x y, p q,
where x, y ∈ D, p, q ∈ P. The system is strongly monotone
on D × P, if it is monotone and if φf (t, x, p) φf (t, y, q)
holds for all t > 0 provided that x y, p q, and either
x ≺ y or p ≺ q holds, where x, y ∈ D and p, q ∈ P.
A certificate for monotonicity is given by the Kamke-Müller
conditions (see e.g [33]). The certificate amounts to checking
the sign pattern of the matrices ∂f (x, p)/∂x and ∂f (x, p)/∂p.
We will also consider the comparison principle [9], which is
typically used to extend some properties of monotone systems
to a class of non-monotone ones.
Proposition 1: Consider the dynamical systems ẋ = f (x)
and ẋ = g(x). Let one of the two systems be monotone on
D. If g(x) f (x) for all x ∈ D then φg (t, x2 ) φf (t, x1 )
for all t ≥ 0 and for all x2 x1 .
III. I NCREASING F UNCTIONS AND O RDER -C ONVEX S ETS
Our subsequent derivations are based on the properties of
the increasing functions in Rn , which were studied in the
context of partial orders, for example, in [34].
Definition 6: We call the set dom(g) = {x ∈ Rn ||g(x)| <
∞} the effective domain of a function g. A function g : Rn →
R ∪ {−∞, +∞} is called increasing with respect to the cone
Rn≥0 if g(x) ≥ g(z) for all x z and x, z ∈ dom(g).
In this
section,
we
study
the
properties
of
the
sublevel
sets
n
o
Aα = x ∈ Rn |g(x)| ≤ α of increasing functions continuous on their effective domain dom(g). We first introduce a
few concepts. Let open and closed intervals induced by the
cone Rn≥0 be defined as [[x, y]] = {z ∈ Rn |x z y} and
[x, y] = {z ∈ Rn |x z y}, respectively.
Definition 7: A set A is called order-convex if, for all x,
y ∈ A, the closed interval [x, y] is a subset of A.
For an order-convex set A, the set of maximal elements
∂+ A (respectively, the set of minimal elements ∂− A) of A
4
is a subset of the boundary ∂A of A such that if y z for
some y ∈ ∂+ A (resp., if y z for some y ∈ ∂− A), then
z 6∈ A. It follows from the definition that for all y, z ∈ ∂+ A
(or y, z ∈ ∂− A), we cannot have y z or z y. We have the
following proposition, which is similar to results from [35].
Proposition 2: Let A ∈ Rn be order-convex. Then
(i) the boundary ∂A of A is the union of ∂+ A and ∂− A;
(ii) the interior of the set A is the union of open intervals
[[x, y]] over all x ∈ ∂− A, y ∈ ∂+ A:
[
int(A) =
[[x, y]].
(6)
x∈∂− A, y∈∂+ A
Proof: (i) It follows from their definition that the sets ∂− A
and ∂+ A are the subsets of the boundary ∂A, if they are not
empty. Hence, we only need to show that ∂A ⊆ ∂− A ∪ ∂+ A.
If for z ∈ ∂A there exist x ∈ ∂− A and y ∈ ∂+ A such that
z ∈ [[x, y]], then z ∈ int(A). If there exists z ∈ ∂A such that
there exists no x ∈ A with x z or x z, then z itself is
the maximal or the minimal element of A.
(ii) If for some point z in A there does not exist a minimal
element x ∈ ∂− A and a maximal element y ∈ ∂+ A such that
x z y, then z ∈ ∂A and z cannot belong to the interior
of A. Hence for all z ∈ int(A), there exist x ∈ ∂− A and
y ∈ ∂+ A such that z ∈ [[x, y]], which proves the claim.
Now we discuss the connection between order-convex sets
and connected sets. Recall that the set A ⊂ Rn is called
connected if for any two points x, y ∈ A, there exists a path
γ(t) (i.e., a continuous curve γ : [0, 1] → Rn ) with γ(0) = x,
γ(1) = y, and such that γ(t) ∈ A for all t ∈ [0, 1]. The set
is called simply connected if it is connected and if every path
between x, y ∈ A can be continuously transformed, staying
within A, into any other such path while preserving the endpoints. Since a union of sets can be disconnected, Proposition 2
does not imply that order-convex sets are simply connected
or even connected. However, order-convex
and connectedosets
n
are related to sublevel sets Aα = x ∈ Rn |g(x)| ≤ α of
increasing functions.
Proposition 3: Let g : D → R be a continuous function,
where D is an open order-convex set. Then:
(i) the function g is increasing with respect to Rn≥0 if
and only if the sublevel sets Aα ⊆ D are order-convex and
connected for any α ≥ 0;
(ii) if D ⊆ R2 and if the function g is increasing with
respect to R2≥0 , then the sets Aα ⊆ D are simply connected
for any α ≥ 0.
Proof: (i) Sufficiency. Connectivity of Aα directly follows
from the continuity of g (see e.g. Proposition 1 in [36]). Let x,
y ∈ Aα for some α ≥ 0 and z be such that x z y. Since
D is order-convex, z ∈ D so that the value g(z) is defined
and we have that g(x) ≤ g(z) ≤ g(y). Clearly, if |g(x)| ≤ α
and |g(y)| ≤ α, then |g(z)| ≤ α and z ∈ Aα . This implies
that the set Aα is order-convex.
Necessity. Assume that x ≺ y with x, y ∈ D and g(x) >
g(y). Consider the different cases:
1) Case g(y) < 0 and |g(x)| < |g(y)|. We have A|g(x)| ⊂
A|g(y)| , so that x ∈ A|g(y)| and y ∈ ∂− A|g(y)| . Since
x ∈ A|g(y)| and y ∈ ∂− A|g(y)| , we cannot have x ≺ y
and we arrive at a contradiction.
2) Case g(y) < 0 and |g(x)| > |g(y)|. We have A|g(x)| ⊃
A|g(y)| , so that y ∈ A|g(x)| . This implies g(x) > 0 and
x ∈ ∂+ A|g(x)| . Since y ∈ A|g(x)| and x ∈ ∂+ A|g(x)| ,
we cannot have x ≺ y and we arrive at a contradiction.
3) Case g(y) ≥ 0. Since g(x) > g(y) ≥ 0, this case is
treated in the same way as the case (2).
It follows that x y implies g(x) ≤ g(y).
(ii) Consider an arbitrary closed curve η : [0, 1] → Rn in
Aα and the set Aη enclosed by the curve η. For all z ∈ Aη ,
there exist points y 1 ∈ η ⊂ Aα and y 2 ∈ η ⊂ Aα such
that y 1 ≺ z ≺ y 2 . Then order-convexity of Aα implies that
z ∈ Aα and it follows that Aη ⊆ Aα . Since the curve η is
in R2 we can shrink this curve continuously to a point which
belongs to the set Aα . Since the curve is an arbitrary closed
curve in Aα , the set Aα is simply connected.
Proposition 3 also holds for functions g : Rn → R ∪
{−∞, +∞} continuous on their effective domain dom(g). In
this case, the function g is increasing if and only if dom(g) and
Aα are connected and order-convex. It also stands to reason
that point (ii) may potentially be extended to the case of Rn
due to the structure of order-convex sets. However, such an
extension has proved to be cumbersome, therefore, we leave
it outside the scope of this paper.
IV. G EOMETRIC P ROPERTIES OF M ONOTONE SYSTEMS
A. Spectral Properties and Lyapunov Functions
We first establish the spectral properties of the Koopman
operator associated with monotone systems.
Proposition 4: Consider the system ẋ = f (x) with a stable
hyperbolic fixed point x∗ with a basin of attraction B. Assume
that <(λ1 ) > <(λj ) for all j ≥ 2. Let v1 be a right eigenvector
of the Jacobian matrix J(x∗ ) and let s1 ∈ C 1 (B) be an
eigenfunction corresponding to λ1 (with v1T ∇s1 (x∗ ) = 1).
(i) if the system is monotone on B, then λ1 is real and
negative, and there exist s1 and v1 such that s1 (x) ≥ s1 (y)
for all x, y ∈ B satisfying x y, and v1 0;
(ii) if the system is strongly monotone on B, then λ1
is simple (i.e., its multiplicity µ1 = 1), real and negative,
s1 (x) > s1 (y) for all x, y ∈ B satisfying x y, v1 0.
The proof of Proposition 4 is almost identical to the proof
of a similar result in [37], and hence it is omitted. In both
cases, the conditions on λ1 , v1 and s1 are only necessary
and not sufficient for monotonicity, which is consistent with
the linear case and necessary conditions for positivity. We
note that s1 ∈ C 1 (B), if for example f ∈ C 2 . We note
that since eigenfunctions define isostables, this proposition
plays a crucial role in the derivation of geometric properties
of monotone systems. Additionally, a maximal Lyapunov
function can be constructed by using s1 .
Proposition 5: Assume that the system ẋ = f (x) is
monotone on the basin of attraction B of a stable hyperbolic
fixed point x∗ . If ∂B0 intersects x∗ + Rn≥0 only in x∗ , where
+ is the Minkowski addition, then
T s1 is a Lyapunov function
of ẋ = f (x) on X = x∗ + Rn≥0 B. Moreover, the function
V : X → R≥0 , equal to s1 on X, can be extended to Rn so
that this extension is a maximal Lyapunov function.
5
Proof: It is clear that the set x∗ +Rn≥0 is forward-invariant
for monotone systems. Moreover, according to (3), we have
ds1 (x)
= (f (x))T ∇s1 (x) = λ1 s1 (x),
dt
where λ1 is real and negative and s1 (x) is equal to zero only
for x = x∗ on x∗ + Rn≥0 . Hence s1 is a Lyapunov function
on X. The second part of the statement is straightforward.
The assumption on ∂B0 is needed, since for some systems
the isostable ∂B0 can intersect x∗ + Rn≥0 along one of the
axes. For planar systems, this may occur if, for example,
∂f1 (x)/∂x2 = 0 on x∗ + R2≥0 ∩ B (that is, the Jacobian
matrix is reducible on x∗ + R2≥0 ∩ B). On the other hand,
∂B0 ∩ x∗ + Rn≥0 = {x∗ } if the system is strongly monotone.
B. Geometry of Basins of Attraction and Isostables
In this section, we study the properties of Bα , including
the limit case α = ∞, where some of the properties of
eigenfunctions may not longer hold. First, we recall the
following result in monotone systems theory. A proof can be
found for example in [29].
Proposition 6: Let the system ẋ = f (x) be monotone on
the basin of attraction B(x∗ ) of an asymptotically stable fixed
point x∗ , then B(x∗ ) is order-convex.
Let s1 be an eigenfunction corresponding to λ1 , which is
increasing on B(x∗ ). Since B(x∗ ) is order-convex and connected, we can extend s1 to Rn by assigning ∞ on Rn \B(x∗ ).
With a slight abuse of notation, we denote the extended
function by s1 as well. In general, for an eigenvalue λ1 with
a multiplicity µ1 , there exist µ1 eigenfunctions corresponding
to λ1 . Therefore, without loss of generality, we define the
isostables as level sets of an increasing eigenfunction s1
corresponding to λ1 .
The role of order-convexity and topology induced by orderintervals has been studied in the literature. Most of the results
were shown for the semiflow case (e.g., [1]). We expand these
results to characterize the geometric properties of sublevel sets
of the eigenfunction s1 as follows:
Proposition 7: Let the system (1) have a stable hyperbolic
fixed point x∗ with a domain of attraction B(x∗ ). If the system
is monotone on B(x∗ ) with s1 ∈ C 1 (B(x∗ )), then
(i) the set Bα ⊆ B(x∗ ) is order-convex and connected for
any nonnegative α including α = ∞;
(ii) the boundary ∂Bα of Bα can be split into two manifolds:
the set of minimal elements equal to ∂− Bα and the set of
maximal elements equal to ∂+ Bα . Moreover, the manifolds
∂− Bα , ∂+ Bα do not contain points x, y such that x y for
any nonnegative α including α = ∞;
(iii) if the system is strongly monotone, then the manifolds
∂− Bα , ∂+ Bα do not contain points x, y such that x y for
any finite nonnegative α.
Proof: (i) By Proposition 4, the eigenfunction s1 is increasing, while its effective domain dom(s1 ) = B(x∗ ) is
order-convex by Proposition 6. Hence, the result follows from
Proposition 3.
(ii) We have shown in the first point that the set Bα is
order-convex for any nonnegative α including +∞. Hence, the
first statement follows by Proposition 2. The second statement
follows by definition of the set of minimal (maximal) elements
and the fact that ∂Bα = ∂− Bα ∪ ∂+ Bα .
(iii) Let there exist x, y ∈ ∂− Bα such that x y. Since
x, y ∈ ∂− Bα , we have that s1 (x) = s1 (y), but according to
Proposition 4, x y implies that s1 (x) > s1 (y). Hence no
such x and y exist. Similarly, the case of ∂+ Bα is shown.
It is important to note that the boundary ∂B can contain
two points x, y such that x y. But in this case these points
belong to different manifolds ∂− Bα , ∂+ Bα .
As we have shown above, some of the geometric properties
of Bα are preserved in the limiting case α = ∞. However,
the third point of Proposition 7 is shown only for the case of
a finite α. This is due to the fact that the notion of orderconvexity does not fully capture the properties of strictly
increasing functions such as the dominant eigenfunction s1
of a strongly monotone system. We discuss this issue under
additional assumptions. We will consider the case of a bistable
monotone system, which allows deriving many geometric
properties of monotone systems. We make the following
assumptions, which we also use in the sequel.
A1. Let the system ẋ = f (x) have two asymptotically stable
fixed points in Df , denoted as x∗ and x• , and let Df =
cl(B(x∗ ) ∪ B(x• ));
A2. Let the fixed points be such that x• x∗ .
The following proposition is a direct corollary of the results
in [1], where the basins of attraction of semiflows were
studied. We prove it for completeness, in order to discuss how
assumptions in [1] translate into our simplified case.
Proposition 8: Let the system satisfy Assumptions A1–A2
and be strongly monotone on Df . Assume
also that for any
S
bounded set A ∈ Df the set O(A) = t≥0 φ(t, A) is bounded.
Then the boundary between the basins B(x∗ ) and B(x• ) does
not contain two points x, y such that x y.
Proof: This result is a corollary of Theorem 2.2 in [1],
hence we need to make sure that all assumptions are satisfied.
Assumption (A1) in [1] states that the semiflow should be an
α-contraction, where α(·) is a Kuratowski measure of noncompactness. Since our flow is in Rn , α(B) for any bounded
set is equal to zero. The strongly order preserving (SOP)
property and strong monotonicity in our case are equivalent.
Furthermore, the operator ∂φ(t, x), which is the fundamental
˙ = J(φ(t, x))δx, is strongly positive for strongly
solution of δx
monotone systems. Finally we complete the proof by applying
Theorem 2.2 in [1].
The assumption on boundedness of O(A) is technical and
is typically made to avoid pathological cases in monotone
systems theory. Hence the only assumption to check is strong
monotonicity, which is valid if the system is monotone and
the Jacobian is irreducible for all x ∈ Df . Proposition 8 offers
a strong theoretical result, however, its practical implications
is limited in comparison with Proposition 7. Even if we
establish that the system is strongly monotone it does not offer
any direct computational advantage in comparison with the
monotone –but not strongly monotone– case.
C. Basins of Attraction of Bistable Systems
In this subsection, we first consider the class of nonmonotone systems with vector fields that are bounded from
6
g(x) f (x) h(x) for x ∈ D.
x•f
6∈
[x∗g , x∗h ].
(8)
boundary of
State 1
Fig. 1. A schematic depiction of the conditions (8) and (9). The condition (8)
ensures that all the fixed points lie in the intersection of the corresponding
domains of attractions (in this case, it is B(x∗h )). The fixed point x•f cannot
lie in the dashed blue box due to condition (9).
(9)
Then the following relations hold:
B(x∗g ) ⊇ B(x∗f ) ⊇ B(x∗h ).
boundary of
(7)
Assume also that the fixed points x∗g , x∗f , x∗h , x•f satisfy
x∗g , x∗f , x∗h ∈ B(x∗g ) ∩ B(x∗h ),
boundary of
State 2
below and above by vector fields of monotone systems. We
show that the basins of attraction of such systems can be
bounded by basins of attraction of the bounding systems.
Theorem 1: Let the systems ẋ = f (x), ẋ = h(x), and ẋ =
g(x) satisfy Assumptions A1–A2. Let D = Dg = Df = Dh ,
the systems ẋ = h(x) and ẋ = g(x) be monotone on D and
(10)
Moreover, the sets B(x∗g ), B(x∗h ) are order-convex.
Proof: i) First we note that the assumption (8) implies that
x∗g x∗f x∗h . Indeed, by Proposition 1 for all t ≥ 0, we
have φg (t, x∗f ) φf (t, x∗f ) φh (t, x∗f ), and thus taking the
limit t → ∞ we get x∗g x∗f x∗h .
ii) Next we show that g(x) f (x) for all x ∈ D implies
that B(x∗g ) ⊇ B(x∗f ). Let y ∈ B(x∗f ). By Proposition 1 we have
that φg (t, y) φf (t, y). Furthermore, limt→∞ φf (t, y) = x∗f
and φg (t, y) converges to x∗g or x•g , which implies that
limt→∞ φg (t, y) x∗g . Hence, there exists a time T such that
x∗g − ε1 φg (t, y) x∗f + ε1
for all t > T and some positive ε. We can pick a small ε such
that x∗f + ε1 and x∗g − ε1 lie in B(x∗g ) (due to (8)). According
to Proposition 6, the flow φg (t, y) lies in B(x∗g ) and hence
y ∈ B(x∗g ), which completes the proof.
iii) Similarly to ii), we have that B(x∗g ) ⊇ B(x∗h ).
iv) Finally, we show that B(x∗f ) ⊇ B(x∗h ). Let y ∈ B(x∗h ).
By Proposition 1, we have that
φg (t, y) φf (t, y) φh (t, y),
for all t ≥ 0. Furthermore, due to iii), we have that y ∈ B(x∗g )
and φg (t, y) converges to x∗g , and there exists a T such that
x∗g − ε1 φf (t, y) x∗h + ε1
for all t > T and some small positive ε. We can also choose an
ε such that x∗h +ε1 and x∗g −ε1 lie in B(x∗g )∩B(x∗h ) due to (8).
Hence the flow φf (t, y) belongs to the set {z|x∗g − ε1 z
x∗h + ε1} for all t > T . Since the system ẋ = f (x) is bistable,
the flow must converge to x•f or x∗f . If the flow converges to
x•f , it violates condition (9). Hence the flow φf (t, y) converges
to x∗f and y ∈ B(x∗f ).
v) Order-convexity of the sets B(x∗g ) and B(x∗h ) follows
from Proposition 6.
The conditions (8), (9) are technical and generally easy to
satisfy. An illustration of these conditions is provided in Figure 1. Checking the condition (9) is equivalent to computing
the stable fixed points. Similarly, condition (8) holds if the
trajectories of the systems ẋ = h(x) and ẋ = g(x) initialized
at x∗f converge to x∗g and x∗h , respectively. This can be verified
by numerical integration.
A key to using this result is the computation of bounding systems, which we discuss below. Systems with vector
fields that can be bounded from above and below by vector
fields of monotone systems are typically such that only a
few interactions between the states are not consistent with
monotonicity. These systems are called near-monotone and are
often observed in biological applications (for a comprehensive
discussion on near-monotonicity, see [38]). Assume that there
exists a single interaction which is not compatible with monotonicity. Namely, let the (i, 1)-th entry in the Jacobian matrix
J(x) be smaller than zero for some i 6= 1 and some x. A
bounding monotone system can be obtained by replacing the
variable x1 with a constant in the function fi (x1 , x2 , . . . , xn ),
which removes the interaction between the states xi and
x1 . Hence, we simply use gj (x) = fj (x) if i 6= j and
gi (x) = fi (δ, x2 , . . . , xn ) with δ small enough. Similarly we
set hj (x) = fj (x) if i 6= j and hi (x) = fi (η, x2 , . . . , xn )
with η large enough. More details on this procedure can be
found in [29]. However, when the bounding systems are found,
we still need to check the other conditions in the premise of
Theorem 1. It can happen that the bounding systems are not
bistable, or not even stable (see, the toxin-antitoxin system
in Section VI-C). Unfortunately, we are not aware of an
algorithm, which can guarantee bistability and monotonicity
in the bounding systems.
A corollary from Theorem 1 allows to estimate the basins of
attraction of monotone systems under parametric uncertainty.
Corollary 1: Consider a family of systems ẋ = f (x, p)
with a vector of parameters p taking values from a set P.
Let the systems ẋ = f (x, p) satisfy Assumptions A1–A2 and
be monotone on D × P, where D = Df (·,q) for all q ∈ P.
Consider an interval [pmin , pmax ] such that
x∗ (pmin ) ∈ B(x∗ (pmax )), x∗ (pmax ) ∈ B(x∗ (pmin )),
•
∗
∗
x (pmin ) 6∈ [x (pmin ), x (pmax )].
(11)
(12)
Then the following relation holds:
B(x∗ (pmin )) ⊇ B(x∗ (p)) ⊇ B(x∗ (pmax ))
(13)
for all p in [pmin , pmax ]. Moreover, the sets B(x∗ (p)) are orderconvex for all p in [pmin , pmax ].
Proof: Let g(x) = f (x, pmin ) and h(x) = f (x, pmax ).
According to the order in the parameter set, we have that
g(x) f (x, p) h(x),
for all (x, p) ∈ D × [pmin , pmax ].
(14)
7
∗
The interval
(pmin ), x∗ (pmax )] is a subset of
T [x
∗
∗
B(x (pmax )) B(x (pmin )), since the endpoints of this
interval belong to this intersection according to (11).
Therefore for all pT ∈ [pmin , pmax ] we have that
x∗ (p) ∈ B(x∗ (pmax )) B(x∗ (pmin )) and (8) in the premise
of Theorem 1 follows.
Due to monotonicity we have that x• (pmin ) x• (p)
for all pmin p. Since x• (pmin ) x∗ (pmin ), the condition x• (p) ∈ [x∗ (pmin ), x∗ (pmax )] implies that x• (pmin ) ∈
[x∗ (pmin ), x∗ (pmax )] and contradicts (12). Hence x• (p) 6∈
[x∗ (pmin ), x∗ (pmax )] and (9) in the premise of Theorem 1
follows, application of which completes the proof.
Corollary 1 implies that we can predict a direction of change
in the basins of attraction subject to parameter variations if
we check a couple of simple conditions. This result may
be valuable for design purposes in some applications. For
instance, in the toxin-antitoxin example (see Section VI-C)
it is desirable to make the basin of attraction of the fixed
point corresponding to the high toxin concentration smaller.
This would increase the likelihood of switching from this fixed
point to another (which corresponds to low toxin concentrations) subject to intrinsic and/or exogenous noise. However,
the considered toxin-antitoxin model is not monotone and
further investigation is required to extend Corollary 1 to a
larger class of systems.
Another design problem is to determine the set of all
parameters for which a monotone system is at least bistable.
We offer the following development of this problem.
Proposition 9: Consider a family of systems ẋ = f (x, p)
with a vector of parameters p taking values from an orderconvex set P. Let the system be strongly monotone on D × P,
where D = Df (·,q) for all q ∈ P, i.e. φ(t, x, p) φ(t, y, q)
if x ≺ y or p ≺ q. Consider an interval [pmin , pmax ] such
that the systems ẋ = f (x, pmin ) and ẋ = f (x, pmax ) satisfy
Assumptions A1–A2 and (11,12) hold. Then the intervals
[x∗ (pmin ), x∗ (pmax )], [x• (pmin ), x• (pmax )] are compact attractors for the system ẋ = f (x, p) for all p ∈ [pmin , pmax ].
Furthermore, both intervals contain at least one fixed point
such that the eigenvalues of the Jacobian at these fixed points
have nonpositive real parts.
Proof: Since φf (t, x, pmin )
φf (t, x, p)
φf (t, x, pmax ) for all p ∈ [pmin , pmax ], the flow φf (t, x, p)
with the initial condition x ∈ [x∗ (pmin ) − ε1, x∗ (pmax ) + ε1]
for a small ε > 0 converges to [x∗ (pmin ), x∗ (pmax )]. The
same result holds for the interval [x• (pmin ), x• (pmax )]. Since
the intervals [x∗ (pmin ), x∗ (pmax )] and [x• (pmin ), x• (pmax )]
do not intersect according to (12), they are compact attractors
for the system ẋ = f (x, p) for every p. According to
Theorem 4.6 in [39], both intervals contain a fixed point such
that the Jacobian at this fixed point has eigenvalues with
nonpositive real parts.
We note that these fixed points are not necessarily stable
in the classical sense. However, using strong monotonicity
and the convergence criterion [9] we can show that e =
lim φ(t, x∗ (pmin ), p) is a fixed point for all p ∈ [pmin , pmax ].
t→∞
Furthermore, all points in [x∗ (pmin ), e] are converging to e
due to strong monotonicity. Fixed points with this property
are called stable from below [9].
V. D ISCUSSION ON C OMPUTATION OF BASINS OF
ATTRACTION AND I SOSTABLES
A. Lyapunov Methods
In control theory, a go-to approach for computing forwardinvariant sets (not only basins of attraction) of dynamical
systems is sum-of-squares (SOS) programming, see e.g. [26]
and [27] and the references within. This approach can be
applied to systems with polynomial vector fields.
Another option is to compute the eigenfunction s1 , which
provides the isostables and the basin of attraction. In the case
of a polynomial vector field, we can formulate the computation
of s1 as an infinite dimensional linear algebraic problem
using (3). Hence, we can provide an approximation of s1
using linear algebra by parameterizing s1 with a finite number
of basis functions [25]. On another hand, we can estimate
s1 directly from data using dynamic mode decomposition
methods [40], [41]. These two options provide extremely
cheap estimates of s1 . In fact, the algebraic methods (as
demonstrated in [25]) also provide good estimates on basins
of attraction. However, we cannot typically compute estimates
with an excellent approximation quality, which comes as a
tradeoff for fast computations. The eigenfunction s1 can also
be computed on a mesh grid by using Laplace averages (5)
and by simulating a number of trajectories with initial points
on this mesh grid. Interpolation or machine learning methods
can then be applied to estimate the dominant eigenfunction.
In [27], the authors combined the maximal Lyapunov function approach with SOS techniques, which resulted in an
iterative semidefinite program. As we have discussed above,
the function s1 can be used to construct a maximal Lyapunov
function. Hence the main difference between using s1 and [27]
is algorithmic. In our point of view, using s1 can be beneficial,
since we compute s1 directly (by linear algebra or Laplace
averages), while in [27] it is required to optimize over the
shape of the maximal Lyapunov function.
In [26], the authors propose an approach that is conceptually
similar to the computation of s1 . They also lift the problem to
an infinite dimensional space, but in the framework of the socalled Perron-Frobenius operator acting on measures, which is
dual to the Koopman operator. The authors propose an infinitedimensional linear program to compute a specific occupation
measure related to the basin of attraction. They then consider
a series of relaxations using the moments of the measures
to formulate their problem as an iterative finite-dimensional
semidefinite program (as in the case of [27]). In some sense,
this is similar to the relaxation of the procedure to compute
s1 from an infinite-dimensional algebraic problem to a finitedimensional one. We note that, in contrast to the computation
of s1 , the methods proposed in [26] do not rely on the spectral
properties of the operator.
Even though recent advances in optimization allowed
solving some semidefinite programs as iterative linear programs [42], SOS approaches still lead to highly dimensional
iterative optimization algorithms. Therefore, the applicability
of SOS-based methods to high dimensional systems is delicate
due to memory and computational power requirements. In
our opinion, there is a tradeoff between two options: com-
8
puting rather cheaply s1 (e.g., using Laplace averages) with
weak guarantees or employing semidefinite programming with
heavy computational requirements and strong guarantees. In
this context, the estimation of s1 is simply one possible option.
B. Data Sampling Algorithms
In the case of monotone systems, it follows from Proposition 7 that for all z1 , z2 ∈ Bα with α > 0 and z1 z2 ,
the set [z1 , z2 ] is an inner approximation of Bα (with a nonzero measure). With {zSi }N
i=1 ∈ Bα , an inner approximation
of Bα is computed as i,j=1,...,N [zi , zj ]. It is also possible
to build an outer approximation. Assume we want to compute
an approximation of Bα on an interval B = [b0 , b1 ], where
o
b0 ∈ Bα and b1 6∈ Bα . Let {zio }N
i=1 be in B, but not in
Bα .SThen the outer approximation is computed as follows
B/ i=1,...,No [zio , b1 ]. This implies that Bα can be estimated
by a data sampled approach in the case of monotone systems.
The data sampling algorithms presented below have a few of
advantages over Lyapunov methods:
• The algorithms can be applied to non-polynomial vector
fields and have low memory requirements;
• The algorithms can be parallelized by generating several
samples z j at the same time;
• Inner and outer approximations of Bα are computed at
the same time;
• It is straightforward to compute estimates on 2-D and 3-D
cross-sections of Bα .
Remark 1: By k-D cross-sections we mean the following.
We fix n − k state components to be equal to a constant, that
is xji = pi for i = 1, . . . n − k and ji ∈ I ⊂ {1, . . . , n}, and
define a cross-section as follows:
n
o
e p, I) = x ∈ B(z) xj = pi , ji ∈ I ,
B(z,
i
Algorithm 1 Computation of the level set of O(z)
1: Inputs: Oracle O, the initial set B
2: Outputs: The sets of points Mmin , Mmax .
3: Set A = B, Mmin = {b0 }, Mmax = {b1 }
4: while stopping criterion is not satisfied do
5:
Generate a sample z ∈ Ã, where à ⊂ int(A)
6:
If O(z) = 0, then add z to Mmin
7:
If O(z) = 1, then add z to Mmax
8:
Update A using Mmin and Mmax
9: end while
O(w) = 1. Therefore, we need to keep track of the largest
(in the order) samples z j with O(z j ) = 0, the set of which
we denote Mmin , and the smallest (in the order) samples z j
with O(z j ) = 1, the set of which we denote Mmax . The set
Mmin lies in {z ∈ Rn |O(z) = 0}, while the set Mmax lies
in {z ∈ Rn |O(z) = 1}. Since O(z) is an increasing function,
the set Mmin (respectively, the set Mmax ) can be used to
build a piecewise constant inner approximation (respectively,
an outer approximation) of the set {z ∈ B|O(z) = 0}. In
order to improve the approximation quality at every step, we
generate new samples z in the set
A = z ∈ Rn |x 6 z, z 6 y, ∀x ∈ Mmin , ∀y ∈ Mmax ,
where z is an asymptotically stable fixed point.
We discuss the approach in an abstract form using oracles.
Let O(z) : B → {0, 1} be an increasing function, that is
O(x) ≤ O(y) for all x y and let B = {z ∈ Rn |b0 z
b1 } be such that O(b0 ) = 0 and O(b1 ) = 1. Our goal is to
compute the set {z ∈ B|O(z) = 0}. For the computation of
Bα we use the following oracle:
(
0 if |s1 (z)| < α and kφ(T, z) − x∗ k < ε,
(15)
O(z) =
1 otherwise,
and update Mmin and Mmax . The approach is summarized
in Algorithm 1, the major parts of which are the stopping
criterion and the generation of new samples.
In [23] it was proposed to use an estimate on the volume
of A to establish a stopping criterion. It was also proposed to
generate part of the samples randomly using a distribution with
the support on à = int(A), and to generate the rest greedily by
finding the areas of A which contain largest intervals. While
the volume of A is large, random sampling helps to learn
the shape of the function O(z). As the algorithm progresses,
the greedy sampling ensures faster convergence of inner and
outer approximations. In this algorithm, it is required to keep
the points in Mmin , Mmin unordered, i.e., x 6 y for all
x, y in Mmin and x 6≺ y for all x, y ∈ Mmax . Therefore
after generation of new samples, we need to prune Mmin
(respectively, Mmax ) by removing all x for which there exists
y ∈ Mmin such that x ≺ y (respectively, y ∈ Mmin such that
x y). In [28], it was proposed to use
max
Ãε = z ∈ Rn |x 6 z, z 6 y, ∀x ∈ Mmin
,
ε , ∀y ∈ Mε
where the value of the eigenfunction s1 (z) is computed using
Laplace averages (5), and T is large enough. In our examples,
we chose the observable g(x) = w1T (x − x∗ ), where w1 is
the right dominant eigenvector of the Jacobian matrix J(x∗ ).
We need the second condition kφ(T, z) − x∗ k < ε in (15)
to make sure that the points also lie in B. Even though it is
unlikely to have a point z with a finite s1 (z) not lying in B,
such situations can occur numerically. If we need to compute
B, then we drop the condition |s1 (z)| < α, since here α = ∞.
The main idea of the algorithm is based on the increasing
property of O. If a sample z j is such that O(z j ) = 0, then
for all w z j we have O(w) = 0. Similarly, if a sample
z j is such that O(z j ) = 1, then for all w z j we have
where ε > 0 is called the learning rate, Mmax
= {y − ε1 ∈
ε
n
min
Rn |y ∈ Mmax }, Mmin
=
{x+ε1
∈
R
|x
∈
M
}. If Ãε is
ε
empty, then the learning rate is adjusted as follows: ε = αε for
some α ∈ (0, 1). The existence of z in Ãε can be established
using the solver [43], which also produces a solution point z,
if it exists. The stopping criterion is a lower bound on ε.
The algorithm from [23] allows controlling where the new
samples are generated and ensures that the samples always
decrease the volume of A in a maximal way according to the
proposed heuristic. In order to do so, we need to sweep through
a large number of points in Mmin , Mmax , which can be
computationally expensive for a large n (recall that z ∈ Rn ).
At the same time, the algorithm from [28] avoids computing an
9
Isostables and Boundaries of Basins of Attraction
100
80
60
State
estimate of the volume of A by solving a feasibility problem
using an efficient solver. Even though the algorithm [23] can
potentially be implemented using efficient search methods over
partially ordered sets, the algorithm [28] seems to be more
appealing due to the off-the-shelf feasibility solver.
We finally note that in numerical experiments the algorithm
from [23] exhibits exponential convergence in the estimation
error versus the number of generated samples (at least in the
cases n = 2, n = 3). In [28] it is argued that, in order
ε
to converge
ε so
n that à is empty, it is required to
to an
maxi |b1i −b0i |
samples in the worst case. Hence,
generate
ε
this approach requires a finite number of samples to converge
and potentially has an exponential convergence.
40
20
0
0
5
15
10
20
25
State
Fig. 2. Isostables ∂B0 (x∗ (q)), ∂B2·103 (x∗ (q)), and boundaries of basins of
attraction ∂B(x∗ (q)) for q equal to qint , qmin , or qmax . The dots represent
the fixed points x∗ (qint ) (blue), x∗ (qmin ) (red), x∗ (qmax ) (green) for
different parameter values
VI. E XAMPLES
A. A Two-State Toggle Switch
We illustrate our methods on a genetic toggle switch model
(see e.g. [44]). We choose this model since it is extensively
studied in synthetic biology so that our results can be verified
with other techniques. We consider the following model:
p12
− p14 x1 ,
ẋ1 = p11 +
1 + xp213
(16)
p22
ẋ2 = p21 +
p23 − p24 x2 ,
1 + x1
where all pij ≥ 0. The states xi represent the concentration
of proteins, whose mutual repression is modeled via a rational
function. The parameters p11 and p21 model the basal synthesis rate of each protein. The parameters p14 and p24 are
degradation rates, and p12 , p22 describe the strength of mutual
repression. The parameters p13 , p23 are called Hill coefficients.
The model is monotone on R2≥0 for all nonnegative
parameter
values with respect to the orthant diag 1 −1 R2≥0 . Moreover, the model is monotone with respect to all parameters
but p13 , p23 . In this setting the fixed point x∗ has the state x2
“switched on” (x∗2 is much larger than x∗1 ), while x• has the
state x1 “switched on” (x•1 is much larger than x•2 ).
First, we check Corollary 1 by considering
q1 q2 4 1
p=
,
q3 q4 3 2
with the set of admissible parameters Q = {q|qmax q
q q qmin }, where qmin = 1.8 950 1.2 1050 ,
2.2 1100 0.7 900 , and q is induced
qmax =
by diag 1 1 −1 −1 R4≥0 . We compute the isostables
∂Bα (x∗ (·)) with α = 0, 2 · 103 , ∞ (where ∂B∞ (x∗ (·))
is the boundary of the basin of attraction) for systems with
parameters qmin , qmax , and qint = 2 1000 1 1000 .
The computational results depicted in Figure 2 suggest that for all parameter values q ∈ Q the manifold
∂B∞ (x∗ (q)) will lie between the manifolds ∂B∞ (x∗ (qmin ))
and ∂B∞ (x∗ (qmax )). It appears that ∂B0 (x∗ (qint )) also lies
between the manifolds ∂B0 (x∗ (qmin )) and ∂B0 (x∗ (qmax )),
however, in a different order. This change of order and continuity of s1 implies that there exists an α such that at least two
manifolds ∂Bα (x∗ (qint )), ∂Bα (x∗ (qmax )), ∂Bα (x∗ (qmin )) intersect. This case is also depicted with α = 2 · 103 . This
observation implies that s1 (x, q) is not an increasing function
in q. This is consistent with the linear case, where changes
in the drift matrix A, will simply rotate the hyperplane w1T x,
where w1 is the left dominant eigenvector of A.
Another feature for a successful design of a genetic toggle
switch is choosing the Hill coefficients. Computing the derivative of the repression term with respect to a Hill coefficient pi3
p12 ln(x)xpi3
. Therefore for positive x the derivative
gives −
(1 + xpi3 )2
changes sign at x = 1, and the partial order with respect
to pi3 cannot be defined on the whole state-space R2>0 and
parameter space R2>0 . Nevertheless, we can study how basins
of attraction change subject to changes in Hill coefficients.
Consider the following parameter values
1 1000 q1 1
p=
.
(17)
1 1000 q2 2
The computation results are depicted in Figure 3. We observe
that the variations in basins of attraction are consistent with the
result of Corollary 1, and the changes in Hill coefficients with
respect to the order diag(1, −1)R2 . It remains to verify if this
result is an evidence of a partial order in Hill coefficients or
simply a coincidence. However, in any case this result holds
when we vary other parameters and hence can serve as a rule
of thumb in designing toggle switches. We finally note that
the curve ∂B with q1 = 5, q2 = 2 is not convex or concave,
which in other examples seems to be the case. We verified this
observation by computing the curves with higher accuracy and
smaller resolutions. The curve ∂B with q1 = 5, q2 = 2 still
separates R2 into two order-convex regions.
Finally, we illustrate the result of Proposition 9 by considering another important problem for a successful design of a
toggle switch: estimating the set of parameters for which the
system is bistable. Consider the following parameter values
2 700 2 d1
p=
.
(18)
1 1000 2 d2
The partial order in the parameter space is induced by the
orthant diag(1, − 1)R2 . We have computed the multistability
region on a mesh grid in [0, 4]2 , while computing the number
of fixed points for given parameter values, and plot the
estimation results in Figure 4. This picture confirms some of
the previous findings (including the ones in [44]).
10
Boundaries of Basins of Attraction
Boundaries of Basins of Attraction
50
600
30
20
400
State
State
40
10
00
20
40
60
80
State
0
200
Fig. 3. The effect of Hill coefficients (here denoted as q1 and q2 ) on
the domain of attractions of the toggle switch system (16). The remaining
parameters are fixed and described in (17).
We note that most of the properties described above are
known. We show, however, that these properties stem from
monotone systems theory and are not limited to a particular
model. Hence they can be extended to other models as long
as they satisfy the premise of our theoretical results.
B. Basin of Attraction of a Non-Monotone System
We consider the following three-state system
ẋ1
ẋ2
ẋ3
1000
− 0.4x1 ,
1 + x23
1000
=
− 4x2 + u,
1 + x41
1
= p1 + p2 x1x+1
+ 5x2 − 0.3x3
=
(19)
which is not monotone with respect to any order for all positive
parameter values p1 and p2 . This model does not have any
biological interpretation and was designed in order to illustrate
Theorem 1 on a simple example. We consider the nominal
system F (with p1 = 0.1, p2 = 1) and two bounding systems
G1 (with p1 = 0.1, p2 = 0) and G2 (p1 = 1.1, p2 = 0).
It can be verified that G1 and G2 are indeed monotone,
1
bistable and bound F, since 0 ≤ x1x+1
≤ 1. In this example,
we illustrate the application of Theorem 1, according to which
we can bound the basin of attraction of the non-monotone system F by the basin of attraction of the monotone systems G1 ,
G2 . We also illustrate the effectiveness of our computational
algorithm. We intentionally do not let the algorithm converge
to an accurate solution. We generate only 1000 samples in
Multistability Region
4
Parameter
3
2
1
1
200
100
2
3
4
Parameter
Fig. 4. Approximation of the multistability region in the toggle switch
system (16) subject to variations in degradation rates d1 , d2 and parameter
values (18). The region of monostability is white, while the region of
multistability is light green.
0
100
100
State
State
0
200
Fig. 5. Illustration to Example VI-B. Boundaries of domains of attraction B
for the non-monotone system F and its bounding monotone systems G1 , G2 .
order to compute every surface. In this case, 550−650 samples
(depending on an example) are used to compute upper and
lower bounds on each surface (we depict only upper bounding
surfaces, which are built using 250-350 samples). We see
in Figure 5 that the green surface related to BG2 does not
bound the blue surface related to BF for all points in the
state-space, which happens since our algorithm did not have
enough samples to converge. However, the algorithm is already
capable of producing surfaces that depict the overall shape of
boundaries of domains of attraction. With more samples, we
obtain a more accurate picture, which validates our theoretical
results. The computational time for this example is between
10 to 20 seconds depending on the computational tolerance of
the differential equation solver.
C. Toxin-Antitoxin System
Consider the toxin-antitoxin system studied in [45]:
σT
1
Ṫ =
−
T
[Af ][Tf ]
(1
+
β
C [Tf ])
(1 + βM [Tf ])
1 + K0
σA
Ȧ =
− ΓA A
[Af ][Tf ]
1 + K0
(1 + βM [Tf ])
[Af ][Tf ] [Af ][Tf ]2
[Ȧf ] = A − [Af ] +
+
KT
KT KT T
[Af ][Tf ]
[Af ][Tf ]2
[Ṫf ] = T − [Tf ] +
+2
,
KT
KT KT T
where A and T is the total number of toxin and antitoxin
proteins, respectively, while [Af ], [Tf ] is the number of free
toxin and antitoxin proteins. In [45] the model was considered
with = 0. Here, we set = 10−9 in order to show that, in
the case of systems admitting a time-scale separation, we can
estimate the basins of attraction of the reduced model without
performing model reduction. For the parameters σT = 166.28,
K0 = 1, βM = βc = 0.16, σA = 102 , ΓA = 0.2, KT =
KT T = 0.3, the system has two stable hyperbolic fixed points:
x• = 27.1517 80.5151 58.4429 0.0877
x∗ = 162.8103 26.2221 0.0002 110.4375 .
Although the full and reduced systems are not monotone
with respect to any orthant, numerical results in [37] indicate
11
Points in Basins of Attraction
80
Protein A
60
40
a bistable toggle switch with two states. We discuss the effect
of different parameters on the shape of the basin of attraction
and provide some simple strategies to predict the possible
shape of a basin without explicitly computing the basin itself.
R EFERENCES
20
100
120
140
160
Protein T
180
200
e • , I, p1 ) of B(x• ),
Fig. 6. The red circles belong to the cross-section B(x
e ∗ , I, p1 ) of B(x∗ ).
while the blue crosses belong to the cross-section B(x
that the basins of attraction of a reduced order system are still
order-convex with respect to diag{[1, − 1]}R2≥0 .
e ∗ , I, pi )
In this example we compute 2-D cross-sections B(x
of a basin of attraction of the full order model with I =
{3, 4} and p1 = [58.4429, 0.0877], p2 = [25, 50], p3 =
[0.0002, 110.4375]. All cross-sections gave results indistinguishable by the naked eye, which validates our approach
for computing basins of attraction of slow dynamics only. In
Figure 6, we present our numerical results for the computation
e ∗ , I, p1 ), B(x
e • , I, p1 ). We plot all the
of cross-sections B(x
points generated by our algorithm, including the ones that
were pruned out during the algorithm. The total number
of generated points was 1000. The red circles belong to
e • , I, p1 ), while the blue crosses belong to B(x
e ∗ , I, p1 ).
B(x
Note that the vast majority of samples is generated near the
separatrix between the basins of attraction. We also observe
that the red circles and the blue crosses do not violate our
assumption on order-convexity of B(x∗ ) and B(x• ).
VII. C ONCLUSION
In this paper, we study geometric properties of monotone
systems. In particular, we investigate the properties of basins of
attraction and relate them to the properties of isostables defined
in the framework of the Koopman operator. We discuss in
detail the relation between these concepts and their properties
under some general assumptions and then focus on properties
of basins of attraction of bistable systems. First, we show that
we can estimate basins of attraction of bistable non-monotone
systems, whose vector fields can be bounded from below and
above by bistable monotone systems. This result uses standard
tools in monotone systems theory and leads to estimation
of basins of attraction of bistable monotone systems under
parametric uncertainty. We also discuss a complementary
problem: finding the set of parameter values for which a
monotone system is (at least) bistable.
We discuss a numerical method for computing inner and
outer approximations of basins of attraction of monotone
systems. This method exploits the geometric properties of
monotone systems and uses the trajectories of the system for
computation. The method is potentially well-suited to highdimensional spaces since it can be easily parallelized and has
lower memory requirements than optimization-based methods.
We also show how our theoretical results can be used to design
[1] J. Jiang, X. Liang, and X.-Q. Zhao, “Saddle-point behavior for monotone
semiflows and reaction–diffusion models,” J. Differential Equations, vol.
203, no. 2, pp. 313–330, 2004.
[2] P. G. Coxson and H. Shapiro, “Positive input reachability and controllability of positive systems,” Linear Algebra and its Applications, vol. 94,
pp. 35–53, 1987.
[3] L. Farina and S. Rinaldi, Positive linear systems: theory and applications. John Wiley & Sons, 2010.
[4] A. Sootla and A. Rantzer, “Scalable positivity preserving model reduction using linear energy functions,” in Proc IEEE Conf Decision Control,
Dec. 2012, pp. 4285–4290.
[5] C. Grussler and T. Damm, “A symmetry approach for balanced truncation of positive linear systems,” in 51st IEEE Conference on Decision
and Control, Maui, Hi, USA, Dec. 2012, pp. 4308–4313.
[6] A. Rantzer, “Scalable control of positive systems,” European Journal of
Control, vol. 24, pp. 72–80, 2015.
[7] T. Tanaka and C. Langbort, “The bounded real lemma for internally
positive systems and h-infinity structured static state feedback,” IEEE
transactions on automatic control, vol. 56, no. 9, pp. 2218–2223, 2011.
[8] A. Berman and R. J. Plemmons, Nonnegative Matrices in the Mathematical Sciences. SIAM, 1994, vol. 9.
[9] M. Hirsch and H. Smith, Monotone dynamical systems. Elsevier BV
Amsterdam, 2005.
[10] G. Dirr, H. Ito, A. Rantzer, and B. Rüffer, “Separable lyapunov functions
for monotone systems: Constructions and limitations,” Discret and
Contin Dyn S, vol. 20, no. 8, pp. 2497–2526, 2015.
[11] J.-L. Gouzé, A. Rapaport, and M. Z. Hadj-Sadok, “Interval observers
for uncertain biological systems,” Ecological modelling, vol. 133, no. 1,
pp. 45–56, 2000.
[12] J. Bergenthum and L. Rüschendorf, “Comparison of semimartingales
and Lévy processes,” Ann. Prob., vol. 35, no. 1, pp. 228–254, 2007.
[13] A. Zlotnik, S. Misra, M. Vuffray, and M. Chertkov, “Monotonicity of actuated flows on dissipative transport networks,” Nov 2015,
arXiv:1511.07585.
[14] P.-J. Meyer, A. Girard, and E. Witrant, “Controllability and invariance
of monotone systems for robust ventilation automation in buildings,” in
Proc. IEEE Conf. Decision Control, 2013, pp. 1289–1294.
[15] D. Angeli, J. Ferrell, and E. Sontag, “Detection of multistability, bifurcations, and hysteresis in a large class of biological positive-feedback
systems,” PNAS, vol. 101, no. 7, 2004.
[16] D. Angeli, P. De Leenheer, and E. Sontag, “Graph-theoretic characterizations of monotonicity of chemical networks in reaction coordinates,”
Journal of mathematical biology, vol. 61, no. 4, pp. 581–616, 2010.
[17] M. A. Rami, V. Bokharaie, O. Mason, and F. Wirth, “Stability criteria
for sis epidemiological models under switching policies,” Discret and
Contin Dyn S, vol. 19, no. 9, pp. 2865–2887, 2014.
[18] M. Budišić, R. Mohr, and I. Mezić, “Applied koopmanism,” Chaos,
vol. 22, no. 4, p. 047510, 2012.
[19] A. Mauroy, I. Mezić, and J. Moehlis, “Isostables, isochrons, and Koopman spectrum for the action–angle representation of stable fixed point
dynamics,” Physica D, vol. 261, pp. 19–30, 2013.
[20] A. Vannelli and M. Vidyasagar, “Maximal lyapunov functions and
domains of attraction for autonomous nonlinear systems,” Automatica,
vol. 21, no. 1, pp. 69–80, 1985.
[21] P. Takác, “Domains of attraction of generic omega-limit sets for strongly
monotone semi-flows,” Zeitschrift fur Analysis und ihre Anwendungen,
vol. 10, pp. 275–317, 1991.
[22] H. Smith and H. Thieme, “Stable coexistence and bi-stability for
competitive systems on ordered banach spaces,” Journal of Differential
Equations, vol. 176, no. 1, pp. 195–222, 2001.
[23] A. Sootla and A. Mauroy, “Properties of isostables and basins of
attraction of monotone systems,” in Am Control Conf, 2016, 2016, pp.
7365–7370.
[24] N. Ramdani, N. Meslem, and Y. Candau, “Computing reachable sets
for uncertain nonlinear monotone systems,” Nonlinear Analysis: Hybrid
Systems, vol. 4, no. 2, pp. 263–278, 2010.
[25] A. Mauroy and I. Mezić, “Global stability analysis using the eigenfunctions of the Koopman operator,” IEEE Tran Autom Control, vol. 61,
no. 11, pp. 3356–3369, 2016.
12
[26] D. Henrion and M. Korda, “Convex computation of the region of
attraction of polynomial control systems,” IEEE Tran Autom Control,,
vol. 59, no. 2, pp. 297–312, 2014.
[27] G. Valmorbida and J. Anderson, “Region of attraction analysis via
invariant sets,” in Proc Am Control Conf, 2014, pp. 3591–3596.
[28] E. S. Kim, M. Arcak, and S. A. Seshia, “Directed specifications and
assumption mining for monotone dynamical systems,” in Proc Conf
Hybrid Systems: Computation Control. ACM, 2016, pp. 21–30.
[29] A. Sootla, D. Oyarzún, D. Angeli, and G.-B. Stan, “Shaping pulses to
control bistable systems: Analysis, computation and counterexamples,”
Automatica, vol. 63, pp. 254–264, Jan. 2016.
[30] A. Sootla, A. Mauroy, and J. Gonçalves, “Shaping pulses to control
monotone bistable systems using Koopman operator,” in In Proc Symposium Nonlinear Control Systems, Aug 2016, pp. 710–715.
[31] I. Mezić, “Spectral properties of dynamical systems, model reduction
and decompositions,” Nonlinear Dynam, vol. 41, no. 1-3, pp. 309–325,
2005.
[32] I. Mezić, “On applications of the spectral theory of the Koopman
operator in dynamical systems and control theory,” in IEEE Conf
Decision Control, 2015, pp. 7034–7041.
[33] D. Angeli and E. Sontag, “Monotone control systems,” IEEE Trans
Autom Control, vol. 48, no. 10, pp. 1684–1698, 2003.
[34] M. Shaked and J. G. Shanthikumar, Stochastic orders. Springer, 2007.
[35] S. P. Franklin, “Some results on order-convexity,” The American Mathematical Monthly, vol. 69, no. 5, pp. 357–359, 1962.
[36] L. Tibilelli, “Connectedness, arcwise-connectedness and convexity for
level-sets of multidimensional distribution functions,” Ratio Math.,
vol. 8, pp. 21 – 27, 1994.
[37] A. Sootla and A. Mauroy, “Operator-theoretic characterization of eventually monotone systems,” March 2016, http://arxiv.org/abs/1510.01149.
[38] E. D. Sontag, “Monotone and near-monotone biochemical networks,”
Syst Synthetic Biol, vol. 1, no. 2, pp. 59–87, 2007.
[39] M. W. Hirsch, “Systems of differential equations that are competitive
or cooperative ii: Convergence almost everywhere,” SIAM Journal on
Mathematical Analysis, vol. 16, no. 3, pp. 423–439, 1985.
[40] P. J. Schmid, “Dynamic mode decomposition of numerical and experimental data,” Journal of Fluid Mechanics, vol. 656, pp. 5–28, 2010.
[41] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N.
Kutz, “On dynamic mode decomposition: Theory and applications,” J
Comput Dynamics, vol. 1, no. 2, pp. 391 – 421, December 2014.
[42] A. A. Ahmadi and G. Hall, “Sum of squares basis pursuit with linear
and second order cone programming,” arXiv preprint arXiv:1510.01597,
2015.
[43] L. De Moura and N. Bjørner, “Z3: An efficient SMT solver,” in
International conference on Tools and Algorithms for the Construction
and Analysis of Systems. Springer, 2008, pp. 337–340.
[44] T. Gardner, C. R. Cantor, and J. J. Collins, “Construction of a genetic
toggle switch in escherichia coli,” Nature, vol. 403, pp. 339–342, 2000.
[45] I. Cataudella, K. Sneppen, K. Gerdes, and N. Mitarai, “Conditional cooperativity of toxin-antitoxin regulation can mediate bistability between
growth and dormancy,” PLoS Comput Biol, vol. 9, no. 8, p. e1003174,
2013.
| 3cs.SY
|
Window-of-interest based Multi-objective
Evolutionary Search for Satisficing Concepts
Eliran Farhi and Amiram Moshaiov
School of Mechanical Engineering
Tel-Aviv University
Tel-Aviv, Israel
Abstract— The set-based concept approach has been
suggested as a means to simultaneously explore different design
concepts, which are meaningful sub-sets of the entire set of
solutions. Previous efforts concerning the suggested approach
focused on either revealing the global front (s-Pareto front), of all
the concepts, or on finding the concepts' fronts, within a
relaxation zone. In contrast, here the aim is to reveal which of the
concepts have at least one solution with a performance vector
within a pre-defined window-of-interest (WOI). This paper
provides the rational for this new concept-based exploration
problem, and suggests a WOI-based rather than Pareto-based
multi-objective evolutionary algorithm. The proposed algorithm,
which simultaneously explores different concepts, is tested using
a recently suggested concept-based benchmarking approach. The
numerical study of this paper shows that the algorithm can cope
with various numerical difficulties in a simultaneous way, which
outperforms a sequential exploration approach.
It is conceivable that, using a meaningful division
procedure, thousands of meaningful concepts may be
constructed for a given real-life engineering problem. For
example, we are currently working on the design of a
propulsion system for unmanned aerial vehicles. For this
particular problem, we have identified over 10,000 meaningful
concepts that are of interest to the designers. It is noted that a
non-interactive exploration run with 100 such concepts, using
the algorithm of [3] may take about a week (on a standard
workstation). In general, with an increasing number of
concepts, finding the fronts of all the predefined concepts
could be computationally prohibitive. This serves as a strong
motivation for the work that is presented here.
Keywords— conceptual design; multi-criteria decision-making;
aspiration levels; set-based concept; design space exploration;
concept-based benchmarking
I.
INTRODUCTION
The aim of this study is to devise and demonstrate a novel
evolutionary search approach, which aims to find satisficing
conceptual solutions. It concerns the Set-Based Concept (SBC)
approach. In this approach, a conceptual solution (or in short –
a concept) is a meaningful set of potential solution alternatives,
which possess some common features [1, 2]. The considered
concepts are predefined by the designers.
Fig. 1 illustrates the SBC approach. Three concepts of
bridge designs are shown. The (generally different) design
spaces of the concepts are marked by ellipses of different gray
levels. As shown in Fig. 1, the associated performance vectors
of particular designs, from all concepts, are to be compared in a
mutual objective space. The most studied SBC approach is
known as the s-Pareto approach [2]. It involves finding which
particular designs, of which concepts, are associated with the
global Pareto-front that is obtained by domination comparisons
among all individual designs from all concepts.
"© 2017 IEEE. Personal use of this material is permitted. Permission from
IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional
purposes, creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works."
Figure 1: Illustration of the SBC approach
In [1], it was suggested that concepts should not be selected
by the s-Pareto approach. A possible alternative has been
suggested in [3], which employs a concept-based relaxedPareto approach.
In contrast to front-related methods, as in [1-4], the current
work takes a Window-Of-Interest (WOI) approach. The WOI
indicates what is considered as acceptable performance vector.
Rather than being interested in finding concepts' fronts, here
the designers are interested in finding which of the considered
concepts have at least one solution with performance vector
within a pre-defined WOI. Concepts that meet this requirement
are considered satisficing. It should be noted that aspiration
levels are commonly used to articulate objective preferences in
multi-objective optimization. Yet, in contrast to studies as
reviewed in [5], here aspiration levels, by way of the WOI,
refer to satisficing concepts rather than to particular solutions.
The significance of the proposed approach, as compared
with the concept-based relaxed-Pareto method of [3], is twofolds. First, in the method of [3], the sought information
involves finding all the concepts' fronts within the relaxation
zone. When the number of concepts within that zone is large,
the computational effort might become prohibitive. Second, the
sought information in the concept-based relaxed-Pareto method
is defined by the relaxation zone, which is unknown in itself.
This causes a computational difficulty in estimating the
dynamic allocation of computational resources among the
considered concepts. These two deficiencies are eliminated in
the current approach. First, in the new problem definition there
is no need to find any front. Second, the WOI, within which
solutions are sought, is pre-defined.
As suggested above, thousands of meaningful concepts can
be envisioned for a real life-problem. With this respect, it is
noted that the proposed problem definition should be
understood as a means to preprocess the concepts. Namely, it
aims to reduce the considered set of concepts into those that
have performances within the WOI. Eventually, depending on
the final goal of the exploration, further search might be
needed when aiming at selection of a concept or a particular
design. Such a search could be done to either all the set of the
obtained satisficing concepts from the preprocessing stage (the
current problem solution), or to a subset of these concepts.
The contribution of this paper is threefold including: a. the
proposed new problem definition and its rationalization, b. the
suggested search algorithm, and c. the study to assess the
proposed algorithm.
(1)
In other words, the decision-makers define
inequality
constraints in the objective space, where each Ck is a
hyperplane in the objective space. These should be defined in a
consistent way with the optimization problem such that there is
no vector in the WOI that is dominated by a vector outside the
WOI.
A satisficing concept is defined as a concept that has at
least one solution with performance vector, which is a member
of the set WOI. The subset of the satisficing concepts,
, is defined as:
(2)
Given a set S of β concepts, the considered problem is to
find within given computational resources, up to l concepts that
are satisficing, where l < β.
III.
This section describes the search methodology. In
subsection III.A the main procedure to search for the
satisficing concepts is described. The other three subsections,
III.B-D, contain the descriptions of the required subprocedures, which are used by the main procedure.
A. Main Procedure
The pseudo-code for the main procedure is described in
Table 1.
TABLE 1: PSEUDO CODE FOR THE ENTIRE SEARCH
1
Initialization
The rest of this paper is organized as follows. In section 0
the considered problem is defined. Next, section III provides
the details on the proposed evolutionary search. Section IV
describes the numerical experiments, their results and analysis.
Finally, section V outlines the conclusions of this paper.
II.
Let no be the dimension of the objective-space
. Also,
let S be the set of the considered concepts, and nc be the
number of concepts (
). Let
be the designspace of the m-th concept, where nm is the dimension of Xm.
Also, define
as the concept's vector of
objective-functions. Furthermore, let s m be any particular
design of the m-th concept, and let
represents the design
vector of s m. Also, let
represents the
performance vector of sm, where
.
Without loss of generality, for a minimization problem let
the WOI of the objective space be defined as the set:
1.1
Define the problem (see section II)
1.2
Set GQ0 ,GQi,, and the total # of generations
Set counters J=0 and Ij=GQ0 for each concept j
PROBLEM DEFINITION
As mentioned above, the main goal of the problem is to
reveal which of the concepts have at least one solution within
the pre-defined WOI. This approach is different from the one
that is described in [3], since it focuses on finding satisficing
concepts rather than finding their Pareto fronts.
THE EVOLUTIONARY SEARCH
2
1.3
Set the GA parameters
1.4
For each concept:
1.4.1
Initialize random population of solutions
1.4.2
Perform WOI-based Individual Evaluation
(see sub-procedure in subsection III.B)
For each concept j, while Ij >0 and J<l, Do:
2.1
Perform Selection and Reproduction (see subprocedure in subsection III.C)
2.2
If the concept is satisficing then set J=J+1
2.3
Set Ij=Ij-1
3
Perform Resource Allocation (see sub-procedure in
subsection III.D)
4
While stopping conditions are not met go to step # 2
The main procedure starts with initializations, including the
generation and initialization quotas. The i-th generation-quota
(GQi) is a predefined number of generations, where i=1, …,
ncat, and ncat nc. These quotas are needed for the resource
allocation procedure (see subsection III.D). Also pre-defined is
an initialization-quota (GQ0), which is a predefined number of
generations to be applied for each concept, before the resource
allocation procedure is to be performed. Next, per each
concept, an initial population is randomly generated. This is
followed by the evaluation of the individuals, as detailed in
subsection III.B. Then, an evolutionary process is carried out
for each concept for a fixed number of generations (Ij=GQ0).
Following this early evolutionary process, an advanced
evolutionary process is taken place in which the resource
allocation procedure becomes effective. The stopping
conditions for the process are met when either the predefined
number of satisficing concepts is obtained, or if the predefined
total number of generations is reached.
B. WOI-based individual evaluation
In this sub-procedure, individuals of a concept population
are evaluated. The j-th concept's population
, is the set of
all the evaluated designs, at the current generation # , which
are associated with concept j. First, the performance vector for
each individual is obtained. Next, a distance,
, is calculated
for each k-th individual of the j-th concept. This distance is the
minimal Euclidean distance between the WOI and the
performance vector of the individual. Next, each such
individual is assigned with a WOI-based rank, denoted as
. This ranking, which is based on a predefined number
of distance ranges, nr, is calculated as follows:
(3)
Where Δ is defined as:
(4) Δ
After assigning a ranking level to each solution, as
illustrated in Fig. 2, a crowding distance is calculated for
individuals having the same rank.
The crowding-distance allows diversity preservation
within the same rank. First, the boundary solutions are
assigned by an infinite distance value, and then all other
solutions are assigned with a calculated crowding distance.
This procedure follows the one in [6]. The role of this
assignment is to keep diversity and prevent premature
convergence.
C. WOI-based Selection and Reproduction
This sub-procedure, which is outlined in Table 2, is
employed for each concept. It starts with tournament selection.
This is based on lexicographic selection, where the rank is
preferred over the crowding. If the ranks of the compared
individuals are the same, then their crowding distance is
employed. Next, the parent and the offspring populations of the
considered concept are united, and the ranking and crowding
distance are being calculated according to the sub-procedure in
subsection III.B. Finally, a new elite population is selected
using the ranks and the crowding distance.
TABLE 2: THE WOI-BASED SELECTION AND REPRODUCTION
1
2
3
4
5
Tournament selection of the parent population
Create offspring population
Create the union population (parents & offspring)
Assign rank and crowding distance (see III.B)
Create the elite population
D. Resource Allocation
The purpose of this procedure is to distribute, during the
search, the computational resources among all the concepts.
The heuristic, which governs the resource allocation for each
concept, is based on the expectation that the concept will be
found to be satisficing. The resource allocation for each
concept depends on its category, where those that belong to a
better category (more promising) will get more resources than
those with a worse category. The concept categorization, which
is updated during the search, depends on the rate by which the
concept solutions are approaching the WOI. Let
be the
distance for the j-th concept, at generation # , as follows:
(5)
The concept-distance-rate is the change of the conceptdistance between generations, namely:
Rank 3
Objective f2
Rank 2
Rank 1
WOI
Objective f1
Figure 2: Illustration of ranking
(6)
The concepts are sorted according to their conceptdistance-rate and each is assigned with a category. Concepts,
which are categorized into the 1st category (most-promising),
receive a larger generation quota (GQi) as compared with those
that belong to the lower categories. Resource allocation is
recalled, for re-assigning the concepts with new generationquotas, after all concepts finish their generation quotas.
IV.
EXPERIMENTS AND RESULTS
This section provides the results of testing the suggested
algorithm and their analyses on min-min bi-objective
problems using a WOI, which is defined by a minimal
performance level for each objective. Subsection IV.A
contains results from running the algorithm with only one
concept at a time. Next, in subsection IV.B, results are
provided concerning a benchmark example of running nine
different concepts simultaneously. It follows the conceptbased benchmarking method of [3]. Finally, in subsection
IV.C, the same example is reworked with a different WOI.
A. Single concept tests
This subsection describes some typical results from runs in
which only one concept participate (one test function at each
run). Namely, it focuses on the performance of the WOI-based
selection, rather than on the performance of the resource
allocation procedure. The latter is dealt with in the next two
subsections.
To test the behavior of the algorithm, each of the nine biobjective test functions, which are detailed in subsection IV.B,
were employed with different population sizes and with
different WOIs. The maximum number of generation was set
to 1000, which is taken to be larger than needed for the
particular examples.
To obtain statistical results, each case was tested for 30
runs. The runs were done by real-coding with crossover and
mutations as in [6]. The crossover probability was taken as 0.9
and the mutation probability was taken as 1/n.
Table 3 provides typical results. The first column lists the
employed test-function. The second and the third columns
contain the WOI limits. The fourth column includes the
population size. The fifth and sixth columns show the average
number of generations, and total number of evaluations, until
finding the concept to be satisficing.
decided to use a population size of 20 for each concept in the
simultaneous searches of subsections IV.B-C.
Figs. 3-4 show typical population progress at different
generations for runs with the ZDT1 test-function. The results
are shown for two different WOI (gray color), using a
population size of 20. In both cases, the same phenomenon is
observed. At the beginning, the performances of the initial
population are widely spread over the objective space. Then,
the process exhibits a strong reduction of the diversity.
Nevertheless, the search is not locked at local optima. This
demonstrates the effectiveness of the suggested WOI-based
selection approach. The fact that the search is successful,
without a large diversity, calls for the use of a small
population size or an adaptive population size.
Initial population
Pop. at
174th gen.
Pop. at 10th. gen
th
Pop. at
100th gen.
Pop. at 50 gen.
Figure 3: Search progress for ZDT1 with WOI = [0.2, 2.0]
Initial population
TABLE 3: AVERAGE # OF GENERATIONS AND EVALUATIONS
Test
function
F1
limit
F2
limit
Pop.
size
Avg. # of
generations
Avg. # of
Evaluations
ZDT1
0.5
0.5
ZDT1
0.2
2.0
ZDT1
2.0
0.2
ZDT3
0.5
0.5
100
50
20
10
100
50
20
10
100
50
20
10
100
50
20
10
244
319
436
556
113
134
188
266
193
210
272
334
94
100
127
161
24,400
15,950
8,720
5,560
11,300
7,200
2,760
2,660
19,300
10,500
5,440
3,340
9,400
5,000
2,540
1,610
Based on the obtained results for all the considered
concepts (test-functions), which are not shown here, it was
Pop. at 10th gen.
Pop. at 50th gen.
Pop. at 100th gen.
Pop. at 222th gen.
Figure 4: Search progress for ZDT1 with WOI = [2.0, 0.2]
Figure 5 describes the results for the case of ZDT1 with
WOI = [0.5 ,0.5] and a population size of 20. It shows the
statistics from 30 runs of the concept-distance vs. the number
of function evaluations (in a boxplot form). In this particular
case, the concept was found to be satisficing (distance reached
a zero value) at around 8000 evaluations (for most runs).
In the resource allocation sub-procedure the concepts were
categorized into three categories. The generation-quotas were
set as GQ = [10,3,1]. In the current study, two of the concepts
received the highest quota (GQ1=10), five concepts received a
medium quota (GQ2=3), and the worst two concepts received
the least quota (GQ3=1).
Figure 5: The statics of the concept-distance for ZDT1; WOI = [0.5 ,0.5]
In this case study the goal was to find just one satisficing
concept, which is actually the only one such concept (ZDT1-2).
Table 5 provides a comparison between sequential runs (10
runs for each concept at a time, with a limit of 1000
generations) and 10 simultaneous runs with the suggested
resource allocation. The table shows the average number of
generations that it took to stop the runs.
TABLE 5: PARALLEL VS. SEQUENTIAL RUNS FOR L=1, WOI = [0.2, 0.5]
B. Simultanous search of concepts- Case 1
This min-min bi-objective test problem involves nine
concepts. Each concept is represented by a different testfunction, based on well-know test-functions and their
transformations (as in [3]). Table 4 shows all the concepts that
were tested. The first column lists the function name and its
version. For example: ZDT1-1 is for the classical ZDT1
problem and ZDT1-2 is a transformed ZDT1. The second and
third columns correspond to the transformation scale and
offset, respectively.
TABLE 4: THE DIFFERENT CONCEPTS DEFINITIONS
Test
Function
ZDT1-1
ZDT1-2
ZDT1-3
SCH1-1
SCH1-2
ZDT2-1
ZDT2-2
ZDT2-3
ZDT3-1
Objectives
Scale
[1,1]
[2,2]
[1,1]
[1,1]
[0.5,1]
[1,1]
[0.7,0.7]
[0.7,0.7]
[1,1]
Objectives
Offset
[0,0]
[-0.5,-0.5]
[0.2,0]
[0,0]
[0,0]
[0,0]
[0,0]
[0.2,0.2]
[1,1]
The Pareto fronts of the concepts are shown in Fig. 6. Also
shown is the chosen WOI for this study case. It can easily be
seen that only one concept (ZDT1-2) has solutions within the
chosen WOI [0.2, 0.5].
Concept #
Average # of
generations for the
simultaneous runs
Average # of
generations in
sequential runs
ZDT1-1
ZDT1-2
ZDT1-3
SCH1-1
SCH1-2
ZDT2-1
ZDT2-2
ZDT2-3
ZDT3-1
Total
450
498
445
142
113
248
269
165
333
2663
1000
499
1000
1000
1000
1000
1000
1000
1000
8499
In all the simultaneous runs, the ZDT1-2 concept was
revealed and the average amount of the total number of
generations till stopping was 2,663 (which is equal to 53,260
evaluations). The sequential searches involved a total of 8,499
generations at most (= 169,980 evaluations). The latter number,
and the comparison, should be understood as follows. In a
random order of the sequential searches, the chance of
choosing a satisficing concept is increasingly small with an
increasing number of non-satisficing concepts. In such a case it
appears that the simultaneous approach has a clear advantage
over the sequential approach.
Fig 7 shows, for each concept, the median of the conceptdistance versus the number of generations, as obtained by the
simultaneous approach.
Figure 7: Medians of concept-distances versus # of generations
Figure 6: Pareto fronts of the concepts functions
The statics of the number of generations till stopping, per
each concept, are displayed in Fig. 8. As seen there, the less
promising concepts received a relative small amount of
resources during the search, whereas the more promising ones
received a larger amount of resources, until the satisficing
concept was found.
Figure 8: The statistics of the # of generations per concept
C. Simultanous search of concepts- Case 2
The second case study is almost the same as the first one.
The only difference is that the WOI was re-set to [0.3, 0.4],
which amounts to having two rather than one satisficing
concepts (ZDT1-2 and ZDT3-1). The goal of this case was set
to find the two satisficing concepts. The same comparison
between a parallel run and a sequence run was made. The
results are summarized in table 6.
as a concept with at least one particular solution within a
window-of-interest in the objective space. An evolutionary
algorithm is suggested for solving this problem by a
simultaneous search with the considered concepts. The
algorithm is examined using concept-based benchmarking
problems, which simultaneously employ various well-known
test functions. The simultaneous approach is compared with a
sequential one. It is concluded from the current case studies
that the algorithm performs well and is capable to reveal the
sought satisficing concepts, while reducing the resources from
non-promising concepts. The comparisons with the sequential
approach suggest, as expected, that when the ratio of the
number of satisficing concepts to that of the non-satisficing
ones is decreased (first case study as compared with the second
case), the benefit of using the simultaneous search approach is
more evident.
A major element in making search algorithms efficient for
the considered type of problems is to have a good on-line
prediction technique that a concept is expected to be
satisficing. With this respect, future studies should suggest and
compare various such predictors. Future studies should also
employ additional benchmark problems with different
arrangements of the concepts fronts, more objectives and
different WOIs. Developing and comparing of other
evolutionary algorithms should also be investigated.
Finally, real-life engineering problems with increasing
number of concepts may require the development of both
parallel processing and interactive approaches. These are left
for future research.
REFERENCES
The average number of the total generations used by the
algorithm was 4,008 (which is equal to 80,160 evaluations),
whereas in the sequential runs, the total number of generation
was 8,036, which is equal to 160,720 evaluations). As
expected, with increasing number of satisficing concepts, as
compared with the previous case, the advantage of the
simultaneous approach is reduced.
[1]
Avigad, G., and Moshaiov, A. Set-based concept selection in multiobjective problems: optimality versus variability approach. J. of Eng.
Design, Vol. 20, No. 3, pp. 217-242, 2009.
[2]
Mattson C. A., Messac, A. Pareto frontier based concept selection under
uncertainty with visualization. Opt. and Eng., 6; p. 85–115, 2005.
[3]
Moshaiov, A., Snir, A., and Samina, B. Concept-based evolutionary
exploration of design spaces by a resolution-relaxation-Pareto approach,
IEEE Congress on Evolutionary Computation, pp. 1845-1852, 2015.
TABLE 6: PARALLEL VS. SEQUENTIAL RUNS FOR L=2, WOI=[0.3,0.4]
[4]
Avigad, G. and Moshaiov, A. Interactive evolutionary multiobjective
search and optimization of set-based concepts. IEEE Transactions on
Systems, Man, and Cybernetics, Part B (Cybernetics), 39(4), pp.10131027, 2009.
[5]
Purshouse, R.C., Deb, K., Mansor, M.M., Mostaghim, S. and Wang, R.
A review of hybrid evolutionary multiple criteria decision making
methods. IEEE Congress on Evolutionary Computation, 1147-1154,
2014.
[6]
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T. A fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE Transactions on
Evolutionary Computation, , vol. 6, pp. 182-197, 2002.
Concept
#
Average # of
generations for the
simultaneous runs
Average # of
generations in
sequential runs
1
2
3
4
5
6
7
8
9
Total
584
530
620
186
167
488
488
473
472
4008
1000
605
1000
1000
1000
1000
1000
1000
431
8036
V.
SUMMARY AND CONCLUSIONS
This paper describes a new type of a concept-based multiobjective problem, in which the decision-makers are interested
in finding some satisficing concepts out of possibly a very
large number of concepts. A satisficing concept is defined here
| 2cs.AI
|
1
Smart Grid Testbed for Demand Focused Energy
Management in End User Environments
arXiv:1603.06756v1 [cs.SY] 22 Mar 2016
Wayes Tushar, Chau Yuen, Bo Chai, Shisheng Huang, Kristin L. Wood, See Gim Kerk and Zaiyue Yang
Abstract—Successful deployment of smart grids necessitates
experimental validities of their state-of-the-art designs in twoway communications, real-time demand response and monitoring
of consumers’ energy usage behavior. The objective is to observe
consumers’ energy usage pattern and exploit this information
to assist the grid in designing incentives, energy management
mechanisms, and real-time demand response protocols; so as
help the grid achieving lower costs and improve energy supply
stability. Further, by feeding the observed information back to
the consumers instantaneously, it is also possible to promote
energy efficient behavior among the users. To this end, this paper
performs a literature survey on smart grid testbeds around the
world, and presents the main accomplishments towards realizing
a smart grid testbed at the Singapore University of Technology
and Design (SUTD). The testbed is able to monitor, analyze and
evaluate smart grid communication network design and control
mechanisms, and test the suitability of various communications
networks for both residential and commercial buildings. The
testbeds are deployed within the SUTD student dormitories
and the main university campus to monitor and record enduser energy consumption in real-time, which will enable us
to design incentives, control algorithms and real-time demand
response schemes. The testbed also provides an effective channel
to evaluate the needs on communication networks to support
various smart grid applications. In addition, our initial results
demonstrate that our testbed can provide an effective platform
to identify energy wastage, and prompt the needs of a secure
communications channel as the energy usage pattern can provide
privacy related information on individual user.
I. I NTRODUCTION
The smart grid is a power network composed of intelligent nodes that can operate, communicate, and interact autonomously to efficiently deliver electricity to their consumers.
It features ubiquitous interconnections of power equipments
to enable two-way flow of information and electricity so as to
shape the demand in order to balance the supply and demand
in real-time. Such pervasive equipment interconnections necessitate a full-fledge communication infrastructure to leverage
a fast, accurate, and reliable information flow in the smart
grid. In this context, the research on different aspects of smart
grid has gained significant attention in the past few years,
e.g., the literatures surveyed in [1]. Although a lot has been
done from theoretical perspectives, it is until recently when the
actual implementation of prototypes has been given deliberate
consideration. Currently, a considerable number of research
W. Tushar, C. Yuen and K. L. Wood are with the Singapore University
of Technology and Design (SUTD), Singapore.
B. Chai is with State Grid Global Energy Interconnection Research
Institute, Beijing, 102211, China.
S. Huang is with the Ministry of Home Affairs, Singapore.
S. G. Kerk is with the Power Automation, Singapore.
Z. Yang is with Zhejiang University, China.
groups are working towards establishing testbeds to validate
designs and implemented protocols related to smart grid. These
testbeds have various aims, scale, limitations and features;
these have been summarized and presented in Table I.
Most of these testbeds conduct experiments either in a
lab environment (e.g., SmartGridLab, VAST, Micro Grid
Lab, cyber-physical), or in isolation, in a residential
(e.g.,PowerMatching City) or a commercial space (e.g., Smart
Microgrid). In the testbeds surveyed, only the JEJU testbed
is comprehensive enough to consider both the residential
and commercial paradigm with user and grid interactions.
Moreover, in spite of considerable on-going studies towards
smart grid prototypes, and the massive efforts by utilities and
local authorities, deployment of smart grids has met with
near customer rejection1 . Therefore, there is much impetus on
designing and implementing well accepted solutions to bolster
consumer-grid interactions.
In this respect, this paper presents the main accomplishments towards a user-centric smart grid testbed design at
the Singapore University of Technology and Design (SUTD).
The highlight of the paper is mainly on the communication
infrastructure that has been implemented in the testbed in
order to provide ICT services to support a greener smart
grid. The testbed is simulated in an approximate real world
scenario, where the student dormitory (a 3 bedroom unit with
6 to 9 resident students) is the approximated residential space,
and the faculty offices and shared meeting rooms are proxy
commercial office spaces. The study focuses on both residential and commercial consumers and their interaction with a
central administrative body, e.g., the grid or the intelligent
energy system (IES), where the IES is an entity that provides
alternative energy management services to the consumers.
The testbed at SUTD consists of two networks: a home area
network (HAN) and a neighborhood area network (NAN). A
HAN network is implemented within each residential unit or
at SUTD office in order to collect different energy related data
via a unified home gateway (UHG). A NAN, on the contrary,
is developed to connect each HAN to the data concentrator
through various communication protocols. We provide further
detail discussion on both HAN and NAN networks that are
implemented at SUTD in Section III.
Due to the interactive nature of the system, an efficient
and reliable communication infrastructure of low delay is very
important. Therefore, this paper mainly discusses the various
communication aspects such as broadband power line com1 http://www.greentechmedia.com/articles/read/illinois-rejects-amerenssmart-grid-plans1.
2
munication cables (BPL), TV white space (TVWS), HAN and
NAN of the testbed at SUTD. Benefits that could be attributed
and examined in this system include lowering of operational
costs, increasing ancillary electricity options and better peak
demand management. We would also design and examine
suitable incentives, which would encourage consumers to
adopt these demand response mechanisms for participation in
the market by exploiting the use of the testbed.
II. T ESTBED C OMPONENTS , F EATURES AND C HALLENGES
The smart grid is envisioned as a future power network
with hundreds of millions of endpoints, where each endpoint
may generate, sense, compute, communicate and actuate. Deployment of such complex systems necessitates sophisticated
validation prior installation, e.g., by establishing smart grid
testbeds. However, the rapid fluctuation in power supply
and demand, voltage and frequency, and active consumers’
interaction with the grid make the practical realization of
testbeds extremely challenging. To this end, we classify the
suitable technologies and features of such testbeds into three
categories: 1) Hardware based components, 2) Software based
elements, and 3) Other features. We give a brief description of
different technologies and elements within each of the defined
categories and their impact on the design of a communication
network as follows.
A. Hardware based components
1) Two-way communication network: In a smart grid, a
large number of sensors, actuators, and communication devices
are deployed at generators, transformers, storage devices,
electric vehicles, smart appliances, along power lines and
in the distributed energy resources. The optimal control and
management of these nodes in real-time is essential for the
successful deployment of smart grids [2], which gives rise to
the need for fast and reliable two-way communication infrastructures. For example, depending on the online requirements
to response, reserve power is required in the smart grid in case
of an unexpected outage of a scheduled energy resource, and
required response has to be deployed within 30 seconds2 .
2) Advanced metering infrastructure (AMI): The AMI measures, collects and analyzes the energy usage, and communicates with metering devices either on requests or on a schedule.
There is an increasing concern regarding their costs, security
and privacy effects, and the remote controllable “kill switch”
included in them. A scalable secured MAC protocol becomes
critical for the AMI system.
3) HAN and NAN: In accordance with [3], we organize
the smart grid network into a HAN and NAN. A HAN
is the core component of establishing a residential energy
management system as part of a demand response management
(DRM) scheme in a smart grid. Establishing the HAN is
required to maintain an internet-of-things protocol for different
sensors/actuators that run on different physical layer communication protocols within a home. This consequently imposes
2 Reserve power can be split into primary (less than 30 seconds),
secondary (less than 15 minutes) and tertiary (15 minutes), depending on
required response times to monitor and control.
the necessity of devising gateways/controllers/central communicators, which are capable of handling these miscellaneous
communication techniques. Depending on the business model,
there could be multiple NANs. For example, AMI is setup and
maintained by the grid, and is connected to one NAN; while
third party IES (e.g. Google NEST) could be on another NAN
that provides additional services such as home automation,
energy management, or security services.
4) Renewables supply: The generation from renewable
sources such as wind and solar is intermittent and unpredictable [4], [5]. Hence, how to design the control architectures, algorithms, and market mechanisms to balance supply
and demand, support voltage, and regulate frequency in realtime are the key challenges in the presence of these volatile
green energy supplies.
5) Smart power electronics: Smart components such as
DC/AC inverters, flexible AC transmission systems and smart
switches will be increasingly deployed on various components
such as the transmission systems, transformers and power
generators in the smart grid. Hence, there are needs of solutions that will enable the design of power electronics and
control appliances to maximize reliability, efficiency and cost
reduction for the power grid.
6) Energy storage system: An energy storage system stores
electricity when the demand is low and provides the grid/IES
with electricity when the demand is high. Energy storage
systems are becoming an essential part of smart grids due to
the increasing integration of intermittent renewables and the
development of micro grids [6]. Innovative and user-centric
solutions are required to determine the type and size of energy
storage according to the application, to determine the optimal
storage location and reduce the cost of storage devices.
7) Electric vehicles (EVs): EVs can act as mobile storage
devices and assist the grid to fulfil various energy management
objectives through grid-to-vehicle (G2V) and vehicle-to-grid
(V2G) settings. Key challenges are: addressing issues like
quick response time, power sharing, EV charging protocol
design and exploiting the use of EVs storage devices for
power compensation [4], and voltage and frequency regulation
services [7].
B. Software based elements
1) DRM: The two-way flow of information and electricity
in smart grids establishes the foundation of DRM; which is in
fact the change of electricity usage patterns by the end-users
in response to the incentives or changes in price of electricity
over time. DRM can be accomplished in either a centralized or
a distributed fashion [8]. A distributed DRM algorithm greatly
reduces the amount of information that has to be transmitted
when compared to a centralized DRM algorithm. Therefore, a
smart grid testbed needs to have the capability to investigate
both DRM schemes, if possible, in order to examine their
effectiveness on the target load management.
2) Dynamic pricing: Dynamic pricing is critical for effective DRM in smart grids as a driver for behavior change [9].
The pricing scheme needs to be beneficial to the grid in
terms of reduction in operational cost, energy peak shaving
3
TABLE I: Attributes of Different Smart Grid Testbeds.
Testbeds
Set-up
Adopted Communication Technology
Main Innovations
Components and
Features
(From Section II)
SmartGridLab
(Washington State University)
In a Lab environment.
IEEE 802.15.4 wireless technology is being used, which
is the basis for ZigBee, ISA100.11a, WirelessHART and
MiWi.
Intelligent power switch control and real-time demand response and
self-healing, energy storage and flow balance algorithms.
A.1, A.2, A.6, B.1,
B.4
Cyber-physical
(University of Illinois at UrbanaChampaign)
In a Lab environment.
Have not used any particular communication technology.
A multi-platform protocol dissector library, an AMI visualization
framework and Advanced Metering Infrastructure (AMI) security.
A.1, A.2, B.4
Enable optimal operation of a
university campus power
system.
IEC-61850 based bi-directional communication link is
used at the BCIT Intelligent micro-grid.
Control strategies for efficient integration and operation of on-site
DER units, energy storage and implementation of fault detection,
isolation and restoration protocols.
A.1, A.4, A.5, A.6
Virtual communication link and Programmable inverter control,
identification of false or incomplete data, and high latency
coordination of response to events such as a fault or
coupling/decoupling to an external grid.
A.1, A.4, A.5, B.4
A generic coordination mechanism, comfort control mechanism, and
monitoring.
A.1, A.3, C.3
Energy monitoring and management in a smart building,
environmental monitoring, demand response, web development and
smartphone application for control energy usage.
A.1, A.2, A.3, B.1,
C.3
Power transmission using IT with solar and wind power, energy
storage, optimizing energy usage by utilizing new and renewable
energy sources, collecting real time data on energy usage and demand
to limit the unnecessary use of electricity.
A.1, A.2, A.3, B.1,
A.4, A.5, A.6, A.7,
B.2, C.3
Fiber-to-home communication system.
Identification of information security threats, energy control in smart
house, use of wind, solar and bio-fuel as energy sources, monitor load
profiles and electricity production in zero energy house and
determining user flexibility.
A.3, A.4, B.4, C.3
Have not used any specific communication technology.
Efficient use of energy storage, integration of distributed generation
and customer engagement, charging of electric vehicle and demand
response management.
A.4, A.6, A.7, B.1,
C.3
For communication within a home area network: UHG
to communicate and control different household
appliances via different communication protocols: WiFi, ZigBee, Z-wave, Bluetooth
For communication from cloud to the home or smart
meter: broadband PLC, TV White Space
Monitoring and recording real-time energy usage data to encourage
and control the energy usage of consumers, designing and evaluating
communication technology for end user environment, centralized and
distributed demand response in real-time, designing incentives for
consumers and dynamic pricing.
A.1, A.2, A.3, B.1,
B.2, B.4, C.1, C.2,
C.3
Smart Microgrid
(British Columbia Institute of
Technology)
VAST
(Georgia Institute of Technology)
In a Lab environment.
PowerMatching City
(The Netherlands)
Demonstration consists of 25
households in the city of
interest.
Smart Building Platform
(International Hellenic University
(IHU), Greece)
Testbeds are set up at IHU
campus.
JEJU testbed
(Korea)
National Smart Grid Lab
(Norway)
Newcastle University testbed
(UK)
Smart grid testbed
(SUTD, Singapore)
A virtual communication link is used that allows the
testbed to implement any type of communication
protocol required. Reprogramming the system using
LABVIEW does the implementation of different
protocols.
Data communication between EV and the grid of
PowerMatchingCity is done through VPN router over
UMTS. Communication between the households and the
central servers is provided by VPN network by a
separate ADSL connection.
Open and encrypted ZigBee, Z-wave, Wi-Fi and Radio
Frequency (RF)
Customer domain- SEP, ZigBee, WiFi, PLC;
Transmission and Distribution: IEC 61850, DNP 3.0;
Testbeds are set up for both
Operations Domain IEC 61968/61970; AMR; Public and
residential and commercial
Private Internet; and Cellular based (CDMA, WCDMA),
spaces at JEJU Island in Korea.
WiMAX (IEEE 802.16)
Testbeds are set up to offer
realistic physical model grids
and interface with generation
and loads.
Trial of new storage
technologies to efficiently and
sustainably deliver energy
across the UK electricity grid.
Experimental set up includes
both student dormitories and
main university campus.
and valley filling. It also needs be economically attractive to
consumers by reducing their electricity bills and should not
cause significant inconvenience to them for changing their
energy consumption behavior, i.e., for scheduling or throttling.
Securing the pricing information is important to protect against
malicious attacks, e.g., an outdated pricing information is
injected to destabilize the smart grid.
3) Outage management: In a time of crisis, the capacity
for fault detection and self-healing with minimal restoration
time and improved efficiency, i.e., outage management [10] is
important. Hence, with real-time insight, the smart grid testbed
should possess automatic key processes to easily locate and
route power around affected spots to reduce unnecessary truck
rolls and save costs.
4) Security and privacy: Cyber security is a pivotal challenge for smart grid. In fact, cyber-based threats to critical
infrastructure are real and increasing in frequency. However,
the testing of potential threats are challenging due to the
general lack of defined methodologies and prescribed ways
to quantify security combined with the constantly evolving
threat landscape. Moreover, as the participation of consumers
in smart grid becomes more prevalent, privacy information,
e.g., their home occupancy, will be more vulnerable. Hence,
applied methods and rigorous privacy and security assessments
are necessary for complex systems where heterogeneous components and their users regularly interact with each other and
the IES.
C. Other features
1) Ancillary Services: Regulation power allows for the grid
to manage second-to-second fluctuations from forecasted demand and reserve markets enable electricity providers to have
instant-on solution to kick in when power delivery problems
emerge. Theoretically, smart grids are able to provide such
services through DRM [11]. Therefore, smart grid solutions
need to enable the grid providers to combine the DRM
schemes with energy storage solutions to make the grid system
more reliable.
2) Scalability: A smart grid system may consists of millions of active nodes, which need to be managed in realtime. Accordingly, optimal economic mechanisms and business models are required for engendering desired global outcomes. However, it is extremely difficult to maintain such a
large scale and complex cyber-physical system in real-time.
Besides, features like automatic fault detection, self-healing,
autonomous distributed demand-side management and disaster
management make the conservation more strenuous. Hence,
innovations are required to improve the scalability of such a
huge system without affecting any of its features. For instance,
instead of using different gateways for communicating with
different equipment in a home (where each devices may run
on different communication protocols), devising a single UHG
could make a HAN considerably scalable. A scalable MAC
protocol would be interesting to explore in dense residential
areas, where over thousands of machine-type devices such as
smart meters communicate at the same time [12].
4
Fig. 1: Overview of SUTD testbed consisting of NAN and HAN.
3) User acceptability and participation: One of the key
challenges for the success of smart grids is to be able to clearly
elucidate the benefits for the consumers [13]. One solution
could be a clear user friendly interface (e.g., smart phone apps)
to demonstrate the ability to visualise electricity usage and
real-time electricity prices, so as to ease energy savings for
the consumers through smarter load shifting.
It can be noticed that advanced ICT technologies and
communications networks play important roles in various
applications, features, and even in the acceptance of a smart
grid. To this end, we now discuss the main features of the
testbed at SUTD in the next section.
grid so as to address some of the challenges mentioned in
Section II is especially critical. The interaction between the
consumers and the grid/IES has been prioritized in this work,
which leads us to divide the whole smart grid testbed into two
networks: 1) NAN and 2) HAN. The IES set up the HAN,
which is connected to the cloud directly though broadband
internet, while the grid will set up the NAN that connects the
smart meters. However, if the IES is provided by the grid, then
the HAN and the UHG would certainly be connected under
the same network. To keep the structure general, we consider
them to be separated in the SUTD testbed. To this end, what
follows are the brief descriptions of the NAN and HAN, and
the associated technologies used in the SUTD testbed.
III. SUTD T ESTBED OVERVIEW
The aims of the SUTD smart grid testbed are to develop
innovative technical approaches and incentives for smart energy management, through the integration of commercial space
strategies, residential space strategies, and the design and
development of supporting technologies. Hence, testing the
suitability of various communications technologies for smart
A. NAN
The NAN of the testbed is composed of a number of HAN
units from either the SUTD dormitories or the university
main campus and the cloud. In our setup, we employ two
technologies for NAN: one is BPL and TVWS for smart
meters setup by the grid, and another is the traditional fiber
5
TABLE II: The technical specifications of TVWS network deployed at SUTD
campus.
/ cable internet, where all the UHGs are connected via WiFi access points, which could be setup by a 3rd party IES.
In Fig. 1, we show a schematic diagram of how the NAN
is set up within the SUTD testbed. Now, we provide a brief
description of the used communications technologies, i.e., BPL
and TVWS, in the following subsections.
1) BPL: In the testbed, the communication between the
smart meter outside each residential unit and the data concentrator at the top of the apartment building is based on
BPL. This choice of communication between them is simple
and obvious as no additional cabling is needed in connecting
all the units within an apartment block. It provides superior
performance across thick walls as compared to any other
types of communications and simplified the entire network
management of smart meters.
2) TVWS: At SUTD, TVWS is used for linking the data
concentrator from every apartment block to the base-station
before data is uploaded to the cloud. In Table II we show
the detail specification of TVWS that we have used at SUTD.
Currently, Infocomm Development Authority (IDA) is setting
up trials and the standardization of TVWS in Singapore.
Please note that TVWS is a freely available spectrum and thus
provides a great opportunity for many new small and virtual
operators to setup their networks for M2M applications at low
cost. In this testbed, building a NAN on TVWS spectrum
also provides more operational flexibility that can easily be
scaled up by adding additional Base Stations (BSs). Therefore,
for NAN wireless access provides good solution and wide
coverage without much installation cost. Furthermore, TVWS
(unlicensed band) can greatly reduce the cost, as compared to
licensed band and fixed access.
Our testbed based on BPL and TVWS provides a dedicated
and separated network from the Internet cabling, which is
also used for web surfing or video streaming. It provides an
opportunity to test the network delay and reliability for smart
grid applications, and can be compared against the HAN that
will be discussed next.
B. HAN
While all participating units from the dormitories and the
campus form the NAN together, each unit is equipped with its
own HAN inside the house as shown in Fig. 2. A UHG is the
central point of attention of a HAN, through which the IES
can monitor the energy usage pattern or control each of the
equipments of the unit within the network. Here, on one hand,
the gateway adopts two-way communication protocols such as
ZigBee, Z-wave and Bluetooth to connect with equipments and
sensors in the house. On the other hand, it is connected to the
IES through Wi-Fi for sending the monitored information to
the IES.
1) UHG: A novel aspect of the SUTD testbed is the
introduction of a UHG in each of the HAN unit of both
residential and commercial space areas, built on a Raspberry
Pi computer (Model-B Rev 1) as shown in Fig. 2. Please note
that the HAN network consists of different type of smart plug
and sensors that may need to communicate with each other. In
this context, Universal home gateway (UHG) is developed and
used in the testbed to facilitate such heterogeneous communication between different sensors and smart plugs. The UHG
is capable of handling a number of communication protocols
including Z-wave, ZigBee and Bluetooth for communicating
with different smart devices such as smart plugs and various
sensors. The UHG directly transmits the monitored data from
a HAN unit to the IES through Wi-Fi. The challenge of such
a UHG is that it needs to communicate with devices (e.g.
smart-plug / sensors) of various protocols. The delay between
the commands from the IES to the UHG and from the UHG
to the end devices could also pose challenges depending on
the applications. It however provides a scalable solution that
enables thousands or even ten thousands of households in the
system. In the SUTD testbed, we have implemented Restful
HTTP and XMPP protocol for the communication between
the UHG and the cloud of the IES. Hence, the UHG serves
as a gateway for many non-Internet Protocol based sensors to
connect to the Internet.
2) ZigBee: ZigBee is a low-cost and low-power wireless
mesh network specification, which is built on the top of the
IEEE 802.15.4 standard for low-rate wireless personal area
networks (WPANs), and operates in the ISM frequency band
2.4 GHz. ZigBee is essentially a non-IP based communication
protocol and thus suitable for communication in testbeds set
up in a non-residential setting. This is due to the fact that
in SUTD campus (which is also the same as many other
commercial/industrial network), the Wi-Fi network has very
strict security, where the password is changed on regular basis
(e.g., in every three months at SUTD campus). However, it is
significantly difficult to change the password of every sensor
in every three months time. Therefore, non-IP based solution
is required and we use ZigBee in the MPN to serve that
purpose. Besides, for a multi-hop extension considering the
office scenario, we require a protocol that can support multihop to extend the coverage to reduce the needs of a gateway
(within a house, however, single hope with UHG is usually
good enough). Furthermore, ZigBee provides security (e.g.
the ZigBee provides basic security where only nodes of the
same network id can join in the network) and the simplicity
of having two-way communication facilities.
The multi-purpose node (MPN) designed in SUTD, as
shown in Fig. 2, consists of a number of low power sensors and
6
Fig. 2: Home Area Network.
actuators. It is equipped with a motion detector, a temperature
and humidity sensor, a noise sensor and a Lux sensor. It
communicates with the UHG through its Xbee Chip, which
is a ZigBee communication module, to send the monitored information. The MPN is capable of controlling devices through
its actuators like an IR Blaster (to control the air conditioning
system (ACS)) and potentiometer (to control the LED light
power supply). The MPN is driven by an Arduino Fio micro
controller, which is installed within the MPN. In every hostel
unit in SUTD, each HAN consists of 4 MPNs, one for each
bedroom, and one for the living room. While in the office,
there are 20 MPNs (one for each faculty office) in one section
of an office block connected through multiple ZigBee relays
to a single UHG. The UHG also fulfils an important role later
on, when distributed DRM is implemented, such that users
occupancy and electric appliance usage info can be stored
locally without sending to the IES to protect users’ privacy.
3) Z-wave: Z-wave is a wireless communication protocol
around 900 MHz that uses a low power technology for home
automation. For instance, to monitor and remotely control the
energy consumption of different equipments in the HAN, we
have used a Z-wave smart-plug for each of the devices in the
unit. In the current set-up, electric appliances to be monitored
are connected to power outlets through Z-wave smart-plugs,
whereby each smart-plug is connected to the UHG through a
RaZberry module. The plug is able to monitor the amount
of energy consumed by the connected device in real-time
and instantaneously send that information to the UHG via Z-
wave communication. Further, it also has a remote actuation
capability e.g., delaying an electric appliances due to peak in
energy demand. Note that the switching signals of smart-plugs
are initiated through either the IES (for centralized control) or
the UHG (for distributed control).
4) Bluetooth: We use a Bluetooth Low Energy (BLE)
sensor based on Texas Instrument’s (TI’s) CC2541 MCU to
monitor the temperature, humidity and pressure level of its
surroundings, and to communicate with the UHG via bluetooth. The BLE sensor is also equipped with an accelerometer,
a gyroscope and a magnetometer that provide the UHG with
information like acceleration, orientation and magnetization of
an object respectively. In the course of the experiments, we
will distribute such TI sensors among the students so that they
can design new Internet-of-Things product that can connect to
our UHG and provide new green energy services in smart
homes.
5) Mobile application: We have developed a mobile application3 to install in students’ and faculties’ cellphones to
engage them in the energy management experiments at SUTD.
For example, they can remotely switch on and off the smart
plug or the ACS, or monitor the energy usage collected by the
testbed. Dynamic pricing information can also be pushed to
them over the application.
3 The
figure is not given due to constraint on the number of figure.
7
C. Challenges
During the experiments in the testbed, we face a number of
challenges including communication delay, subject recruitment
for case studies, modeling non-homogeneous scenarios and
investigating the mismatch between the theoretical results and
the outcomes from the experiments.
First, communication delay is one of the challenges that
we have encountered while running the experiments. We
understand that there is a trade-off between the sampling rate
and communication delay, and the delay can be reduced by
increasing the rate of sampling. Delays can also be caused
by the time required for data savings and retrieval at the
server. For example, if a server needs to retrieve data from
large number of sensors at the same time, it may incur some
delays [14]. However, one potential way to reduce such delay
is to efficiently design the database and use more customized
script for increasing the processing speed. Nevertheless, such
delay is beyond the scope of this paper. Now considering the
response times of primary, secondary and tertiary reserves,
our implemented centralized monitor and control scheme can
comfortably use for secondary (less than 15 minutes) and
tertiary reserve (15 minutes). However, it might not be suitable
for primary reserve requiring very fast response time, e.g., less
than one minute. Implementing distributed control instead of
centralized and thus enable each node under observation to
respond very fast, if necessary, could be a potential way to
resolve this issue.
Second challenge is to recruit the subjects for experiments
who will allow us to install sensors and communication
modules in their rooms for monitoring and control purposes.
Although the experiments are completely Institutional Review
Board (IRB) approved and are designed to conduct in an
academic environment, it is hard to find participants for the
experiments. This is due to the fact that the participants are
concerned about their privacy and worry about the inconvenience that may create due to energy management. Further,
students lifestyle is significantly different from the lifestyle of
typical residents. In this context, as a token of encouragement
for participation, we have provided monetary incentives to
each of the participants at the end of the study.
After the rooms were chosen to set up the testbed, the
third challenge was to interpret the readings from the sensors
installed within the rooms. This is due to the fact that each
user has different preferences and hence sets the sensors at
different locations of the rooms. For instance, some may keep
the sensors near the window whereby some prefer to keep
them far from the windows. Similarly, some prefer to always
keep their window curtains open whereas some prefer them to
remain close most of the time. As a consequence, the readings
on temperature, light, noise, and motion from the sensors have
the possibility to be very different although they may represent
the same environment. Hence, there is a need to consider such
behavioral differences in our experiments, and we have tried
to differentiate the sensors data based on the context. For
example, we perform some sort of learning before interpreting
the readings from sensors.
Finally, we find it difficult to match the theoretical load
consumption model with the model that we derive from the
experiments. For example, let us consider the case of air
conditioning systems (ACSs). Although there are many studies
on the energy consumptions of ACS in the literature, it is
extremely difficult to find a study that has used a setup as
same as the one we use, e.g., in terms of ACS type, room size
and weather. Therefore, algorithms that are designed based on
the theoretical model may not behave as expected in practical
scenarios. As a result, additional steps are needed to build
a model based on practical system setup, and building such
model could be time consuming and very customized to a
practical setting.
After discussing various challenges, now we will discuss
the experiments that are currently (or will be) implemented at
SUTD in the next section to give a glimpse of the competence
of this extensive testbed setup.
IV. O N - GOING E XPERIMENTS
Most of the experiments using the testbed are deliberately
related to electricity management either through direct control
of electricity from the IES or by designing incentives that will
encourage behavioral modification towards efficient energy
use. Our experiments at the SUTD testbed are conducted
1) within the student hostels, and 2) at the SUTD campus.
Based on their electricity usage pattern, the SUTD campus
is considered as the commercial space in our experiments,
whereby student hostels are thought of as residences.
A. SUTD student hostels
The energy consumed by residences differ based on the
set of home appliances, standard of living, climate, social
awareness and residence type. We use our testbed to develop
solutions for residential energy management to attain different
objectives.
1) Ancillary electricity markets: In the SUTD testbed, we
are interested in determining the potential of residential smart
grid participation in ancillary markets. Real-time ancillary
markets allow participants to supply both regulation and
reserve power. Reserve power is generation capacity that is required during an unexpected outage of a scheduled generating
plant and we are concerned with whether the residential users
can provide the grid with either primary or secondary reserve
power. In most systems, discretionary loads can voluntarily
participate in these markets as interruptible loads, willing to
“load-shed” in an exceptional event; however, current market
rules only allow participation of big loads. In our system, the
IES can aggregate smaller distributed loads and participate
in these markets as a unified entity. Although theoretically
possible, the practical implementation with consideration of
practical limitations have not been fully examined in full
scale testbeds, especially the communication constraints. With
the SUTD testbed, we will verify its feasibility, and the
requirements on the communication network in supporting
such an application.
8
Energy
consumption
40
Energy
consumption
Time
Total power (kW)
35
30
25
Total load profile wihout DRM
Base load profile
Total load profile with DRM
Peak limit
20
Time
15
0
3
6
9
12
15
18
21
24
Energy
consumption
Time (Hours)
Fig. 4: Illustration of energy management scheme in the 10 hostel units in
order to reduce peak load.
Time
Fig. 3: Demonstration of the real-time energy consumption by different home
appliances as obtained from the deployed testbed. The horizontal axis refer
to different time of the day whereby the vertical axis represent the energy
consumption by different appliances.
2) Dynamic pricing: Using the implemented testbed, we
also focus on designing dynamic pricing schemes for residential users. Different design objectives such as penetration
rates [8], flexibility thresholds and flexibility constraints will
be considered. As an initial step towards this development, a
mobile application is being developed for the residential users
at SUTD that can periodically notify them about their realtime energy consumption amount and the associated price per
unit of energy. We will use such dynamic pricing information
to achieve certain DRM goals and to promote energy saving
awareness. The smart meter is installed in parallel with original
meter, such that the smart meter can be used to simulate the
electricity price based on dynamic pricing, while the original
meter by the grid can continue provide meter readings for
actual electricity billing.
3) DRM under communication impairment: Real-time
DRM can be affected significantly by the quality of the
communication network of the smart grid. In this context,
we are conducting experiments to observe the effect of the
status of communication network on DRM. For instance, if
the communication network is congested, it would result in
packet losses or delay. We are interested in investigating how
this congestion of the network affects the DRM in terms of
delay and consequently the cost, reduction of peak load and
energy savings.
4) DRM design based on preferences and device characteristics: The installed smart-plugs and MPNs in each HAN unit
at the SUTD student hostel enables real-time monitoring of
energy consumption of different appliances as shown in Fig 3.
This provides insights on the students personal preferences
and allows to feed the energy consumption data back to the
consumers in real-time, which along with the incentive based
DRM policies would enable the users to modify their usage
pattern towards more efficient energy management.
Remark 1. It is important to note that our testbed is to
monitor the energy usage. However, each appliance has a
unique characteristic profile as shown in Fig. 3, which can be
used to identify what appliance that is, and in return, speculate
the activity and number of users. This prompts the need for
secure communication.
Now to demonstrate how the proposed testbed can be used
to perform energy management under practical communication
impairment, we run experiment for peak load shaving in
our testbed. To do so, we have considered 10 home units
where each unit is equipped with two flexible loads (we
use lights as flexible loads that provide on/off control). We
model the daily base load according to the reports of National
Electricity Market of Singapore (NEMS) and scale according
to the energy rating of light bulbs such that we can obtain a
significant percentage of flexible load. We assume a threshold
for total maximum allowable load consumption by all the
units and design an algorithm to control the flexible load of
each home units in order to keep the energy consumption
always below the threshold. The algorithm is based on a
centralized control scheme in which the energy management
service provider conduct demand management by switching
off several appliances of the users to maintain the total
power demands below the given threshold. As designed, the
energy management service provider considered not only the
engagement of users in the demand response but also the
inconvenience of users when the demand response protocol is
conducted. We find that the demand response can be affected
by communication delay. The detail of the algorithm and
assumption of the experiment can be found in [15].
The results from the experiment are demonstrated in Fig. 4.
In Fig. 4, the green zone and blue zone of the figure denote
the load profile with and without demand management respectively whereby grey zone indicate the base load and red line
show the maximum allowable peak demand threshold (which
is 33 kW in the considered case). Now, as demonstrated in
Fig. 4, when the total demand is under peak limit, there is no
need to do any controlling and hence no difference is observed
between green zone and the blue line. However, once the
total demand exceed the peak limit, the algorithm is executed
and thus controls the flexible loads in each home unit and
turns some of them off to reduce demand load promptly as
9
TABLE III: Demonstration of how the monitoring of users occupancy and
energy use of lights and ACS in eight office rooms at SUTD are done through
the implemented testbed in order to reduce energy wastage. The total wastage
is calculated in terms of both kWh and SGD.
indicated by the blue line. It is important to note that the
algorithm continuously run in the backend and monitors for
situations when the total demand may exceed the threshold.
Once such situation arises, the algorithm switches off some
of the appliances to keep the demand within the threshold.
However, the algorithm, as it is designed, considers not only
the demand control but also the associated inconvenience
that the users may experience for such control. Therefore,
control of appliances is always kept at as low as possible in
maintaining the demand. Furthermore, while controlling some
appliances for handling excess demand, there could be new
loads switched on by the users that can also contribute to the
overall demand. As a consequence, there is no sudden decrease
is observed in the blue line as can be seen from Fig. 4. Thus,
the result in Fig. 4 clearly shows that our developed testbed
is effective to preform energy management applications.
B. SUTD campus
We are also interested in studying energy management in
commercial buildings with a view to reduce energy costs for
all stakeholders. To this end, we are currently conducting the
following experiments at the SUTD campus.
1) DRM in office environments: One of our ongoing experiments is to implement DRM schemes for offices that considers
the real-time control of ACSs. The objectives are mainly
two-fold: 1) real-time thermostat control to manage peak
demand, and 2) optimal management of ACS demand under
dynamic pricing for energy cost management. To support
this, we are developing energy consumption models and realtime control algorithms for ACSs. The testbed allows us to
monitor and collect the data on consumed power by the ACSs
under different permutations of indoor and outdoor conditions.
These are pivotal for algorithm development and ACS energy
consumption modeling.
2) Meeting scheduling: In offices, usually there exists a
large number of meeting rooms that are used only occasionally.
Hence, the scheduling of their usage is a possible way to
shape the energy consumption of commercial buildings. We
formulate a comprehensive study, where for a given set of
meeting room requests over a fixed time period and a set
of meeting rooms in a dynamic pricing market, the meeting
scheduler finds a feasible routine that minimizes the total
energy consumption (and/or cost). The SUTD testbed is used
for obtaining the real-data, and to test the scheduler in order
to verify the effectiveness of the optimal meeting scheduling
protocol.
Now, in order to save energy and associated costs of
individual office rooms, we experiment to identify the potential
wastage in an office environment. To do so, we have deployed
a number of sensors, i.e., the MPN of Fig. 2 in each of the
offices. For instance, if the sensors do not detect any motion
or noise in a room for a predefined period of time, we can
assume that there is no one inside the room. A record of such
observation is shown in Table III, which shows the duration of
turned-on lights and ACSs of eight different office rooms when
there are no occupants. According to Table III, the energy
wastage in most of the considered rooms is significant in
absence of the occupants, which is on average 6.376 kWh
and 232.025 kWh per room from lights and ACS respectively.
Now, if we assume that the average wastage is the same
for all 200 office rooms of a typical medium size university
campus, the total energy that wasted during the considered
time duration of 54 days can be estimated as 47, 680.2 kWh.
Now considering the electricity rate in Singapore, which is
calculated via Home Electricity Audit Form available online4 ,
this wastage can be translated into a total of around 11, 100
SGD from the considered 200 rooms during a period of 54
days. Therefore, if the proposed testbed can be used in offices
for the considered setting, about 11, 100 SGD can be saved
from wastage. The saving will be even more if the electricity
price is dynamic, as the office hour is typically overlap with
day time peak period. One can then perform on/off control to
the ACSs (such as in Fig. 4) to achieve energy saving or peak
load shaving.
V. C ONCLUSION
In this paper, we have discussed some aspects on the development of the smart grid testbed at the Singapore University
of Technology and Design (SUTD). These testbeds within
the university campus have been setup to approximate both
residential and commercial spaces. We have discussed the
general components, features and related challenges of such
a testbed implementation with emphasis on the creation of
the communication system architecture. Ongoing experiments
at the testbed include the harnessing of detailed data streams
from the testbed for different customized energy management
programs such as demand response. The flexibility and extensivity of the deployed testbed allows for the research team
to explore and implement effective and practical communications technologies and smart grid applications, which could
ultimately increase the acceptance and adoption rate of these
systems.
ACKNOWLEDGEMENTS
This work is supported in part by the Singapore University
of Technology and Design through the Energy Innovation
Research Program Singapore under Grant NRF2012EWTEIRP002-045, in part by the SUTD-MIT International Design
4 The cost can be calculated by using the Home Electricity Audit Form available online for public use in https://services. spservices.sg/elect audit frameset.asp.
10
Center, Singapore under Grant IDG31500106 and in part by
NSFC-61550110244.
[8]
R EFERENCES
[1] X. Fang, S. Misra, G. Xue, and D. Yang, “Smart grid - The new and
improved power grid: A survey,” IEEE Commun. Surveys Tuts., vol. 14,
no. 4, pp. 944–980, Oct 2012.
[2] S. Maharjan, Q. Zhu, Y. Zhang, S. Gjessing, and T. Başar, “Dependable
demand response management in the smart grid: A Stackelberg game
approach,” IEEE Trans. Smart Grid, vol. 4, no. 1, pp. 120–132, Mar
2013.
[3] R. Yu, Y. Zhang, S. Gjessing, C. Yuen, S. Xie, and M. Guizani,
“Cognitive radio based hierarchical communications infrastructure for
smart grid,” IEEE Netw., vol. 25, no. 5, pp. 6–14, Oct 2011.
[4] W. Tushar, C. Yuen, S. Huang, D. Smith, and H. V. Poor, “Cost
minimization of charging stations with photovoltaic: An approach with
EV classification,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 1, pp.
156–169, Jan. 2016.
[5] A. Naeem, A. Shabbir, N. U. Hassan, C. Yuen, A. Ahmed, and
W. Tushar, “Understanding customer behavior in multi-tier demand
response management program,” IEEE Access (Special issue on Smart
Grids: A Hub of Interdisciplinary Research), vol. 3, pp. 2613–2625,
Nov. 2015.
[6] I. Atzeni, L. G. Ordóñez, G. Scutari, D. P. Palomar, and J. R. Fonollosa,
“Demand-side management via distributed energy generation and storage optimization,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 866–876,
June 2013.
[7] Y. Liu, C. Yuen, S. Huang, N. Ul Hassan, X. Wang, and S. Xie, “Peakto-average ratio constrained demand-side management with consumer’s
[9]
[10]
[11]
[12]
[13]
[14]
[15]
preference in residential smart grid,” IEEE J. Sel. Topics Signal Process.,
vol. 8, no. 6, pp. 1084–1097, Dec 2014.
N. U. Hassan, M. A. Pasha, C. Yuen, S. Huang, and X. Wang, “Impact of
scheduling flexibility on demand profile flatness and user inconvenience
in residential smart grid system,” Energies, vol. 6, no. 12, pp. 6608–
6635, Dec 2012.
T. Jiang, Y. Cao, L. Yu, and Z. Wang, “Load shaping strategy based on
energy storage and dynamic pricing in smart grid,” IEEE Trans. Smart
Grid, vol. 5, no. 6, pp. 2868–2876, Nov 2014.
L. Yu, T. Jiang, and Y. Cao, “Energy cost minimization for distributed
internet data centers in smart microgrids considering power outages,”
IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 1, pp. 120–130, Jan
2015.
W. Tushar, B. Chai, C. Yuen, D. B. Smith, K. L. Wood, Z. Yang, and
H. V. Poor, “Three-party energy management with distributed energy
resources in smart grid,” IEEE Trans. Ind. Electron., vol. 62, no. 4, pp.
2487–2498, Apr. 2015.
Y. Liu, C. Yuen, X. Cao, N. Hassan, and J. Chen, “Design of a scalable
hybrid mac protocol for heterogeneous M2M networks,” IEEE Internet
Things J., vol. 1, no. 1, pp. 99–111, Feb 2014.
W. Tushar, J. A. Zhang, D. B. Smith, H. V. Poor, and S. Thiébaux,
“Prioritizing consumers in smart grid: A game theoretic approach,” IEEE
Trans. Smart Grid, vol. 5, no. 3, pp. 1429–1438, May 2014.
Y. Cao, T. Jiang, M. He, and J. Zhang, “Device-to-device communications for energy management: A smart grid case,” IEEE J. Sel. Areas
Commun., vol. 34, no. 1, pp. 190–201, 2016.
W.-T. Li, C. Yuen, N. U. Hassan, W. Tushar, and C.-K. Wen, “Demand response management for residential smart grid: From theory
to practice,” IEEE Access (Special issue on Smart Grids: A Hub of
Interdisciplinary Research), vol. 3, pp. 2431–2440, Nov. 2015.
| 3cs.SY
|
Repairing Ontologies via Axiom Weakening
Nicolas Troquard, Roberto Confalonieri, Pietro Galliani, Rafael Penaloza, Daniele Porello, Oliver Kutz
arXiv:1711.03430v1 [cs.AI] 9 Nov 2017
KRDB, Unibz, Bozen-Bolzano, Italy
Abstract
Ontology engineering is a hard and error-prone task, in which
small changes may lead to errors, or even produce an inconsistent ontology. As ontologies grow in size, the need for automated methods for repairing inconsistencies while preserving as much of the original knowledge as possible increases.
Most previous approaches to this task are based on removing a few axioms from the ontology to regain consistency.
We propose a new method based on weakening these axioms
to make them less restrictive, employing the use of refinement operators. We introduce the theoretical framework for
weakening DL ontologies, propose algorithms to repair ontologies based on the framework, and provide an analysis of
the computational complexity. Through an empirical analysis made over real-life ontologies, we show that our approach
preserves significantly more of the original knowledge of the
ontology than removing axioms.
Introduction
Ontology engineering is a hard and error-prone task, where
even small changes may lead to unforeseen errors, in
particular to inconsistency. Ontologies are not only growing
in size, they are also increasingly being used in a variety
of AI and NLP applications, e.g., (Bateman et al. 2010;
Prestes et al. 2013). At the same time, methods to generate
ontologies through automated methods gain popularity:
e.g., ontology learning (Lehmann and Hitzler 2010;
Sazonau, Sattler, and Brown 2015),
extraction
from
web resources such as DBpedia (Auer et al. 2007), or
the combination of knowledge from different sources
(Stuckenschmidt, Parent, and Spaccapietra 2009).
Such ontology generation methods are all likely to require
ontology repair and refinement steps, and trying to repair an
ontology containing hundreds, or even thousands of axioms
by hand is infeasible. For these reasons, it has become fundamental to develop automated methods for repairing ontologies while preserving as much of the original knowledge as
possible.
Most
existing
ontology
repair
approaches
are based on removing a few axioms to expel
the
errors
(Schlobach and Cornet 2003;
Kalyanpur et al. 2005;
Kalyanpur et al. 2006;
Baader, Peñaloza, and Suntisrivaraporn 2007).
While
these methods are effective, and have been used in practice,
they have the side effect of removing also many potentially
wanted implicit consequences. In this paper, we propose
a more fine-grained method for ontology repair based
on weakening axioms, thus making them more general.
The idea is that, through this weakening, more of the
original knowledge is preserved; that is, our method is less
destructive.
We show, both theoretically and empirically, that axiom
weakening is a powerful approach for repairing ontologies.
On the theoretical side, we prove that the computational
complexity of this task is not greater than that of the standard reasoning tasks in description logics. Empirically, we
compare the results of weakening axioms against deleting
them, over existing ontologies developed in the life sciences. This comparison shows that our approach preserves
significantly more of the original ontological knowledge
than removing axioms, based on an evaluation measure inspecting the preservation of taxonomic structure (see e.g.,
(Alani, Brewster, and Shadbolt 2006; Resnik 1999) for related measures).
The main result of this paper is to present a new ontology repair methodology capable of preserving most of
the original knowledge, without incurring any additional
costs in terms of computational complexity. By thereby preserving more implicit consequences of the ontology, our
methodology also provides a contribution to the ontology
development cycle (Neuhaus et al. 2013). Indeed, it can be
a useful tool for test-driven ontology development, where
the preservation of the entailment of competency questions from the weakened ontology can be seen as a measure for the quality of the repair (Grüninger and Fox 1995;
Ren et al. 2014).
We begin by outlining formal preliminaries, including the
introduction of refinement operators and a basic analysis of
properties of both, specialisation and generalisation operators. This is followed by a complexity analysis of the problem of computing weakened axioms in our approach. We
then present several variations of repair algorithms, a detailed empirical evaluation of their performance, and a quality analysis of the returned ontologies. We close with a discussion of related work and an outlook to future extensions
and refinements of the presented ideas.
Preliminaries
From a formal point of view, an ontology is a set of formulas in an appropriate logical language with the purpose of describing a particular domain of interest. The precise logic used is in fact not crucial for our approach as
most techniques introduced apply to a variety of logics;
however, for the sake of clarity we use description logics (DLs) as well-known examples of ontology languages.
We briefly introduce the basic DL ALC; for full details
see (Baader et al. 2003). The syntax of ALC is based on two
disjoint sets NC and NR of concept names and role names,
respectively. The set of ALC concepts is generated by the
grammar
C
::=
A | ¬C | C ⊓ C | C ⊔ C | ∀R.C | ∃R.C ,
where A ∈ NC and R ∈ NR . A TBox is a finite set of
concept inclusions (GCIs) of the form C ⊑ D where C and
D are concepts. It is used to store terminological knowledge
regarding the relationships between concepts. An ABox is a
finite set of formulas of the form C(a) and R(a, b), which
express knowledge about objects in the knowledge domain.
The semantics of ALC is defined through interpretations
I = (∆I , ·I ), where ∆I is a non-empty domain, and ·I is
a function mapping every individual name to an element of
∆I , each concept name to a subset of the domain, and each
role name to a binary relation on the domain. The interpretation I is a model of the TBox T if it satisfies all the GCIs
in T . Given two concepts C and D, we say that C is subsumed by D w.r.t. the TBox T (C ⊑T D) if C I ⊆ DI for
every model I of T . We write C ≡T D when C ⊑T D and
D ⊑T C. C is strictly subsumed by D w.r.t. T (C ⊏T D) if
C ⊑T D and C 6≡T D.
EL is the restriction of ALC allowing only conjunctions
and existential restrictions (Baader, Brandt, and Lutz 2005).
It is widely used in biomedical ontologies for describing
large terminologies since classification can be computed in
polynomial time.1 In the following, DL denotes either ALC
or EL, and L(DL, NC , NR ) denotes the set of (complex)
concepts that can be built over NC and NR in DL.
Definition 1. Let T be a DL TBox with concept names
from NC . The set of subconcepts of T is given by
[
sub(T ) = {⊤, ⊥} ∪
sub(C) ∪ sub(D) ,
C⊑D∈T
where for C ∈ NC ∪ {⊤, ⊥}, sub(C) = {C}, and
sub(¬C)
sub(C ⊓ D)
sub(C ⊔ D)
sub(∀R.C)
sub(∃R.C)
=
=
=
=
=
{¬C} ∪ sub(C)
{C ⊓ D} ∪ sub(C) ∪ sub(D)
{C ⊔ D} ∪ sub(C) ∪ sub(D)
{∀R.C} ∪ sub(C)
{∃R.C} ∪ sub(C) .
The size |C| of a concept C is the size of its syntactic tree
where for every role R, ∃R. and ∀R. are individual nodes.
1
The OWL 2 EL profile significantly extends the basic EL logic
whilst maintaining its desirable polynomial time complexity, see
https://www.w3.org/TR/owl2-profiles/.
Definition 2. The size |C| of a concept C is inductively defined as follows. For C ∈ NC ∪ {⊤, ⊥}, |C| = 1. Then,
|¬C| = 1 + |C|; |C ⊓ D| = |C ⊔ D| = 1 + |C| + |D|; and
|∃R.C| = |∀R.C| = 1 + |C|.
P
The size |T | of the TBox T is C⊑D∈T (|C| + |D|).
Clearly, for every C we have card(sub(C)) ≤ |C| and for
every TBox T we have card(sub(T )) ≤ |T | + 2.
We now define the upward and downward cover sets of
concept names. Intuitively, the upward set of the concept C
collects the most specific subconcepts of the TBox T that
subsume C; conversely, the downward set of C collects the
most general subconcepts from T subsumed by C. The concepts in sub(T ) are some concepts that are relevant in the
context of T, and that are used as building blocks for generalisations and specialisations. The properties of sub(T ) guarantee that the upward and downward cover sets are finite.
Definition 3. Let T be a DL TBox and C a concept. The
upward cover and downward cover of C w.r.t. T are:
UpCovT (C) := {D ∈ sub(T ) | C ⊑T D and
∄.D′ ∈ sub(T ) with C ⊏T D′ ⊏T D},
DownCovT (C) := {D ∈ sub(T ) | D ⊑T C and
∄.D′ ∈ sub(T ) with D ⊏T D′ ⊏T C}.
Observe that UpCovT and DownCovT miss interesting
refinements. Note also that this definition only returns meaningful results when used with a consistent ontology; otherwise it returns the whole set sub(T ). Hence, when dealing
with the repair problem of an inconsistent ontology O, we
need a derived, consistent ‘reference ontology’ Oref to steer
the repair process; this is outlined in greater detail in the section on repairing ontologies.
Example 4. Let A, B, C ∈ NC and T = {A ⊑ B}.
We have UpCovT (A ⊓ C) = {A}. Iterating, we get
UpCovT (A) = {A, B} and UpCovT (B) = {B, ⊤}. We
could reasonably expect B ⊓ C to be also a generalisation of
A ⊓ C w.r.t. T but it will be missed by the iterated application of UpCovT . Similarly, UpCovT (∃R.A) = {⊤}, while
we can expect ∃R.B to be a generalisation of ∃R.A.
To take care of these omissions, we introduce a generalisation and specialisation operator. We denote as nnf(C)
the negation normal form of the concept C. Let ↑ and ↓
be two functions from L(DL, NC , NR ) to the powerset of
L(DL, NC , NR ). We define ζ↑,↓ , the abstract refinement operator, by induction on the structure of concept descriptions
as shown in Table 1. Complying with the previous observation, we define two concrete refinement operators from the
abstract operator ζ↑,↓ .
Definition 5. The generalisation operator and specialisation operator are defined, respectively, as
γT = ζUpCovT ,DownCovT , and
ρT = ζDowCovT ,UpCovT .
Returning to our example, notice that for T = {A ⊑ B},
we now have γT (A ⊓ C) = {B ⊓ C, A ⊓ ⊤, A}.
Table 1: Abstract refinement operator
ζ↑,↓ (A) = ↑(A)
ζ↑,↓ (¬A) = {nnf(¬C) | C ∈ ↓(A)} ∪ ↑(¬A)
ζ↑,↓ (⊤) = ↑(⊤)
ζ↑,↓ (⊥) = ↑(⊥)
ζ↑,↓ (C ⊓ D) = {C ′ ⊓ D | C ′ ∈ ζ↑,↓ (C)}∪
{C ⊓ D′ | D′ ∈ ζ↑,↓ (D)} ∪ ↑(C ⊓ D)
ζ↑,↓ (C ⊔ D) = {C ′ ⊔ D | C ′ ∈ ζ↑,↓ (C)}∪
{C ⊔ D′ | D′ ∈ ζ↑,↓ (D)} ∪ ↑(C ⊔ D)
ζ↑,↓ (∀R.C) = {∀R.C ′ | C ′ ∈ ζ↑,↓ (C)} ∪ ↑(∀R.C)
ζ↑,↓ (∃R.C) = {∃R.C ′ | C ′ ∈ ζ↑,↓ (C)} ∪ ↑(∃R.C)
Example 8. If T = {A⊑∃r.A}, then γT (A) = {A, ∃r.A}.
Thus γT (∃r.A) = {∃r.A, ∃r.∃r.A} ∪ {⊤} (notice that
⊤ ∈ γT2 (A)). Continuing the iteration of γT on A, we get
(∃r.)k A ∈ γTk (A) for every k ≥ 0.
This is not a feature caused by the existential quantification alone. Similar examples exist that involve universal quantification, disjunction, and conjunction.2 Notice
that although the covers of two provably equivalent concepts are the same (Lemma 7.3), it is not the case that
γT (C1 ) = γT (C2 ) whenever C1 ≡T C2 . For example, with
the TBox T = {A ⊑ B}, we have γT (A) = {A, B} and
γT (⊤ ⊓ A) = {⊤ ⊓ A, ⊤ ⊓ B, A, B}.
Complexity
We now analyse the computational aspects of the refinement
operators.
Definition 6. Given a DL concept C, its i-th refinement iteri
ation by means of ζ↑,↓ (viz., ζ↑,↓
(C)) is inductively defined
as follows:
Definition 9. Given a TBox T and concepts C, D, the problems γT - MEMBERSHIP and ρT - MEMBERSHIP ask whether
D ∈ γT (C) and D ∈ ρT (C), respectively.
0
• ζ↑,↓
(C) = {C};
We show that γT and ρT are efficient refinement operators, in the sense that deciding γT - MEMBERSHIP and
ρT - MEMBERSHIP is not harder than deciding (atomic)
concept subsumption in the underlying logic. Recall that
subsumption is ExpTime-complete in ALC and PTimecomplete in EL. We show that the same complexity bounds
hold for γT - MEMBERSHIP.
For proving hardness, we first show that deciding whether
C ′ ∈ UpCoverT (C) is as hard as atomic concept subsumption (Theorem 13). Then we show that γT - MEMBERSHIP
is just as hard (Theorem 17). For the upper bounds,
we first establish the complexity of computing the set
UpCoverT (C) (Theorem 15). We then show that we can decide γT - MEMBERSHIP resorting to at most a linear number
of computations UpCoverT (C ′ ) (Theorem 18). Combining
Theorem 17 and Theorem 18, we obtain the result.
•
j+1
ζ↑,↓
(C)
=
j
ζ↑,↓
(C)
∪
S
j
C ′ ∈ζ↑,↓
(C) ζ↑,↓ (C
′
),
j ≥ 0.
The set of all concepts reachable from C
of ζ↑,↓
S by means
i
∗
(C).
in a finite number of steps is ζ↑,↓
(C) = i≥0 ζ↑,↓
Some basic properties about γT and ρT follow.
Lemma 7. For every TBox T :
1. generalisation: if X ∈ γT (C) then C ⊑T X
specialisation: if X ∈ ρT (C) then X ⊑T C
2. reflexivity: if C ∈ sub(T ) then C ∈ UpCovT (C) and
C ∈ DownCovT (C)
3. semantic stability of cover: if C1 ≡T C2 then
C1 ∈ UpCovT (C) iff C2 ∈ UpCovT (C) and
C1 ∈ DownCovT (C) iff C2 ∈ DownCovT (C)
4. relevant completeness: UpCovT (C) ⊆ γT (C) and
DownCovT (C) ⊆ ρT (C)
5. generalisability: if C, D ∈ sub(T ) and C ⊑T D then
D ∈ γT∗ (C)
specialisability: if C, D ∈ sub(T ) and D ⊑T C then
D ∈ ρ∗T (C)
6. trivial generalisability: ⊤ ∈ γT∗ (C)
falsehood specialisability: ⊥ ∈ ρ∗T (C)
7. generalisation finiteness: γT (C) is finite
specialisation finiteness: ρT (C) is finite
Proof. See the appendix. Item 1 is Lemma 24. Item 2,
item 3, and item 4 are simple consequences of Definition 3
and Definition 5. Item 5 is Corollary 29. Item 6 is in turn a
corollary of item 5. Item 7 is Lemma 25.
Although γT (C) and ρT (C) are always finite (see
Lemma 7.7), this is not the case for γT∗ (C) and ρ∗T (C). Indeed, their iterated application can produce an infinite chain
of refinements.
Theorem 10. γT - MEMBERSHIP is ExpTime-complete for
ALC and PTime-complete for EL.
Similar arguments can be used to establish the same complexities for ρT - MEMBERSHIP.
Corollary 11. ρT - MEMBERSHIP is ExpTime-complete for
ALC and PTime-complete for EL.
The remainder of this section provides the details. We first
prove a technical lemma used in the reduction from concept
subsumption to deciding whether C ′ ∈ UpCoverT (C).
Lemma 12. Let T be a DL TBox and X ∈
/ sub(T ). Then,
for every model I of T ′ := T ∪ {X ⊓ B ⊑ ⊤} there is a
model J of T ′ such that
1. X J = ∅, and
2. for every C ∈ sub(T ), C I = C J .
2
From the perspective of ontology repair, infinite refinement
chains are not an issue since there are always finite chains
(Lemma 7.6). If needed, it can be simply circumvented by imposing a bound on the size of the considered refinements.
Proof. We define the interpretation J where all role names
are interpreted as in I, and for every concept name A ∈ NC
∅
if A = X
J
A :=
AI otherwise.
Since X only appears in a tautology, J is also a model of T ′ .
Using induction on the structure of the concepts, it is easy to
show that the second condition of the lemma holds.
The following theorem is instrumental in the proof of
Theorem 17.
Theorem 13. Let T be a DL TBox and let C be an arbitrary
DL concept. Deciding whether D ∈ UpCovT (C) is as hard
as deciding atomic subsumption w.r.t. a TBox over DL.
Proof. We propose a reduction from the problem of deciding atomic subsumption w.r.t. a TBox. Let T be a DL TBox,
and A, B be two concept names. We assume w.l.o.g. that
{A, B} ⊆ sub(T ). Define the new TBox
T ′ := T ∪ {X ⊓ B ⊑ ⊤} ,
where X is a new concept name (not appearing in T ).3 We
show that A ⊑T B iff X ⊓ B ∈ UpCovT ′ (X ⊓ A).
[⇒] If A ⊑T B, then X ⊓ A ⊑T X ⊓ B, and hence it
also holds X ⊓ A ⊑T ′ X ⊓ B. Assume that there is some
E ∈ sub(T ′ ) with X ⊓ A ⊏T ′ E ⊏T ′ X ⊓ B. Then E
cannot be X ⊓ B, nor X. Hence E ∈ sub(T ). Let I be
an arbitrary model of T ′ . By Lemma 12, there is a model
J with E I = E J and X J = ∅. But since by assumption
E ⊑T ′ X ⊓B, it must be that E J = ∅, and hence E I = ∅. It
then follows that for every model I of T ′ , we have E I = ∅,
which is a contradiction with the assumption X ⊓ A ⊏T ′ E.
We conclude that X ⊓ B ∈ UpCovT ′ (X ⊓ A).
[⇐] If A 6⊑T B, there is a model I of T with AI 6⊆ B I . We
can extend this interpretation to a model J of T ′ by setting
X J = ∆I , and AJ = AI for all other concept names; and
rJ = rI for all role names. Then (X ⊓ A)J 6⊆ (X ⊓ B)J ,
and hence X ⊓ B ∈
/ UpCovT ′ (X ⊓ A).
Theorem 14. Let T be a DL TBox and let C be an arbitrary DL concept. Deciding whether D ∈ UpCovT (C) can
be done in exponential time when DL = ALC and in polynomial time when DL = EL.
Proof. An algorithm goes as follows. If D 6∈ sub(T ) or
C 6⊑T D, return false. Then, for every E ∈ sub(T ), check
whether: (1) C ⊑T E, (2) E ⊑T D, (3) E 6⊑T C, and
(4) D 6⊑T E. If conditions (1)–(4) are all satisfied, return
false. Return true after trying all E ∈ sub(T ). The routine
requires at most 1+4×card(sub(T )) calls to the subroutine
for DL concept subsumption. Since card(sub(T )) is linear in |T |, the overall routine runs in exponential time when
DL = ALC and in polynomial time when DL = EL.
The following theorem is instrumental in the proof of
Theorem 18.
3
We use this tautology only to ensure that X ⊓ B ∈ sub(T ′ ),
to satisfy the restriction on the definition of the upward cover.
Theorem 15. Let T be a DL TBox and let C be a DL concept. UpCovT (C) is computable in exponential time when
DL = ALC and in polynomial time when DL = EL.
Proof. It suffices to check for every D ∈ sub(T ) whether
D ∈ UpCovT (C) and collect those concepts for which the
answer is positive. Since card(sub(T )) is linear in the size
of T , the result holds.
Lemma 16. Let T be a DL TBox, C a DL concept, and
X∈
/ sub(T ). Define T ′ := T ∪ {X ≡ C}. If D ∈ sub(T )
then D ∈ UpCovT (C) iff D ∈ UpCovT ′ (C).
Proof. We have sub(T ′ )=sub(T ) ∪ {X}. Let D ∈ sub(T ).
Suppose D ∈ UpCovT ′ (C). Then C ⊑T ′ D and there is
no E ∈ sub(T ′ ) such that C ⊏T ′ E ⊏T ′ D. We thus have
C ⊑T D. Since sub(T ) ⊂ sub(T ′ ) there is no E ∈ sub(T )
such that C ⊏T E ⊏T D.
Let D ∈ UpCovT (C). Then C ⊑T D and C ⊑T ′ D.
Moreover, there is no E ∈ sub(T ) with C ⊏T E ⊏T D. So
there is no E ∈ sub(T ) such that C ⊏T ′ E ⊏T ′ D.
Since X ≡T ′ C, it is not the case that C ⊏T ′ X ⊏T ′ D.
Since sub(T ′ ) = sub(T ) ∪ {X}, there is no E ∈ sub(T ′ )
such that C ⊏T ′ E ⊏T ′ D. Then D ∈ UpCovT ′ (C).
Theorem 17. Deciding γT - MEMBERSHIP is as hard as deciding whether D ∈ UpCovT (C).
Proof. Let T be a DL TBox, C a concept, and X ∈
/ sub(T ).
Define T ′ := T ∪ {X ≡ C}. For every concept D 6= X, we
show that D ∈ UpCovT (C) iff D ∈ γT ′ (X).
By Lemma 16, D ∈ UpCovT (C) iff D ∈ UpCovT ′ (C).
Since X ≡T ′ C, Lemma 7.3 yields D ∈ UpCovT ′ (C) iff
D ∈ UpCovT ′ (X). As X is a concept name, by definition
of γ we have D ∈ UpCovT ′ (X) iff D ∈ γT ′ (X).
Theorem 18. Let T be a DL TBox and C a concept.
γT - MEMBERSHIP can be decided in exponential time when
DL = ALC and in polynomial time when DL = EL.
Proof. We can decide whether γT (C) contains a particular concept by computing only a linear number of times
UpCovT (C ′ ), where |C ′ | is linearly bounded by |C ′ | + |T |.
Theorem 15 tells us that each of these computations can be
done in exponential time when DL = ALC and in polynomial time when DL = EL. This yields an exponential time
procedure when DL = ALC and a polynomial time procedure when DL = EL.
Repairing Ontologies
Our refinement operators can be used as components of a
method for repairing inconsistent ontologies by weakening,
instead of removing, problematic axioms.
Given an inconsistent ontology O, we proceed as described in Algorithm 1. Briefly, we first need to find a
consistent subontology Oref of O to serve as reference ontology in order to be able to compute a non-trivial upcover and downcover. The brave approach (which we use
in our evaluation) picks a random maximally consistent
subset of O and chooses it as reference ontology Oref .
The cautious approach takes as Oref the intersection of all
Algorithm 1 RepairOntologyWeaken(O)
ref
O ← MaximallyConsistent(O)
while O is inconsistent do
BadAx ← FindBadAxiom(O)
WeakerAx ← WeakenAxiom(BadAx, Oref )
O ← O \ {BadAx} ∪ {WeakerAx}
end while
Return O
Algorithm 2 RepairOntologyRemove(O)
while O is inconsistent do
BadAx ← FindBadAxiom(O)
O ← O \ {BadAx}
end while
Return O
Table 2: BioPortal ontologies considered for experimental
validation
maximally consistent subsets (Ludwig and Peñaloza 2014;
Lembo et al. 2010). While the brave approach is faster to
compute and still guarantees to find solutions, the cautious
approach has the advantage of not excluding certain repairs
a priori. However, it also returns, e.g., a much impoverished
upcover.
Once a reference ontology Oref has been chosen, and as
long as O is inconsistent, we select a “bad axiom” and replace it with a random weakening of it with respect to Oref .
In view of evaluation, we consider two variants of the subprocedure FindBadAxiom(O). The first variant (‘mis’) randomly samples a number of minimally inconsistent subsets I1 , I2 , . . . Ik ⊆ O and returns one axiom from the
ones occurring the most often, i.e., an axiom from the set
argmaxφ∈O (card({j | φ ∈ Ij and 1 ≤ j ≤ k})). The second variant (‘rand’) of FindBadAxiom(O) merely returns an
axiom in O at random.
The set of all weakenings of an axiom with respect to a
reference ontology is defined as follows:
Definition 19 (Axiom weakening). Given a subsumption
axiom C ⊑ D of O, the set of (least) weakenings of C ⊑ D
w.r.t. O, denoted by gO (C ⊑ D) is the set of all axioms
C ′ ⊑ D′ such that
C ′ ∈ ρO (C) and D′ ∈ γO (D) .
Given an assertional axiom C(a) of O, the set of (least)
weakenings of C(a), denoted gO (C(a)) is the set of all axioms C ′ (a) such that
C ′ ∈ γO (C) .
The subprocedure WeakenAxiom(φ, Oref ) randomly returns one axiom in gO (φ). For every subsumption or assertional axiom φ, the axioms in the set gO (φ) are indeed
weaker than φ.
Lemma 20. For every subsumption or assertional axiom φ,
if φ′ ∈ gO (φ), then φ |=O φ′ .
′
′
Proof. Suppose C ⊑ D ∈ gO (C ⊑ D). Then, by definition of gO and Lemma 7.1, C ′ ⊑ C and D ⊑ D′ are inferred
from O. Thus, by transitivity of subsumption, we obtain that
C ⊑ D |=O C ′ ⊑ D′ . For the weakening of assertions, the
result follows immediately from Lemma 7.1 again.
Clearly, substituting an axiom φ with one axiom from
gO (φ) cannot diminish the set of interpretations of an ontology. By Lemma 7.6, any subsumption axiom is a finite number of refinement steps away from the trivial axiom ⊥ ⊑ ⊤.
Any assertional axiom C(a) is also a finite number of generalisations away from the trivial assertion ⊤(a). It follows
Abbreviation
bctt
co-wheat
elig
hom
icd11
ofsmr
ogr
pe
taxrank
xeo
Name
Behaviour Change Technique Taxonomy
Wheat Trait Ontology
Eligibility Feature Hierarchy
Homology and Related Concepts in Biology
Body System Terms from ICD11
Open Food Safety Model Repository
Ontology of Geographical Region
Pulmonary Embolism Ontology
Taxonomic Rank Vocabulary
XEML Environment Ontology
that by repeatedly replacing an axiom with one of its weakenings, the weakening procedure will eventually obtain an
ontology with some interpretations. Hence, the algorithm
will eventually terminate.
In the next section, we compare Algorithm 1 with Algorithm 2, which merely removes bad axioms until an ontology
becomes consistent. We do so for both variants ‘mis’ and
‘rand’ of FindBadAxiom(O). As we will see, Algorithm 1
generally allows us to obtain consistent ontologies which retain significantly more of the informational content of the
axioms of the original (and inconsistent) ontology than the
ones obtained through Algorithm 2. This is most significant
with the ‘mis’ variant of FindBadAxiom(O) which reliably
pinpoints the problematic axioms.
Evaluation
The question of which one of two consistent repairs O1 and
O2 of a given inconsistent ontology O is preferable is not, in
general, well-defined. In this work, we compare two such repairs by taking into account the corresponding inferred class
hierarchies. To this end, we define:
Inf(Oi ) = {A ⊑ B : A, B ∈ NC , Oi |= A ⊑ B} .
The intuition behind the choice of measure is that if
card(Inf(O1 ) \ Inf(O2 )) > card(Inf(O2 ) \ Inf(O1 )) (that
is, if there exist more subsumptions between classes which
can be inferred in O1 but not in O2 than vice versa) then O1
is to be preferred to O2 . Furthermore, class subsumptions,
which can be inferred from both O1 or O2 , should be of no
consequence to determine which repaired ontology is preferable. That is, whenever Inf(O1 ) ⊆ Inf(O1′ ), Inf(O2 ) ⊆
Inf(O2′ ) and Inf(O1′ )\Inf(O1 ) = Inf(O2′ )\Inf(O2 ) it should
hold that the quality of O1 with respect to O2 is the same
as the quality of O1′ with respect to O2′ . Thus, we define
the following measure to compare the inferable information
content of two ontologies.
Axiom Weakening VS Axiom Removal
Repair quality (left=random, right=MIS)
0.80
Mean IIC
0.75
0.70
0.65
bctt
co-wheat
elig
hom
icd11
ofsmr
ogr
pe
taxrank
xeo
Random
0.55 (0.35)
0.69 (0.29)
0.61 (0.30)
0.68 (0.26)
0.60 (0.30)
0.65 (0.31)
0.56 (0.32)
0.56 (0.33)
0.56 (0.31)
0.67 (0.29)
MIS
0.72 (0.36)
0.76 (0.31)
0.72 (0.27)
0.71 (0.31)
0.71 (0.40)
0.76 (0.29)
0.70 (0.35)
0.67 (0.41)
0.82 (0.36)
0.67 (0.34)
0.60
0.55
bctt
co
elig hom icd11 ofsmr ogr
pe taxrank xeo
Figure 1: Comparing weakening-based ontology repair with
removal-based ontology repair. Mean IIC of weakeningbased against removal-based repair for each ontology, when
choosing axioms at random (left) or by sampling minimally
inconsistent sets (right).
Definition 21. Let O1 and O2 be two consistent ontologies.
If Inf(O1 ) 6= Inf(O2 ), we define the inferable information
content IIC(O1 , O2 ) of O1 w.r.t. O2 as IIC(O1 , O2 ) =
card(Inf(O1 ) \ Inf(O2 ))
;
card(Inf(O1 ) \ Inf(O2 )) + card(Inf(O2 ) \ Inf(O1 ))
if instead Inf(O1 ) = Inf(O2 ), we set IIC(O1 , O2 ) = 0.5.
It is readily seen that this definition satisfies the two conditions mentioned above. Furthermore, the following properties hold:
1. IIC(O1 , O2 ) ∈ [0, 1];
2. IIC(O1 , O2 ) = 1 − IIC(O2 , O1 );
3. IIC(O1 , O2 ) = 0.5 if and only if card(Inf(O1 )) =
card(Inf(O2 ));
4. IIC(O1 , O2 ) = 1 if and only if Inf(O2 ) ⊂ Inf(O1 );
5. IIC(O1 , O2 ) > 0.5 if and only if card(Inf(O1 ) \
Inf(O2 )) > card(Inf(O2 ) \ Inf(O1 )).
Although this is by no means the only possible
measure
for
comparing
two
ontologies (Tartir et al. 2005; Alani, Brewster, and Shadbolt 2006;
Vrandečić and Sure 2007; Vrandečić 2009), these properties suggest that our definition captures a notion of “quality”
that is meaningful for our intended application: in particular, if for two proposed repairs O1 , O2 of an inconsistent
ontology O we have IIC(O1 , O2 ) > 0.5, then there are
more class subsumptions which can be inferred in O1 but
not in O2 than vice versa, and hence—all other things being
equal—O1 is a better repair of O than O2 .
One possible criticism of our definition of IIC(O1 , O2 )
is that its value depends only on Inf(O1 ) and Inf(O2 ): if
O1 and O2 differ only w.r.t. subsumptions between complex
Table 3: Mean and standard deviation (in parentheses) of IIC
between RepairOntologyWeaken and RepairOntologyRemove, both when choosing axioms at random (left column)
and by sampling minimally inconsistent sets (right). Bolded
values are significant (p < 0.05) with respect to both
Wilcoxon and T-test with Holm-Bonferroni correction; nonbolded values were not significant for either.
concepts, then IIC(O1 , O2 ) = 0.5 (even though the implications of O1 might still be considerably richer than those of
O2 ). On the other hand, focusing on atomic subsumptions
makes also conceptual sense, as these are the ones that our
inconsistent ontology—as well as the proposed repairs—
discuss about. It is, in any case, certainly true that our measure is fairly coarse: if IIC(O1 , O2 ) is significantly greater
than 0.5 there are good grounds to claim that O1 is a better
repair of O than O2 is, but it may easily be that repair candidates between which our measure cannot discriminate are
nonetheless of different quality.
To empirically test whether weakening axioms is a
better approach to ontology repair than removing them,
we tested our approach on ten ontologies from BioPortal (Matentzoglu and Parsia 2017), expressed in ALC (see
Table 2). On average the ontologies have 105 logical axioms and 90 classes. We compared the performance of RepairOntologyWeaken (Algorithm 1) with the one of the non
weakening-based RepairOntologyRemove (Algorithm 2) by
first making the ontologies inconsistent through the addition
of random axioms, then attempting to repair them through
the two algorithms (using the original ontology as the reference), and then computing IIC.4 This procedure has the
following rationale: one may think that the axioms added
constitute some new claims made concerning the relationships between the classes of the ontology, which however
unfortunately made it inconsistent. It is thus desirable to fix
this inconsistency while preserving as much as possible of
the informational content of these axioms and of the other
axioms in the ontology.
The procedure was repeated one hundred times per ontology, selecting the axioms to weaken or remove by sampling minimally inconsistent sets, and one further hundred
4
Limited to the ‘mis’ variant of FindBadAxiom, the
prototype
implementation
of
RepairOntologyWeaken,
RepairOntologyRemove, and the IIC measure is provided as
supplemental material.
times selecting the axioms to remove or weaken completely
randomly. We tested the significance of our results through
both Wilcoxon signed-rank tests and T-tests, applying the
Holm-Bonferroni correction for multiple comparison, with
a p-value threshold of 0.05.
Figure 1 and Table 3 summarise the results of our experiments. When choosing the axioms to weaken or remove
through sampling minimally inconsistent sets, the means
(in the case of the T-test) and medians (in the case of the
Wilcoxon test) of the IIC for RepairOntologyWeaken against
RepairOntologyRemove were all significantly greater than
0.5 for all ontologies. This confirms that our repair-byweakening technique is able to preserve more of the informational content of axioms than repair-by-removal techniques. When selecting the axioms to repair randomly, on
the other hand, this was not always the case, as shown in
Table 3. This illustrates how our weakening-approach on
ontology repair reliably constitutes an improvement over
removal-based ontology repair only when problematic axioms can be reliably pinpointed.
Figure 1 highlights the effect of choosing the axioms to
repair or remove randomly rather than through sampling inconsistent sets. While the difference is not statistically significant for all ontologies,5 we observe that the quality of the
repair compared to the corresponding removal is always improved by choosing the axioms to repair via sampling. The
natural next step in this line of investigation would consist in
evaluating the effect of varying the number of minimally inconsistent sets sampled by FindBadAxiom, which for these
experiments was set to one tenth of the ontology size.
To summarise, the main conclusion of our experiments is
that, when problematic axioms can be reliably identified, our
approach is better able to preserve the informational content
of inconsistent ontologies than the corresponding repair-byremoval method.
Related Work
Refinements
operators
were
also
discussed
in the context of inductive logic programming
(van der Laag and Nienhuys-Cheng 1998),
and
formalised in description logics for concept learning (Lehmann and Hitzler 2010). The refinement operators used by our weakening approach were introduced in (Confalonieri et al. 2016), and further analysed
in (Confalonieri et al. 2017) in relation to incoherence
detection. They were not previously applied to ontology
repair.
The problem of identifying and repairing inconsistencies in ontologies has received much attention in recent
years. Our approach differs from many other works in
the area, see for instance (Schlobach and Cornet 2003;
Kalyanpur et al. 2005;
Kalyanpur et al. 2006;
Baader, Peñaloza, and Suntisrivaraporn 2007;
Haase and Qi 2007), in that—rather than removing the
problematic axioms altogether—we attempt to repair the
5
For instance, w.r.t. the Wilcoxon test it is statistically significant for bctt, elig, ogr, pe and taxrank, but not for the other five
ontologies.
ontology by replacing them with weakened rewritings.
On the one hand, our method requires the choice of a
(consistent) reference ontology with respect to which one
can compute the weakenings; on the other hand, it allows
us to perform a softer, more fine-grained form of ontology
repair.
A different approach for repairing ontologies through
weakening was discussed in (Lam et al. 2008). Our own approach is, however, quite different from it: while the repair
algorithm of (Lam et al. 2008) operates by pinpointing (and
subsequently removing) the subcomponents of the axioms
responsible for the contradiction, ours is based on a refinement operator, which combines both semantic (via the cover
operators) and syntactic (via the compositional definitions
of generalisations and specialisations of complex formulas)
information in order to identify candidates for the replacement of the offending axiom(s). In particular, this implies—
using the terminology of (Ji et al. 2014)—that our repair algorithm, in contrast to (Lam et al. 2008), is ‘black box’ in
that it treats the reasoner as an oracle, and can thus be more
easily combined with different choices of reasoner (or, with
slightly more effort, applied to different logics).
Another influential approach to ontology repair is discussed in (Qi, Liu, and Bell 2006a) and
in (Qi, Liu, and Bell 2006b). That approach, like ours,
attempts to weaken problematic axioms; but it does so by
adding exceptions to value restrictions ∀R.C(a),6 rather
than by means of a more general-purpose transformation.
We leave to future work the evaluation of our approach in
comparison to other state-of-the-art ontology repair frameworks. As already stated, this is not an entirely well-posed
problem; but if, as in this work, we accept that a suggested
repair O1 is preferable to another suggested repair O2 whenever card(Inf(O1 ) \ Inf(O2 )) > card(Inf(O2 ) \ Inf(O1 ))
then the question becomes amenable to analysis. Possibly, complementary metrics for further evaluations can be
chosen from (Alani, Brewster, and Shadbolt 2006). Experiments involving user evaluation could be also considered in
this context.
Conclusions
We have proposed a new strategy for repairing ontologies
based on the idea of weakening terminological and assertional axioms. Axiom weakening is a way to improve the
balance between regaining consistency and keeping as much
information from the original ontology as possible.
We have investigated the theoretical properties of the refinement operators that are required in the definition of axiom weakening and analysed the computational complexity
of employing them. Furthermore, the empirical evaluation
shows that our weakening-based approach to repairing ontologies performs significantly better, in terms of preservation of information, than the removal-based approach.
Future work will concentrate on the following two directions. Firstly, we plan to extend our evaluation to further cor6
Another difference is that we are also interested in repairing
TBoxes, whereas the approach of (Qi, Liu, and Bell 2006b) operates only over ABoxes.
pora of ontologies and measures of information, including
in particular measures to reflect the syntactic complexity of
repaired ontologies and measures that reflect the preservation of entailments of competency questions. Secondly, we
plan to extend the presented approach to axiom weakening
to more expressive DL languages, including SROIQ underlying OWL 2 DL (Horrocks, Kutz, and Sattler 2006) and
to full first-order logic, for which debugging is a particularly
challenging problem (Kutz and Mossakowski 2011). We expect that, for more complex languages, the weakening-based
strategy will likewise significantly improve on the removalbased strategy, and indeed be even more appropriate by exploiting the higher syntactic complexity.
References
[Alani, Brewster, and Shadbolt 2006] Alani, H.; Brewster,
C.; and Shadbolt, N. 2006. Ranking ontologies with AKTiveRank. In Proceedings of ISWC’06, LNCS, 1–15. Berlin,
Heidelberg: Springer-Verlag.
[Auer et al. 2007] Auer, S.; Bizer, C.; Kobilarov, G.;
Lehmann, J.; Cyganiak, R.; and Ives, Z. 2007. DBpedia:
A Nucleus for a Web of Open Data. Berlin, Heidelberg:
Springer. 722–735.
[Baader et al. 2003] Baader, F.; Calvanese, D.; McGuinness,
D. L.; Nardi, D.; and Patel-Schneider, P. F., eds. 2003. The
Description Logic Handbook: Theory, Implementation, and
Applications. New York, NY, USA: Cambridge University
Press.
[Baader, Brandt, and Lutz 2005] Baader, F.; Brandt, S.; and
Lutz, C. 2005. Pushing the EL envelope. In IJCAI, 364–
369.
[Baader, Peñaloza, and Suntisrivaraporn 2007] Baader, F.;
Peñaloza, R.; and Suntisrivaraporn, B. 2007. Pinpointing
in the description logic EL+ . In Proc. of KI 2007, volume
4667 of LNCS, 52–67. Springer.
[Bateman et al. 2010] Bateman, J.; Hois, J.; Ross, R.; and
Tenbrink, T. 2010. A Linguistic Ontology of Space
for Natural Language Processing. Artificial Intelligence
174(14):1027–1071.
[Confalonieri et al. 2016] Confalonieri, R.; Eppe, M.; Schorlemmer, M.; Kutz, O.; Peñaloza, R.; and Plaza, E. 2016.
Upward refinement operators for conceptual blending in the
description logic EL++ . Annals of Mathematics and Artificial Intelligence.
[Confalonieri et al. 2017] Confalonieri, R.; Kutz, O.; Troquard, N.; Galliani, P.; Porello, D.; Peñaloza, R.; and Schorlemmer, M. 2017. Coherence, Similarity, and Concept Generalisation. In Proc. of DL 2017. CEUR.
[Grüninger and Fox 1995] Grüninger, M., and Fox, M. S.
1995. The role of competency questions in enterprise engineering. In Benchmarking—Theory and Practice. Springer.
22–31.
[Haase and Qi 2007] Haase, P., and Qi, G. 2007. An analysis of approaches to resolving inconsistencies in DL-based
ontologies. In Proc. of IWOD-07, 97–109.
[Horrocks, Kutz, and Sattler 2006] Horrocks, I.; Kutz, O.;
and Sattler, U. 2006. The Even More Irresistible SROIQ.
In Proc. of KR’06, 57–67. AAAI Press.
[Ji et al. 2014] Ji, Q.; Gao, Z.; Huang, Z.; and Zhu, M. 2014.
Measuring effectiveness of ontology debugging systems.
Knowledge-Based Systems 71:169–186.
[Kalyanpur et al. 2005] Kalyanpur, A.; Parsia, B.; Sirin, E.;
and Hendler, J. 2005. Debugging unsatisfiable classes in
OWL ontologies. Web Semantics: Science, Services and
Agents on the World Wide Web 3(4):268–293.
[Kalyanpur et al. 2006] Kalyanpur, A.; Parsia, B.; Sirin, E.;
and Grau, B. C. 2006. Repairing unsatisfiable concepts in
OWL ontologies. In ESWC, volume 6, 170–184. Springer.
[Kutz and Mossakowski 2011] Kutz, O., and Mossakowski,
T. 2011. A Modular Consistency Proof for Dolce. In Proc.
of AAAI-11. AAAI Press.
[Lam et al. 2008] Lam, J. S. C.; Sleeman, D.; Pan, J. Z.; and
Vasconcelos, W. 2008. A fine-grained approach to resolving
unsatisfiable ontologies. In Journal on Data Semantics X.
Springer. 62–95.
[Lehmann and Hitzler 2010] Lehmann, J., and Hitzler, P.
2010. Concept learning in description logics using refinement operators. Machine Learning 78(1-2):203–250.
[Lembo et al. 2010] Lembo, D.; Lenzerini, M.; Rosati, R.;
Ruzzi, M.; and Savo, D. F. 2010. Inconsistency-tolerant semantics for description logics. In Proc. of RR 2010, volume
6333 of LNCS, 103–117. Springer.
[Ludwig and Peñaloza 2014] Ludwig, M., and Peñaloza, R.
2014. Error-tolerant reasoning in the description logic el.
In Proc. of JELIA 2014, volume 8761 of LNCS, 107–121.
Springer.
[Matentzoglu and Parsia 2017] Matentzoglu, N., and Parsia,
B. 2017. BioPortal snapshot 30.03.2017. last accessed,
2017/08/04.
[Neuhaus et al. 2013] Neuhaus, F.; Vizedom, A.; Baclawski,
K.; Bennett, M.; Dean, M.; Denny, M.; Grüninger, M.;
Hashemi, A.; Longstreth, T.; Obrst, L.; et al. 2013. Towards
ontology evaluation across the life cycle – the ontology summit 2013. Applied Ontology 8(3):179–194.
[Prestes et al. 2013] Prestes, E.; Carbonera, J. L.; Fiorini,
S. R.; Jorge, V. A. M.; Abel, M.; Madhavan, R.; Locoro,
A.; Goncalves, P.; Barreto, M. E.; Habib, M.; Chibani, A.;
Gérard, S.; Amirat, Y.; and Schlenoff, C. 2013. Towards
a core ontology for robotics and automation. Robotics
and Autonomous Systems 61(11):1193 – 1204. Ubiquitous
Robotics.
[Qi, Liu, and Bell 2006a] Qi, G.; Liu, W.; and Bell, D.
2006a. A revision-based approach to handling inconsistency in description logics. Artificial Intelligence Review
26(1):115–128.
[Qi, Liu, and Bell 2006b] Qi, G.; Liu, W.; and Bell, D. A.
2006b. Knowledge base revision in description logics.
In European Workshop on Logics in Artificial Intelligence,
386–398. Springer.
[Ren et al. 2014] Ren, Y.; Parvizi, A.; Mellish, C.; Pan, J. Z.;
van Deemter, K.; and Stevens, R. 2014. Towards compe-
tency question-driven ontology authoring. In The Semantic
Web: Trends and Challenges. Springer. 752–767.
[Resnik 1999] Resnik, P. 1999. Semantic similarity in a taxonomy: An information-based measure and its application
to problems of ambiguity in natural language. Journal of
Artificial Intelligence Research 11:95–130.
[Sazonau, Sattler, and Brown 2015] Sazonau, V.; Sattler, U.;
and Brown, G. 2015. General terminology induction in
OWL. In Proc. ISWC 2015, Part I, volume 9366 of LNCS,
533–550. Springer.
[Schlobach and Cornet 2003] Schlobach, S., and Cornet, R.
2003. Non-standard reasoning services for the debugging of
description logic terminologies. In Proc. of IJCAI-03, 355–
362. Morgan Kaufmann.
[Stuckenschmidt, Parent, and Spaccapietra 2009]
Stuckenschmidt, H.; Parent, C.; and Spaccapietra, S.,
eds. 2009. Modular Ontologies—Concepts, Theories and
Techniques for Knowledge Modularization, volume 5445 of
LNCS. Springer.
[Tartir et al. 2005] Tartir, S.; Arpinar, I. B.; Moore, M.;
Sheth, A. P.; and Aleman-Meza, B. 2005. Ontoqa: Metricbased ontology quality analysis.
[van der Laag and Nienhuys-Cheng 1998] van der Laag,
P. R., and Nienhuys-Cheng, S.-H. 1998. Completeness and
properness of refinement operators in inductive logic programming. The Journal of Logic Programming 34(3):201 –
225.
[Vrandečić and Sure 2007] Vrandečić, D., and Sure, Y.
2007. How to design better ontology metrics. The Semantic
Web: Research and Applications 311–325.
[Vrandečić 2009] Vrandečić, D. 2009. Ontology evaluation.
In Handbook on Ontologies. Springer. 293–313.
Appendix
Proposition 22. Let T be an ALC TBox. We have
X
card(sub(T )) ≤
(| C | + | D |) + 2 .
C⊑D
This is a consequence of Lemma 23.
Lemma 23. Let C be an ALC concept. We have
card(sub(C)) ≤| C |.
Proof. If C ∈ NC , or C = ⊤, or C = ⊥, then the result
holds. Hence, it holds for all concepts of size 1.
If C = ¬A, with A ∈ NC , then the result holds. (We have
| C |= card(sub(C)) = 2.) Hence, it holds for all concepts
of size 2.
Let k ≥ 1, and suppose for induction that for every concept C, if | C |≤ k, then card(sub(C)) ≤| C |.
Let C be a concept such that | C |= k + 1.
Case C = ¬A. The result holds.
Case C = E ⊓ F . Clearly, | E |≤ k and | F |≤ k. Thus,
by IH, card(sub(E)) ≤| E | and card(sub(F )) ≤| F |.
Hence, card(sub(C)) ≤| E | + | F | +1 =| C |.
Case C = E ⊓ F . Idem.
Case C = ∃R.D. Clearly, | D |= k. Thus, by IH,
cardsub(D)) ≤| D |. Hence, card(sub(C)) ≤| D | +1 =|
C |.
Case C = ∀R.D. Idem.
Lemma 24. For every ALC concept C, and for every C ′ ∈
γ(C), we have C ⊑ C ′ .
Proof. By induction on the complexity of the concept C.
Case C ∈ NC , or C = ⊤, or C = ⊥. γ(C) = UpCov(C).
By definition of UpCov, for every C ′ ∈ UpCov(C) we have
C ⊑ C′.
Case C = ¬A, with A ∈ NC . Let C ′ ∈ γ(C). Either
C ∈ UpCov(C), for which the result is immediate, or there
is E ≡ ¬C ′ such that E ∈ DownCov(A). By definition of
DownCov, we have E ⊑ A, and thus ¬C ′ ⊑ ¬C. Hence,
C ⊑ C′.
If C ′ ∈ γ(C) ∩ UpCov(C), it is clear from the definition
of UpCov that C ⊑ C ′ . We’ll ignore this subcase in the
cases.
′
Now, suppose for induction that the property holds for the
concepts E and F . Let R be an arbitrary role in NR .
Case C = E ⊓ F . Subcase C ′ is of the form E ′ ⊓ F ,
with E ′ ∈ γ(E). By IH, we have E ⊑ E ′ . Indeed, C ⊑ C ′ .
Subcase C ′ is of the form E ⊓ F ′ , with F ′ ∈ γ(F ). By IH,
we have F ⊑ F ′ . Indeed, C ⊑ C ′ .
Case C = E ⊔ F . Subcase C ′ is of the form E ′ ⊔ F ,
with E ′ ∈ γ(E). By IH, we have E ⊑ E ′ . Indeed, C ⊑ C ′ .
Subcase C ′ is of the form E ⊔ F ′ , with F ′ ∈ γ(F ). By IH,
we have F ⊑ F ′ . Indeed, C ⊑ C ′ .
Case C = ∀R.E. Let C ′ ∈ γ(C). Thus, C ′ is of the form
∀R.E ′ . By IH, we have E ⊑ E ′ . Indeed, C ⊑ C ′ .
Case C = ∃R.E. Let C ′ ∈ γ(C). Thus, C ′ is of the form
∃R.E ′ . By IH, we have E ⊑ E ′ . Indeed, C ⊑ C ′ .
Lemma 25. For every finite TBox T and for every concept
C, the set of concepts γT (C) is finite.
Proof. The set sub(T ) is finite; In fact linear in the size of
T , Lemma 22.
For every concept C, we have UpCovT (C) ⊆ sub(T ). So
UpCovT (C) is linearly bounded by |T |.
For every concept C, we have DownCovT (C) ⊆ sub(T ).
So DownCovT (C) is linearly bounded by |T |.
For every concept C, the size |nnf(C)| is linearly bounded
by |C|.
Finally, all recursive calls of γT are done on concepts of
strictly decreasing size.
We can do better, and bound the size of γT (C) linearly.
Lemma 26. For every finite TBox T and for every concept
C, we have card(γT (C)) ≤ (|T | + 2) · |C|.
Proof. The proof is done by induction on the complexity of
the concept C.
When C ∈ NC , or C = ⊤, or C = ⊥,
γT (C) = UpCovT (C). Moreover card(UpCovT (C)) ≤
card(sub(T )). Proposition 22 ensures that the result holds.
When C
=
¬A,
card(γT (C))
=
card(DownCovT (A)) + card(UpCovT (C)). Since
card(DownCovT (A)) ≤ card(sub(T )), Proposition 22
ensures the result again.
Now suppose for induction that the result holds for the
concepts D and E.
When C = D ⊓ E or C = D ⊔ E, card(γT (C)) ≤
card(γT (D)) + card(γT (E)) + card(UpCovT (C)). By
I.H., card(γT (C)) ≤ (|D| · (|T | + 2)) + (|E| · (|T | + 2)) +
(|T | + 2), which yields the result.
When C = ∃r.D or C = ∀r.D, card(γT (C)) ≤
card(γT (D)) + card(UpCovT (C)).
By
I.H.,
card(γT (C)) ≤ (|D| · (|T | + 2)) + (|T | + 2), which yields
the result.
Proof of Generalisability We prove in detail the generalisability property:
If C, D ∈ sub(T ) and C ⊑T D then D ∈ γT∗ (C) .
We will do so by proving a similar statement about the upwards cover operator:
If C, D ∈ sub(T ) and C ⊑T D then D ∈ UpCov∗T (C) ,
where
• UpCov0T (C) = {C};
• S
UpCovj+1
T (C)
=
UpCovjT (C)
′
C ′ ∈UpCovjT (C) UpCovT (C ) , j ≥ 0;
S
• UpCov∗T (C) = i≥0 UpCoviT (C).
∪
First, we will prove the latter statement; then, we will use it
(as well as the definition of γT ) to prove the former.
Lemma 27. Let C, D ∈ sub(T ), C ⊑T D and D 6∈
UpCovT (C) then there exists some C ′ ∈ UpCovT (C) such
that C ⊏T C ′ ⊏T D.
Proof. We prove this by contradiction. Assume that C, D ∈
sub(T ), C ⊑T D, D 6∈ UpCovT (C), and that for all E ∈
sub(T ) with C ⊏T E ⊏ D it is the case that E 6∈ UpCovT .
Then we will prove, by induction on n ∈ N, that for any
integer n there exists a chain C ⊏T Dn ⊏T Dn−1 ⊏T
. . . ⊏T D1 ⊏T D of concepts D1 . . . Dn ∈ sub(T ).
This will imply a contradiction, because sub(T ) is finite7
and because these D1 . . . Dn will be pairwise distinct (since
Di ⊏T Dj whenever i > j).
Let us proceed with the induction.
• Base case: Let n = 1. Since C ⊑T D but D 6∈
UpCovT (C), by definition of the upward cover set there
exists some D1 ∈ sub(T ) such that C ⊏T D1 ⊏T D, as
required.
• Inductive case: Suppose that the statement holds for n,
i.e., there exist D1 . . . Dn ∈ sub(T ) such that C ⊏T
Dn ⊏T Dn−1 ⊏T . . . ⊏T D1 ⊏T D.
Now by hypothesis, since C ⊏T Dn ⊏T D, it must
be the case that Dn 6∈ UpCovT (C). But then, again
by definition of upward cover set, there must exist some
Dn+1 ∈ sub(T ) with C ⊏T Dn+1 ⊏T Dn , as required.
Using this lemma, we can now prove the generalisability
property of the upward cover:
Theorem 28. Let C, D ∈ sub(T ) be such that C ⊑T D.
Then D ∈ UpCov∗T (C), that is, there exist n ∈ N and
C1 , . . . Cn ∈ sub(T ) such that
1. C1 ∈ UpCovT (C);
2. For all i < n, Ci+1 ∈ UpCovT (Ci );
3. Cn = D.
Proof. Assume that C, D ∈ sub(T ), C ⊑T D but D 6∈
UpCov∗T (C). We prove, by induction over n ∈ N, that
for any integer n there exists a chain C ⊏T C1 ⊏T
C2 ⊏T . . . ⊏T Cn ⊏T D of pairwise distinct concepts
C1 . . . Cn ∈ sub(T ) such that C1 ∈ UpCovT (C) and
Ci+1 ∈ UpCovT (Ci ) for i < n. As in Lemma 27, this is
a contradiction, because sub(T ) is finite.
• Base case: Let n = 1. Since D 6∈ UpCov∗T (C), in particular it is the case that D 6∈ UpCovT (C).8 Then by
Lemma 27, there exists some C1 ∈ UpCovT (C) such that
C ⊏T C1 ⊏T D, as required.
• Inductive case: Suppose that the statement holds for n,
i.e., there exist C1 . . . Cn ∈ sub(T ) such that C ⊏T
C1 . . . ⊏T Cn ⊏T D, C1 ∈ UpCovT (C), and Ci+1 ∈
UpCovT (Ci ) for all i < n.
Then it must be the case that D 6∈ UpCovT (Cn ), since
otherwise it would be true that D ∈ UpCov∗T (C) by
means of the chain C1 C2 . . . Cn D. But then, again by
Lemma 27, it is also true that there exists some Cn+1 ∈
UpCovT (Cn ) such that Cn ⊏T Cn+1 ⊏T D, as required.
A straightforward consequence of this result is that our
generalisation operator also satisfies the generalisability
property:
Corollary 29. Let C, D ∈ sub(T ) be such that C ⊑T D.
Then D ∈ γT∗ (C).
Proof. By Theorem 28 we have at once that D ∈
UpCov∗T (C), i.e., there exist concepts C1 . . . Cn such that
C1 ∈ UpCovT (C), Ci+1 ∈ UpCovT (Ci ) for all i <
n, and Cn = D. But by Lemma 7.4, UpCovT (E) ⊆
γT (E) for every concept E; thus, the same chain C →γ
C1 →γ C2 . . . →γ Cn = D also demonstrates that
D ∈ γT∗ (C).
7
It would suffice to know that the length of subsumption chains
A1 ⊏T A2 ⊏T . . . ⊏T An of elements in sub(T ) is finite and
bounded.
8
Indeed, otherwise the trivial chain C1 = D would prove that
D ∈ UpCovT (C).
| 2cs.AI
|
Robust Causal Transform Coding for LQG
Systems with Delay Loss in Communications
Mohammad Kazem Izadinasab*, Amir Homayoun Bahari Sani**, Farshad Lahouti+, and Babak Hassibi+
*
Department of Electrical and Computer Engineering, University of Waterloo
**
School of Electrical and Computer Engineering, College of Engineering, University of Tehran
+
Department of Electrical Engineering, California Institute of Technology
[email protected], [email protected], [lahouti, hassibi]@caltech.edu
Abstract— A networked controlled system (NCS) in which the
plant communicates to the controller over a channel with
random delay loss is considered. The channel model is motivated
by recent development of tree codes for NCS, which effectively
translates an erasure channel to one with random delay. A causal
transform coding scheme is presented which exploits the plant
state memory for efficient communications (compression) and
provides robustness to channel delay loss. In this setting, we
analyze the performance of linear quadratic Gaussian (LQG)
closed-loop systems and the design of the optimal controller. The
design of the transform code for LQG systems is posed as a
channel optimized source coding problem of minimizing a
weighted mean squared error over the channel. The solution is
characterized in two steps of obtaining the optimized causal
encoding and decoding transforms and rate allocation across a
set of transform coding quantizers. Numerical and simulation
results for Gauss-Markov sources and an LQG system
demonstrate the effectiveness of the proposed schemes.
I.
I
INTRODUCTION
t is now widely recognized that many of the future
applications of systems and control, such as those that
arise in the context of cyber-physical systems, will pertain
to problems of distributed estimation and control of multiple
agents such as sensors and actuators over unreliable networks.
While substantial research has been done on ways of
constructing channel codes to facilitate reliable
communications in network control systems (NCS), the
design of source codes for rate efficient communications in
such systems is much less explored. This paper presents the
design of a transform-based source coding scheme for
networked control systems communicating over channels
with delay loss.
In NCS, one needs to deal with both the typical real-time
constraints in control systems and the underlying unreliability
of communications in a simultaneous and systematic way. A
general framework to study such problems is introduced in [5]
in the context of interactive communications over noisy
channels. To solve the problem a new coding paradigm called
“tree coding” is presented [5]-[7]. Subsequently in [8] the
problem of stabilizing a plant over noisy links using tree
codes is considered. To this end, they recognized that to
stabilize an unstable plant over a noisy channel one needs
real-time encoding and decoding, and a reliability that
increases exponentially with delay. In [9], the existence with
“high probability” of “linear” tree codes is proven and explicit
codes with efficient encoding and decoding for the erasure
channel are presented. The tree codes effectively replace a
lossy link with a (asymptotically lossless) link with random
delay.
In [4], a linear quadratic Gaussian (LQG) control problem
where there is a lossy channel between the plant and the
controller is considered. Under some rather general
conditions, they show that the optimal LQG controller
“separates” into a state-feedback controller and a “causal
source coder” where the objective is to causally minimize a
weighted mean-square-error of the plant state (the particular
weighting coming from the LQG cost and plant dynamics).
While this is quite nice, little is known about constructing
appropriate causal source coding schemes.
Transform source coders ideally whiten the source signals
and produces uncorrelated transform coefficients, which are
more appropriate for quantization. Such a transform, which is
known as the Karhunen Loeve transform (KLT), has the
eigenvectors of the autocorrelation matrix of the source as its
bases. Taking advantage of transform codes in controls
applications calls for their equivalent causal counterparts. An
interesting point is that one may still design a causal transform
that optimally de-correlates the signal and hence achieves a
coding gain similar to what is expected of KLT. The penalty
though is that the transform will no longer be orthogonal and
we require a more delicate quantization architecture[10]-[12].
In this paper, a causal transform code with vector
quantization is presented for LQG closed-loop systems where
the plant state is communicated over a channel with random
delay. We analyze the system performance and the design of
the optimal controller in this setting. The code is designed to
minimize a weighted squared error measure and hence
amounts to a joint source and channel coding or a channel
optimized source coding scheme. To the best knowledge of
the authors, this is the first work in this direction.
This paper is organized as follows. Section II presents the
preliminaries including the system and channel models.
Section III presents the structure of the proposed robust causal
transform coding schemes. Section IV is devoted to the
optimized design of controller and robust source coding.
Performance evaluations of the proposed schemes in isolation
and in the context of an LQG system are presented in Section
V. Section VI concludes the paper.
II.
PRELIMINARIES
Figure 1 depicts the system model under consideration. We
consider the following linear quadratic Gaussian plant:
𝒙𝑡+1 = 𝐹𝒙𝑡 + 𝐺𝒖𝑡 + 𝒘𝑡
𝒚𝑡 = 𝐶𝒙𝑡 + 𝒗𝑡
(1)
where 𝒙𝑡 ∈ ℝ𝑑 is a vector of states, 𝒖𝑡 ∈ ℝ 𝑛 is a vector of
control signal, and 𝒚𝑡 ∈ ℝ𝑚 is a vector of plant output. The
terms 𝒘𝑡 ∈ ℝ𝑑 and 𝒗𝑡 ∈ ℝ𝑚 are independent and identically
distributed (iid) zero-mean Gaussian vectors modeling the
plant and measurement noise, respectively, with covariance
matrices 𝐾𝒘 and 𝐾𝒗 . We have 𝐹 ∈ ℝ𝑑×𝑑 , 𝐺 ∈ ℝ𝑑×𝑛 , 𝐶 ∈
ℝ𝑚×𝑑 and the pair of (𝐹, 𝐺) is controllable.
Assuming the system state is to be communicated over the
channel, the encoder maps each continuous plant state vector
to an index in a finite alphabet. More specifically, the plant
state is sampled with period 𝑇𝑠 and is encoded with a causal
transform code 𝐴 of length 𝑁 and quantizer set 𝑸. We have
𝒙𝒄 = ℰ𝐴,𝑸 (𝒙)
(2)
where 𝒙𝒄 is the encoder output that represents a reconstruction
codevector, whose index is communicated over a lossy
channel. In Section IV, we elaborate on the optimized design
of the causal transform code.
We consider a channel model with random independent
delay. Each encoder output index when communicated over
the channel is subject to a time varying random independent
delay 𝛿 which follows an exponential distribution with mean
1/𝜆,
(3)
𝑓𝛿 (𝛿) = 𝜆𝑒 −𝜆𝛿 .
Other models of channels with delay are available in [13].
The decoder maps back the indices received in time over
the channel to a signal in ℝ𝑚 , i.e., 𝒟: ℐ ⋃ {Erasure} ⟶ ℝ𝑚 .
In parallel to encoding, the decoding involves a reverse
quantization that uses the reproduction codebook 𝒞 and a
transformation 𝐴̂ which is not necessarily equal to 𝐴 because
of the channel. Hence, we have
̂ = 𝒟𝐴̂,𝒞 (𝒙
̃𝑐 )
𝒙
(4)
̃𝑐 is the channel output.
where as shown in Figure 1, 𝒙
If a decoding delay of Δ is acceptable, any communicated
index that is not available at the time of signal reconstruction
is assumed erased. Since decoding of each sample relies on
(up to) the past 𝑁 indexes, while a given index may be
determined erased at a given decoding time, it may arrive for
the subsequent decoding. The channel state at time 𝑖 is then
𝑁
described by a binary 𝑁 dimensional vector [𝑏𝑖𝑗 ]𝑗=1 , which
indicates the availability or unavailability of each of the
indices at the receiver with a one or zero, respectively.
When there is no channel between plant and controller, i.e.,
under full state observation, the optimal control policy is the
Fig. 1. System Model
certainty equivalent control. The addition of communication
channel to the system changes the problem to partially
observed problem [4]. The LQG cost when plant states are
sent over the channel turns into the following performance
metric
𝐾−1
lim sup
𝐾→∞
1
̂𝑡 ′𝑅𝒙
̂𝑡 + 𝒖′𝑡 𝑆𝒖𝑡 ] + 𝐸𝑋,𝐵 [𝒆′𝑡 𝑅𝒆𝑡 ]
∑ 𝐸𝑋,𝐵 [𝒙
𝐾
𝑡=0
(5)
where 𝑅 and 𝑆 are positive definite matrices with appropriate
1
dimensions and the pair of (𝐹, 𝑅2 ) is observable [4].
̂𝑡 is the decoder output, and the vector 𝒆𝑡 = 𝒙𝑡 −
Moreover, 𝒙
̂𝑡 indicates the state estimation error. Note that if the plant
𝒙
output is sent over the channel instead of the states, 𝑅 is
replaced with 𝑅̅ = 𝐶′𝑅𝐶 in (5), where 𝐶 describes how states
and plant outputs are related in (1).
III.
CAUSAL TRANSFORM CODING STRUCTURE
In this Section, we elaborate on the structure of causal
transform code in an LQG networked control system. We
consider a causal transform coding scheme of length 𝑁. The
efficiency in such settings is characterized by the trade-off of
rate and distortion that is typically quantified by mean squared
error.
Figure 2 depicts the structure of the causal transform code.
Within this structure, the encoder and decoder are designed
by identifying the encoding matrix 𝐴, the decoding matrix 𝐴̂,
and 𝑁 vector quantizers and their rates 𝑟𝑖 , 𝑖 = 1, … , 𝑁. The
encoding matrix 𝐴 is an 𝑚𝑁 × 𝑚𝑁 unit diagonal lower
triangular coding matrix
𝐼𝑚×𝑚
0
0
0
⋯
𝐴21 𝐼𝑚×𝑚
0
0
⋮
⋱
⋮
𝐴=
,
𝐴(𝑁−1)1 𝐴(𝑁−1)2
𝐼𝑚×𝑚
0
⋯ 𝐴
𝐴𝑁2
𝑁(𝑁−1) 𝐼𝑚×𝑚 )
( 𝐴𝑁1
where
𝛼𝑗𝑖,1
0
… 0
0
𝛼𝑗𝑖,2
… 0
𝐴𝑗𝑖 =
.
⋮
⋱
⋮
0
… 𝛼𝑗𝑖,𝑚 )
( 0
0
𝐼𝑚×𝑚 is an 𝑚 × 𝑚 identity matrix. The decoding matrix 𝐴̂ is
another 𝑚𝑁 × 𝑚𝑁 unit diagonal lower triangular matrix with
the same structure as 𝐴 where:
𝛼′𝑗𝑖,1
0
𝐴̂𝑗𝑖 =
⋮
0
(
0
𝛼′𝑗𝑖,2
0
0
…
…
⋱
…
0
0
⋮
𝛼′𝑗𝑖,𝑚
)
random variable indicating the (un)availability of transform
coding element 𝑗 at the decoder for the reconstruction of
element 𝑖; 𝑏𝑖𝑗 = 0, 1 ≤ 𝑖 < 𝑗 ≤ 𝑁.
It is straightforward to verify that the 𝑚𝑁-dimensional
reconstructed vector at the receiver may now be described as
follows
(10)
̂ = (𝐴̂ ∘ 𝐵)𝒙𝑐 = 𝐻𝒙𝑐 = 𝐻𝐴−1 (𝒙 + 𝒒)
𝒙
where 𝐻 = 𝐴̂ ∘ 𝐵 and operator ∘ is element-wise Hadamard
production.
Fig. 2. Structure of encoder, channel and decoder.
The quantization noise may be considered as an additive noise
vector 𝒒 of size 𝑚𝑁 × 1. The encoding operation in Figure 2
may be described by
𝒙𝑐 = 𝐴−1 (𝒙 + 𝒒),
(6)
where 𝒙 and 𝒙𝑐 are the 𝑚𝑁-dimensional vectors of successive
encoder input and output symbols, respectively.
Each vector quantizer is allocated 𝑟𝑖 bits satisfying
𝑁
1
∑ 𝑟𝑖 = 𝑟,
𝑁
(7)
𝑖=1
where 𝑟 is the average encoder bitrate.
At the decoder, the maximum allowable delay tolerated for
reconstruction of each component is defined based on the
requirements of the control system. Because of the ladder
structure, the construction and reconstruction of each
component affects the subsequent ones. At the time that the
𝑖’th element of the transform coding frame is to be
reconstructed, the 𝑗'th element (𝑗 ≤ 𝑖) has a maximum
allowable delay of ∆ + (𝑖 − 𝑗)𝑇𝑠 , and as a result 𝑝𝑟{𝑏𝑖𝑗 =
0} = 𝑝𝑟{𝛿𝑗𝑖 > ∆ + (𝑖 − 𝑗)𝑇𝑠 }. For our channel model, based
on (3) we have:
𝑃𝑟{𝛿𝑗𝑖 > ∆ + (𝑖 − 𝑗)𝑇𝑠 } = e−𝜆(∆+(𝑖−𝑗)𝑇𝑠) .
(8)
We define 𝑝𝑟{𝑏𝑖𝑖 = 0} = 𝑝𝑟{𝛿𝑖𝑖 > ∆} = 𝑝 and refer to it as
delay violation probability in simulations. Note that for the
reconstruction of elements that appear later in the transform
coding frame, greater delay in reception of earlier elements
could be tolerated.
To describe the decoder output, we present the channel
state information (discussed in Section II) within a transform
coding frame with the following matrix:
𝐵=
𝐵11
𝐵21
𝐵(𝑁−1)1
( 𝐵𝑁1
where
⋮
0
𝐵22
𝐵(𝑁−1)2
𝐵𝑁2
𝑏𝑖𝑗
0
𝐵𝑖𝑗 =
⋮
(0
⋯
⋱
⋯
0
𝑏𝑖𝑗
0
0
0
0
⋮
𝐵(𝑁−1)(𝑁−1)
𝐵𝑁(𝑁−1)
0
0
0
,
𝐵𝑁𝑁 )
… 0
… 0
⋱ ⋮
… 𝑏𝑖𝑗 )
where as stated 𝑏𝑖𝑗 ∈ {0,1}, 1 ≤ 𝑗 ≤ 𝑖 ≤ 𝑁 is a Bernoulli
(9)
IV.
CONTROLLER AND ROBUST SOURCE CODING
DESIGNS
In this section, the design of controller and robust causal
transform coding for closed-loop LQG problem is presented.
A.
Optimal Control Policy
Before dealing with the design, we consider the following
definition.
Definition: The control is said to have no-dual effect if for
all {𝒖𝑡 } and ∀𝑡, we have
̃𝑡𝑐 , 𝒖𝑡−1 } = 𝐸{𝒆′𝑡 𝒆𝑡 |𝒙
̃𝑡𝑐 }
𝐸{𝒆′𝑡 𝒆𝑡 |𝒙
(11)
which implies that control has no role in reducing state
uncertainty at the decoder or equivalently control has no
effect on the second term of (5). In (11), 𝒖𝑡−1 =
(𝒖𝟏 , 𝒖𝟐 , … , 𝒖𝑡−1 ). The following proposition elaborates on
optimal control policy when the robust causal transform code
is utilized over the delay loss channel model presented.
Proposition 1: For a closed-loop LQG system with causal
transform coding accompanying fine quantization and
decoder reconstruction in (10), the subsystem from encoder to
the decoder could be modeled by a block fading channel with
additive independent noise and the optimal control policy is
certainty equivalent control law.
Proof: Considering equation (9), it is evident that the block
consisting of the encoder, the channel and the decoder is a
fading channel with additive noise
̂ = 𝐻𝐴−1 𝒙 + 𝐻𝐴−1 𝒒 = 𝐻𝑒𝑞 𝒙 + 𝒏𝑒𝑞
𝒙
(12)
−1
where 𝐻𝑒𝑞 = (𝐴̂ ∘ 𝐵)𝐴 is a lower triangular fading matrix
which is constant over a transform coding frame, 𝑁, but varies
frame to frame. When fine quantization assumption is
invoked, the quantization noise, 𝒒, is independent of the input
signal, 𝒙. Hence, the RHS of (12) indicates a block fading
equivalent channel model for the encoder, lossy channel and
the decoder. Such a model satisfies the requirements of
lemma 5.2.1 in [16], and as a result the no-dual effect holds.
And hence, the optimal control policy is certainty equivalent
̂𝑡 [4] where,
control law 𝒖𝑡 = 𝐿𝒙
𝐿 = −(𝐺 ′ 𝑃𝐺 + 𝑆)−1 𝐺′𝑃𝐹
(13)
in which 𝑃 is a matrix that satisfies the following Riccati
equation
𝑃 = 𝐹 ′ (𝑃 − 𝑃𝐺(𝐺 ′ 𝑃𝐺 + 𝑆)−1 𝐺′𝑃)𝐹 + 𝑅.
(14)
∎
Corollary 1: For small delay violation probability, 𝑝 → 0, the
robust causal transform encoder, the lossy channel and the
corresponding decoder can be modeled as an equivalent
channel with additive independent noise, i.e.,
transform coding frame are equally weighted (over time), the
elements within each vector may be weighted differently.
Formally for a given channel and channel code the desired
robust transform code design problem is formulated as a
distortion-rate optimization problem as follows
1
̂‖𝑀 2 ]
min
𝐸 [‖𝒙 − 𝒙
(19)
𝐴,𝐴̂,{𝑟1 ,𝑟2 ,...,𝑟𝑁 }
𝑚𝑁 𝑋,𝐵
𝑁
̂ = 𝒙 + 𝒒.
𝒙
Proof: In Proposition 1, as 𝑝 → 0, 𝐵𝑖𝑗 ≈ 𝐼, 1 ≤ 𝑗 ≤ 𝑖 ≤ 𝑁 in
(9) and 𝐴̂ ≈ 𝐴 and 𝐻𝑒𝑞 ≈ 𝐼.
∎
B.
Code Design Problem Formulation and Analysis
We here investigate the design criteria for the robust causal
transform code based on the control policy prescribed in
Proposition 1. We also analyze the performance of the closedloop LQG system with robust causal transform coding.
We consider the partially observed LQG problem of (1). As
stated in [4], this could be converted to a fully observed LQ
̃𝑡+1
problem with 𝒘
̅ = 𝐸[𝐹𝒆𝑡 + 𝒘𝑡 |𝒖, 𝒙
Hence, the
𝑐 ].
covariance of this equivalent matrix is
𝐾𝒘̅ = 𝐹 ′ Λ𝐹 + 𝐾𝒘 − Λ
(15)
where Λ = Λ𝑡 = Cov(𝒆𝑡 ) and the optimal LQ cost is equal to
𝑡𝑟(𝑃𝐾𝒘̅ ) .
Considering (5) and (15) we have
𝐾−1
lim sup
𝐾→∞
1
̂𝑡 ′𝑅𝒙
̂𝑡 + 𝒖′𝑡 𝑆𝒖𝑡 ]
∑ 𝐸𝑋,𝐵 [𝒙
𝐾
(16.a)
𝑡=0
+ 𝐸𝑋,𝐵 [𝒆′𝑡 𝑅𝒆𝑡 ]
= 𝑡𝑟(𝑃𝐾𝒘̅ ) + 𝑡𝑟(𝑅Λ)
𝑠. 𝑡.:
1
∑ 𝑟𝑖 = 𝑟
𝑁
𝑖=1
This means specifically identifying the optimized causal
transform 𝐴𝑚𝑁×𝑚𝑁 , the corresponding decoding transform
𝐴̂𝑚𝑁×𝑚𝑁 , and the rates of the associated 𝑁 quantizers
{𝑟1 , 𝑟2 , . . . , 𝑟𝑁 } over the lossy channel.
The analysis in (16) and the Proposition 1 enable the analysis
of the LQG closed-loop system performance with robust
causal transform coding and in presence of delay loss in
communications. The following Proposition presents this
result.
Proposition 2: For a closed-loop LQG system with causal
transform coding accompanying fine quantization and
decoder reconstruction in (10), the LQG cost when the plant
state is sent over the channel is given by
𝑡𝑟(𝑃𝐾𝒘 ) + 𝑡𝑟(𝐸𝐵 [(𝐼 − 𝐻𝑒𝑞 )′𝑅𝑒𝑞 (𝐼 − 𝐻𝑒𝑞 )]𝐾𝒙 )
+ 𝑡𝑟(𝐸𝐵 [𝐻′ 𝑒𝑞 𝑅𝑒𝑞 𝐻𝑒𝑞 ]𝐾𝒒 )
(20)
where, 𝐾𝒙 and 𝐾𝒒 are plant state and quantization error
covariance matrices.
Proof: From Proposition 1, we have
𝒆 = (𝐼 − 𝐻𝑒𝑞 )𝒙 − 𝒏𝑒𝑞
(16.b)
And
= 𝑡𝑟(𝑃(𝐹 ′ Λ𝐹 + 𝐾𝒘 − Λ)) + 𝑡𝑟(𝑅Λ)
(16.c)
𝐸𝑋,𝐵 [𝒆′𝑅𝑒𝑞 𝒆] = 𝐸𝑋,𝐵 [((𝐼 − 𝐻𝑒𝑞 )𝒙 − 𝒏𝒆𝒒 ) 𝑅𝑒𝑞 ((𝐼 − 𝐻𝑒𝑞 )𝒙 − 𝒏𝒆𝒒 )]
= 𝑡𝑟(𝑃𝐾𝒘 ) + 𝑡𝑟((𝐹 ′ 𝑃𝐹 − 𝑃 + 𝑅)Λ).
(16.d)
′
′
= 𝐸𝑋,𝐵 [𝒙′ (𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 (𝐼 − 𝐻𝑒𝑞 )𝒙]
′
In fact, the effect of coding scheme on both terms of (16.a)
has been collected in the second term of (16.d) to provide us
with a criterion for code design. Based on this analysis, we
consider the arithmetic mean of the weighted mean square
error between the plant states and the decoder outputs as the
joint source channel coding design criteria
1
(17)
̂‖𝑀 2 ]
AM − WMSE =
𝐸𝑋,𝐵 [‖𝒙 − 𝒙
=
𝑚𝑁
1
𝑚𝑁
Where
𝑅𝑒𝑞 = 𝐹 ′ 𝑃𝐹 − 𝑃 + 𝑅.
′
−2𝐸𝑋,𝐵 [𝒙′ (𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 𝒏𝑒𝑞 ] + 𝐸𝑋,𝐵 [𝒒′ 𝐻′𝑒𝑞 𝑅𝑒𝑞 𝐻𝑒𝑞 𝒒]
The second term of above with fine quantization assumption
is zero, hence
1
′
𝑡𝑟(𝑅𝑒𝑞 Λ) = lim sup ∑𝐾−1
𝑡=0 𝐸𝑋,𝐵 [𝒆𝑡 𝑅𝑒𝑞 𝒆]
𝐾
′
= 𝑡𝑟(𝐸𝐵 [(𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 (𝐼 − 𝐻𝑒𝑞 )]𝐾𝒙 )
Comparing equations (16.d) and (17), according to causal
transform coding structure the weight matrix is 𝑀 is given by
0
𝑅𝑒𝑞
0
0
′
= 𝐸𝑋,𝐵 [𝒙′ (𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 (𝐼 − 𝐻𝑒𝑞 )𝒙]
𝐾→∞
̂)′𝑀(𝒙 − 𝒙
̂)]
𝐸𝑋,𝐵 [(𝒙 − 𝒙
𝑅𝑒𝑞
0
𝑀=
⋮
( 0
−2𝐸𝑋,𝐵 [𝒙′ (𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 𝒏𝑒𝑞 ] + 𝐸𝑋,𝐵 [𝒏′ 𝑒𝑞 𝑅𝑒𝑞 𝒏𝑒𝑞 ]
… 0
… 0
⋮
⋱
… 𝑅𝑒𝑞 )
+ 𝑡𝑟(𝐸𝐵 [𝐻′ 𝑒𝑞 𝑅𝑒𝑞 𝐻𝑒𝑞 ]𝐾𝒒 )
where, 𝐾𝒙 and 𝐾𝒒 are the plant state and quantization error
covariance matrices. Hence , the optimal LQG cost will be
′
𝑡𝑟(𝑃𝐾𝒘 ) + 𝑡𝑟(𝐸𝐵 [(𝐼 − 𝐻𝑒𝑞 ) 𝑅𝑒𝑞 (𝐼 − 𝐻𝑒𝑞 )]𝐾𝒙 )
+𝑡𝑟(𝐸𝐵 [𝐻′ 𝑒𝑞 𝑅𝑒𝑞 𝐻𝑒𝑞 ]𝐾𝒒 )
(18)
The weight matrix 𝑀 depends on the LQG cost and system
matrices and indicates that while different components of a
∎
Note that using Proposition 2, given the transform coding
matrices, the LQG system parameters and the covariance of
the quantization noise, one may numerically analyze the
system performance. The statistics of the quantization error
with fine quantization (asymptotic regimes) have been studied
in, e.g., [17].
Remark: In the setting of Corollary 1, when the delay
violation probability is small, the LQG cost in Proposition 2
simplifies as follows
𝑡𝑟(𝑃𝐾𝒘 ) + 𝑡𝑟(𝑅𝑒𝑞 𝐾𝒒 )
C.
Code Design Scheme
The robust source coding design optimization problem in
(19) does not lend itself to a direct closed-from solution, due
to the nonlinearity of the quantizers and the closed-loop
structure of the system. Instead, we take a two-step
approximate solution. To obtain 𝐴∗ and 𝐴̂∗ , we consider fine
quantization and invoke a numerical optimization approach in
line with the non-gradient direct search method of Hooks and
Jeeves [14]. Next, given the optimized transform matrices, we
obtain the optimized rate allocation scheme to the set of
quantizers.
A very good initial point for the search algorithm is the
causal transform corresponding to the ideal (lossless)
communication aka the Prediction-based lower triangular
transform (PLT) scheme [12]. Once optimized solutions of 𝐴∗
and 𝐴̂∗ are available in (19), it is straight forward to verify that
the quantization rate allocation problem may be posed as a
weighted quantization error [3] problem. We have
min
{𝑟1 ,…,𝑟𝑁 }
1
𝐸 {𝒒′𝑊𝒒}
𝑚𝑁 𝐵
𝑁
𝑠. 𝑡.:
1
∑ 𝑟𝑖 = 𝑟
𝑁
D.
Structured Encoding and Decoding Matrices
In addition to optimized matrices for Robust Causal
Transform Code (RC-TC), structured coding and decoding
matrices could also be considered for reduced design and
implementation complexity. One interesting structure is the
unit diagonal lower triangular and Toeplitz matrices for
coding and decoding. Such an encoding matrix is given by
𝐼𝑚×𝑚
0
…
0
0
𝐴21 𝐼𝑚×𝑚 …
0
0
𝐴21
⋱
0
0
𝐴 𝑇𝑜𝑒𝑝𝑙𝑖𝑡𝑧 = 𝐴31
⋮
𝐴21 𝐼𝑚×𝑚
0
𝐴31
𝐴31 𝐴21 𝐼𝑚×𝑚 )
…
( 𝐴𝑁1
We refer to this as the Robust Toeplitz Causal Transform
Code (RTC-TC). In RC-TC, because of the unit diagonal
lower triangular structure there are 𝑚(𝑁 2 − 𝑁) optimization
variables whereas in the case of RTC-TC there are only
2𝑚(𝑁 − 1) variables. This difference in complexity is
considerable for large 𝑁.
V.
PERFORMANCE RESULTS
In this Section, we study the performance of the proposed
robust causal transform code (RC-TC) and the robust Toeplitz
causal transform code (RTC-TC) for communication of a
Gauss-Markov source over a lossy channel. We also
investigate the performance of the proposed schemes in an
LQG system. For comparisons, we consider plain
quantization (no transform coding), and PLT [12] schemes.
(21)
A.
Robust Causal Transform Coding
𝑖=1
We experimented with two scalar (𝑚 = 1) GM sources with
where, 𝑊 = (𝐻𝐴−1 )′ 𝑀(𝐻𝐴−1 ). For a symmetric and orders 1 and 10 [15]. We here report the first order case where
positive definite 𝑊, decomposing 𝑊 = 𝑍′𝑍 with 𝑍 a lower the source is zero mean with unit variance and coefficient
triangular matrix, (21) coverts to an unweighted quantization equal to 0.9. We compare the AM-MSE performance (𝑀 =
error problem. The optimal bit allocation with fine 𝐼) for the four schemes. In the analysis, the covariance matrix
quantization assumption using Karush–Kuhn–Tucker (KKT) 𝐾 is set based on 𝜎 2 = 𝑐2−2𝑟𝑖 𝜎2 for a Gaussian source,
𝒒
𝑞𝑖
𝑑𝑖
conditions is
where 𝑐 is a constant depending on the type of quantizer [1].
1
𝜎̂𝑖2
Figure 3 shows the analytic and simulated AM-MSE
𝑟𝑖∗ = 𝑟 + log 2
1 , 𝑖 = 1,2, … , 𝑁
(22) performance results for the source. One sees that the
2
𝑁
2
(∏𝑘=1 𝜎̂𝑘 )𝑁
performance of PLT structure is better than that without
1
coding. In addition, the proposed RC-TC and RTC-TC
𝑚
where, 𝜎̂𝑖2 = [det (𝐸𝐵 [Cov (𝒅𝑖 𝑒𝑞 )])] , 𝒅𝑖𝑒𝑞 = 𝑍𝒅𝑖 and 𝒅𝑖
schemes still improve the performance. According to the
is the vector of inputs of the 𝑖-th quantizer. To derive (22), it conducted simulations, the transform coding schemes provide
has been assumed that the quantizer dimensions and input a higher gain for coding of sources with higher orders, which
distributions are the same (here Gaussian). Note that in may be attributed to their stronger memory structure. It is
presence of the channel, even if we have 𝑀 = 𝐼, we still face noteworthy that the RTC-TC provides a performance very
similar to that of the RC-TC, albeit with much smaller
a weighted quantization error problem as 𝑊 ≠ 𝐼.
Note that in the lossless scalar case, the rate allocation in (design) complexity especially for large values of 𝑁. herefore,
(22) coincides with what was obtained in [12] for the PLT in subsequent simulations we will focus on the RTC-TC.
Our experiments (not reported here) reveal that RTC-TC
structure. It is noteworthy that the optimized rate allocations
in (22) are obtained without regard to the non-negativity or and RC-TC with higher transform dimensions have better
possible integer constraints on quantization rates. In practice, performance. A larger 𝑁 allows the transform to exploit the
one may resort to heuristic techniques to impose such source dependencies more effectively and provides a greater
constraints on the obtained results. In [2], related algorithms opportunity for the delayed arrival of previous symbols at the
for the case of classic transform codes (non-causal) are cost of increased complexity.
reported.
Fig. 3. Performance of different schemes for 𝑁 = 6, 𝑟 = 5, ∆= 50
∆
msec and 𝑇𝑠 = .
4
B.
Robust Causal Transform Coding in NCS
In this part, we examine the performance of the proposed
robust causal transform coding schemes in a closed-loop LQG
system. We consider the following LQG system
𝑥𝑛+1 = 1.49𝑥𝑛 + 0.05𝑢𝑛 + 𝑤𝑛
(23)
𝑦𝑛 = 𝑥𝑛 + 𝑣𝑛
For this system, 𝑅 = 1, 𝑀 = 1 and variances of 𝑤𝑛 and 𝑣𝑛
are 0.01 and 0.001, respectively. For numerical optimization
of causal transform coding matrices, the plant states may be
modeled approximately by a first order GM source with
coefficient 0.8677.
Figure 4 depicts the LQG cost of the system in (23) for the
∆
four schemes and 𝑁 = 8, 𝑟 = 5, ∆= 0.05sec and 𝑇𝑠 = . As
4
evident, the proposed schemes significantly reduce the LQG
cost of the system and enhance the resilience of the system to
the delay loss over the channel.
{
VI.
CONCLUSIONS
The main purpose of this paper has been to design robust
causal transform codes for networked control systems
communicating over a channel with delay loss. Two channel
optimized coding scheme were proposed which provide a
trade-off of performance and complexity of design and
implementation. The suggested schemes exploit the
dependencies in the plant state for efficient communications
(compression) and provide robustness to channel loss. We
also analyzed the performance and the design of optimal
controllers with fine quantization. Numerical results and
simulations demonstrate the effectiveness of the presented
schemes for closed-loop LQG systems and the effect of
design parameters.
ACKNOWLEDGMENT
This work was supported in part by the National Science
Foundation under grants CCF-1423663 and CCF-1409204,
by a grant from Qualcomm Inc., by NASA’s Jet Propulsion
Laboratory through the President and Director’s Fund, and by
King Abdullah University of Science and Technology.
Fig. 4. LQG performance of different schemes as a function of delay
violation probability.
REFERENCES
[1]
N. S. Jayant and P. Noll, Digital Coding of Waveforms, Englewood
Cliffs, NJ:Prentice-Hall, 1954.
[2] K. Sayood, Introduction to Data Compression. San Francisco, CA:
Morgan Kaufmann, 2000.
[3] A. Gersho and R. M. Gray, Vector quantization and signal
compression, Springer, 1992.
[4] S. Tatikonda , A. Sahai and S. M. Mitter, "Stochastic linear control
over a communication channel," IEEE Trans. Autom. Control, vol.
49, no. 9, pp.1549 -1561, 2004.
[5] L. Schulman, “Coding for interactive communication,” IEEE
Transactions on Information Theory, Vol. 42, pp. 1745–1756, Dec.
1996.
[6] M. Braverman and A. Rao, “Towards coding for maximum errors in
interactive communication,” in Proc. of the 43rd annual ACM
Symposium on Theory of Computing, 2011.
[7] R. Gelles and A. Sahai, “Potent tree codes and their applications:
Coding for interactive communication, revisited,” CoRR, vol.
abs/1104.0739, 2011.
[8] A. Sahai and S. Mitter, “The necessity and sufficiency of anytime
capacity for stabilization of a linear system over a noisy communication
link–part i: Scalar systems,” IEEE Transactions on Information
Theory, 52:3369–3395, August 2006.
[9] R. T. Sukhavasi and B. Hassibi, “Error correcting codes for
distributed control,” http://arxiv.org/abs/1112.4236l, February 2012.
[10] F. Lahouti and A. K. Khandani, “Sequential vector decorrelation
technique,” Technical Report UW-ECE-2001-3, 2001.
[11] S. M. Phoong and Y. P. Lin, “Prediction-based lower triangular
transform,” IEEE Trans. Signal Process., Vol. 48, No. 7, pp. 1947–
1955, July 2000.
[12] C.-C. Weng, C.-Y. Chen, and P. P. Vaidyanathan, “Generalized
triangular decomposition in transform coding,” IEEE Trans. Signal
Process., Vol. 58, No. 2, February 2010.
[13] J. Nilsson, Real-time control systems with delays, 1998 :Dept.
Automatic Control, Lund Inst. Technol.
[14] R. Hooke and T.A. Jeeves, “Direct search solution of numerical and
statistical problems,” Journal of the Association for Computing
Machinery (ACM), 8(2):212–229, 1961.
[15] F. Lahouti and A. K. Khandani, "Reconstruction of predictively
encoded signals over noisy channels using a sequence MMSE
decoder," IEEE Trans. Commun., pp.1292 -1301, 2004.
[16] D. P. Bertsekas. Dynamic Programming and Optimal Control. Athena
Scientific ,2nd ed. , 2000.
[17] D. H. Lee and D. L. Neuhoff, "Asymptotic distribution of the errors in
scalar and vector quantizers", IEEE Trans. Inform. Theory, vol. 42,
pp.446 -460 1996.
| 3cs.SY
|
arXiv:cs/0602098v2 [cs.PL] 7 May 2006
Compositional Semantics
for the Procedural Interpretation of Logic
Research Report DCS-307-IR∗
Department of Computer Science
University of Victoria
M.H. van Emden
Abstract
Semantics of logic programs has been given by proof theory, model
theory and by fixpoint of the immediate-consequence operator. If clausal
logic is a programming language, then it should also have a compositional
semantics. Compositional semantics for programming languages follows
the abstract syntax of programs, composing the meaning of a unit by a
mathematical operation on the meanings of its constituent units. The
procedural interpretation of logic has only yielded an incomplete abstract
syntax for logic programs. We complete it and use the result as basis of a
compositional semantics. We present for comparison Tarski’s algebraization of first-order predicate logic, which is in substance the compositional
semantics for his choice of syntax. We characterize our semantics by
equivalence with the immediate-consequence operator.
1
Introduction
This paper concerns the semantics of the part of Prolog that remains
when the built-in predicates have been removed and when unification
is enhanced by the occurs check. Let us call this part “pure Prolog”.
It can be regarded as the result of Kowalski’s procedural interpretation
of positive Horn clauses [14, 13, 8]. The semantics of pure Prolog has
been given by proof theory, by model-theory, and by a fixpoint method
[22, 16, 1]. All three approaches follow the syntax of clausal form. As
a result, the procedural interpretation has been ignored. The purpose of
the present paper is to remedy this defect.
One of the symptoms of the current deficiency in the semantics of
Prolog is that procedures can only be recognized in an informal way.
As it stands, the procedural interpretation does not provide procedurevalued expressions that can be substituted for the procedure symbol in a
∗ Shortened version to appear in the Proceedings of the 2006 International Conference on
Logic Programming.
1
procedure call. Procedures are not “first-class citizens” the way functions
can be in functional programming [20].
Compositional semantics does provide this possibility. According to
this method, programs are expressions, consisting, if composite, of an
operation and its operand(s). The value of the composite expression is the
result of the operation on the values of its operands. The method is taken
for granted when doing school-room sums: the value of (4 ÷ 2) × (1 + 1) is
4 because the value of 4 ÷ 2 and 1 + 1 are both 2 and because 2 × 2 = 4.
In the late sixties Landin [15] and Scott and Strachey [19] applied the
method to expressions that are programs.
In logic programming, compositional semantics seems to have been
used only for elucidating how the union of two logic programs affects the
definition of a predicate [6, 4]. In this paper we identify the compositions
that occur within a clause and give a compositional semantics for these.
There are several advantages to a compositional semantics for a programming language. One is that it guides implementation. In fact,
“syntax-directed compilation” [12], a widely used implementation technique, is compositional semantics avant la lettre. The compositional semantics presented here decomposes logic programs down to single procedure symbols, which take relations as value. This accommodates relations
that are not defined in the logic program itself.
Another advantage of compositional semantics is that it forces a language to be modular. For example, in a functional language with compositional semantics E0 E1 is the result of applying the value of E0 , which must
be a function, to the value of E1 , which may or may not be a function.
The result can be a function, but need not be.
Compositionality requires that the value of E0 E1 does not change
when E0 is replaced by a different expression with the same value. This
forces modularity in the sense that names of auxiliary functions occurring
in E0 do not affect its value, hence are local. Compositional semantics
endows logic programs with the same property. The value of a procedure
call p(t0 , . . . , tn−1 ) is obtained by an operation on the value of p (which
is a relation) and the argument tuple ht0 , . . . , tn−1 i. Again, the result
depends on the value of the relational expression substituted for p, not on
the expression itself.
Contributions of this paper When one attempts a compositional
semantics for the procedural interpretation of logic, it becomes apparent
that it needs development beyond Kowalski’s original formulation. This
is done in Section 3.
Section 4 contains no contributions. It needs to be included because
cylindric set algebras are a compositional semantics for first-order predicate logic and hence are a candidate for compositional semantics for the
procedural interpretation of logic. This section includes enough to show
why these algebras are not suitable. We do find, however, an interesting
connection between the tables introduced here and the cylinders of Tarski
(see Theorem 2).
Tables, their operations and some of their properties are described
in Section 5. This is the basis on which the compositional semantics of
2
Section 6 rests. Implications for modularity are discussed in Section 7.
2
Notation and terminology
In this section we collect terminology and notation that may differ between
authors.
2.1
General terminology
Definition 1 (tuple, function, index set, type, restriction, subtuple)
– A tuple is a function t that maps every index i to t(i), which is called
the tuple’s component at i.
– A function is a triple consisting of a set that is its domain, a set that is
its co-domain, and a mapping that associates with every element of the
domain a unique element of the co-domain.
– If the function is a tuple, then the domain is usually called “index set”.
– The set of all functions with domain S and co-domain T is denoted
S → T . This set is often referred to as the type of the functions belonging
to it.
– Let f be a function in S → T and let S ′ be a subset of S. f ↓ S ′ is
the restriction of f to S ′ . It has S ′ as domain, T as co-domain and its
mapping associates f (x) ∈ T with every x ∈ S ′ .
– If t is a tuple with index set I and if I ′ is a subset of I, then t ↓ I ′ is
the subtuple of t defined by I ′ .
Definition 2 (relation) A relation with index set I and co-domain T
is a set of tuples that have I as index set and T as co-domain. An nary relation is a relation that has as index set the set {0, . . . , n − 1} of
integers.
Note that a relation need not be an n-ary relation. Indeed, any set can
be the index set of a relation.
Definition 3 (projection, cylindrification) Let r be a relation that
has I as index set. Let I ′ be a subset of I. The projection πI ′ (r) of r on
I ′ is {t ↓ I ′ | t ∈ r}.
The cylinder in I on a relation r ′ with index set I ′ is denoted πI−1 (r ′ )
and is the greatest relation with index set I and co-domain T that has r ′
as its projection on I ′ ; that is
πI−1 (r ′ ) = ∪{ρ | πI ′ (ρ) = r ′ and ρ has index set I and co-domain T }
2.2 Mathematical objects arising in connection
with the semantics of logic programs
To serve as semantic objects, three basic objects are defined independently
of another; all three are mutually disjoint sets:
• H, an Herbrand universe
• V, a set of variables
• P, a set of predicate symbols, also called “procedure symbols”
3
From the three basic objects the following are derived:
• TV , the set of terms that contain no function symbols or constants
other than those occurring in H and no variables other than those
occurring in a subset V of V. We write T for TV .
• Substitutions, each of which is a tuple of type V → TV , for some
subset V of V. If θ is a substitution and θ(x) = t, then we say that
θ substitutes t for x. We may equate θ with the set {x = t | θ(x) =
t and x ∈ V } of term equations.
• Term equations are equations of the form t0 = t1 , where t0 and t1 are
terms belonging to T . A set of term equations is said to be in solved
form if every left-hand side is a variable, and if all these variables
are different, and if all variables in the right-hand sides also occur
as a left-hand side. If a set of term equations has a solution, then it
has a solution in solved form.
We will not distinguish between term equations in solved form, substitutions, and tuples of elements of T with a subset of V as index
set.
• Relations consisting of tuples of elements of H that are indexed by
{0, . . . , n − 1}. To distinguish these from the next item, we refer to
them as integer-indexed relations.
• Relations consisting of tuples of elements of H that are indexed by
a subset V of V that is characteristic of the relation. We refer to
these as variable-indexed relations.
• The Herbrand base, which is the set of ground atoms.
• Herbrand interpretations, which are subsets of the Herbrand base.
• Relational interpretations, which are tuples of integer-indexed relations indexed by P.
2.3
Compositional semantics
Compositional semantics assigns the value M(E) to the expression E. We
are interested in expressions that are programs. In this case the value is
the behaviour of the program. As “value” and “behaviour” do not match
very well, we often use “meaning” instead of “value” as a more neutral
term. It also happens to fit well with “semantics”.
Compositionality of the semantics means that if E is composed of
subexpressions E0 and E1 , then M(E) is the result of an operation on
M(E0 ) and M(E1 ). A well-known example illustrates the compositional
semantics of binary numerals. It specifies how integers are assigned as
meanings to binary numerals:
M(0) = 0; M(1) = 1; M(N0) = 2M(N); M(N1) = 2M(N) + 1
4
3 The procedural interpretation of positive Horn clauses
3.1
The original procedural interpretation
Kowalski [14] gives the procedural interpretation of positive Horn clauses
as follows:
“A Horn clause B ← A1 , . . . , Am , with m ≥ 0, is interpreted
as a procedure whose body {A1 , . . . , Am } is a set of procedure
calls Ai . Top-down derivations are computations. Generation of a new goal statement from an old one by matching
the selected procedure call with the name B of a procedure
B ← A1 , . . . , Am is a procedure invocation.
A logic program consists of a set of Horn clause procedures and
is activated by an initial goal statement.”
Its semantics can be given by the fact that a ground substitution θ
is included in a result of activating program P with goal G iff P ∪ {Gθ}
is false in all Herbrand interpretations. A more general characterization
exists.
3.2
A complete procedural interpretation
The procedural interpretation of logic can be formalized by expressing
it as an abstract procedural syntax. Kowalski proposed, in effect, B ←
A0 , . . . , Am−1 as an alternative syntax in the form of a decomposition of
{B, ¬A0 , . . . , ¬Am−1 } into a procedure heading and a procedure body.
This omits several decomposition steps: (1) the clause may be but one of
several several that can respond to the same procedure call, so it is really
a partial procedure, (2) a body needs to be decomposed into calls, and
(3) each call needs to be decomposed into its predicate symbol and its
argument tuple. To make the procedural interpretation not only formal,
but also to complete it, we propose Definition 4 as the abstract syntax
needed for compositional semantics.
Definition 4 (procedural program)
1. A procedural program is a
tuple of procedures with index set P 1 .
2. A n-ary procedure is a set of n-ary clauses.
3. An n-ary clause is a pair consisting of a parameter tuple of order n
and a procedure body.
4. A procedure body is a set of procedure calls.
5. A procedure call is a pair consisting of an n-ary procedure symbol
and an argument tuple of order n.
6. A parameter tuple of order n and an argument tuple of order n are
both n-tuples of terms.
1 The procedure symbols in P index only one procedure. This differs from Prolog where
predicate symbols include an arity.
5
app(nil,y,y).
app(u.x,y,u.z) :- app(x,y,z).
{app{(nil,y,y) :- {}
,(u.x,y,u.z) :- {app(x,y,z)}
}
,mem{(x,y) :- {app(u,x.v,y)}}
}
mem(x,y) :- app(u,x.v,y).
Figure 1: A Prolog program (left) and an equivalent procedural program (right).
Let us consider as example a set P = {app,mem} of procedure symbols
and the procedural program in Figure 1; let us call it p. As p is a tuple
with P as index set, and as a tuple is a function, p can be specified by
p(app) = {(nil,y,y) :- {}, (u.x,y,u.z) :- {app(x,y,z)}}
p(mem) = {(x,y) :- {app(u,x.v,y)}}
By itself, Definition 4 defines some procedural language. It is only of
interest in so far as it is related to clausal logic. Similarly, the relational
interpretations for procedural programs need to be related to Herbrand
interpretations. Hence the following definition.
Definition 5 (correspondence between logic and procedural programs)
An Herbrand interpretation I and a relational interpretation R correspond
to each other (I ∼ R) iff the following holds:
R(p) = {ha0 , . . . , an−1 i | p(a0 , . . . , an−1 ) ∈ I} for all p ∈ P and
I = {p(a0 , . . . , an−1 ) | p ∈ P and ha0 , . . . , an−1 i ∈ R(p)}
Let S be a sentence consisting of positive Horn clauses (for which we
assume Kowalski’s notation). Let P be a procedural program. S and P
correspond to each other (S ∼ P ) iff the following holds:
P (p) = {partuple :- body | p(partuple) ← body ∈ S} for all p ∈ P and
S = {p(partuple) ← body | ∃p ∈ P such that partuple :- body ∈ P (p)}.
Each of the syntactical rules of Definition 4 specifies that a certain type
of expression is composed of sub-expressions. Compositional semantics
then assigns to each of syntactical rules a semantical rule that specifies the
corresponding operation on meanings of the constituent sub-expressions.
The next section introduces the mathematical objects that are suitable
meanings. Section 6 describes the semantical rules.
Before starting on this we give an informal idea of what is involved.
Let us work through the items in Definition 4, starting at the bottom.
Rule 6 Here we have very little to add: a term denotes the set of its ground
instances; a tuple of terms denotes a tuple of sets of ground instances.
Rule 5 Consider the atoms p(x, v, w) and p(u, w, y). Although both involve
the same relation p, they are different calls and typically have different meanings. These meanings are the result of a binary operation
with the relation p and the tuple of arguments as operands.
6
The meaning of the entire call can be viewed as a selection from the
tuples that constitute relation p. The selection is specified by the
argument tuple, and selects the tuples from the relation that match
the argument tuple. Each such match takes the form of a substitution for the variables in the argument tuple. Therefore the result of
the operation, which we call filtering, is a set of such substitutions.
As such sets are best presented in tabular form, we call the result of
the filtering operation on a relation and an argument tuple a table
(see Definition 6).
Rule 4 We define the product operation on tables (see Definition 8) by means
of which procedure bodies obtain values. These values are tables.
Theorem 2 shows how product is related to the semantic counterpart
of conjunction in Tarski’s cylindric set algebra.
Rule 3 The meaning of a clause is the n-ary relation that results from an
operation on the meanings of the constituents of the clause: the
parameter tuple and the body. As a parameter tuple has itself as
meaning, we define an operation, which we call projection, on a
parameter tuple of order n and a table (see Definition 11). The
operation yields an n-ary relation.
This completes the preview of the novel semantic operations: filtering,
product, and projection. The remaining operations, those arising from
Rules 1 and 2, will not require any explanation beyond the following few
lines. In Rule 2, a procedure symbol is combined with a set of clauses. As
the meaning of a clause is an n-ary relation, a set of such clauses denotes
the union of these relations, that is, an n-ary relation again. Rule 2 merely
creates a pair consisting of a procedure symbol and a relation.
Rule 1 combines into a set a number of procedures, each of which
is a pair of a procedure symbol and a relation. The semantic object
corresponding to a program is therefore a tuple of procedures indexed by
P, the set of procedure symbols.
4
Compositional semantics for logic
Though there does not seem to exist any compositional semantics for the
procedural interpretation of logic, one does exist for logic that is parsed
in the conventional way. It is called algebraic logic, which would be called
compositional semantics if it would concern a programming language. It
is therefore a good starting point for a compositional semantics of logic
programs.
Algebraic logic assigns elements of an algebra as meanings to formulas
of logic; it assigns operations of the algebra as meaning to the connectives that compose logical formulas. The more widely known approach to
algebraic is based on the cylindric set algebras of Tarski [9, 21] of which
we give a brief sketch here. Tarski’s approach is based on the algebraic
interpretation of propositional logic due to Boole [2].
7
4.1
Propositional logic and Boolean algebra
In general, a Boolean algebra is any algebra that satisfies certain defining
axioms. A Boolean set algebra is a special case. It is described as the
tuple hS, ∪, ∩, ∼, ∅, U i where S is a set of subsets of U that contains ∅ and
U and is closed under union, intersection, and complementation (here
denoted as ∼).
A special case of a Boolean set algebra is the one where U is the
Cartesian product Dn , for some given non-empty set D. Recall that the
Cartesian product Dn is the set of all n-tuples of elements of D. We
can further specify the Boolean set algebra by choosing U = D0 = {hi}
and S = {{}, {hi}}. As a result, the algebra has two elements: {} and
{hi}. Boolean addition, multiplication, and complementation then become
set union, set intersection, and set complement, respectively. Let M be
the mapping from propositional formulas to the elements of the Boolean
algebra. We have that M(p0 ∨ p1 ) = M(p0 ) ∪ M(p1 ), M(p0 ∧ p1 ) =
M(p0 ) ∩ M(p1 ), and M(¬p) = ∼ M(p) when we define M(true) = {hi}
and M(f alse) = {}.
4.2
Predicate logic and cylindric set algebra
Tarski sought an algebra that would do for first-order predicate logic what
Boolean algebra does for propositional logic. The result was cylindric set
algebra [21, 9].
In model theory, formulas correspond to relations. If this intuitively
attractive feature is to be retained, a puzzle needs to be solved. Consider
M(p(x, y) ∧ p(y, z)). As the formula has three free variables, this should
be a ternary relation. As conjunction means the same in predicate logic
as in propositional logic, this ternary relation should be the result of set
intersection. But the arguments of the set intersection are derived from
binary predicates.
Another part of the puzzle is that p(x, y) and p(y, z) should both
denote binary relations, but these should be different and cannot both be
the relation denoted by p.
Tarski solved these conundrums by mapping every formula to a relation
consisting tuples indexed by all the variables in the language. He assumed
a countable infinity of variables in the language, in a given order. In this
way he could identify each variable with a natural number. Thus this
meaning algebra has as elements relations that are subsets of the Cartesian
product Dω .
The choice of the two 0-ary relations on D for the two elements of
the Boolean algebra for propositional logic is now clear: the number of
variables in a propositional formula is 0.
A first-order predicate logic formula without free variables is either
true or false. It is mapped accordingly to the full or empty ω-ary relation
over D; that is, to Dω or ∅. At first sight it might seem right to map
a formula F [x0 , . . . , xn−1 ] with free variables x0 , . . . , xn−1 to the relation
that consists of all the tuples ha0 , . . . , an−1 i such that F [a0 , . . . , an−1 ] is
true. By mapping instead this formula to the cylinder on this relation with
respect to all variables, Tarski ensured that M(p0 ∨ p1 ) = M(p0 ) ∪ M(p1 )
8
and M(p0 ∧ p1 ) = M(p0 ) ∩ M(p1 ), just as in the case of propositional
logic.
Going back to the above puzzle, we see that M(p(x, y)) and M(p(y, z))
are not binary relations but ω-ary relations that are cylinders on a binary
relation. Though the binary relation denoted by p in these formulas is the
same, the cylinders on M(p(x, y)) and M(p(y, z)) are different. In this
way M(p(x, y)) ∩ M(p(y, z)) is a cylinder on a ternary relation.
Thus Tarski devised a compositional semantics for first-order predicate
logic. He simplified the language to contain as connectives only conjunction, disjunction, and negation. The presence of the negation connective
makes it possible to do with a single quantifier, the existential one. There
are no function symbols. An atomic formula can be of the form x = y.
For this language a suitable algebra for a compositional semantics is
the cylindric set algebra hS, ∪, ∩, ∼, ∅, Dω , Ck , δi,j i for all natural numbers
i, j, and k. This algebra is a Boolean algebra (for the first six items). In
addition, there are δi,j , the (i, j) diagonal relations: the subsets of Dω
consisting of the tuples where the elements indexed by i and j are equal.
The specification of cylindric set algebras also includes for all k ∈ ω the
cylindrification operations Ck , which are defined by Ck r being the subset
of Dω consisting of the tuples that differ from a tuple in r in at most the
k-th component.
S is the set that contains ∅, Dω , as well as all the diagonal relations
δi,j and that is closed under the Boolean operations as well as under Ck .
4.3 Cylindric set algebra for the compositional
semantics of procedural programs?
Cylindric set algebra interprets formulas as relations; relations are a suitable model for the procedures of a procedure-oriented language. These
facts might suggest that cylindric set algebras be used for a compositional
semantics for the procedural interpretation of logic.
The following are reasons not to do so.
• Tarski’s choice of language for first-order predicate logic is no more
procedure-oriented than clausal form is.
• Tarski’s semantics does not specify by what operation, for example,
the binary relation M(p(x, y, x)) arises from the ternary relation p
and the argument tuple hx, y, xi. That is, his compositionality stops
short of the atomic formula.
Accordingly, we create an independent alternative, centered around the
concept of table. Surprisingly, one of the operations on tables reflects the
way Tarski uses cylinders to algebraize the logical connectives.
5
Tables
Some of the semantic objects for the procedural programs of Definition 4
are familiar; they have been introduced in Section 2. This section is
devoted to the one novel type of semantic object.
9
Definition 6 (table) A table on a subset V of V is a set of tuples each
of which has type V → TV . If the set of tuples is empty, then we have
the null table, which we write as ⊥. If V is empty and the set of tuples
is not, then the table is the unit table, which we write as ⊤.
As there is only one function of type {} → TV for any subset V of V, we
have that ⊤ = {hi}.
To every table there corresponds a unique variable-indexed relation,
which we call the result of grounding the table.
Definition 7 (grounding, table equivalence) Let t be a table with tuples of type V → TV . Γ(t), the result of grounding t, is the variableindexed relation consisting of the tuples of type V → H each of which is a
ground instance of a tuple in t.
Tables t0 and t1 are equivalent if Γ(t0 ) = Γ(t1 ).
In this section we define and discuss the product, filtering, and projection operations. These operations are adapted from [10], where filtering
is called “application”.
5.1
Product
As we will see, compositional semantics assigns tables as values to the calls
in a procedure body as well as to the body itself. The co-occurrence of
calls in a body corresponds to the product operation of the corresponding
tables. An example will be given in Section 6.1.
Definition 8 (product) Let τ0 and τ1 be tables consisting of tuples with
index sets V0 and V1 , respectively. The product τ0 ∗ τ1 of these tables is
defined as a table with V0 ∪ V1 as index set. The product table τ0 ∗ τ1
contains a tuple t if and only if there is a tuple t0 in τ0 and a tuple t1
in τ1 such that the set of equations t0 ∪ t1 is solvable and has t as solved
form.
Theorem 1 – Product is commutative and associative.
– The null table ⊥ is an absorbing element: ⊥ ∗ τ = τ ∗ ⊥ = ⊥ for all
tables τ .
– The top table ⊤ is a unit: ⊤ ∗ τ = τ ∗ ⊤ = τ for all tables t.
– τ ∗ τ and τ are equivalent.
Commutativity and associativity give the obvious meaning to ∗S,
where S is a set of tables, assuming that ∗{} = ⊤.
Definition 9 (cylinder on table) The cylinder π −1 (T ) on a table T
with index set V ∈ V is a table where V is the index set and where every
tuple t′ is obtained from a tuple t in T by defining t′ (v) = t(v) for every
v ∈ V and t′ (v) = v for every v ∈ V \ V .
This definition of “cylinder” is independent of Tarski’s notion, which is
the one in Definition 3. The two notions are connected as follows.
Lemma 1 Let T be a table with index set V , a subset of V. We have that
Γ(π −1 (T )) = π −1 (Γ(T )). The first occurrence of π −1 is the cylindrification on tables from Definition 9; the second one is the cylindrification on
relations in Definition 3.
10
The distinguishing feature of Tarski’s use of cylindric set algebra as semantics for first-order predicate logic is that conjunction in logic simply
translates to intersection in the algebra. And this is the case even though
the conjunction may be between two formulas with sets V0 and V1 of free
variables. There is no restriction on these sets: they may be disjoint, one
may be a subset of the other, or neither may be the case. Tarski’s device
works because the intersection is not between relations with V0 and V1 as
index sets, but between cylinders on these relations in the set of all variables. This crucial idea reappears in the product of tables defined here.
The connection is made apparent by the following theorem.
Theorem 2 Let τi be a table with set Vi of variables, for i ∈ {0, 1}.
Γ(τ0 ∗ τ1 ) = πV0 ∪V1 (πV−1 (Γ(τ0 )) ∩ πV−1 (Γ(τ1 ))).
5.2
Filtering: from relations to tables
Just as in a functional programming language a function is applied to
the n-tuple of its arguments, we think of the combination of a procedure
symbol with its argument tuple as a binary operation. Consider therefore
a call consisting of a procedure symbol and an argument tuple of order n.
The procedure symbol has as value an integer-indexed relation of order
n. It combines with the argument tuple to produce a table. This is the
operation we call filtering. An example of this operation can be found in
Section 6.1.
Definition 10 (filtering) Let p be an integer-indexed relation of order
n and let t be an n-tuple of terms with V as set of variables. The result of
the filtering p : t is a table where V is the index set of the tuples. For every
tuple ha0 , . . . , an−1 i in p for which the set {t0 = a0 , . . . , tn−1 = an−1 } of
equations is solvable, the table contains a tuple that is the solved form of
these equations.
In functional programming, an expression E0 E1 denotes function application. Here E0 is an expression that evaluates to a function, and it is this
function that is applied. Filtering is the relational counterpart: in p : t
the first operand p has a relation as value; it is filtered by the tuple t; the
result is a table.
5.3 Projection: from tables to integer-indexed relations
Finally, a clause is a contribution to a procedure, which is an integerindexed relation of order n. This relation, which is the clause’s value,
is somehow produced by a combination of the parameter tuple of the
clause and the table that is the value of its body. We call this operation
projection. An example of this operation can be found in Section 6.1.
Definition 11 (projection) Let T be a table consisting of tuples whose
index set is a subset V of V . The result of projecting T on an n-tuple of
terms, denoted ht0 , . . . , tn−1 i/T , is an integer-indexed relation consisting
of n-tuples of ground terms. The relation contains such a tuple if and only
if it is a ground instance of ht0 θ, . . . , tn−1 θi, for some θ in T .
11
Usually every variable in the parameter tuple of a clause occurs also
in the body of that clause. It often happens that the parameter tuple contains fewer variables than the body. For example, mem(x,y) :- app(u,x.v,y).
Any operation that yields an integer-indexed relation consisting of ntuples from a body that contains more than n variables is reminiscent
of a projection operation. Hence the name.
5.4
Are projection and filtering inverses?
Now that we have operations from tables to relations and vice versa, one
may wonder whether these are each other’s inverses. The short answer is,
in general, “no”, because
((t0 , . . . , tn−1 )/T ) : (t0 , . . . , tn−1 )
is not always the table T . Take, for example, the case that t0 , . . . , tn−1
have no variables. Then the above expression is ⊤ whenever T is not ⊥.
But the absence of variables in t0 , . . . , tn−1 is a rather pathological case.
When we add certain restrictions, we can say that, in a sense, “/” and “:”
are each other’s inverses, as shown by the following theorems.
Theorem 3 For all tables T with a subset V of V as index set and all
terms t0 , . . . , tn−1 with V as set of variables ((t0 , . . . , tn−1 )/T ) : (t0 , . . . , tn−1 )
is equivalent to T .
For an inverse in the other direction, compare the n-ary relation r with
(t0 , . . . , tn−1 )/(r : (t0 , . . . , tn−1 )).
That this expression does not always equal r is shown by (c, d)/({(a, b)} :
(c, d)) = {}, where a, b, c and d are constants. This example suggests:
Theorem 4 For all n-ary relations r and all terms t0 , . . . , tn−1 , we have
(t0 , . . . , tn−1 )/(r : (t0 , . . . , tn−1 )) ⊆ r.
However, by strengthening the restrictions, we can have equality instead of inclusion, as shown in the following theorem.
Theorem 5 For all n-ary relations r and all distinct variables x0 , . . . , xn−1 ,
we have
(x0 , . . . , xn−1 )/(r : (x0 , . . . , xn−1 )) = r.
6
Compositional semantics
The operations of product, filtering, and projection are intended to be the
semantical counterparts of the way in which procedural programs are put
together syntactically. But so far only the intention exists.
The definition below formalizes this intention. It defines the meaning M(P ) of a procedural program P , where P is regarded as a tuple
with index set P of integer-indexed relations. This meaning depends on
a relational interpretation I (Definition 5) that assigns relations to the
12
procedure symbols in P. We indicate this dependence by a subscript, as
in MI .
Definition 12 gives the compositional semantics for procedural programs. As Definition 5 shows, a procedural program is just another way
of writing a set of positive Horn clauses. The semantics of these has been
defined in three equivalent ways: model-theoretically, proof-theoretically,
and by means of fixpoints. The main theorem (6) of this paper relates
the compositional semantics of procedural programs to the established
semantics of the corresponding clausal sentences.
Definition 12
1. For every procedural program prog, MI (prog) is the
tuple with index set P such that for every prsym ∈ P the prsymcomponent is MI (prog(prsym)).
2. For every procedure proc, MI (proc) = ∪{MI (clause) | clause ∈
proc}
3. For every clause with pars as parameter tuple and B as body,
MI (pars :- B) = pars/MI (B) (use of projection)
4. For every procedure body B, we have MI (B) = ∗{MI (call) | call ∈
B} (use of product)
5. For every call with prsym as procedure symbol and args as argument
tuple, MI (prsym args) = MI (prsym) : args (use of filtering)
6. For every prsym ∈ P we have that MI (prsym) = I(prsym)
Here the numbering follows that of the syntactical rules of Definition 4.
Theorem 6 Let I be a relational interpretation and I ′ the corresponding
(Definition 5) Herbrand interpretation. Let P be a procedural program
and P ′ the corresponding (Definition 5) set of positive Horn clauses. We
have
TP ′ (I ′ ) ∼ MI (P ),
where T is the immediate-consequence operator for logic programs.
We only know a cumbersome, though straightforward, proof of this
theorem.
T has a unique least fixpoint [22, 16, 1]. The partial order among Herbrand interpretations (set inclusion) translates according to the correspondence in Definition 5 to a partial order among relational interpretations
(component-wise inclusion). Hence there is, for each procedural program
P , a unique least relational interpretation I such that I = MI (P ).
Definition 13 M(P ) = MIm (P ) where Im is the least relational interpretation I such that I = MI (P ).
Theorem 7 Let P ′ be a logic program and let P be the corresponding
procedural program. Then we have lfp(TP ′ ) ∼ M(P ).
This relates the compositional semantics of procedural programs to the
mutually equivalent least fixpoint, proof-theoretical, and model-theoretical
semantics of logic programs.
13
6.1
An example
Consider the procedural program clause (f (y), z) :- {p(x, f (y)), p(f (x), z)}.
Here M(p) is an integer-indexed relation with {0, 1} as index set. Let us
0
a
f (a) f (a) f (b)
assume that M(p) =
. Here the four
1
f (b)
b
f (b) f (a)
2-tuples, indexed by {0, 1}, are displayed vertically.
The value of a call is a table; that is, a variable-indexed relation.
x
a f (a) f (b)
M(p(x, f (y))) = M(p) : hx, f (y)i =
.
y
b
b
a
a
a
b
x
Similarly, M(p(f (x), z)) = M(p) : hf (x), zi =
z
b f (b) f (a)
.
The value of the body is the product of the above two tables:
a
a
x
b
b
M(p(x, f (y)), p(f (x), z)) = M(p(x, f (y)))∗M(p(f (x), z)) = y
z
b f (b)
.
Finally, the meaning of the entire clause
(f (y), z) :- p(x, f (y)), p(f (x), z) is obtained by projection:
0
f (b) f (b)
hf (y), zi/M(p(x, f (y)), p(f (x), z)) =
.
1
b
f (b)
7
Implications for modularity
Suppose P and P ′ are procedural programs with the same Herbrand universe. If p(t0 , . . . , tn−1 ) is a call in P , then the meaning of p is (M(P ))(p).
But p is a special case of an expression that has an n-ary integer-indexed
relation as value. Such an expression could also be (M(P ′ ))(p′ ) if p′ is
a procedure symbol in P ′ paired with an n-ary procedure. The value of
this expression is a set of n-tuples of ground terms. This value is independent of the procedure symbols occurring in P ′ . Hence these symbols
are “encapsulated” in the expression (M(P ′ ))(p′ ).
This only addresses the semantics of a module mechanisms. It leaves
open the syntax that indicates which set of clauses is a module and which
procedure symbol is exported.
8
Related work
Modules for logic programs can be obtained via proof theory [18, 17] or via
higher-order logic [5]. A different approach is to base it on decompositions
of the immediate-consequence operator as done by Brogi et al. [3]. It is
baffling that the various approaches to modularity are so difficult to relate.
Several more are mentioned by Brogi et al. [3], who also seem at a loss in
relating them to their own work.
Additional details about the operations on tables and relations, there
called “table-relation algebra”, can be found in [10, 11].
14
9
Concluding remarks
The procedural programs of Definition 4 are the result of the desire to
give a procedural interpretation not only of an entire clause, but also of
the composition of head and body within a clause as well as of the compositions that can be recognized in the body. Thus procedural programs
are but another way of parsing a set of positive Horn clauses.
But suppose that in 1972 one had never heard of clausal logic and that
the motivation was to characterize in what way languages with procedures,
such as Algol, are of a higher level than their predecessors. A higher
level of programming in such languages is achieved by using procedure
calls as much as possible. That suggests the ultimate altitude in level of
programming: procedure bodies contain procedure calls only.
What about data structures for a pure procedural language? Just as
Lisp simplified by standardizing all data structures to lists, one could make
a similar choice by standardizing on trees. In this way a pure procedureoriented language would arise that coincides with the procedural programs
of Definition 4.
Functional programming languages have an obvious semantics in the
form of functions as defined in mathematics. The semantics of Algol-like
languages is defined in terms of transitions between computational states.
These transitions are specified directly or indirectly in terms of assignments. In this way one might think that procedure-oriented programming languages are of inherently lower level than functional programming
languages.
It is not necessary to specify procedures in terms of state transitions. A
procedure is more directly specified as the set of all possible combinations
of values of the arguments of a call. That is, as a set of tuples of the same
arity, which is a relation.
In this way the procedural programs of Definition 4 become as highlevel as functional programs and obtain a semantics that is as mathematical.
One might argue that this gives procedural programs a significance
that extends beyond logic programming. For example, they may be a
way to describe Colmerauer’s view [7] that Prolog is not necessarily a
logic programming language. In the procedural interpretation described
here, the Herbrand universe can be replaced by a sufficiently similar data
structure, such as the rational trees.
10
Acknowledgements
I am grateful to Belaid Moa and three anonymous reviewers for their suggestions for improvement. This research was supported by the University
of Victoria and by the Natural Science and Engineering Research Council
of Canada.
15
References
[1] K.R. Apt. Logic programming. In Handbook of Theoretical Computer
Science, 1990.
[2] George Boole. An Investigation of the Laws of Thought. Dover, 1854.
Dover edition not dated.
[3] Antonio Brogi, Paolo Mancarella, Dino Pedreschi, and Franco Turini.
Modular logic programming. ACM Transactions on Programming
Languages and Systems, 16:1361– 1398, 1994.
[4] François Bry. A compositional semantics for logic programs and deductive databases. In Proc. Joint Int. Conf. Symp. Logic Programming, 1996.
[5] W. Chen, M. Kifer, and D.S. Warren. HILOG: A foundation for
higher-order logic programming. Journal of Logic Programming,
15:187 – 230, 1993.
[6] Michael Codish, Saumya K. Debray, and Roberto Giacobazzi. Compositional analysis of modular logic programs. In Proc. 20th ACM
Symposium on Principles of Programming Languages (POPL), 1993.
[7] A. Colmerauer. Sur les bases théoriques de Prolog. Technical report,
Groupe d’Intelligence Artificielle, Université d’Aix-Marseille II, 1979.
[8] A. Colmerauer, H. Kanoui, R. Paséro, and P. Roussel. Un système
de communication homme-machine en français. Technical report,
Groupe d’Intelligence Artificielle, Université d’Aix-Marseille II, 1972.
[9] Leon Henkin, J. Donald Monk, and Alfred Tarski. Cylindric Algebras,
Parts I, II. Studies in Logic and the Foundations of Mathematics.
North-Holland, 1985.
[10] H. Ibrahim and M.H. van Emden. Towards applicative relational
programming. Unpublished draft; available from the Computing Research Repository (http://arxiv.org/corr/home), 1991.
[11] Husain Ibrahim. Applicative expressions for relational programming.
Master’s thesis, University of Victoria, 1992.
[12] Edgar T. Irons. A syntax-directed compiler for Algol 60. Communications of the ACM, 4:51–55, 1961.
[13] R.A. Kowalski. Predicate logic as programming language. In Proc.
IFIP 74, pages 569–574, 1974.
[14] R.A. Kowalski. Logic for Problem-Solving. Elsevier North-Holland,
1979.
[15] P. Landin. The mechanical evaluation of expressions. Computer
Journal, 6:308–320, 1963.
[16] J.W. Lloyd. Foundations of Logic Programming. Springer-Verlag,
2nd edition, 1987.
[17] Francis G. McCabe. Logic and Objects. Prentice Hall, 1992.
[18] Dale Miller. A logical analysis of modules in logic programming.
Journal of Logic Programming, 6:79 – 108, 1989.
16
[19] Dana Scott and Christopher Strachey. Toward a mathematical semantics for computer languages. In Proceedings of the Symposium
on Computers and Automata, pages 19–46, Polytechnic Institute of
Brooklyn, 1971.
[20] Joseph E. Stoy. Denotational Semantics: The Scott-Strachey approach to Programming Language Theory. MIT Press, 1977.
[21] A. Tarski. A representation theorem for cylindric algebras. Bull.
Amer. Math. Soc., 58:65 – 66, 1952.
[22] M.H. van Emden and R.A. Kowalski. The semantics of predicate
logic as a programming language. Journal of the ACM, 23(4):733–
742, 1976.
17
| 6cs.PL
|
Statistical Analysis of Synchrosqueezed Transforms
Haizhao Yang
arXiv:1410.5939v3 [math.ST] 8 Aug 2016
Department of Mathematics, Duke University
October 2014, Revised: August 2016
Abstract
Synchrosqueezed transforms are non-linear processes for a sharpened time-frequency
representation of wave-like components. They are efficient tools for identifying and analyzing wave-like components from their superposition. This paper is concerned with the
statistical properties of compactly supported synchrosqueezed transforms for wave-like
components embedded in a generalized Gaussian random process in multidimensional
spaces. Guided by the theoretical analysis of these properties, new numerical implementations are proposed to reduce the noise fluctuations of these transforms on noisy
data. A MATLAB package SynLab together with several heavily noisy examples is
provided to support these theoretical claims.
Keywords. Wave-like components, instantaneous (local) properties, synchrosqueezed
transforms, noise robustness, generalized Gaussian random process.
AMS subject classifications: 42A99 and 65T99.
1
Introduction
Non-linear and non-stationary wave-like signals (also termed as chirp signals in 1D) are
ubiquitous in science and engineering, e.g., clinical data [43, 44], seismic data [19, 22,
34], climate data [36, 45], astronomical data [7, 11], materials science [26, 33, 50], and
art investigation [49]. Analyzing instantaneous properties (e.g., instantaneous frequencies,
instantaneous amplitudes and instantaneous phases [6, 32]) or local properties (concepts for
2D signals similar to “instantaneous” in 1D) of signals has been an important topic for over
two decades. In many applications [15, 23, 41, 42, 51, 52], these signals can be modeled as a
superposition of several wave-like components with slowly varying amplitudes, frequencies
or wave vectors, contaminated by noise. For example, a complex signal
f (x) =
K
X
αk (x)e2πiNk φk (x) + e(x),
(1)
k=1
where αk (x) is the instantaneous (local) amplitude, Nk φk (x) is the instantaneous (local)
phase, Nk φ0k (x) is the instantaneous frequency (or Nk ∇φk (x) as the local wave vector), and
e(x) is a noisy perturbation term.
A powerful tool for analyzing signal (1) is the synchrosqueezed transform (SST) consisting of a linear time-frequency analysis tool and a synchrosqueezing technique. It belongs
to more generally time-frequency reassignment techniques [3, 8, 9, 16] (see also [4] for a
1
recent review). The SST was initialized in [16] and further analyzed in [15] for the 1D
wavelet transform. Suppose Wf (ξ, x) is the wavelet transform of a 1D wave-like component f (x) = α(x)e2πiN φ(x) . The wavelet time-frequency representation has a wide support
spreading around the instantaneous frequency curve (N φ0 (x), x). It was proved that the
∂x Wf (ξ,x)
instantaneous frequency information function vf (ξ, x) = 2πiW
is able to approximate
f (ξ,x)
0
N φ (x). Hence, the synchrosqueezing technique shifts the value of Wf (ξ, x) from (ξ, x) to
(vf (ξ, x), x), generating a sharpened time-frequency representation Tf (v, x) with a support
concentrating around the curve (N φ0 (x), x). The localization of the new representation not
only improves the resolution of the original spectral analysis due to the uncertainty principle
but also make it easier to decompose the superposition in (1) into individual components.
A variety of SSTs have been proposed after the 1D synchrosqueezed wavelet transform (SSWT) in [16], e.g., the synchrosqueezed short time Fourier transform (SSSTFT) in
[37], the synchrosqueezed wave packet transform (SSWPT) in [48, 51], the synchrosqueezed
curvelet transform (SSCT) in [52] and the 2D monogenic synchrosqueezed wavelet transform in [14]. Rigorous analysis has proved that these transforms can accurately decompose
a class of superpositions of wave-like components and estimate their instantaneous (local)
properties if the given signal is noiseless. To improve the synchrosqueezing operator in
the presence of strongly non-linear instantaneous frequencies, some further methods have
been proposed in [4, 25, 28] based on an extra investigation of the higher order derivatives of the phase of a wave-like component. All these SSTs are compactly supported in
the frequency domain to ensure accurate estimations. To better analyze signals with a
trend or big data sets that need to be handled in real-time computation, a recent paper
[13] proposes a new synchrosqueezing method based on carefully designed wavelets with
sufficiently many vanishing moments and a minimum support in the time domain. Previous synchrosqueezed transforms need to access every data sample of f (x) to compute the
synchrosqueezed transform at a specific time or location x. However, compactly supported
synchrosqueezed transforms only need a small portion of the data samples in a neighborhood of x. Hence, compactly supported synchrosqueezed transforms are computationally
more efficient and are better tools for modern big data analysis. However, mathematical analysis on the accuracy of this compactly supported SSWT is still under development.
This paper addresses this problem in the framework of the SSWPT that includes the SSWT
as a special case.
Another important topic in the study of SSTs is the statistical analysis of the synchrosqueezing operator, since noise is ubiquitous in real applications. A pioneer paper
[10] in this direction studied the statistical properties of the 1D spectrogram reassignment
method by calculating the probability density function of white Gaussian noise after reassignment. A recent paper [36] focused on the statistical analysis of the 1D SSWT for
white Gaussian noise. It estimated the probability of a good estimation of the instantaneous frequency provided by the instantaneous frequency information function vf (ξ, x). A
following paper [12] generalized its results to generalized stationary Gaussian process. To
support the application of SSTs to real-time problems and multidimensional problems, this
paper analyzes the statistical properties of multidimensional SSTs that can be compactly
supported in the time domain.
Turning to the robustness issue in a numerical sense, it is of interest to design an efficient
implementation of a sharpened time-frequency representation with reduced noise fluctuations. The idea of multitapering, first proposed in [38] for stationary signals and further
extended in [5, 20] for non-stationary signals, attempted to improve the statistical stability
of spectral analysis by generating multiple windowed trials of the noisy signal and averaging
2
the spectral analysis of these trials. By combining the multitapering and time-frequency
reassignment techniques, [46] proposed the 1D multitapering time-frequency reassignment
for a sharpened time-frequency representation with reduced noise fluctuations. Since its implementation is based on Hermite functions, its efficient generalization in multidimensional
spaces is not straightforward. Guided by the theoretical analysis of the statistical properties
of SSTs, this paper proposes efficient numerical implementations of multidimensional SSTs
based on highly redundant frames. Since the SST of a noiseless signal is frame-independent,
averaging the SSTs from multiple time-frequency frames reduces the noise fluctuation while
keeping the localization of the synchrosqueezed time-frequency representation.
The rest of this paper is organized as follows. In Section 2, the main theorems for the
compactly supported SSTs in multidimensional spaces and their statistical properties are
presented. In Section 3, a few algorithms and their implementations are introduced in detail
to improve the statistical stability of SSTs. Several numerical examples with heavy noise
are provided to demonstrate the proposed properties. We conclude this paper in Section 4.
2
Theory for synchrosqueezed transforms (SSTs)
Let us briefly introduce the basics and assumptions of compactly supported SSTs in Section
2.1 before discussing their statistical properties in Section 2.2. While the synchrosqueezing
technique can be applied to a wide range of time-frequency transforms, the discussion here
is restricted to the framework of multidimensional wave packet transforms to save space.
It is easy to extend these results to other transforms (see [47] for the example of the 2D
synchrosqueezed curvelet transform). Due to the space limitation, only the main ideas of
the proofs are presented. Readers are referred to [47] for more details.
2.1
Compactly supported SSTs
Previously, the synchrosqueezed wave packet transform (SSWPT) was built using mother
wave packets compactly supported in the frequency domain. This paper studies a wider
class of mother wave packets defined below.
Definition 2.1. An n-dimensional mother wave packet w(x) ∈ C m (Rn ) is of type (, m)
for some ≥ 0, and some non-negative integer m, if w(ξ)
b
is a real-valued smooth function
with a support that covers the ball B1 (0) centered at the origin with a radius 1 satisfying
that:
|w(ξ)|
b
≤
,
(1 + |ξ|)m
for |ξ| > 1 and |w(ξ)|
b
> 0, for |ξ| < 1.
Since w ∈ C m (Rn ), the above decaying requirement is easy to satisfy. Actually, we can
further assume w(ξ)
b
is essentially supported in a ball Bd (0) with d ∈ (0, 1] to adapt signals
with close instantaneous frequencies, i.e., |w(ξ)|
b
is approximately zero outside this support
up to an truncation error. However, d is just a constant in later asymptotic analysis.
Hence, we omit this discussion and consider it as 1 in the analysis but implement it in the
numerical tool. Similarly to the discussion in [48, 51], we can use this mother wave packet
w(x) to define a family of n-dimensional compactly supported wave packets
wab (x) = |a|ns/2 w (|a|s (x − b)) e2πi(x−b)·a ,
3
through scaling, modulation, and translation, controlled by a geometric parameter s, for a,
b ∈ Rn . With this family of wave packets ready, we define the wave packet transform via
Z
Wf (a, b) = hf, wab i = f (x)wab (x)dx.
Previously in [48, 51], the synchrosqueezed wave packet transform (SSWPT) was proposed to analyze a class of intrinsic mode type functions as defined below.
Definition 2.2. A function f (x) = α(x)e2πiN φ(x) in Rn is an intrinsic mode type function
(IMT) of type (M, N ) if α(x) and φ(x) satisfy
α(x) ∈ C ∞ ,
∞
φ(x) ∈ C ,
|∇α(x)| ≤ M,
1/M ≤ α(x) ≤ M,
1/M ≤ |∇φ(x)| ≤ M,
|∇2 φ(x)| ≤ M.
It has been proved that the instantaneous frequency (or local wave vector when n > 1)
∇b Wf (a,b)
information function vf (a, b) = 2πiW
of an IMT f (x) can approximate N ∇φ(b) if
f (a,b)
the mother wave packet w(x) is of type (0, ∞) and N is sufficiently large. A careful
inspection of previous proofs shows that the approximation vf (a, b) ≈ N ∇φ(b) is still
valid up to an relative error if the mother wave packet is of type (, m) for any positive
integer. See Theorem 2.2.7 in [47] for a detailed proof. Hence, if we squeeze the coefficients
Wf (a, b) together based upon the same information function vf (a, b), then we would obtain
a sharpened time-frequency representation of f (x). This motivates the definition of the
synchrosqueezed energy distribution
Z
Tf (v, b) =
|Wf (a, b)|2 δ (<vf (a, b) − v) da
Rn
for v, b ∈ Rn . Here δ denotes the Dirac delta function and <vf (a, b) means the real part of
vf (a, b).
P
2πiNk φk (x) , the synchrosqueezed enFor a multi-component signal f (x) = K
k=1 αk (x)e
ergy of each component will also concentrate around each Nk ∇φk (x) if these components
satisfy the well-separation condition defined below.
Definition 2.3. A function f (x) is a well-separated superposition of type (M, N, K, s) if
f (x) =
K
X
fk (x),
k=1
where each fk (x) = αk (x)e2πiNk φk (x) is an IMT of type (M, Nk ) with Nk ≥ N and the phase
functions satisfy the separation condition: for any (a, b) ∈ R2n , there exists at most one fk
satisfying that
|a|−s |a − Nk ∇φk (b)| ≤ 1.
We denote by F (M, N, K, s) the set of all such functions.
In real applications, this well-separation condition might not be valid for a multicomponent signal at every x. However, the SST will work wherever the well-separation
condition is satisfied locally.
The key analysis of the SSWPT is how well the information function vf (a, b) approximates the instantaneous frequencies or local wave vectors. If the approximation is accurate
4
enough, the synchrosqueezed energy distribution Tf (a, b) gives a sharpened time-frequency
representation of f (x). We close this section with the following theorem that summarizes
the main analysis of the n-dimensional SSWPT for a superposition of IMTs without noise
or perturbation. In what follows, when we write O (·), ., or &, the implicit constants may
depend on M , m and K. Readers are referred to Theorem 2.2.7 in [47] for the proof of
Theorem 2.4 here.
Theorem 2.4. Suppose the n-dimensional mother wave packet is of type (, m), for any
fixed ∈ (0, 1) and any fixed integer m ≥ 0. For a function f (x), we define
√
R = {(a, b) : |Wf (a, b)| ≥ |a|−ns/2 },
√
S = {(a, b) : |Wf (a, b)| ≥ },
and
Zk = {(a, b) : |a − Nk ∇φk (b)| ≤ |a|s }
for 1n≤ k ≤ K. o
For fixed M , m, K, s, and , there exists a constant N0 (M, m, K, s, ) '
−2
−1
2s−1
1−s
max
such that for any N > N0 and f (x) ∈ F (M, N, K, s) the following
,
statements hold.
S
(i) {Zk : 1 ≤ k ≤ K} are disjoint and S ⊂ R ⊂ 1≤k≤K Zk ;
(ii) For any (a, b) ∈ R ∩ Zk ,
|vf (a, b) − Nk ∇φk (b)| √
. ;
|Nk ∇φk (b)|
(iii) For any (a, b) ∈ S ∩ Zk ,
|vf (a, b) − Nk ∇φk (b)|
−ns/2 √
.
. Nk
|Nk ∇φk (b)|
2.2
Statistical Properties of SSTs
Similarly to the noiseless case, we will analyze how well the information function vf (a, b)
approximates instantaneous frequencies or local wave vectors in the case when a superposition of IMTs is contaminated by random noise. To simplify the discussion, we will
sketch out the proofs in the one-dimensional case and refer the readers to [47] for higher
dimensional cases.
Let us start with a simple case in which the superposition is perturbed slightly by a
contaminant, Theorem 2.5 below shows that the information function vf (a, b) can approximate instantaneous frequencies with a reasonable error determined by the magnitude of
the perturbation.
Theorem 2.5. Suppose the mother wave packet is of type (, m), for any fixed ∈ (0, 1)
and any fixed integer m ≥ 0. Suppose g(x) = f (x) + e(x), where e(x) ∈ L∞ is a small error
1
√
√
−p
term that satisfies kekL∞ ≤ 1 for some 1 > 0. For any p ∈ 0, 12 , let δ = + 12 .
Define
Rδ = {(a, b) : |Wg (a, b)| ≥ |a|−s/2 δ},
Sδ = {(a, b) : |Wg (a, b)| ≥ δ},
5
and
Zk = {(a, b) : |a − Nk φ0k (b)| ≤ |a|s }
for 1n≤ k ≤ K. o
For fixed M , m, K, s, and , there exists a constant N0 (M, m, K, s, ) '
−1
−1
max 2s−1 , 1−s such that for any N > N0 (M, m, K, s, ) and f (x) ∈ F (M, N, K, s) the
following statements hold.
S
(i) {Zk : 1 ≤ k ≤ K} are disjoint and Sδ ⊂ Rδ ⊂ 1≤k≤K Zk ;
(ii) For any (a, b) ∈ Rδ ∩ Zk ,
|vg (a, b) − Nk φ0k (b)| √
. + p1 ;
|Nk φ0k (b)|
(iii) For any (a, b) ∈ Sδ ∩ Zk ,
√
|vg (a, b) − Nk φ0k (b)|
+ p1
.
.
s/2
|Nk φ0k (b)|
Nk
We introduce the parameter p to clarify the relation among the perturbation, the threshold and the accuracy for better understanding the influence of perturbation or noise. For
the same purpose, a parameter
p 1 q will be introduced in the coming theorems. If the threshold δ is larger, e.g., δ ≥
, the relative estimate errors in (ii) and (iii) are bounded
√
√
by and s/2 , respectively. Similarly, one can show that the information function comNk
puted from the wave packet coefficient with a larger magnitude can better approximate the
instantaneous frequency.
Below is a sketch of the proof of Theorem 2.5. See the proof of Theorem 3.2.1 in [47]
for a detailed proof.
Proof. We only need to discuss the case when a > 0. By the definition of the wave packet
transform, we have
√
√
|We (a, b)| . 1 a−s/2 and |∂b We (a, b)| . 1 a1−s/2 + as/2 .
(2)
If (a, b) ∈ Rδ , then |Wg (a, b)| ≥ a−s/2 δ. Together with Equation (2), it holds that
√
√
(3)
|Wf (a, b)| ≥ |Wg (a, b)| − |We (a, b)| ≥ a−s/2 (δ − 1 ) ≥ a−s/2 .
S
Hence, Sδ ⊂ Rδ ⊂ R , where R is defined in Theorem 2.4 and is a subset of 1≤k≤K Zk .
So, (i) is true by Theorem 2.4.
Since Rδ ⊂ R , (a, b) ∈ Rδ ∩ Zk implies (a, b) ∈ R ∩ Zk . Hence, by Theorem 2.4, it
holds that
|vf (a, b) − Nk φ0k (b)| √
. ,
(4)
|Nk φ0k (b)|
n −1
o
−1
when N is larger than a constant N0 (M, m, K, s, ) ' max 2s−1 , 1−s . Notice that
(a, b) ∈ Zk implies a ' Nk . Hence, by Equation (2) to (4),
∂ W (a,b)
∂ W (a,b)
b f
b g
|vf (a, b) − Nk φ0k (b)| | 2πiWf (a,b) − 2πiWg (a,b) | √
|vg (a, b) − Nk φ0k (b)|
≤
+
≤ + p1 ,
|Nk φ0k (b)|
|Nk φ0k (b)|
|Nk φ0k (b)|
when N > N0 . Hence, (ii) is proved. The proof of (iii) is similar.
6
Next, we will discuss the case when the contamination e is a random perturbation.
[21, 24, 27, 35, 39] are referred to for basic facts about generalized random fields and
complex Gaussian processes. To warm up, we start with additive white Gaussian process
in Theorem 2.6 and extend it to a generalized zero mean stationary Gaussian process in
Theorem 2.7. We assume that e has an explicit power spectral function denoted by eb(ξ).
k · k represents the L2 norm and h·, ·i is the standard inner product.
Theorem 2.6. Suppose the mother wave packet is of type (, m), for any fixed ∈ (0, 1)
2
and any fixed integer m ≥ 1−s
+ 4. Suppose g(x) = f (x) + e, where e is zero mean white
Gaussian process with a variance 1+q
for some q > 0 and some 1 > 0. For any p ∈ 0, 21 ,
1
1
√
−p
let δ = + 12 . Define
Rδ = {(a, b) : |Wg (a, b)| ≥ a−s/2 δ},
Sδ = {(a, b) : |Wg (a, b)| ≥ δ},
and
Zk = {(a, b) : |a − Nk φ0k (b)| ≤ as }
for 1n≤ k ≤ K. o
For fixed M , m, K, s, and , there exists a constant N0 (M, m, K, s, ) '
−1
−1
2s−1
1−s
,
such that for any N > N0 (M, m, K, s, ) and f (x) ∈ F (M, N, K, s) the
max
following statements hold.
(i) {Zk : 1 ≤ k ≤ K} are disjoint.
S
(ii) If (a, b) ∈ Rδ , then (a, b) ∈ 1≤k≤K Zk with a probability at least
!
−s −q
1 − e−O(Nk 1 ) + O
.
m(1−s)
Nk
S
(iii) If (a, b) ∈ Sδ , then (a, b) ∈ 1≤k≤K Zk with a probability at least
!
−q
−2
1 − e−1 kwk + O
.
m(1−s)
Nk
(iv) If (a, b) ∈ Rδ ∩ Zk for some k, then
|vg (a, b) − Nk φ0k (b)| √
. + p1
|Nk φ0k (b)|
is true with a probability at least
1 − e−O(
Nk2−3s −q
1
!
) 1 − e−O(Nk−s−2 −q
1 )
+O
.
(m−4)(1−s)−2
Nk
(v) If (a, b) ∈ Sδ ∩ Zk , then
√
|vg (a, b) − Nk φ0k (b)|
+ p1
.
s/2
|Nk φ0k (b)|
Nk
is true with a probability at least
1 − e−O(
Nk2−2s −q
1
) 1 − e−O(Nk−2 −q
1 )
+O
(m−4)(1−s)−2
Nk
7
!
.
We only sketch the proof of the above theorem. See the proof of Theorem 3.2.2 in [47]
for a long proof.
Proof. Step 1: we prove this theorem when the mother wave packet is of type (0, m) first,
i.e., compactly supported in the frequency domain.
Since wab and ∂b wab are in L1 ∩ C m−1 , We (a, b) and ∂b We (a, b) are Gaussian variables.
Hence, Wg (a, b) = Wf (a, b) + We (a, b) and ∂b Wg (a, b) = ∂b Wf (a, b) + ∂b We (a, b) can be
understood as Gaussian variables. Furthermore, We (a, b) and (We (a, b), ∂b We (a, b)) are
circularly symmetric Gaussian variables by checking that their pseudo-covariance matrices
are zero. Therefore, the distribution of We (a, b) is determined by its variance
−(1+q)
e−1
|z1 |2 kwk−2
2
π1+q
1 kwk
If we define
.
kwk
b 2
hd
wab , 2πiξd
wab i
,
h2πiξd
wab , w
d
wab , 2πiξd
wab i
ab i h2πiξd
V =
then 1+q
1 V is the covariance matrix of (We (a, b), ∂b We (a, b)) and its distribution is described
by the joint probability density
−(1+q) ∗ −1
z V z
e−1
,
2(1+q)
π 2 1
det V
where z = (z1 , z2 )T , T and ∗ denote the transpose operator and conjugate transpose operator. V is an invertible and self-adjoint matrix, since We (a, b) and ∂b We (a, b) are linearly
independent. Hence, there exist a diagonal matrix D and a unitary matrix U such that
V −1 = U ∗ DU .
Part (i) is true by previous theorems. Define the following events
n
√ o
G1 = |We (a, b)| < a−s/2 1 ,
√
G2 = {|We (a, b)| < 1 } ,
n
o
√
G3 = |∂b We (a, b)| < 1 a1−s/2 + as/2 ,
|vg (a, b) − Nk φ0k (b)| √
p
Hk =
.
+
1 ,
|Nk φ0k (b)|
and
(
Jk =
)
√
|vg (a, b) − Nk φ0k (b)|
+ p1
.
,
s/2
|Nk φ0k (b)|
Nk
for 1 ≤ k ≤ K. To conclude Part (ii) to (v), we need to estimate the probability P (G1 ),
P (G2 ), P (G1 ∩ G3 ), P (G2 ∩ G3 ), P (Hk ) and P (Jk ). Algebraic calculations show that
−s −q kwk−2
1
P (G1 ) = 1 − e−a
and
−q
P (G2 ) = 1 − e−1
kwk−2
.
We are ready to summarize and conclude (ii) and (iii). If (a, b) ∈ Rδ , then
√
1/2−p
|We (a, b) + Wf (a, b)| ≥ a−s/2 1
+ .
8
(5)
If (a, b) ∈
/
S
1≤k≤K
Zk , then
|Wf (a, b)| ≤ a−s/2 .
√
Equation (5) and (6) lead to |We (a, b)| ≥ a−s/2 1 . Hence,
[
√
P (a, b) ∈
/
Zk ≤ P |We (a, b)| ≥ a−s/2 1 = 1 − P (G1 ) .
(6)
1≤k≤K
S
This means that if (a, b) ∈ Rδ , then (a, b) ∈ 1≤k≤K Zk with a probability at least P (G1 ) =
−s −q
−2
−s −q
1 − e−a 1 kwk = 1 − e−O(Nk 1 ) , since a ' Nk ifS(a, b) ∈ Zk . So, (ii) is true. A similar
argument applied to (a, b) ∈ Sδ shows that (a, b) ∈ 1≤k≤K Zk with a probability at least
−q
−2
P (G2 ) = 1 − e−1 kwk . Hence, (iii) is proved.
Recall that V −1 = U ∗ DU . By the change of variables z 0 = U z, we can show that
2
2
−
P (G1 ∩ G3 ) ≥ 1 − e
D11 d1
1+q
1
−
1 − e
D22 d1
1+q
1
,
and
P (G2 ∩ G3 ) ≥ 1 − e
−
D11 d2
2
1+q
1
1 − e
−
D22 d2
2
1+q
1
.
We can further estimate that D11 ' a2(1−s) and D22 ' a−2 . Therefore,
2−3s −q
−s−2 −q
1 )
1 ) ,
P (G1 ∩ G3 ) ≥ 1 − e−O(a
1 − e−O(a
and
−2 −q
2−2s −q
1 )
1 − e−O(a 1 ) .
P (G2 ∩ G3 ) ≥ 1 − e−O(a
By Theorem 2.5, if (a, b) ∈ Rδ ∩ Zk for some k, then
−s−2 −q
2−3s −q
1 ) .
1 )
1 − e−O(a
P (Hk ) ≥ P (Hk |G1 ∩ G3 ) P (G1 ∩ G3 ) = P (G1 ∩ G3 ) ≥ 1 − e−O(a
Note that a ' Nk when a ∈ Zk , then
2−3s −q
−s−2 −q
P (Hk ) ≥ 1 − e−O(Nk 1 ) 1 − e−O(Nk 1 ) .
Similarly, if (a, b) ∈ Sδ ∩ Zk for some k, then
2−2s −q
−2 −q
P (Jk ) ≥ P (Jk |G2 ∩ G3 ) P (G2 ∩ G3 ) = P (G2 ∩ G3 ) ≥ 1 − e−O(Nk 1 ) 1 − e−O(Nk 1 ) .
These arguments prove (iv) and (v).
Step 2: we go on to prove this theorem when the mother wave packet is of type (, m)
2
with m ≥ 1−s
+ 4. We would like to emphasize that the requirement is crucial to the
following asymptotic analysis and it keeps the error caused by the non-compact support of
w
b reasonably small.
The sketch of the proof is similar to the first step. We (a, b) and (We (a, b), ∂b We (a, b))
are still Gaussian variables but in general not circularly symmetric, because they would
9
not have zero pseudo-covariance matrices. Suppose they have covariance matrices C1 and
C2 , pseudo-covariance matrices P1 and P2 , respectively. We can still check that they have
1+q
2
zero mean, C1 = 1+q
1 kwk and C2 = 1 V , where V is defined in the first step. By the
definition of the mother
wavepacket of type (, m), the magnitude of every entry in P1 and
P2 is bounded by O
1+q
1
m(1−s)
a
. Notice that the covariance matrix of (We (a, b), We∗ (a, b)) is
V1 =
C1 P1
.
P1∗ C1∗
By Equation (27) in [27] and the Taylor expansion, the distribution of We (a, b) is described
by the following distribution
!!
−(1+q)
−1
1 ∗
∗ T
|z1 |2 kwk−2
e−1
e− 2 (z1 ,z1 )V1 (z1 ,z1 )
|z1 |2
√
=
.
1+O
m(1−s)
2
π det V1
π1+q
1+q
1 kwk
1 a
By the same argument, the covariance matrix of (We (a, b), ∂b We (a, b), We∗ (a, b), ∂b We∗ (a, b))
is
C2 P 2
V2 =
.
P2∗ C2∗
Let z = (z1 , z2 )T . Then the distribution of (We (a, b), ∂b We (a, b)) is described by the joint
probability density
−1
1 ∗ ∗
∗ ∗ T
e− 2 (z1 ,z2 ,z1 ,z2 )V2 (z1 ,z2 ,z1 ,z2 )
√
.
(7)
π 2 det V2
2
2(s−1) determined by
Notice that C2 = 1+p
1 V and V has eigenvalues of order a and a
estimating the diagonal entries of the matrix D in the diagonalization V −1 = U ∗ DU .
1+p 2(s−1)
2
. Recall that the magnitude of
Hence, C2 has eigenvalues of order1+p
1 a and
1 a
every entry in P2 is bounded by O
1+q
1
am(1−s)
. This means that V2 is nearly dominated by
diagonal blocks C2 and C2∗ . Basic spectral theory for linear transforms shows that
−1
C2
−1
V2 =
+ P ,
(C2∗ )−1
−(1+q) (m−4)(s−1)
where P is a matrix with 2-norm bounded by O 1
a
. m−6
m−4 ≥ s is crucial
1+q
1
to the above spectral analysis. Since every entry of P2 is bounded by O am(1−s) ,
4(1+q)
2
det V2 = (det C2 ) + O
1
am−2−(m+2)s
!
,
where the residual comes from the entry bound and the eigenvalues of C2 . Hence (7) is
actually
−(1+q) ∗ −1
z V z e− 21 (z1∗ ,z2∗ ,z1 ,z2 )P (z1 ,z2 ,z1∗ ,z2∗ )T
e−1
r
.
2(1+q)
2
2
π 1
(det V ) + O am−2−(m+2)s)
By the same argument as in the first step, we can show that there exist a diagonal matrix
D = diag{a2(1−s) , a−2 } and a unitary matrix U such that V −1 = U ∗ DU . Part (i) is
10
still true by previous theorems. To conclude Part (ii) to (v), we still need to estimate
the probability of those events defined in the first step, i.e., P (G1 ), P (G2 ), P (G1 ∩ G3 ),
P (G2 ∩ G3 ), P (Hk ) and P (Jk ). By the estimations above, one can show that
−s −q
−2
P (G1 ) = 1 − e−a 1 kwk + O m(1−s) ,
a
and
−q
P (G2 ) = 1 − e−1
kwk−2
+O
am(1−s)
.
Hence, we can conclude (ii) and (iii) follows the same proof in the first step. Next, we
look at the last two part of this theorem.
√
√
√
Let us introduce notations δ1 = a−s/2 1 , δ2 = 1 , δ3 = a1−s/2 + as/2
1 , d1 =
min{ √δ12 , √δ32 }, and d2 = min{ √δ22 , √δ32 }. By previous estimations, we have
P (G1 ∩ G3 )
Z
=
{|z1 |<δ1 ,|z2 |<δ3 }
∗
−1
∗
−(1+q) ∗ −1
z V z
=
{|z1 |<δ1 ,|z2 |<δ3 }
∗
∗ T
1
∗
∗
∗
∗ T
e− 2 (z1 ,z2 ,z1 ,z2 )P (z1 ,z2 ,z1 ,z2 )
r
dz1 dz2 .
2(1+q)
π 2 1
(det V )2 + O am−2−(m+2)s)
e−1
Z
Since
1
e− 2 (z1 ,z2 ,z1 ,z2 )V2 (z1 ,z2 ,z1 ,z2 )
√
dz1 dz2
π 2 det V2
det V
r
= 1 + O a(m−2)(1−s)) ,
(det V )2 + O am−2−(m+2)s)
(8)
(9)
we can drop out the term O am−2−(m+2)s)
in (8), which would generate an absolute error
no more than O a(m−2)(1−s))
in the estimate of P (G1 ∩ G3 ). Let
1
g(z) = − (z1∗ , z2∗ , z1 , z2 )P (z1 , z2 , z1∗ , z2∗ )T ,
2
11
then by the change of variables we have
P (G1 ∩ G3 )
Z
−(1+q) ∗ −1
z V z
e−1
≈
2(1+q)
π 2 1
{|z1 |<δ1 ,|z2 |<δ3 }
−(1+q)
e−1
Z
=
−(1+q)
=
=
Z
2(1+q)
π 2 1
det V
2(1+q)
π 2 1
det V
D11 d2
1
− 1+q
1
+ 1 − e
d1
Z
(D11 |z10 |2 +D22 |z20 |2 ) eg(U ∗ z 0 )
d1
2π
Z
2π
Z
−
r1 r2 e
0
Z
1
dz10 dz20
2(1+q)
π 2 1
det V
{|z10 |<d1 ,|z20 |<d1 }
1
dz1 dz2
2(1+q)
π 2 1
det V
e−1
≥
det V
(D11 |z10 |2 +D22 |z20 |2 ) eg(U ∗ z 0 )
{|z1 |<δ1 ,|z2 |<δ3 }
Z
eg(z)
0
d1
Z
0
d1
2π
Z
−
r1 r2 e
0
0
0
−
2
D22 r2
1+q
1
−
2
D22 r2
1+q
1
e
eg̃(r1 ,θ1 ,r2 ,θ2 ) dθ1 dθ2 dr1 dr2
0
2π
Z
2
D11 r1
1+q
1
dz10 dz20
2
D11 r1
1+q
1
e
eg̃(r1 ,θ1 ,r2 ,θ2 ) − 1 dθ1 dθ2 dr1 dr2
0
D22 d2
1
− 1+q
1
1 − e
,
(10)
g(U ∗ z 0 ).
−(1+q) (m−4)(s−1)
1
a
where g̃(r1 , θ1 , r2 , θ2 ) =
Recall that the 2-norm of P is bounded by O
Hence,
−(1+q) (m−4)(s−1)
−(1+q) (m−4)(s−1)
|g̃(r1 , θ1 , r2 , θ2 )| ≤ O 1
a
|z1 |2 + |z2 |2 = O 1
a
r12 + r22 .
Therefore, the first term in (10) is bounded by
O a(m−4)(s−1)
Z
≤ O
Z
d1
2
D11 r1
1+q
1
−
r1 r2 e
3(1+q)
1
det V
d1
0
Z
D22 a(m−4)(1−s)
= O (m−4)(1−s)−2 .
a
0
∞Z ∞
0
−
e
2
D22 r2
1+q
1
r12 + r22 dr1 dr2
2
2
r1 r2 r12 + r22 e−r1 e−r2 dr1 dr2
0
(11)
The analysis in (9) and (11) implies that
2
−
P (G1 ∪ G3 ) ≥ 1 − e
D11 d1
1+q
1
−
1 − e
D22 d2
1
1+q
1
+O
a(m−4)(1−s)−2
.
and similarly
−
P (G2 ∩ G3 ) ≥ 1 − e
D11 d2
2
1+q
1
−
1 − e
D22 d2
2
1+q
1
+O
a(m−4)(1−s)−2
.
The rest of the proof is exactly the same as the one in the first step and consequently we
know this theorem is also true for a mother wave packets of type (, m) with m satisfying
2
m ≥ 1−s
+ 4.
12
.
Thus far, we have considered the analysis for small perturbation and white Gaussian
process. Next, Theorem 2.6 is extended to a broader class of colored random processes.
Theorem 2.7. Suppose the mother wave packet is of type (, m), for any fixed ∈ (0, 1)
2
and any fixed integer m ≥ 1−s
+ 4. Suppose g(x) = f (x) + e, where e is a zero mean
stationary Gaussian process. Let eb(ξ) denote the spectrum of e, maxξ |b
e(ξ)| ≤ −1 and
1
√
( −p)/(1+q)
Ma = max|ξ|<1 eb(as ξ + a). For any p ∈ (0, 21 ] and q > 0, let δa = Ma 2
+ ,
Rδa = {(a, b) : |Wg (a, b)| ≥ a−s/2 δa },
Sδa = {(a, b) : |Wg (a, b)| ≥ δa },
and
Zk = {(a, b) : |a − Nk φ0k (b)| ≤ as }
for 1n≤ k ≤ K. o
For fixed M , m, K, s, and , there exists a constant N0 (M, m, K, s, ) '
−1
−1
2s−1
1−s
such that for any N > N0 (M, m, K, s, ) and f (x) ∈ F (M, N, K, s) the
max
,
following statements hold.
(i) {Zk : 1 ≤ k ≤ K} are disjoint.
S
(ii) If (a, b) ∈ Rδa , then (a, b) ∈ 1≤k≤K Zk with a probability at least
!
−q/(1+q)
−O Nk−s Ma
.
1−e
+O
m(1−s)
Nk
(iii) If (a, b) ∈ Sδa , then (a, b) ∈
S
Zk with a probability at least
!
−q/(1+q)
−O(Ma
)
1−e
+O
.
m(1−s)
Nk
1≤k≤K
(iv) If (a, b) ∈ Rδa ∩ Zk for some k, then
|vg (a, b) − Nk φ0k (b)| √
. + Map/(1+q)
|Nk φ0k (b)|
is true with a probability at least
−q/(1+q)
−q/(1+q)
−O Nk2−3s Ma
−O Nk−s−2 Ma
1−e
1−e
+O
!
.
(m−4)(1−s)−2
Nk
(v) If (a, b) ∈ Sδa ∩ Zk , then
|vg (a, b) − Nk φ0k (b)|
−s/2 √
p/(1+q)
.
N
+
M
a
k
|Nk φ0k (b)|
is true with a probability at least
−q/(1+q)
−q/(1+q)
−O Nk2−2s Ma
−O Nk−2 Ma
1−e
1−e
+O
(m−4)(1−s)−2
Nk
13
!
.
Proof. The proof of this theorem is nearly identical to Theorem 2.6. Although calculations
are more cumbersome, the derivation proceeds along the same line (see the proof of Theorem
3.2.3 in [47] for details).
Theorem 2.6 and 2.7 provide a new insight that a smaller s yields a synchrosqueezed
transform that can provide a good estimation with higher probability. The parameter m in
2
the mother wave packet is also important, e.g., satisfying m ≥ 1−s
+ 4. In a special case,
if a compactly supported synchrosqueezed wavelet transform (corresponding to s = 1) is
2
preferable, then we require that m ≥ 1−s
+ 4 = ∞. Hence, the mother wavelet is better to
∞
be C .
We have not optimized the requirement of the variance of the Gaussian process e and
the probability bound in Theorem 2.6 and Theorem 2.7. According to the numerical performance of the SSTs, the requirement of the variance could be weakened and the probability
estimation could be improved. A key step is to improve the estimate of P (G1 ∩ G3 ) and
P (G2 ∩ G3 ) in the above proofs. This is left as future work.
The above statistical property of the 1D SSWPT can be extended to higher dimensional
cases, e.g. the 2D SSWPT and SSCT. However, the notations are much heavier and the
calculations are more tedious. We close this section with the theorem for the 2D SSWPT.
See the proof of Theorem 3.4.4 and a similar theorem for the SSCT in [47].
Theorem 2.8. Suppose the 2Dnmother wave packet
is of type (, m), for any fixed ∈ (0, 1)
o
2(1+s)
2
and any fixed integer m ≥ max 1−s , 1−s + 4 . Suppose g(x) = f (x)+e, where e is a zero
mean stationary Gaussian process with a spectrum denoted by eb(ξ) and maxξ |b
e(ξ)| ≤ −1 .
1
( −p)/(1+q) √
Define Ma = max|ξ|<1 eb(|a|s ξ +a). For any p ∈ (0, 12 ] and q > 0, let δa = Ma 2
+ ,
Rδa = {(a, b) : |Wg (a, b)| ≥ |a|−s δa },
Sδa = {(a, b) : |Wg (a, b)| ≥ δa },
and
Zk = {(a, b) : |a − Nk ∇b φk (b)| ≤ |a|s }
for 1n≤ k ≤ K. oFor fixed M , m, s, and K, there exists a constant N0 (M, m, K, s, ) '
−2
−1
max 2s−1 , 1−s such that for any N > N0 and f (x) ∈ F (M, N, K, s) the following
statements hold.
(i) {Zk : 1 ≤ k ≤ K} are disjoint.
S
(ii) If (a, b) ∈ Rδa , then (a, b) ∈ 1≤k≤K Zk with a probability at least
1−e
−q/(1+q)
−O Nk−2s Ma
!
+O
m(1−s)
.
Nk
(iii) If (a, b) ∈ Sδa , then (a, b) ∈
S
1−e
1≤k≤K
Zk with a probability at least
−q/(1+q)
−O(Ma
)
+O
m(1−s)
Nk
14
!
.
(iv) If (a, b) ∈ Rδa ∩ Zk for some k, then
|vg (a, b) − Nk ∇φk (b)| √
. + Map/(1+q)
|Nk ∇φk (b)|
is true with a probability at least
−q/(1+q)
−q/(1+q)
−q/(1+q)
−O Nk−2−2s Ma
−O Nk−4s Ma
−O Nk2−4s Ma
1−e
1−e
1−e
+O
+O
.
(m−4)(1−s)−2
m−2−(m+2)s
Nk
Nk
(v) If (a, b) ∈ Sδa ∩ Zk for some k, then
√
|vg (a, b) − Nk ∇φk (b)|
+ Map/(1+q)
. Nk−s
|Nk ∇φk (b)|
is true with a probability at least
−q/(1+q)
−q/(1+q)
−q/(1+q)
−O Nk−2s Ma
−O Nk−2 Ma
−O Nk2−2s Ma
1−e
1−e
1−e
+O
+
O
.
(m−4)(1−s)−2
m−2−(m+2)s
Nk
3
Nk
Implementation and numerical results
In this section, we provide numerical examples to demonstrate some statistical properties
discussed in Section 2. Guided by these properties, we explore several new ideas to improve
the statistical stability of discrete SSTs in the presence of heavy noise. We have developed
SynLab, a collection of MATLAB implementation for various SSTs that has been publicly available at: https://github.com/HaizhaoYang/SynLab. Most numerical examples
presented in this paper can be found in this toolbox.
3.1
Implementation for better statistical stability
From the discussion in Section 2, we can summarize several observations for designing better
implementation of SSTs to reduce the fluctuation of noise.
1. Smaller geometric parameter s
As we can see in the theorems for noisy data, a smaller geometric parameter s results
in a higher probability of a good estimation of the instantaneous frequency or the
local wave vector via the information function vf (a, b). Hence, in the case when noise
is heavy, it is better to apply an SST with a smaller geometric parameter s.
2. Highly redundant frames
Previously in the synchrosqueezed wavelet transform (SSWT) [15], the synchrosqueezed
short-time Fourier transform (SSSTFT) [37], and the synchrosqueezed wave packet
transform (SSWPT) [48], various SSTs were generated from time-frequency transforms with low redundancy to obtain efficient forward and inverse transforms. However, the resultant transforms are not reliable when noise is heavy as we can see in the
example in Figure 1. In this example, one single IMT f (x) = e60πi(x+0.05 cos(2πx)) is
15
120
100
100
80
80
freq
freq
120
60
60
40
40
20
20
0
0
0.2
0.4
0.6
0.8
1
0.2
0.4
120
100
100
80
80
freq
freq
120
60
40
20
20
0.2
0.4
0.8
1
0.6
0.8
1
60
40
0
0.6
time
time
0.6
0.8
1
time
0
0
0.2
0.4
time
Figure 1: Top-left: The real instantaneous frequency. Top-right: The standard SSWT
available in [1]. Bottom-left: The standard SSSTFT available in [1]. Bottom-right: The
SSWPT with only one frame, i.e., the redundancy parameter red = 1 in SynLab, which
only one wave packet frame is used to compute the synchrosqueezed transform. Parameters
in these transforms have been tuned for reasonably good visualization.
embedded in white Gaussian noise with a distribution 2N (0, 1). The signal is sampled
in a time interval [0, 1] with a sampling rate 1024 Hz. We apply the standard SSWT,
SSSTFT in [1] and the SSWPT with s = 0.75 in SynLab and visualize their results
using the same discrete grid in the time-frequency domain. Although we could identify a rough curve in these results to estimate the instantaneous frequency of f (x),
the accuracy is poor in some areas and there are many misleading curves coming from
the noise.
As we can see in the theory for SSTs, the accuracy of the estimation provided by
the information function vf (a, b) is essentially independent of a and the mother wave
packet w(x). This motivates us to synchrosqueeze a highly redundant time-frequency
transform with over-complete samples in a and different mother wave packets w(x).
This generates many samples of estimations from a single realization of the noisy data.
Averaging these estimation samples leads to a better result. It can be understood
in the point of view that the contribution of IMTs to the synchrosqueezed energy
distribution Tf (a, b) will remain the same due to the coherent averaging, but the
contribution of the noise will be smoothed out because of the incoherent averaging.
Applying different mother wave packets is essentially the same as applying multitapers
in the multitaper time-frequency reassignment in [46]. In this paper, we only focus on
16
oversampling the variable a and leave the design of different multidimensional mother
wave packets as future work. In the numerical examples later, for a fixed geometric
parameter s, we synchrosqueeze a union of wave packet frames generated by timefrequency shifting, dilation (and rotation in multidimensional spaces). The number
of frames is denoted as red in SynLab.
3. Selective time-frequency coefficients
As we can see in the proofs of previous theorems, a larger time-frequency coefficient
results in a higher probability for a good estimation provided by the information
function vf (a, b). This inspires two ideas: 1) applying an adaptive time-frequency
transform before synchrosqueezing; 2) only reassigning the largest coefficient in the
domain where there is at most one IMT. The first idea aims at generating large
coefficients while the second idea avoids incorrect reassignment as much as possible.
Selecting the best coefficients to reassign, in some sense, is similar to the idea of
sparse matching pursuit in a highly redundant frame, but we avoid its expensive
optimization.
3.2
Numerical examples
Here we provide numerical examples to demonstrate the efficiency of the proposed implementation in Section 3.1. In all examples, we assume the given data g(x) = f (x) + e(x)
is defined in [0, 1]n , where f (x) is the target signal, e(x) is white Gaussian noise with a
distribution σ 2 N (0, 1), and n is the dimension.
As we have seen in the theorems in Section 2, a proper threshold adaptive to the noise
level after the wave packet transform is important to obtain an accurate instantaneous
frequency/local wave vector estimate. We refer to [17, 18] for estimating noise level and
[36] for designing thresholds for the SSWT. The generalization of these techniques for
the transforms here is straightforward. In this paper, we use a small uniform threshold
δ = 10−2 (rather than a threshold adaptive to noise level) and set σ 2 such that the noise is
overwhelming the original signal. We refer the reader to [51, 52] for detailed implementation
using these parameters.
3.2.1
Visual illustrations for statistical properties
To support the theoretical analysis in Section 2 and the proposals in Section3.1, we compare
the performance of the SSWPT with different redundancy parameter red and s = 1/2+k/8,
where k = 1, 2 and 3, in both noiseless cases and highly noisy cases.
One-dimensional examples:
We start with the 1D SSWPT. In some real applications, e.g., seismic data analysis
[34, 52], wave-like components are only supported in a bounded domain or they have sharp
changes in instantaneous frequencies. Hence, we would like to test a benchmark signal f (x)
in which there is a component with a bounded support and an oscillatory instantaneous
frequency, and a component with an exponential instantaneous frequency (see Figure 2).
Of a special interest to test the performance of synchrosqueezed transforms for impulsive
waves, a wavelet component is added in this signal at x = 0.2. The synthetic benchmark
17
1
3
10
2
0.5
1
5
0
0
−1
−5
−2
−10
0
−0.5
−1
0
0.2
0.4
0.6
0.8
0
0.2
0.4
0.6
0.8
0
0.2
0.4
0.6
0.8
x
x
x
Figure 2: Left: A 1D synthetic benchmark signal. It is normalized using L∞ norm. Middle:
A noisy version generated with white Gaussian noise 0.75N (0, 1). Right: A noisy version
contaminated by an α stable random noise [2] with parameters α = 1, dispersion= 0.9,
δ = 1, N = 8192. The noise is re-scaled to have a L∞ -norm equal to 15 by dividing a
constant factor.
signal1 f (x) is defined as
f (x) = χ[0,0.6] (x)f1 (x) + χ[0,0.6] (x)f2 (x) + χ[0.6,1] (x)f4 (x) + χ[0.4,0.8] (x)f3 (x) + f5 (x),
where
f1 (x) = 0.6 cos(700πx),
f2 (x) = 0.8 cos(300πx),
f3 (x) = 0.7 cos(1300πx + 5 sin(20πx)),
!
80π1005x/4
f4 (x) = sin
,
ln(100)
2
f5 (x) = 3e−50(x−0.2) cos(50(x − 0.2)),
and χ(x) is the indicator function. f (x) is sampled in [0, 1] with a sampling rate 8192 Hz
and the range of instantaneous frequencies is 150 − 1600 Hz. The white Gaussian noise in
this example is 0.75N (0, 1).
Although we are not aware of the optimal value of the scaling parameter s, it is clear
from Theorem 2.6 and 2.7 that the synchrosqueezed transform with a smaller s is more
suitable for noisy signals. As shown in the second and the third rows in Figure 3, in the
noisy cases, the synchrosqueezed energy distribution with s = 0.625 (in the first column)
is better than the one with s = 0.75 (in the second column), which is better than the one
with s = 0.875 (in the last column). This agrees with the conclusion in Theorem 2.6 and
2.7 that a smaller s results in a higher probability to obtain a good instantaneous frequency
estimate.
Another key point is that a wave packet coefficient with a larger magnitude has a
higher probability to give a good instantaneous frequency estimate. A highly redundant
wave packet transform would have wave packets better fitting the local oscillation of wavelike components. In another word, there would be more coefficients with large magnitudes.
The resulting synchrosqueezed energy distribution has higher non-zero energy concentrating
1
Prepared by Mirko van der Baan and available in [34, 40].
18
around instantaneous frequencies. This is also validated in Figure 3. The results in the
third row obtained with red = 16 is better than those in the second row obtained with
red = 1.
It is also interesting to observe that the SST with a smaller s is better at capturing the
component boundaries, e.g. at x = 0.39, 0.59 and 0.77 and is more robust to an impulsive
perturbation (see Figure 2 and 3 at x = 0.2 and an example of α stable noise in Figure
2 and 4). Boundaries and impulse perturbations would produce frequency aliasing. The
SSWPT with a smaller s has wave packets with a smaller support in frequency and a
larger support in space. Hence, it is more robust to frequency aliasing in the sense that
the influence of impulsive perturbations is smoothed out and the synchrosqueezed energy
of the target components might not get dispersed when it meets the frequency aliasing, as
shown in Figure 4.
However, if s is small, the instantaneous frequency estimate might be smoothed out and
it is difficult to observe detailed information of instantaneous frequencies. As shown in the
first row of Figure 3, when the input signal is noiseless, the synchrosqueezed transforms
with s = 0.75 and 0.875 have better accuracy than the one with s = 0.625. In short, it
is important to tune scaling parameters for data-dependent synchrosqueezed transforms,
which has been implemented in the SynLab toolbox.
Two-dimensional examples:
We now explore the performance of the 2D SSWPT using a single wave-like component
in Figure 5. The function
f (x) = e2πi(60(x1 +0.05 sin(2πx1 ))+60(x2 +0.05 sin(2πx2 )))
(12)
is uniformly sampled in [0, 1]2 with a sampling rate 512 Hz and is disturbed by additive
white Gaussian noise 5N (0, 1). The 2D SSWPTs with s = 0.625, 0.75 and 0.875 are applied
to this noisy example and their results are shown in Figure 6. Since the synchrosqueezed
energy distribution Tf (v1 , v2 , xR1 , x2 ) of an image is a function in R4 , we fix x2 = 0, stack
the results in v2 , and visualize R Tf (v1 , v2 , x1 , 0)dv2 . The results in Figure 6 again validate
the theoretical conclusion in Theorem 2.8 that a smaller scaling parameter s and a higher
redundancy yield to a better SST for noisy data.
It is interesting that a band-limited SST can also provide better statistical stability if
the range of instantaneous frequencies/local wave vectors is known a priori. Let us justify
this idea with the follow example. We apply the band-limited SSWPT to the 2D noisy
image in Figure 6 and present the results in the last row of Figure 6. Comparing to the
results in the second row of Figure 6, the band-limited SSWPT clearly outperforms the
original SSWPT.
Component test
Here we present an example to validate the last observation in Section 3.1. Suppose
we look at a region in the time-frequency or phase space domain and we know there might
be only one IMT in this region. This assumption is reasonable because after the SST
people might be interested in the synchrosqueezed energy in a particular region: is this
corresponding to a component or just heavy noise? A straightforward solution is that, at
19
2000
1500
1500
1500
1000
500
0
0
freq
2000
freq
freq
2000
1000
500
0.2
0.4
0.6
0.8
0
0
1
500
0.2
0.4
0.6
0.8
0
0
1
2000
1500
1500
1500
1000
500
freq
2000
1000
500
0.2
0.4
0.6
0.8
0
0
1
0.2
0.4
0.6
0.8
0
0
1
1500
1500
1500
freq
2000
freq
2000
1000
500
0.2
0.4
0.6
0.8
1
0
0
1
0.2
0.4
0.6
0.8
1
0.6
0.8
1
time
2000
500
0.8
1000
time
1000
0.6
500
time
0
0
0.4
time
2000
0
0
0.2
time
freq
freq
time
freq
1000
1000
500
0.2
0.4
time
0.6
time
0.8
1
0
0
0.2
0.4
time
Figure 3: Synchrosqueezed energy distributions with s = 0.625 (left column), s = 0.75
(middle column) and s = 0.875 (right column). In the first row, we apply the SSWPT to
clean data. In the second row, the SSWPT with a smaller redundancy is applied to the
noisy data with 0.75N (0, 1) noise in Figure 2. In the last row, a highly redundant SSWPT
is applied to the same noisy data.
each time or space grid point, we only reassign those coefficients with the largest magnitude.
By Theorem 2.8, if there is an IMT, we can obtain a sketch of its instantaneous frequency
or local wave vector with a high probability. If there was only noise, we would obtain
random reassigned energy with a high probability. Using this idea, we apply the bandlimited SSWPT with s = 0.625 and red = 10 to a noisy version of the image in Figure
5 left. From left to right, Figure 7 shows the results of a noisy image (12) with 5N (0, 1)
noise, a noisy image (12) with 10N (0, 1) noise, and an image with only noise, respectively.
A reliable sketch of the local wave vector is still visible even if the input image is highly
noisy.
20
2000
1500
1500
1500
1000
500
0
0
freq
2000
freq
freq
2000
1000
1000
500
0.2
0.4
0.6
0.8
0
0
1
500
0.2
0.4
time
0.6
0.8
0
0
1
0.2
time
0.4
0.6
0.8
1
time
Figure 4: Synchrosqueezed energy distributions with s = 0.625 (left), s = 0.75 (middle)
and s = 0.875 (right) using highly redundant SSWPTs. The synchrosqueezed energy with
a smaller s is smoother and the influence of impulsive noise is weaker.
0.8
0.8
0.6
0.6
x2
1
x2
1
0.4
0.4
0.2
0.2
0
0
0.2
0.4
0.6
0.8
0
0
1
0.2
x1
0.4
0.6
0.8
1
x1
Figure 5: Left: A 2D noiseless wave-like component. Right: A noisy wave-like component
generated with white Gaussian noise 5N (0, 1).
3.2.2
Quantitative performance analysis
We quantitatively analyze the performance of the highly redundant SSWPT and compare
it with other methods in this subsection in terms of statistical stability and computational
efficiency. Let us revisit the wave-like component in Figure 1, where
f (x) = e60πi(x+0.05 cos(2πx))
is sampled in [0, 1] with a sampling rate 1024 Hz. Its instantaneous frequency is q(x) =
30(1 − 0.1π sin(2πx)). Let g = f + e be the noisy data, where e is white Gaussian noise
with a distribution σ 2 N (0, 1). To measure noise, we introduce the following signal-to-noise
ratio (SNR) of the input data g = f + e:
Var(f )
SNR[dB](g) = 10 log10
.
Var(e)
An ideal time-frequency distribution of f (x) is a function D(v, x) such that D(v, x) = δ(v −
q(x)), where δ(v) is a Dirac delta function of v. To quantify the numerical performance of
various methods, we introduce the Earth mover’s distance (EMD) [29, 30, 31] to measure the
distance between a resultant time-frequency distribution T (v, x) and the ideal distribution
21
200
100
100
100
0
v1
200
v1
v1
200
0
−100
−100
−100
−200
−200
−200
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
200
100
100
100
0
v1
200
0
−100
−100
−200
−200
−200
0.4
0.6
0.8
1
0
0.2
0.4
x1
0.6
0.8
1
0
100
100
100
v1
200
v1
200
−100
0
−100
−200
0.4
0.6
0.8
1
0
1
0.4
0.6
0.8
1
0.6
0.8
1
0
−100
−200
0.2
0.8
x1
200
0
0.2
x1
0
0.6
0
−100
0.2
0.4
x1
200
0
0.2
x1
v1
v1
x1
v1
0
−200
0.2
0.4
x1
0.6
x1
0.8
1
0
0.2
0.4
x1
R
Figure 6: Stacked synchrosqueezed energy distribution R Tf (v1 , v2 , x1 , 0)dv2 of the noisy
2D signal in Figure 5. From left to right, s = 0.625, 0.75 and 0.875. From top to bottom:
red = 1, red = 10, and red = 10 with a restricted frequency band from 20 to 120 Hz.
D(v, x). At each x, after the discretization and normalization of T (v, x), we obtain a 1D
discrete distribution T̃ (v, x). Similarly, we compute the discrete version D̃(v, x) of D(v, x).
At each x, we compute the 1D EMD between T̃ (v, x) and D̃(v, x). The average EMD at
all x is defined as the distance (also denoted as EMD) between T (v, x) and D(v, x) in this
paper. A smaller EMD means a better time-frequency concentration to the ground true
instantaneous frequency and fewer noise fluctuations.
First, we compare the performance of the SSWPT with different redundancy parameters
red for noisy data with different SNRs. We compare the EMD between their resultant timefrequency distributions and the ideal one. Figure 9 (top-left) shows the EMD as functions
in the variable of red. The EMD functions decrease fast when the red increases to 10. All
the SSWPTs with red ≥ 10 have almost the same efficiency.
Second, we compare the performance of the SSWPT with the same redundancy parameters red = 10 but different geometric scaling parameter s for noisy data with different
noise level, e.g. σ 2 ranging from 0 to 4, i.e., SNR from ∞ to −15. We compare the EMD
between their resultant time-frequency distributions and the ideal one. Figure 9 (top-right)
shows the EMD as functions in the variable of σ 2 for different s. It shows that when the
22
100
100
100
0
v1
200
v1
200
v1
200
0
0
−100
−100
−100
−200
−200
−200
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
x1
0.6
0.8
1
0
100
100
0
v1
100
v1
200
v1
200
0
−100
−100
−200
−200
−200
0.4
0.6
0.8
1
0
0.2
0.4
x1
0.6
0.8
1
0.6
0.8
1
0
−100
0.2
0.4
x1
200
0
0.2
x1
0.6
0.8
1
0
0.2
0.4
x1
x1
120
120
100
100
100
80
80
80
60
freq
120
freq
freq
Figure 7: Top row: The synchrosqueezed energy distribution of the highly redundant bandlimited SSWPT with a frequency band 20 to 120 Hz. Bottom row: reassigned wave packet
coefficients with the largest magnitude at a space location. Left column: 5N (0, 1) noise.
Middle column: 10N (0, 1) noise. Right column: noise only.
60
60
40
40
40
20
20
20
0
0
0.2
0.4
0.6
time
0.8
1
0
0
0.2
0.4
0.6
time
0.8
1
0
0
0.2
0.4
0.6
0.8
1
time
Figure 8: Left and middle: The multitaper time-frequency reassignment using an arithmetic
mean (MRSA) and a geometric mean (MRSG) with 10 tapers. Right: The highly redundant
SSWPT with s = 0.75 and red = 10.
data is clean, the SSWPT with s = 1, which is essentially the SSWT with high redundancy,
has the best performance. However, once the data is noisy, the SSWPT with s = 0.75 has
the best overall performance. This observation agrees with our theorems in Section 2. If
s is larger, the instantaneous frequency estimation is more accurate if there is no noise.
But the probability for a good estimation is smaller if there is noise. If s is smaller , the
estimation accuracy is worse but the probability for a good estimation is higher. In theory,
it is interesting to find an optimal s that makes a good balance. In practice, a heuristic
choice for the optimal s is 0.75.
Third, we compare the performance of the standard SSWT, SSSTFT in [1] and the
23
highly redundant SSWPT. Again, we choose the parameters in the example of Figure 1
because they result in good visualization of time-frequency distribution. These transforms
are applied to noisy data with different noise level. We compare the EMD between their
resultant time-frequency distributions and the ideal one. The comparison is shown in Figure
9 (bottom-left). The highly redundant SSWPT has better time-frequency concentration
than the standard SSWT and SSSTFT in all examples, even if in the noiseless case.
Finally, we compare the multitaper time-frequency reassignment using an arithmetic
mean (MRSA) and a geometric mean (MRSG) in[46] with the highly redundant SSWPT.
A MATLAB package of the MRSA and the MRSG is available on the authors’ homepage.
The time-frequency distribution of these three methods are visualized in Figure 8. We
visualize their results using the same discrete grid in the time-frequency domain. To make
a fair comparison, the number of tapers are chosen to be 10 for the MRSA and the MRSG,
and the redundancy parameter red is 10 for the SSWPT with s = 0.75. For one realization
of this experiment, the running time for the MRSA or the MRSG is 155.13 seconds, while
the running time for the SSWPT is only 0.43 seconds. A visual inspection of Figure 8 shows
that the SSWPT also outperforms the MRSA and the MRSG: clear spectral energy at the
boundary and better time-frequency concentration. To quantify this comparison, we apply
the EMD again and the results are shown in Figure 9 (bottom-right). In most cases, the
SSWPT with s = 0.75 gives a smaller EMD. When σ 2 is near 4, i.e., the SNR is near −15,
the MRSG with 10 tapers and the SSWPT have comparable performance.
4
Conclusion
In theory, the statistical analysis in this paper has analyzed the statistical properties of a
wide range of compactly supported synchrosqueezed transforms in multidimensional spaces,
considering zero mean stationary Gaussian random process and small perturbation. Guided
by these properties, this paper has presented several approaches to improve the performance
of these synchrosqueezed transforms under heavy noise. A MATLAB package SynLab for
these algorithms is available at https://github.com/HaizhaoYang/SynLab.
Acknowledgments. This work was partially supported by the National Science Foundation under award DMS-0846501 and the U.S. Department of Energys Advanced Scientific
Computing Research program under award DE-FC02-13ER26134/DE-SC0009409 for Lexing Ying. H. Yang is also supported by a AMS Simons travel grant. He thanks Jean-Baptiste
Tary and Mirko van der Baan for the discussion of benchmark signals, thanks Lexing Ying,
Charles K. Chui, Jianfeng Lu and Segey Fomel for the discussion of the implementation
and the application of synchrosqueezed transforms.
References
[1] https://sites.google.com/site/hautiengwu/home/download.
[2] http://users.ece.utexas.edu/~bevans/projects/rfi/software/.
[3] F. Auger and P. Flandrin. Improving the readability of time-frequency and time-scale
representations by the reassignment method. Signal Processing, IEEE Transactions
on, 43(5):1068 –1089, 1995.
24
1.7
3
SNR=40
SNR=30
SNR=20
SNR=10
1.65
2.5
2
EMD
EMD
1.6
1.5
1.55
1
1.5
1.45
0
0
10
20
30
Reduncancy
40
0
50
3
3
2.5
2.5
2
2
EMD
EMD
SSWPT,red=10,s=0.5
SSWPT,red=10,s=0.625
SSWPT,red=10,s=0.75
SSWPT,red=10,s=0.875
SSWPT,red=10,s=1
0.5
1.5
1
1
0.5
0
1
2
σ2
3
0
4
4
SSWPT,red=10,s=0.75
SSWPT,red=10,s=1
MRSA,#taper=5
MRSA,#taper=10
MRSG,#taper5
MRSG,#taper=10
0.5
0
3
1.5
1
HSST
SSWT
SSWPT,red=10,s=0.75
SSWPT,red=10,s=1
2
σ2
0
1
2
σ2
3
4
Figure 9: Top-left: The EMD as functions in the variable of red for different SNRs. Bottomleft: Comparison of the standard SSWT, SSSTFT and the SSWPT. Bottom-right: Comparison of the multitaper time-frequency reassignment (the numbers of tapers are 5 and
10) and the SSWPT. Data plotted above is the averaging of 20 independent realizations.
[4] F. Auger, P. Flandrin, Y.-T. Lin, S. McLaughlin, S. Meignen, T. Oberlin, and H. tieng
Wu. Time-frequency reassignment and synchrosqueezing: An overview. Signal Processing Magazine, IEEE, 30(6):32–41, Nov 2013.
[5] M. Bayram and R. Baraniuk. Multiple window time-frequency analysis. In TimeFrequency and Time-Scale Analysis, 1996., Proceedings of the IEEE-SP International
Symposium on, pages 173–176, Jun 1996.
[6] B. Boashash and S. Member. Estimating and interpreting the instantaneous frequency
of a signal. In Proceedings of the IEEE, pages 520–538, 1992.
[7] E. J. Candès, P. R. Charlton, and H. Helgason. Detecting highly oscillatory signals by
chirplet path pursuit. Applied and Computational Harmonic Analysis, 24(1):14 – 40,
2008.
[8] E. Chassande-Mottin, F. Auger, and P. Flandrin. Time-frequency/time-scale reassignment. In Wavelets and signal processing, Appl. Numer. Harmon. Anal., pages 233–267.
Birkhäuser Boston, Boston, MA, 2003.
[9] E. Chassande-Mottin, I. Daubechies, F. Auger, and P. Flandrin. Differential reassignment. Signal Processing Letters, IEEE, 4(10):293 –294, 1997.
25
[10] E. Chassande-Mottin, P. Flandrin, and F. Auger. On the statistics of spectrogram
reassignment vectors. Multidimensional Systems and Signal Processing, 9(4):355–362,
1998.
[11] E. Chassande-Mottin and A. Pai. Best chirplet chain: Near-optimal detection of gravitational wave chirps. Physical Review D, 73(4), 2006.
[12] Y.-C. Chen, M.-Y. Cheng, and H.-T. Wu. Non-parametric and adaptive modelling of
dynamic periodicity and trend with heteroscedastic and dependent errors. Journal of
the Royal Statistical Society: Series B (Statistical Methodology), 76(3):651–682, 2014.
[13] C. K. Chui, Y.-T. Lin, and H.-T. Wu. Real-time dynamics acquisition from irregular samples – with application to anesthesia evaluation. Analysis and Applications,
Accepted.
[14] M. Clausel, T. Oberlin, and V. Perrier. The monogenic synchrosqueezed wavelet transform: a tool for the decomposition/demodulation of amfm images. Applied and Computational Harmonic Analysis, 39(3):450 – 486, 2015.
[15] I. Daubechies, J. Lu, and H.-T. Wu. Synchrosqueezed wavelet transforms: an empirical
mode decomposition-like tool. Appl. Comput. Harmon. Anal., 30(2):243–261, 2011.
[16] I. Daubechies and S. Maes. A nonlinear squeezing of the continuous wavelet transform
based on auditory nerve models. In Wavelets in Medicine and Biology, pages 527–546.
CRC Press, 1996.
[17] D. L. Donoho. De-noising by soft-thresholding. IEEE Trans. Inf. Theor., 41(3):613–
627, may 1995.
[18] D. L. Donoho and J. M. Johnstone. Ideal spatial adaptation by wavelet shrinkage.
Biometrika, 81(3):425–455, 1994.
[19] S. Fomel. Seismic data decomposition into spectral components using regularized
nonstationary autoregression. Geophysics, 78(6):O69–O76, 2013.
[20] G. Fraser and B. Boashash. Multiple window spectrogram and time-frequency distributions. In Acoustics, Speech, and Signal Processing, 1994. ICASSP-94., 1994 IEEE
International Conference on, volume iv, pages IV/293–IV/296 vol.4, Apr 1994.
[21] R. G. Gallager. Circularly-symmetric gaussian random vectors. preprint, 2008.
[22] R. Herrera, J. Han, and M. van der Baan. Applications of the synchrosqueezing
transform in seismic time-frequency analysis. Geophysics, 79(3):V55–V64, 2014.
[23] W. Huang, Z. Shen, N. E. Huang, and Y. C. Fung. Engineering analysis of biological
variables: An example of blood pressure over 1 day. Proc. Natl. Acad. Sci., 95, 1998.
[24] H.-H. Kuo. White noise distribution theory. Probability and stochastics series. CRC
Press, Boca Raton (USA), 1996.
[25] C. Li and M. Liang. A generalized synchrosqueezing transform for enhancing signal
timefrequency representation. Signal Processing, 92(9):2264 – 2274, 2012.
26
[26] J. Lu, B. Wirth, and H. Yang. Combining 2D synchrosqueezed wave packet transform
with optimization for crystal image analysis. arXiv:1501.06254, 2015.
[27] F. Neeser and J. Massey. Proper complex random processes with applications to
information theory. Information Theory, IEEE Transactions on, 39(4):1293–1302, Jul
1993.
[28] T. Oberlin, S. Meignen, and V. Perrier. Second-order synchrosqueezing transform or
invertible reassignment? towards ideal time-frequency representations. Signal Processing, IEEE Transactions on, 63(5):1335–1344, March 2015.
[29] O. Pele and M. Werman. A linear time histogram metric for improved sift matching.
In D. Forsyth, P. Torr, and A. Zisserman, editors, Computer Vision ECCV 2008,
volume 5304 of Lecture Notes in Computer Science, pages 495–508. Springer Berlin
Heidelberg, 2008.
[30] O. Pele and M. Werman. Fast and robust earth mover’s distances. In Computer Vision,
2009 IEEE 12th International Conference on, pages 460–467, Sept 2009.
[31] S. Peleg, M. Werman, and H. Rom. A unified approach to the change of resolution:
space and gray-level. Pattern Analysis and Machine Intelligence, IEEE Transactions
on, 11(7):739–742, Jul 1989.
[32] B. Picinbono. On instantaneous amplitude and phase of signals. IEEE Trans. Signal
Processing, pages 552–560, 1997.
[33] H. M. Singer and I. Singer. Analysis and visualization of multiply oriented lattice
structures by a two-dimensional continuous wavelet transform. Physical Review E,
74:031103, 2006.
[34] J. B. Tary, R. H. Herrera, J. Han, and M. van der Baan. Spectral estimation-What is
new? What is next? Review of Geophysics, 52(4):723–749, Dec. 2014.
[35] M. Taylor. Random fields: stationarity, ergodicity, and spectral behavior, http://
www.unc.edu/math/Faculty/met/rndfcn.pdf.
[36] G. Thakur, E. Brevdo, N. S. Fučkar, and H.-T. Wu. The synchrosqueezing algorithm
for time-varying spectral analysis: robustness properties and new paleoclimate applications. Signal Processing, 93(5):1079–1094, 2013.
[37] G. Thakur and H.-T. Wu. Synchrosqueezing-based recovery of instantaneous frequency
from nonuniform samples. SIAM J. Math. Analysis, 43(5):2078–2095, 2011.
[38] D. Thomson. Spectrum estimation and harmonic analysis. Proceedings of the IEEE,
70(9):1055–1096, Sept 1982.
[39] A. Van Den Bos. The multivariate complex normal distribution – a generalization.
Information Theory, IEEE Transactions on, 41(2):537–539, Mar 1995.
[40] M. van der Baan. http://www.ualberta.ca/~vanderba/ftp/benchmark_signals.
zip.
[41] A. D. Veltcheva. Wave and group transformation by a hilbert spectrum. Coastal
Engineering Journal, 44(4), 2002.
27
[42] H.-T. Wu. Instantaneous frequency and wave shape functions (i). Applied and Computational Harmonic Analysis, 35(2):181 – 199, 2013.
[43] H.-T. Wu, Y.-H. Chan, Y.-T. Lin, and Y.-H. Yeh. Using synchrosqueezing transform to
discover breathing dynamics from ECG signals. Applied and Computational Harmonic
Analysis, 36(2):354 – 359, 2014.
[44] H.-T. Wu, S.-S. Hseu, M.-Y. Bien, Y. R. Kou, and I. Daubechies. Evaluating physiological dynamics via synchrosqueezing: Prediction of ventilator weaning. IEEE Transactions on Biomedical Engineering, 61(3):736–744, March 2014.
[45] Z. Wu, N. E. Huang, and X. Chen. Some considerations on physical analysis of data.
Advances in Adaptive Data Analysis, 3(1-2):95–113, 2011.
[46] J. Xiao and P. Flandrin. Multitaper time-frequency reassignment for nonstationary
spectrum estimation and chirp enhancement. Signal Processing, IEEE Transactions
on, 55(6):2851–2860, June 2007.
[47] H. Yang.
Oscillatory data analysis and fast algorithms for integral operators.
PhD thesis, Stanford University, 2015.
This dissertation is online at: http://purl.stanford.edu/fq061ny3299 or http://web.stanford.edu/
~haizhao/publications/ThesisHaizhao.pdf.
[48] H. Yang. Synchrosqueezed wave packet transforms and diffeomorphism based spectral
analysis for 1d general mode decompositions. Applied and Computational Harmonic
Analysis, 39(1):33 – 66, 2015.
[49] H. Yang, J. Lu, W. Brown, I. Daubechies, and L. Ying. Quantitative canvas weave
analysis using 2-D synchrosqueezed transforms: Application of time-frequency analysis
to art investigation. Signal Processing Magazine, IEEE, 32(4):55–63, July 2015.
[50] H. Yang, J. Lu, and L. Ying. Crystal image analysis using 2D synchrosqueezed transforms. Multiscale Modeling & Simulation. A SIAM Interdisciplinary Journal, 2015.
[51] H. Yang and L. Ying. Synchrosqueezed wave packet transform for 2d mode decomposition. SIAM Journal on Imaging Sciences, 6(4):1979–2009, 2013.
[52] H. Yang and L. Ying. Synchrosqueezed curvelet transform for two-dimensional mode
decomposition. SIAM Journal on Mathematical Analysis, 46(3):2052–2083, 2014.
28
| 10math.ST
|
Formal Guarantees on the Robustness of a
Classifier against Adversarial Manipulation
arXiv:1705.08475v2 [cs.LG] 5 Nov 2017
Matthias Hein and Maksym Andriushchenko
Department of Mathematics and Computer Science
Saarland University, Saarbrücken Informatics Campus, Germany
Abstract
Recent work has shown that state-of-the-art classifiers are quite brittle,
in the sense that a small adversarial change of an originally with high
confidence correctly classified input leads to a wrong classification again
with high confidence. This raises concerns that such classifiers are vulnerable
to attacks and calls into question their usage in safety-critical systems. We
show in this paper for the first time formal guarantees on the robustness
of a classifier by giving instance-specific lower bounds on the norm of the
input manipulation required to change the classifier decision. Based on
this analysis we propose the Cross-Lipschitz regularization functional. We
show that using this form of regularization in kernel methods resp. neural
networks improves the robustness of the classifier with no or small loss in
prediction performance.
1
Introduction
The problem of adversarial manipulation of classifiers has been addressed initially in the area
of spam email detection, see e.g. [5, 16]. The goal of the spammer is to manipulate the spam
email (the input of the classifier) in such a way that it is not detected by the classifier. In deep
learning the problem was brought up in the seminal paper by [24]. They showed for state-ofthe-art deep neural networks, that one can manipulate an originally correctly classified input
image with a non-perceivable small transformation so that the classifier now misclassifies
this image with high confidence, see [7] or Figure 4 for an illustration. This property calls
into question the usage of neural networks and other classifiers showing this behavior in
safety critical systems, as they are vulnerable to attacks. On the other hand this also shows
that the concepts learned by a classifier are still quite far away from the visual perception
of humans. Subsequent research has found fast ways to generate adversarial samples with
high probability [7, 12, 19] and suggested to use them during training as a form of data
augmentation to gain more robustness. However, it turns out that the so-called adversarial
training does not settle the problem as one can yet again construct adversarial examples
for the final classifier. Interestingly, it has recently been shown that there exist universal
adversarial changes which when applied lead, for every image, to a wrong classification with
high probability [17]. While one needs access to the neural network model for the generation
of adversarial changes, it has been shown that adversarial manipulations generalize across
neural networks [18, 15, 14], which means that neural network classifiers can be attacked
even as a black-box method. The most extreme case has been shown recently [15], where
they attack the commercial system Clarifai, which is a black-box system as neither the
underlying classifier nor the training data are known. Nevertheless, they could successfully
generate adversarial images with an existing network and fool this commercial system. This
emphasizes that there are indeed severe security issues with modern neural networks. While
countermeasures have been proposed [8, 7, 26, 18, 12, 2], none of them provides a guarantee
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
of preventing this behavior [3]. One might think that generative adversarial neural networks
should be resistant to this problem, but it has recently been shown [13] that they can also
be attacked by adversarial manipulation of input images.
In this paper we show for the first time instance-specific formal guarantees on the robustness
of a classifier against adversarial manipulation. That means we provide lower bounds on the
norm of the change of the input required to alter the classifier decision or said otherwise: we
provide a guarantee that the classifier decision does not change in a certain ball around the
considered instance. We exemplify our technique for two widely used family of classifiers:
kernel methods and neural networks. Based on the analysis we propose a new regularization
functional, which we call Cross-Lipschitz Regularization. This regularization functional
can be used in kernel methods and neural networks. We show that using Cross-Lipschitz
regularization improves both the formal guarantees of the resulting classifier (lower bounds)
as well as the change required for adversarial manipulation (upper bounds) while maintaining
similar prediction performance achievable with other forms of regularization. While there
exist fast ways to generate adversarial samples [7, 12, 19] without constraints, we provide
algorithms based on the first order approximation of the classifier which generate adversarial
samples satisfying box constraints in O(d log d), where d is the input dimension.
2
Formal Robustness Guarantees for Classifiers
In the following we consider the multi-class setting for K classes and d features where one
has a classifier f : Rd → RK and a point x is classified via c = arg max fj (x). We call a
j=1,...,K
classifier robust at x if small changes of the input do not alter the decision. Formally, the
problem can be described as follows [24]. Suppose that the classifier outputs class c for
input x, that is fc (x) > fj (x) for j =
6 c (we assume the decision is unique). The problem of
generating an input x + δ such that the classifier decision changes, can be formulated as
min kδkp ,
δ∈Rd
s.th.
max fl (x + δ) ≥ fc (x + δ) and x + δ ∈ C,
l6=c
(1)
where C is a constraint set specifying certain requirements on the generated input x + δ, e.g.,
an image has to be in [0, 1]d . Typically, the optimization problem (1) is non-convex and thus
intractable. The so generated points x + δ are called adversarial samples. Depending on
the p-norm the perturbations have different characteristics: for p = ∞ the perturbations are
small and affect all features, whereas for p = 1 one gets sparse solutions up to the extreme
case that only a single feature is changed. In [24] they used p = 2 which leads to more spread
but still localized perturbations. The striking result of [24, 7] was that for most instances
in computer vision datasets, the change δ necessary to alter the decision is astonishingly
small and thus clearly the label should not change. However, we will see later that our new
regularizer leads to robust classifiers in the sense that the required adversarial change is so
large that now also the class label changes (we have found the correct decision boundary),
see Fig 4. Already in [24] it is suggested to add the generated adversarial samples as a form
of data augmentation during the training of neural networks in order to achieve robustness.
This is denoted as adversarial training. Later on fast ways to approximately solve (1) were
proposed in order to speed up the adversarial training process [7, 12, 19]. However, in this
way, given that the approximation is successful, that is arg max fj (x + δ) 6= c, one gets just
j
upper bounds on the perturbation necessary to change the classifier decision. Also it was
noted early on, that the final classifier achieved by adversarial training is again vulnerable
to adversarial samples [7]. Robust optimization has been suggested as a measure against
adversarial manipulation [12, 21] which effectively boils down to adversarial training in
practice. It is thus fair to say that up to date no mechanism exists which prevents the
generation of adversarial samples nor can defend against it [3].
In this paper we focus instead on robustness guarantees, that is we show that the classifier
decision does not change in a small ball around the instance. Thus our guarantees hold for
any method to generate adversarial samples or input transformations due to noise or sensor
failure etc. Such formal guarantees are in our point of view absolutely necessary when a
classifier becomes part of a safety-critical technical system such as autonomous driving. In
the following we will first show how one can achieve such a guarantee and then explicitly
2
derive bounds for kernel methods and neural networks. We think that such formal guarantees
on robustness should be investigated further and it should become standard to report them
for different classifiers alongside the usual performance measures.
2.1
Formal Robustness Guarantee against Adversarial Manipulation
The following guarantee holds for any classifier which is continuously differentiable with
respect to the input in each output component. It is instance-specific and depends to some
extent on the confidence in the decision, at least if we measure confidence by the relative
difference fc (x) − maxj6=c fj (x) as it is typical for the cross-entropy loss and other multi-class
losses. In the following we use the notation Bp (x, R) = {y ∈ Rd | kx − ykp ≤ R}.
Theorem 2.1. Let x ∈ Rd and f : Rd → RK be a multi-class classifier with continuously
differentiable components and let c = arg max fj (x) be the class which f predicts for x. Let
j=1,...,K
kδkp
1
p
1
q
= 1, then for all δ ∈ Rd with
fc (x) − fj (x)
≤ max min min
, R := α,
R>0
j6=c max k∇fc (y) − ∇fj (y)kq
q ∈ R be defined as
+
y∈Bp (x,R)
it holds c = arg max fj (x + δ), that is the classifier decision does not change on Bp (x, α).
j=1,...,K
Proof. By the main theorem of calculus, it holds that
Z 1
fj (x + δ) = fj (x) +
h∇fj (x + tδ), δi dt,
for j = 1, . . . , K.
0
Thus, in order to achieve fj (x + δ) ≥ fc (x + δ), it has to hold that
Z 1
0 ≤ fc (x) − fj (x) ≤
h∇fj (x + tδ) − ∇fc (x + tδ), δi dt
0
Z
≤ kδkp
0
1
k∇fj (x + tδ) − ∇fc (x + tδ)kq dt,
where the first inequality holds as fc (x) ≥ fj (x) for all j = 1, . . . , K and in the last step we
have used Hölder inequality together with the fact that the q-norm is dual to the p-norm,
where q is defined via p1 + 1q = 1. Thus the minimal norm of the change δ required to change
the classifier decision from c to j satisfies
kδkp ≥ R 1
0
fc (x) − fj (x)
k∇fj (x + tδ) − ∇fc (x + tδ)kq dt
.
We upper bound the denominator over some fixed ball Bp (x, R). Note that by doing this,
we can only make assertions of perturbations δ ∈ Bp (0, R) and thus the upper bound in the
guarantee is at most R. It holds
Z 1
sup
k∇fj (x + tδ) − ∇fc (x + tδ)kq dt ≤ maxy∈Bp (x,R) k∇fj (y) − ∇fc (y)kq .
δ∈Bp (0,R)
0
Thus we get the lower bound for the minimal norm of the change δ required to change the
classifier decision from c to j,
(
)
fc (x) − fj (x)
kδkp ≥ min R,
:= α.
maxy∈Bp (x,R) k∇fj (y) − ∇fc (y)kq
As we are interested in the worst case, we take the minimum over all j 6= c. Finally, the result
holds for any fixed R > 0 so that we can maximize over R which yields the final result.
3
Note that the bound requires in the denominator a bound on the local Lipschitz constant
of all cross terms fc − fj , which we call local cross-Lipschitz constant in the following.
However, we do not require to have a global bound. The problem with a global bound is
that the ideal robust classifier is basically piecewise constant on larger regions with sharp
transitions between the classes. However, the global Lipschitz constant would then just be
influenced by the sharp transition zones and would not yield a good bound, whereas the
local bound can adapt to regions where the classifier is approximately constant and then
yields good guarantees. In [24, 4] they suggest to study the global Lipschitz constant1 of
each fj , j = 1, . . . , K. A small global Lipschitz constant for all fj implies a good bound as
k∇fj (y) − ∇fc (y)kq ≤ k∇fj (y)kq + k∇fc (y)kq ,
(2)
but the converse does not hold. As discussed below it turns out that our local estimates are
significantly better than the suggested global estimates which implies also better robustness
guarantees. In turn we want to emphasize that our bound is tight, that is the bound is
attained, for linear classifiers fj (x) = hwj , xi, j = 1, . . . , K. It holds
kδkp = min
j6=c
hwc − wj , xi
.
kwc − wj kq
In Section 4 we refine this result for the case when the input is constrained to [0, 1]d . In
general, it is possible to integrate constraints on the input by simply doing the maximum
over the intersection of Bp (x, R) with the constraint set e.g. [0, 1]d for gray-scale images.
2.2
Evaluation of the Bound for Kernel Methods
Next, we discuss how the bound can be evaluated for different classifier models. For simplicity
we restrict ourselves to the case p = 2 (which implies q = 2) and leave the other cases to
future work. We consider the class of kernel methods, that is the classifier has the form
n
X
fj (x) =
αjr k(xr , x),
r=1
where (xr )nr=1 are the n training points, k : Rd × Rd → R is a positive definite kernel function
and α ∈ RK×n are the trained parameters e.g. of a SVM. The goal is to upper bound the
term maxy∈B2 (x,R) k∇fj (y) − ∇fc (y)k2 for this classifier model. A simple calculation shows
2
0 ≤ k∇fj (y) − ∇fc (y)k2 =
n
X
(αjr − αcr )(αjs − αcs ) h∇y k(xr , y), ∇y k(xs , y)i
(3)
r,s=1
It has been reported that kernel methods with a Gaussian kernel are robust to noise. Thus
2
we specialize now to this class, that is k(x, y) = e−γkx−yk2 . In this case
2
2
h∇y k(xr , y), ∇y k(xs , y)i = 4γ 2 hy − xr , y − xs i e−γkxr −yk2 e−γkxs −yk2 .
We will now derive lower and upper bounds on this term uniformly over B2 (x, R) which
allows us to derive the guarantee.
n
o
k2x−xr −xs k2
Lemma 2.1. Let M = min
,
R
, then
2
max
y∈B2 (x,R)
min
y∈B2 (x,R)
max
hy − xr , y − xs i = hx − xr , x − xs i + R k2x − xr − xs k2 + R2
hy − xr , y − xs i = hx − xr , x − xs i − M k2x − xr − xs k2 + M 2
2
2
−γkxr −yk22
−γkxs −yk22
e−γkxr −yk2 e−γkxs −yk2 = e
−γ kx−xr k22 +kx−xs k22 −2M k2x−xr −xs k2 +2M 2
y∈B2 (x,R)
min
e
e
=e
−γ kx−xr k22 +kx−xs k22 +2Rk2x−xr −xs k2 +2R2
y∈B2 (x,R)
1
The Lipschitz constant L wrt to p-norm of a piecewise continuously differentiable function is
given as L = supx∈Rd k∇f (x)kq . Then it holds, |f (x) − f (y)| ≤ L kx − ykp .
4
Proof. For the first part we use
maxy∈B2 (x,R) hy − xr , y − xs i = maxh∈B2 (0,R) hx − xr + h, x − xs + hi
2
= hx − xr , x − xs i + maxh∈B2 (0,R) hh, 2x − xr − xs i + khk2
= hx − xr , x − xs i + R k2x − xr − xs k2 + R2 ,
where the last equality follows by Cauchy-Schwarz and noting that equality is attained as
we maximize over the Euclidean unit ball. For the second part we consider
miny∈B2 (x,R) hy − xr , y − xs i = minh∈B2 (0,R) hx − xr + h, x − xs + hi
2
= hx − xr , x − xs i + minh∈B2 (0,R) hh, 2x − xr − xs i + khk2
= hx − xr , x − xs i + min0≤α≤R −α k2x − xr − xs k2 + α2
n k2x − x − x k
o
r
s 2
= hx − xr , x − xs i − min
, R k2x − xr − xs k2
2
2
n k2x − x − x k
r
s 2
, R} ,
+ min
2
where in the second step we have separated direction and norm of the vector, optimization
over the direction yields with Cauchy-Schwarz the result. Finally, the constrained convex
k2x−xr −xs k2
one-dimensional optimization problem can be solved explicitly as α = min{
, R}.
2
The proof of the other results follows analogously noting that
2
2
e−γkx+h−xr k2 e−γkx+h−xs k2 = e
−γ kx−xr k22 +kx−xs k22 +2hh,2x−xr −xs i+2khk22
.
Using this lemma it is easy to derive the final result.
Proposition 2.1. Let βr = αjr − αcr , r = 1, . . . , n and define M = min
n
k2x−xr −xs k2
,R
2
o
and S = k2x − xr − xs k2 . Then
n
X
maxy∈B2 (x,R) k∇fj (y) − ∇fc (y)k2 ≤ 2γ
h
βr βs max{hx − xr , x − xs i + RS + R2 , 0}e−γ
kx−xr k22 +kx−xs k22 −2M S+2M 2
r,s=1
βr βs ≥0
2
+ min{hx − xr , x − xs i + RS + R , 0}e
+
n
X
−γ kx−xr k22 +kx−xs k22 +2RS+2R2
h
βr βs max{hx − xr , x − xs i − M S + M 2 , 0}e−γ
i
kx−xr k22 +kx−xs k22 +2RS+2R2
r,s=1
βr βs <0
+ min{hx − xr , x − xs i − M S + M 2 , 0}e−γ
kx−xr k22 +kx−xs k22 −2M S+2M 2
1
i! 2
Proof. We bound each term in the sum in Equation 3 separately using that ac ≤ bd if b, c ≥ 0,
a ≤ b and c ≤ d or b ≤ 0, d ≥ 0, a ≤ b and c ≥ d, where c, d correspond to the exponential
terms and a, b to the upper bounds of the inner product. Similarly, ab ≥ cd if b, c ≥ 0, a ≥ b
and c ≥ d or a ≤ 0, d ≥ 0, a ≥ b and c ≤ d. The individual upper and lower bounds are
taken from Lemma 2.1.
While the bound leads to non-trivial estimates as seen in Section 5, the bound is not very
tight. The reason is that the sum is bounded elementwise, which is quite pessimistic. We
think that better bounds are possible but have to postpone this to future work.
5
2.3
Evaluation of the Bound for Neural Networks
We derive the bound for a neural network with one hidden layer. In principle, the technique
we apply below can be used for arbitrary layers but the computational complexity increases
rapidly. The problem is that in the directed network topology one has to consider almost
each path separately to derive the bound. Let U be the number of hidden units and w, u
are the weight matrices of the output resp. input layer. We assume that the activation
function σ is continuously differentiable and assume that the derivative σ 0 is monotonically
increasing. Our prototype activation function we have in mind and which we use later
on in the experiment is the differentiable approximation, σα (x) = α1 log(1 + eαx ) of the
ReLU activation function σReLU (x) = max{0, x}. Note that limα→∞ σα (x) = σReLU (x) and
σα0 (x) = 1+e1−αx . The output of the neural network can be written as
fj (x) =
U
X
wjr σ
d
X
r=1
urs xs ,
j = 1, . . . , K,
s=1
where for simplicity we omit any bias terms, but it is straightforward to consider also models
with bias. A direct computation shows that
U
X
2
k∇fj (y) − ∇fc (y)k2 =
(wjr − wcr )(wjm − wcm )σ 0 (hur , yi)σ 0 (hum , yi)
r,m=1
d
X
url uml , (4)
l=1
where ur ∈ Rd is the r-th row of the weight matrix u ∈ RU ×d . The resulting bound is given
in the following proposition.
Proposition 2.2. Let σ be a continuously differentiable activation function with σ 0 monoPd
tonically increasing. Define βrm = (wjr − wcr )(wjm − wcm ) l=1 url uml . Then
maxy∈B2 (x,R) k∇fj (y) − ∇fc (y)k2
U
h X
≤
max{βrm , 0}σ 0 hur , xi + R kur k2 σ 0 hum , xi + R kum k2
r,m=1
i 12
+ min{βrm , 0}σ 0 hur , xi − R kur k2 σ 0 hum , xi − R kum k2
Proof. The proof is based on the fact that due the monotonicity of σ 0 and with CauchySchwarz,
maxy∈B2 (x,R) σ 0 (hur , yi) = maxh∈B2 (0,R) σ 0 (hur , xi + hur , hi) = σ 0 (hur , xi + R kur k2 ).
Similarly, one gets
miny∈B2 (x,R) σ 0 (hur , yi) = minh∈B2 (0,R) σ 0 (hur , xi + hur , hi) = σ 0 (hur , xi − R kur k2 ).
The rest of the result follows by element-wise bounding the terms in the sum in Equation
4.
As discussed above the global Lipschitz bounds of the individual classifier outputs, see (2),
lead to an upper bound of our desired local cross-Lipschitz constant. In the experiments
below our local bounds on the Lipschitz constant are up to 8 times smaller, than what one
would achieve via the global Lipschitz bounds of [24]. This shows that their global approach
is much too rough to get meaningful robustness guarantees.
3
The Cross-Lipschitz Regularization Functional
We have seen in Section 2 that if
max
max
j6=c y∈Bp (x,R)
k∇fc (y) − ∇fj (y)kq ,
(5)
is small and fc (x) − fj (x) is large, then we get good robustness guarantees. The latter
property is typically already optimized in a multi-class loss function. We consider for all
6
methods in this paper the cross-entropy loss so that the differences in the results only come
from the chosen function class (kernel methods versus neural networks) and the chosen
regularization functional. The cross-entropy loss L : {1, . . . , K} × RK → R is given as
K
efy (x)
X
L(y, f (x)) = − log PK
= log 1 +
efk (x)−fy (x) .
fk (x)
k=1 e
k6=y
In the latter formulation it becomes apparent that the loss tries to make the difference
fy (x) − fk (x) as large as possible for all k = 1, . . . , K.
As our goal are good robustness guarantees it is natural to consider a proxy of the quantity
in (5) for regularization. We define the Cross-Lipschitz Regularization functional as
Ω(f ) =
n
K
1 X X
2
k∇fl (xi ) − ∇fm (xi )k2 ,
nK 2 i=1
(6)
l,m=1
where the (xi )ni=1 are the training points. The goal of this regularization functional is to
make the differences of the classifier functions at the data points as constant as possible. In
total by minimizing
n
1X
L yi , f (xi ) + λΩ(f ),
n i=1
(7)
over some function class we thus try to maximize fc (xi ) − fj (xi ) and at the same time
2
keep k∇fl (xi ) − ∇fm (xi )k2 small uniformly over all classes. This automatically enforces
robustness of the resulting classifier. It is important to note that this regularization functional
is coherent with the loss as it shares the same degrees of freedom, that is adding the same
function g to all outputs: fj0 (x) = fj (x) + g(x) leaves loss and regularization functional
invariant. This is the main difference to [4], where they enforce the global Lipschitz constant
to be smaller than one.
3.1
Cross-Lipschitz Regularization in Kernel Methods
In kernel methods one uses typically the regularization P
functional induced by the kernel which
n
is given as the squared norm of the function, f (x) = i=1 αi k(xi , x), in the corresponding
Pn
2
reproducing kernel Hilbert space Hk , kf kHk =
i,j=1 αi αj k(xi , xj ). In particular, for
translation invariant kernels one can make directly a connection to penalization of derivatives
of the function f via the Fourier transform, see [20]. However, penalizing higher-order
derivatives is irrelevant for achieving robustness. Given the kernel expansion of f , one can
write the Cross-Lipschitz regularization function as
Ω(f ) =
n
K
n
1 X X X
(αlr − αmr )(αls − αms ) h∇y k(xr , xi ), ∇y k(xs , xi )i
nK 2 i,j=1
r,s=1
l,m=1
Ω is convex in α ∈ RK×n as k 0 (xr , xs ) = h∇y k(xr , xi ), ∇y k(xs , xi )i is a positive definite
kernel for any xi and with the convex cross-entropy loss the learning problem in (7) is convex.
3.2
Cross-Lipschitz Regularization in Neural Networks
The standard way to regularize neural networks is weight decay; that is, the squared
Euclidean norm of all weights is added to the objective. More recently dropout [22], which
can be seen as a form of stochastic regularization, has been introduced. Dropout can
also be interpreted as a form of regularization of the weights [22, 10]. It is interesting
to note that classical regularization functionals which penalize derivatives of the resulting
classifier function are not typically used in deep learning, but see [6, 11]. As noted above
we restrict ourselves to one hidden
layer neural networks to simplify notation, that is,
PU
Pd
fj (x) = r=1 wjr σ
j = 1, . . . , K. Then we can write the Cross-Lipschitz
s=1 urs xs ,
7
regularization as
Ω(f ) =
U
K
K
K
n
d
X
X
X
X
2 X X
0
0
w
w
−
w
w
σ
(hu
,
x
i)σ
(hu
,
x
i)
url usl
lr ls
lr
ms
r
i
s
i
nK 2 r,s=1
m=1
i,j=1
l=1
l=1
l=1
which leads to an expression which can be fast evaluated using vectorization. Obviously, one
can also implement the Cross-Lipschitz Regularization also for all standard deep networks.
4
Box Constrained Adversarial Sample Generation
The main emphasis of this paper are robustness guarantees without resorting to particular
ways how to generate adversarial samples. On the other hand while Theorem 2.1 gives
lower bounds on the required input transformation, efficient ways to approximately solve
the adversarial sample generation in (1) are helpful to get upper bounds on the required
change. Upper bounds allow us to check how tight our derived lower bounds are. As all of
our experiments will be concerned with images, it is reasonable that our adversarial samples
are also images. However, up to our knowledge, the current main techniques to generate
adversarial samples [7, 12, 19] integrate box constraints by clipping the results to [0, 1]d . We
provide in the following fast algorithms to generate adversarial samples which lie in [0, 1]d .
The strategy is similar to [12], where they use a linear approximation of the classifier to
derive adversarial samples with respect to different norms. Formally,
fj (x + δ) ≈ fj (x) + h∇fj (x), δi ,
j = 1, . . . , K.
Assuming that the linear approximation holds, the optimization problem (1) integrating box
constraints for changing class c into j becomes
minδ∈Rd kδkp
(8)
sbj. to: fj (x) − fc (x) ≥ h∇fc (x) − ∇fj (x), δi
0 ≤ xj + δj ≤ 1
In order to get the minimal adversarial sample we have to solve this for all j 6= c and take
the one with minimal kδkp . This yields the minimal adversarial change for linear classiifers.
Note that (8) is a convex optimization problem, which can be reduced to a one-parameter
problem in the dual. This allows to derive the following result (proofs and algorithms are in
the supplement).
Proposition 4.1. Let p ∈ {1, 2, ∞}, then (8) can be solved in O(d log d) time.
We start with problem for p = 2 which is given as:
minδ∈Rd kδk2
sbj. to: fj (x) − fc (x) ≥ h∇fc (x) − ∇fj (x), δi
0 ≤ xj + δ j ≤ 1
(9)
Lemma 4.1. Let m = arg max fj (x) and define v = ∇fm (x) − ∇fj (x) and 0 > c =
j
fj (x) − fm (x). If a solution of problem
−λvr ,
δ r = 1 − xr ,
−xr ,
(9) exists, then it is given as
if − xr ≤ −λvr ≤ 1 − xr ,
if − λvr ≥ 1 − xr ,
if − xr ≥ −λvr .
The optimal λ ≥ 0 can be obtained by solving
X
c = hv, δi = −λ
vr2 +
−xr ≤−λvr ≤1−xr
X
−λvr >1−xr
vr (1 − xr ) −
X
vr x r .
−λvr <−xr
If Problem 9 is infeasible, then this equation has no solution for λ ≥ 0. In both the feasible and
infeasible case the solution can be found in O(d log d). The algorithm is given in Algorithm 1.
8
Proof. The Lagrangian is given by
1
2
L(δ, λ, α, β) = kδk2 + λ hv, δi − c + hα, x + δ − 1i − hβ, x + δi .
2
The KKT conditions become
δ + λv + α − β = 0
αr (xr + δr − 1) = 0, ∀r = 1, . . . , d
βr (xr + δr ) = 0, ∀r = 1, . . . , d
λ(hv, δi − c) = 0
αr ≥ 0, ∀r = 1, . . . , d
βr ≥ 0, ∀r = 1, . . . , d
λ ≥ 0.
We deduce that if βr > 0 then αr = 0 which implies
δr = −xr = −λvr + βr =⇒ βr = max{0, −xr + λvr }.
Similarly, if αr > 0 then βr = 0 which implies
δr = 1 − xr = −λvr − αr =⇒ αr = max{0, xr − 1 − λvr }.
It follows
if − xr < −λvr < 1 − xr
−λvr
δr = 1 − xr if − λvr > 1 − xr
.
−xr
if − λvr < −xr
We can determine λ by inspecting hv, δi which is given as
X
X
X
hv, δi = −λ
vr2 +
vr (1 − xr ) −
vr xr .
−xr ≤−λvr ≤1−xr
−λvr >1−xr
−λvr <−xr
Note that λ ≥ 0 and 1 − xr ≥ 0 and thus −λvr > 1 − xr implies vr < 0, thus vr (1 − xr ) ≤ 0
and similarly −λvr < −xr implies vr > 0 and thus also −vr xr < 0. Note that the term
, xvrr } in
hv, δi is monotonically decreasing as λ is increasing. Thus one can sort max{ xrv−1
r
increasing order and they represent the thresholds when the summation changes. Then we
compute hv, δi for all of these thresholds and determine the largest threshold λ∗ such that
hv, δi ≤ c. This fixes the index sets of all sums. Then we determine λ by
P
P
vr (1 − xr ) −
vr x r − c
−λ∗ vr <−xr
−λ∗ vr >1−xr
P
λ=
.
vr2
−xr ≤−λ∗ vr ≤1−xr
In total sorting takes time O(d log d) and solving for λ∗ has complexity O(d).
Next we consider the case p = 1.
minδ∈Rd kδk1
(10)
sbj. to: fj (x) − fc (x) ≥ h∇fc (x) − ∇fj (x), δi
0 ≤ xj + δj ≤ 1
Lemma 4.2. Let m = arg max fj (x) and define v = ∇fm (x) − ∇fj (x) and c = fj (x) −
j
fm (x) < 0, then the solution of the problem (10) can be found by Algorithm 2.
Proof. The result is basically obvious but we derive it formally. First of all we rewrite (10)
as a linear program.
d
X
min
ti
(11)
δ∈Rd
i=1
sbj. to: c ≥ hv, δi
− xj ≤ δ j ≤ 1 − xj
− ti ≤ δi ≤ ti
ti ≥ 0
9
Algorithm 1 Computation of box-constrained adversarial samples wrt to k·k2 -norm.
INPUT: c = fj (x) − fm (x) and v = ∇fm (x) − ∇fj (x) (j, desired class, m original class)
, xvrr } in increasing order π
sort γr = max{ xrv−1
r
s=0; ρ = 0
while ρ > c do
s←s+1
compute ρ = hv, δ(γπs )i, where δ(λ) is the function defined in Lemma 4.1
end while
if ρ ≤ c then
compute Im = {r | − xr ≤ −γπs−1 ) vr ≤ 1 − xr },
compute Iu = {r | − γπs−1 vr > 1 − xr },
compute
− γπs−1 vr < −xr }
P Il = {r | P
vr xr −c
vr (1−xr )−
λ=
r∈Il
r∈Iu
P
vr2
.
r∈Im
δ = max{−xr , min{−λvr , 1 − xr }}
else
Problem has no feasible solution
end if
Algorithm 2 Computation of box-constrained adversarial samples wrt to k·k1 -norm.
INPUT: c = fj (x) − fm (x) and v = ∇fm (x) − ∇fj (x) (j, desired class, m original class)
sort |vi | in decreasing order
s:=0, g:=0,
while g > c do
s←s+1
if vπs > 0 then
δπs = −xπs
else
δπs = 1 − xπs
end if
g ← g + δπs vπs
end while
if g ≤ c then
g ← g − vπs δπs
δπs ← c−g
v πs
else
Problem has no feasible solution
end if
10
The Lagrangian of this problem is
L(δ, t, α, β, γ, θ, κ, λ) = ht, 1i + λ(hv, δi − c) − hα, δ + xi + hβ, δ − 1 + xi − hγ, δ + ti + hθ, δ − ti − hκ, ti
= ht, 1 − γ − θ − κi + hδ, β − α + λv + θ − γi + hβ − α, xi − hβ, 1i − λc
Minimization of the Lagrangian over t resp. δ leads only to a non-trivial result if
1−γ−θ−κ=0
β − α + λv + θ − γ = 0
We get the dual problem
max
α,β,θ,γ,κ,λ
hβ − α, xi − hβ, 1i − λc
sbj. to: 1 − γ − θ − κ = 0
β − α + λv + θ − γ = 0
α ≥ 0, β ≥ 0, θ ≥ 0, γ ≥ 0, κ ≥ 0, λ ≥ 0
(12)
(13)
Using the equalities we can now simplify the problem by replacing α and using the fact that
κ is not part of the objective, the positivity just induces an additional constraint. We get
α = β + λv + θ − γ.
Plugging this into the problem (16) we get
max − hλv + θ − γ, xi − hβ, 1i − λc
β,θ,γ,λ
sbj. to: 1 − γ − θ ≥ 0
β + λv + θ − γ ≥ 0
β ≥ 0, θ ≥ 0, γ ≥ 0, λ ≥ 0
(14)
(15)
We get the constraint β ≥ max{0, −λv − θ + γ} (all the inequalities and functions are taken
here componentwise) and thus we can explicitly maximize over β
max − hλv + θ − γ, xi −
θ,γ,λ
d
X
max{0, −λvi − θi + γi } − λc
(16)
i=1
sbj. to: γ + θ ≤ 1
θ ≥ 0, γ ≥ 0, λ ≥ 0
(17)
As 0 ≤ xi ≤ 1 it holds for all θ ≥ 0, γ ≥ 0
h−λv − θ + γ, xi −
d
X
max{0, −λvi − θi + γi } ≤ 0.
i=1
The maximum is attained if γi − θi − λvi = 0 resp. with the constraints on γi , θi this is
equivalent to −1 ≤ λvi ≤ 1. Suppose that λvi > 1 then the maximum is attained for γi = 1
and θi = 0, and for λvi < −1 the maximum is attained for γi = 0 and θi = 1. Thus by
solving explicitly for θ and γ we obtain
X
X
max
(1 − λi vi )xi +
(1 − λi v)(xi − 1) − λc
(18)
λ
λvi >1
λvi <−1
sbj. to: λ ≥ 0
(19)
Note that the first two terms are decreasing with λ and the last term is increasing with λ.
Let λ∗ be the optimum, then we have the following characterization
−1 < λvi < 1
λvi > 1
λvi < −1
=⇒
=⇒
=⇒
γi + θi < 1 =⇒ κi > 0 =⇒ ti = 0 =⇒ δi = 0
γi = 1, θi = 0, βi = 0, αi > 0 =⇒ δi = −xi ,
γi = 0, θi = 1, βi > 0 =⇒ δi = 1 − xi ,
The cases |λvi | = 1 are undetermined but given that λ > 0 the remaining values can be fixed
by solving for c = hv, δi. The time complexity is again determined by the initial sorting step
of O(d log d). The following linear scan requires O(d).
11
Finally, we consider the case p = ∞.
min kδk∞
(20)
δ∈Rd
sbj. to: fj (x) − fc (x) ≥ h∇fc (x) − ∇fj (x), δi
0 ≤ xj + δj ≤ 1
Lemma 4.3. Let m = arg max fj (x) and define v = ∇fm (x) − ∇fj (x) and c = fj (x) −
j
fm (x) < 0, then the solution of (20) can be found by solving for t ≥ 0,
X
X
vi max{−t, −xr } +
vi min{t, 1 − xr } = c.
vi >0
vi <0
which is done in Algorithm 3.
Proof. We can rewrite the optimization problem into a linear program
min
t∈R,δ∈Rd
t
(21)
sbj. to: c ≥ hv, δi
− xj ≤ δj ≤ 1 − xj , j = 1, . . . , d
− t ≤ δj ≤ t, j = 1, . . . , d
t≥0
(22)
Thus we have max{−t, −xr } ≤ δr ≤ min{t, 1 − xr } for r = 1, . . . , d. Then it holds
X
X
hv, δi ≥
vr max{−t, −xr } +
vr min{t, 1 − xr }
vr >0
=−
vr <0
X
vr >0,t≥xr
vr x r − t
X
vr + t
vr >0,t<xr
X
vr <0,t<1−xr
vr +
X
vr (1 − xr ),
vr <0,t≥1−xr
and the lower bound can be attained if δr = min{t, 1−xr } for vr < 0 and δr = max{−t, −xr }
for vr > 0, δr = 0 if vr = 0. Note that both terms are monotonically decreasing with t. The
algorithm 3 has complexity O(d log d) due to the initial sorting step followed by steps of
complexity O(d).
For nonlinear classifiers a change of the decision is not guaranteed and thus we use later on
a binary search with a variable c instead of fc (x) − fj (x).
5
Experiments
The goal of the experiments is the evaluation of the robustness of the resulting classifiers
and not necessarily state-of-the-art results in terms of test error. In all cases we compute the
robustness guarantees from Theorem 2.1 (lower bound on the norm of the minimal change
required to change the classifier decision), where we optimize over R using binary search,
and adversarial samples with the algorithm for the 2-norm from Section 4 (upper bound
on the norm of the minimal change required to change the classifier decision), where we do
a binary search in the classifier output difference in order to find a point on the decision
boundary. Additional experiments can be found in the supplementary material.
Kernel methods: We optimize the cross-entropy loss once with the standard regularization
(Kernel-LogReg) and with Cross-Lipschitz regularization (Kernel-CL). Both are convex
optimization problems and we use L-BFGS to solve them. We use the Gaussian kernel
2
k(x, y) = e−γkx−yk where γ = ρ2 α
and ρKNN40 is the mean of the 40 nearest neighbor
KNN40
distances on the training set and α ∈ {0.5, 1, 2, 4}. We show the results for MNIST (60000
training and 10000 test samples). However, we have checked that parameter selection
using a subset of 50000 images from the training set and evaluating on the rest yields
indeed the parameters which give the best test errors when trained on the full set. The
12
Algorithm 3 Computation of box-constrained adversarial samples wrt to k·k∞ -norm.
INPUT: c = fj (x) − fm (x) and v = ∇fm (x) − ∇fj (x) (j, desired class, m original class)
d+ := |{l | vl > 0}| and d− := |{l | vl < 0}|
sort {xl | vl > 0} in increasingP
order π, sort {1P
− xl | vl < 0} in increasing order ρ
s:=1, r:=1, g:=0, t:=0, κ+ = vl >0 vl , κ− = vl <0 vl , γ+ = γ− = 0.
while g > c AND (s ≤ d+ OR t ≤ d− ) do
if xπs < 1 − xρr then
t = xπs , κ+ ← κ+ − vπs , γ+ ← γ+ − vπs xπs ,
s←s+1
else
t = 1 − xρr , κ− ← κ− − vρr , γ− ← γ− − vρr (1 − xρr ),
t←t+1
end if
g = γ+ + γ− + t (κ+ − κ− )
end while
if g ≤ c then
undo last step
t = (c − γ+ − γ− )/(κ− − κ+ )
min{t, 1 − xr } for vr > 0
compute δr =
max{−t, −xr } for vr < 0
else
Problem has no feasible solution
end if
regularization parameter is chosen in λ ∈ {10−k |k ∈ {5, 6, 7, 8}} for Kernel-SVM and
λ ∈ {10−k | k ∈ {0, 1, 2, 3}} for our Kernel-CL. The results of the optimal parameters are
given in the following table and the performance of all parameters is shown in Figure 1. Note
that due to the high computational complexity we could evaluate the robustness guarantees
only for the optimal parameters.
No Reg.
(λ = 0)
K-SVM
K-CL
test
error
2.23%
avg. k·k2
adv.
samples
2.39
avg.k·k2
rob.
guar.
0.037
1.48%
1.44%
1.91
3.12
0.058
0.045
Figure 1: Kernel Methods: Cross-Lipschitz regularization achieves both better test error and robustness against
adversarial samples (upper bounds, larger is better) compared to the standard regularization. The robustness
guarantee is weaker than for neural networks but this is most likely due to the relatively loose bound.
Neural Networks: Before we demonstrate how upper and lower bounds improve using
cross-Lipschitz regularization, we first want to highlight the importance of the usage of the
local cross-Lipschitz constant in Theorem 2.1 for our robustness guarantee.
Local versus global Cross-Lipschitz constant: While no robustness guarantee has
been proven before, it has been discussed in [24] that penalization of the global Lipschitz
constant should improve robustness, see also [4]. For that purpose they derive the Lipschitz
constants of several different layers and use the fact that the Lipschitz constant of a
composition of functions is upper bounded by the product of the Lipschitz constants of
the functions. In analogy, this would mean that the term supy∈B(x,R) k∇fc (y) − ∇fj (y)k2 ,
which we have upper bounded in Proposition 2.2, in the denominator in Theorem 2.1 could
be replaced2 by the global Lipschitz constant of g(x) := fc (x) − fj (x). which is given as
supy∈Rd k∇g(x)k2 = supx6=y |g(x)−g(y)|
. We have with kU k2,2 being the largest singular value
kx−yk
2
2
Note that then the optimization of R in Theorem 2.1 would be unnecessary.
13
MNIST (plain)
CIFAR10 (plain)
None Dropout Weight Dec. Cross Lip. None Dropout Weight Dec. Cross Lip.
0.69
0.48
0.68
0.21
0.22
0.13
0.24
0.17
α
Table 1: We show the average ratio αglobal of the robustness guarantees αglobal , αlocal from Theorem 2.1 on
local
the test data for MNIST and CIFAR10 and different regularizers. The guarantees using the local Cross-Lipschitz
constant are up to eight times better than with the global one.
of U ,
|g(x) − g(y)| = hwc − wj , σ(U x) − σ(U y)i ≤ kwc − wj k2 kσ(U x) − σ(U y)k2
≤ kwc − wj k2 kU (x − y)k2 ≤ kwc − wj k2 kU k2,2 kx − yk2 ,
where we used that σ is contractive as σ 0 (z) =
1
1+e−αz
and thus we get
sup k∇fc (x) − ∇fj (x)k2 ≤ kwc − wj k2 kU k2,2 .
y∈Rd
The advantage is clearly that this global Cross-Lipschitz constant can just be computed
once and by using it in Theorem 2.1 one can evaluate the guarantees very quickly. However,
it turns out that one gets significantly better robustness guarantees by using the local
Cross-Lipschitz constant in terms of the bound derived in Proposition 2.2 instead of the just
derived global Lipschitz constant. Note that the optimization over R in Theorem 2.1 is done
using a binary search, noting that the bound of the local Lipschitz constant in Proposition
2.2 is monotonically decreasing in R. We have the following comparison in Table 1. We
want to highlight that the robustness guarantee with the global Cross-Lipschitz constant
was always worse than when using the local Cross-Lipschitz constant across all regularizers
and data sets. Table 1 shows that the guarantees using the local Cross-Lipschitz can be up
to eight times better than for the global one. As these are just one hidden layer networks, it
is obvious that robustness guarantees for deep neural networks based on the global Lipschitz
constants will be too coarse to be useful.
Experiments: We use a one hidden layer network with 1024 hidden units and the softplus
activation function with α = 10. Thus the resulting classifier is continuously differentiable.
We compare three different regularization techniques: weight decay, dropout and our CrossLipschitz regularization. Training is done with SGD. For each method we have adapted the
learning rate (two per method) and regularization parameters (4 per method) so that all
methods achieve good performance. We do experiments for MNIST and CIFAR10 in three
settings: plain, data augmentation and adversarial training. The exact settings of the parameters and the augmentation techniques are described below.The results for MNIST are shown
in Figure 2 and the results for CIFAR10 are in Figure 3.For MNIST there is a clear trend that
our Cross-Lipschitz regularization improves the robustness of the resulting classifier while
having competitive resp. better test error. It is surprising that data augmentation does not
lead to more robust models. However, adversarial training improves the guarantees as well as
adversarial resistance. For CIFAR10 the picture is mixed, our CL-Regularization performs
well for the augmented task in test error and upper bounds but is not significantly better in
the robustness guarantees. The problem might be that the overall bad performance due to the
simple model is preventing a better behavior. Data augmentation leads to better test error
but the robustness properties (upper and lower bounds) are basically unchanged. Adversarial
training slightly improves performance compared to the plain setting and improves upper
and lower bounds in terms of robustness. We want to highlight that our guarantees (lower
bounds) and the upper bounds from the adversarial samples are not too far away.
For MNIST (all settings) the learning rate is for all methods chosen from {0.2, 0.5}.
The regularization parameters for weight decay are chosen from {10−5 , 10−4 , 10−3 , 10−2 },
for Cross-Lipschitz from {10−5 , 10−4 , 5 ∗ 10−4 , 10−3 } and the dropout probabilities are
taken from {0.4, 0.5, 0.6, 0.7}. For CIFAR10 the learning rate is for all methods chosen
from {0.04, 0.1}, the regularization parameters for weight decay and Cross-Lipschitz are
{10−5 , 10−4 , 5 ∗ 10−4 , 10−3 } and dropout probabilities are taken from {0.5, 0.6, 0.7, 0.8}.
For CIFAR10 with data augmentation we choose the learning rate for all methods from
{0.04, 0.1}, the regularization parameters for weight decay are {10−6 , 10−5 , 10−4 , 10−3 } and
14
Adversarial Resistance (Upper Bound)
wrt to L2 -norm
Robustness Guarantee (Lower Bound)
wrt to L2 -norm
Figure 2: Neural Networks, Left: Adversarial resistance wrt to L2 -norm on MNIST. Right: Average robustness guarantee wrt to L2 -norm on MNIST for different neural networks (one hidden layer, 1024 HU) and
hyperparameters. The Cross-Lipschitz regularization leads to better robustness with similar or better prediction
performance. Top row: plain MNIST, Middle: Data Augmentation, Bottom: Adv. Training
for Cross-Lipschitz {10−5 , 10−4 , 5 ∗ 10−4 , 10−3 } and the dropout probabilities are taken from
{0.5, 0.6, 0.7, 0.8}. Data augmentation for MNIST means that we apply random rotations in
π π
the angle [− 20
, 20 ] and random crop from 28x28 to 24x24. For CIFAR-10 we apply the same
and additionally we mirror the image (left to right) with probability 0.5 and apply random
brightness [−0.1, 0.1] and random contrast change [0.6, 1.4]. In each substep we ensure that
we get an image in [0, 1]d by clipping. We implemented adversarial training by generating
adversarial samples wrt to the infinity norm with the code from Section 4 and replaced 50%
of each batch as adversarial samples. Finally, we use for SGD batchsize 64 in all experiments.
Illustration of adversarial samples: we take one test image from MNIST and apply the
adversarial generation from Section 4 wrt to the 2-norm to generate the adversarial samples for
the different kernel methods and neural networks (plain setting), where we use for each method
the parameters leading to best test performance. All classifiers change their originally correct
decision to a “wrong” one. It is interesting to note that for Cross-Lipschitz regularization
(both kernel method and neural network) the “adversarial” sample is really at the decision
boundary between 1 and 8 (as predicted) and thus the new decision is actually correct.
This effect is strongest for our Kernel-CL, which also requires the strongest modification to
generate the adversarial sample. The situation is different for neural networks, where the
classifiers obtained from the two standard regularization techniques are still vulnerable, as
the adversarial sample is still clearly a 1 for dropout and weight decay. We show further
examples below.
15
Adversarial Resistance (Upper Bound)
wrt to L2 -norm
Robustness Guarantee (Lower Bound)
wrt to L2 -norm
Figure 3: Left: Adversarial resistance wrt to L2 -norm on test set of CIFAR10. Right: Average robustness
guarantee on the test set wrt to L2 -norm for the test set of CIFAR10 for different neural networks (one hidden
layer, 1024 HU) and hyperparameters. While Cross-Lipschitz regularization yields good test errors, the guarantees
are not necessarily stronger. Top row: CIFAR10 (plain), Middle: CIFAR10 trained with data augmentation,
Bottom: Adversarial Training.
Original, Class 6
K-SVM, Pred:0, kδk2 = 3.0
K-CL, Pred:0, kδk2 = 4.7
NN-WD, Pred:4, kδk2 = 1.4
NN-DO, Pred:4, kδk2 = 1.9
NN-CL, Pred:0, kδk2 = 3.2
Original, Class 1
K-SVM, Pred:7, kδk2 = 1.2
K-CL, Pred:8, kδk2 = 3.5
16
NN-WD, Pred:8, kδk2 = 1.2 NN-DO, Pred:7, kδk2 = 1.1 NN-CL, Pred:8, kδk2 = 2.6
Figure 4: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for Cross-Lipschitz regularization this new decision
makes (often) sense, whereas for the neural network models (weight decay/dropout) the change is so small that
Figure 5: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes sense,
whereas for all neural network models the change is so small that the new decision is clearly wrong.
Original, Class 4
K-SVM, Pred:9, kδk2 = 1.4
K-CL, Pred:9, kδk2 = 2.2
NN-WD, Pred:9, kδk2 = 1.3
NN-DO, Pred:9, kδk2 = 1.5
NN-CL, Pred:9, kδk2 = 2.2
Figure 6: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes sense,
whereas for all neural network models the change is so small that the new decision is clearly wrong.
Original, Class 2
K-SVM, Pred:3, kδk2 = 2.5
K-CL, Pred:3, kδk2 = 4.4
NN-WD, Pred:3, kδk2 = 1.1
NN-DO, Pred:3, kδk2 = 1.4
NN-CL, Pred:3, kδk2 = 2.0
Figure 7: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes sense,
whereas for all neural network models the change is so small that the new decision is clearly wrong.
Original, Class 8
K-SVM, Pred:3, kδk2 = 2.2
K-CL, Pred:5, kδk2 = 4.2
NN-WD, Pred:3, kδk2 = 1.4
NN-DO, Pred:5, kδk2 = 1.6
NN-CL, Pred:3, kδk2 = 2.8
Figure 8: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes sense,
whereas for all neural network models the change is so small that the new decision is clearly wrong.
17
Original, Class 1
K-SVM, Pred:8, kδk2 = 1.2
K-CL, Pred:2, kδk2 = 3.7
NN-WD, Pred:2, kδk2 = 1.1
NN-DO, Pred:2, kδk2 = 1.1
NN-CL, Pred:8, kδk2 = 2.7
Figure 9: Top left: original test image, for each classifier we generate the corresponding adversarial sample which
changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes sense,
whereas for all neural network models the change is so small that the new decision is clearly wrong.
Original, Class 3
K-SVM, Pred:8, kδk2 = 2.1
K-CL, Pred:8, kδk2 = 3.3
NN-WD, Pred:8, kδk2 = 1.7
NN-DO, Pred:8, kδk2 = 1.4
NN-CL, Pred:5, kδk2 = 3.2
Figure 10: Top left: original test image, for each classifier we generate the corresponding adversarial sample
which changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes
sense, whereas for all neural network models the change is so small that the new decision is clearly wrong.
Original, Class 8
K-SVM, Pred:9, kδk2 = 2.1
K-CL, Pred:5, kδk2 = 2.6
NN-WD, Pred:9, kδk2 = 1.4
NN-DO, Pred:9, kδk2 = 1.8
NN-CL, Pred:9, kδk2 = 1.8
Figure 11: Top left: original test image, for each classifier we generate the corresponding adversarial sample
which changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes
sense, whereas for all neural network models the change is so small that the new decision is clearly wrong.
18
Adversarial Resistance (Upper Bound)
wrt to L2 -norm
Robustness Guarantee (Lower Bound)
wrt to L2 -norm
Figure 13: Left: Adversarial resistance wrt to L2 -norm on test set of the german traffic sign benchmark (GTSB)
in the plain setting. Right: Average robustness guarantee on the test set wrt to L2 -norm for the test set of GTSB
for different neural networks (one hidden layer, 1024 HU) and hyperparameters. Here dropout performs very well
both in terms of performance and robustness.
Original, Class 5
K-SVM, Pred:9, kδk2 = 1.5
K-CL, Pred:9, kδk2 = 2.2
NN-WD, Pred:9, kδk2 = 1.1
NN-DO, Pred:9, kδk2 = 1.0
NN-CL, Pred:9, kδk2 = 1.4
Figure 12: Top left: original test image, for each classifier we generate the corresponding adversarial sample
which changes the classifier decision (denoted as Pred). Note that for the kernel methods this new decision makes
sense, whereas for all neural network models the change is so small that the new decision is clearly wrong.
German Traffic Sign Benchmark: As a third dataset we used the German Traffic Sign
Benchmark (GTSB) [23], which consists of images of german traffic signs, which has 43
classes with 34209 training and 12630 test samples. The results are shown in Figure 13. For
this dataset Cross-Lipschitz regularization improves the upper bounds compared to weight
decay but dropout achieves significantly better prediction performance and has similar upper
bounds. The robustness guarantees for weight decay and Cross-Lipschitz are slightly better
than for dropout.
Residual Networks: All experiments so far were done with one hidden layer neural
networks so that we can evaluate lower and upper bounds. Now we want to demonstrate
that Cross-Lipschitz regularization can also successfully be used for deep networks. We
use residual networks proposed in [9] with 32 parameter layers and non-bottleneck residual
blocks. We follow basically their setting, apart from that we did not subtract the per-pixel
mean so that all images are in [0, 1]d and use random crop but without any padding as in [9].
Similar to [9], we train for 160 epochs, and the learning rate is divided by 10 on the 115-th
and 140-th epochs. For the experiments with dropout we followed the recommendation of
[25], inserting a dropout layer between convolutional layers inside each residual block. For
Cross-Lipschitz regularization we use automatic differentiation in TensorFlow [1] to calculate
the derivative with respect to the input, which slows done the training by a factor of 10.
For the plain setting the learning rate for all methods is chosen from {0.2, 0.5}, except for the
runs without regularization, for which it is from {0.08, 0.1, 0.2, 0.4, 0.6, 0.8}. For weight decay
the regularization parameter is chosen from {10−5 , 10−4 , 10−3 , 10−2 }, for Cross-Lipschitz
from {10−4 , 10−3 , 10−2 , 10−1 }, and for dropout the probabilities are from {0.5, 0.6, 0.7, 0.8}.
For the data augmentation setting the only difference was in the higher learning rates: no
19
Adversarial Resistance (Upper Bound)
wrt to L2 -norm (ResNets)
Adversarial Resistance (Upper Bound)
wrt to L2 -norm (ResNets)
Figure 14: Results on CIFAR10 for a residual network with different regularizers. As we only have lower
bounds for one hidden layer networks, we can only show upper bounds for adversarial resistance. Left: with data
augmentation similar to [25] Right: plain setting
regularization - {0.2, 0.5, 0.8, 1.0, 1.5, 2.0, 3.0, 4.0}, weight decay - {0.1, 0.4}, Cross-Lipschitz
- {0.2, 1.0}. The results are shown in Figure 14. Cross-Lipschitz regularization improves
the upper bounds on the robustness against adversarial manipulation compare to weight
decay and dropout by a factor of 2 to 3 both in the plain setting (right) and with data
augmentation (left). This comes at a price of a slightly worse test performance. However,
it shows that Cross-Lipschitz regularization is also effective for deep neural networks. It
remains interesting future work to come up also with interesting instance-specific lower
bounds (robustness guarantees) for deep neural networks.
Outlook Formal guarantees on machine learning systems are becoming increasingly more
important as they are used in safety-critical systems. We think that there should be more
research on robustness guarantees (lower bounds), whereas current research is focused on
new attacks (upper bounds). We have argued that our instance-specific guarantees using our
local Cross-Lipschitz constant is more effective than using a global one and leads to lower
bounds which are up to 8 times better. A major open problem is to come up with tight
lower bounds for deep networks.
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis,
J. Dean, M. Devin, S. Ghemawat, I. J. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia,
R. Józefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. G.
Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. A. Tucker,
V. Vanhoucke, V. Vasudevan, F. B. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke,
Y. Yu, and X. Zheng. Tensorflow: Large-scale machine learning on heterogeneous distributed
systems, 2016.
[2] O. Bastani, Y. Ioannou, L. Lampropoulos, D. Vytiniotis, A. Nori, and A. Criminisi. Measuring
neural net robustness with constraints. In NIPS, 2016.
[3] N. Carlini and D. Wagner. Adversarial examples are not easily detected: Bypassing ten
detection methods. In ACM Workshop on Artificial Intelligence and Security, 2017.
[4] M. Cisse, P. Bojanowksi, E. Grave, Y. Dauphin, and N. Usunier. Parseval networks: Improving
robustness to adversarial examples. In ICML, 2017.
[5] N. Dalvi, P. Domingos, Mausam, S. Sanghai, and D. Verma. Adversarial classification. In KDD,
2004.
[6] H. Drucker and Y. Le Cun. Double backpropagation increasing generalization performance. In
IJCNN, 1992.
[7] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples.
In ICLR, 2015.
[8] S. Gu and L. Rigazio. Towards deep neural network architectures robust to adversarial examples.
In ICLR Workshop, 2015.
20
[9] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR,
pages 770–778, 2016.
[10] D. P. Helmbold and P. Long. On the inductive bias of dropout. Journal of Machine Learning
Research, 16:3403–3454, 2015.
[11] S. Hochreiter and J. Schmidhuber. Simplifying neural nets by discovering flat minima. In NIPS,
1995.
[12] R. Huang, B. Xu, D. Schuurmans, and C. Szepesvari. Learning with a strong adversary. In
ICLR, 2016.
[13] J. Kos, I. Fischer, and D. Song. Adversarial examples for generative models. In ICLR Workshop,
2017.
[14] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In
ICLR Workshop, 2017.
[15] Y. Liu, X. Chen, C. Liu, and D. Song. Delving into transferable adversarial examples and
black-box attacks. In ICLR, 2017.
[16] D. Lowd and C. Meek. Adversarial learning. In KDD, 2005.
[17] S.M. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard. Universal adversarial perturbations.
In CVPR, 2017.
[18] N. Papernot, P. McDonald, X. Wu, S. Jha, and A. Swami. Distillation as a defense to adversarial
perturbations against deep networks. In IEEE Symposium on Security & Privacy, 2016.
[19] P. Frossard S.-M. Moosavi-Dezfooli, A. Fawzi. Deepfool: a simple and accurate method to fool
deep neural networks. In CVPR, pages 2574–2582, 2016.
[20] B. Schölkopf and A. J. Smola. Learning with Kernels. MIT Press, Cambridge, MA, 2002.
[21] U. Shaham, Y. Yamada, and S. Negahban. Understanding adversarial training: Increasing local
stability of neural nets through robust optimization. In NIPS, 2016.
[22] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A
simple way to prevent neural networks from overfitting. Journal of Machine Learning Research,
15:1929–1958, 2014.
[23] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine
learning algorithms for traffic sign recognition. Neural Networks, 32:323–332, 2012.
[24] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus.
Intriguing properties of neural networks. In ICLR, pages 2503–2511, 2014.
[25] S. Zagoruyko and N. Komodakis. Wide residual networks. In BMVC, pages 87.1–87.12.
[26] S. Zheng, Y. Song, T. Leung, and I. J. Goodfellow. Improving the robustness of deep neural
networks via stability training. In CVPR, 2016.
21
| 1cs.CV
|
Classification Accuracy Improvement for Neuromorphic Computing
Systems with One-level Precision Synapses
arXiv:1701.01791v1 [cs.NE] 7 Jan 2017
Yandan Wang, Wei Wen, Linghao Song, and Hai (Helen) Li
Dept. of Electrical & Computer Engineering, University of Pittsburgh, Pittsburgh, PA, USA
{yaw46,wew57,linghao.song,hal66}@pitt.edu
method for IBM TrueNorth platform which biases the
learned connection probability to binary states (0/1) to
hinder accuracy loss [5]. Neural networks with binary resolution are more suitable for generic platforms [6][7][8].
BinaryConnect [7] as an example can achieve comparable accuracy in deep neural networks. However, neither
TrueNorth nor BinaryConnect are pure binary neural networks: TrueNorth relies on the ensemble averaging layer
in floating-point precision while the last layer of BinaryConnect is a floating-point L2-SVM.
In this work, we focus on the pure binary (1-level precision1 ) neural networks. While the realization of continuous analogue resistance states is still challenging, the
1-level precision is well supported by most of memory
materials and architectures. Three orthogonal methods
of leaning 1-level precision synapses and tuning bias to
improve image classification accuracy are proposed:
Abstract—Brain inspired neuromorphic computing
has demonstrated remarkable advantages over traditional von Neumann architecture for its high energy
efficiency and parallel data processing. However, the
limited resolution of synaptic weights degrades system accuracy and thus impedes the use of neuromorphic systems. In this work, we propose three orthogonal methods to learn synapses with one-level precision, namely, distribution-aware quantization, quantization regularization and bias tuning, to make image
classification accuracy comparable to the state-of-theart. Experiments on both multi-layer perception and
convolutional neural networks show that the accuracy
drop can be well controlled within 0.19% (5.53%) for
MNIST (CIFAR-10) database, compared to an ideal
system without quantization.
• Distribution-aware quantization discretizes weights
in different layers to different values. The method
is proposed based on the observation that the weight
distributions of a network by layers.
• Quantization regularization directly learns a network
with discrete weights during training process. The
regularization can reduce the distance between a
weight and its nearest quantization level with a constant gradient.
• Bias tuning dynamically learns the best bias compensation to minimize the impact of quantization. It
can also alleviate the impact of synaptic variation in
memristor based neuromorphic systems.
I. Introduction
In recent years, brain-inspired neuromorphic computing systems have been extensively studied. For example,
IBM TrueNorth has demonstrated many important features including high computing efficiency, extremely low
power consumption, and compact volume [1]. Integrating
emerging technologies potentially enables a more compact and energy-efficient platform for information processing [2]. For instance, the two-terminal nonlinear memristor presents a series of advantages of good scalability, high endurance and ultra-low power consumption [3].
Thus it is taken as a promising candidate for neuromorphic computing system development.
Neuromorphic hardware implementations usually face
a major challenge on system accuracy. TrueNorth, for example, allows only a few synaptic weights (e.g., 0, ±1, ±2).
Accuracy degradation is inevitable when directly deploying a learned model to the system with limited precision [1]. The situation remains in memristor (or RRAM)
based design. Theoretically, nanoscale memristor can obtain continuously analog resistance. While, a real device
often can achieve only several stable resistance states [4].
The distinction between theoretical and actual properties
results in significant accuracy loss.
Extensive studies on learning low-resolution synapses
have been performed to improve the accuracy of neuromorphic systems. Wen et al. presented a new learning
II. Preliminary
A. Neural Network Models
Neural networks (NNs) are a series of models inspired
by biological neuron networks. The function can be formulated as:
y = W · x + b and z = h(y),
(1)
where the output neuron vector z is determined by the
input neuron vector x, the weight matrix of connections
1 We define n-level quantization as quantifying each weight to
the nearest value of zero, n positive and n negative values with the
same absolute values.
1
1.E+09
Array: W
Memristance(Ω)
1.E+08
1.E+07
LRS
HRS
1.E+06
1.E+05
Output: y
1.E+04
1.E+00
5.E+03
1.E+04
Cycle
Fig. 1. Statistical memristance distributions of a T iO2 device.
W and the bias vector b. Usually, h(·) is a non-linear
activation function and all the data in (1) are in floatingpoint precision.
B. Memristor Technology
Memristor, firstly introduced by Professor Leon Chua
in 1971, is regarded as the fourth fundamental circuit element, representing the dynamic relationship between the
charge q(t) and the flux ϕ(t) [9]. Most significantly, the
total electric flux flowing through a memristor device can
be “remembered” by recording it as its memristance (M ).
In 2008, HP Lab demonstrated the first actual memristor
through a T iO2 thin-film device and realized the memristive property by moving its doping front [10].
Theoretically, a memristor device can achieve continuous analog resistance states. However, the imperfection of
fabrication process causes variations and therefore memristance varies from device to device. Even worse, the
memristance of a single memristor changes from time to
time [11]. In most system designs, only two stable resistance states, high- and low-resistance state (HRS and
LRS), are adopted. As the real statistical measurement
data of a T iO2 memristor in Fig. 1 shows, the distribution of HRS (LRS) follows an approximated lognormal
probability density function (PDF) [4].
C. Neuromorphic Computing Systems
Neuromorphic computing systems (NCS) represents
the hardware implementations of NNs by mimicking
the neuro-biological architectures. For example, IBM
TrueNorth chip is made of a network of neuro-synaptic
cores, each of which includes a configurable synaptic crossbar connecting 256 axons and 256 neurons in close proximity [1]. The synaptic weight in the crossbar can be selected
from 4 possible integers. Memristor based NCS has also
be investigated [12]. Matrix-vector multiplication, the key
operation in NNs, can be realized by memristor crossbar
arrays as illustrated in Fig. 2 [13]. The conductance matrix of memristor crossbar array is utilized as the weight
matrix of NNs [12].
The synaptic weights in these neuromorphic computing
systems usually have a limited precision, constrained either by design cost (e.g., the SRAM cells for each weight
representation in TrueNorth) or current technology process (e.g., two or only a few resistance levels of memristor
devices). As such, the classification accuracy loss could
...
Fig. 2. Mapping neural networks to memristor crossbar array.
be very significant in NCS. To improve the classification
accuracy, lots of research has been done [7][8][14]. Even
though, some of them have floating-point layers and some
ignore circuit design constraints. In this work, we focus on
pure binary neural networks considering the constraints
in NCS hardware implementation.
III. Methodology
This paper aims at improving the classification accuracy of pure binary neural networks in all layers. Such
neural networks can be naturally implemented on NCS,
such as TrueNorth chip and memristor based design.
Three novel classification accuracy improving methods are
proposed in the work, namely, distribution-aware quantization, quantization regularization and bias tuning. The
implementation of convolutional neural network (CNN)
convolution operation in memristor crossbar array and
a crossbar variation demo for accuracy improvement are
also presented.
To explain our methodologies, in this section, we take
LeNet [15] as the example of CNN trained on MNIST – a
28×28 handwritten digit database [16]. Experiments and
analysis on more neural networks and databases shall be
presented in Section IV.
A. Distribution-aware Quantization
In training of neural networks, `2 -norm regularization
is commonly adopted to avoid over-fitting. With `2 -norm
regularization, the final distribution of learned weights
in a layer approximately follows the normal distribution
[17]. A naive quantization method in implementation is
to quantify all weights to the same group of level selection. However, as shown in Fig. 3 (blue bars) by taking
LeNet as an example, the weight distribution varies from
layer to layer: The first convolutional layer (conv1) has
the most scattered distribution with a wider range scope,
while the weights of second convolutional layer (conv2)
and two fully connected layers (ip1, ip2) have concentrated to a relatively narrow scope. The data implies
that a quantization optimized for one layer may result in
a large information loss in another layer.
Here, we propose a heuristic method – distributionaware quantization (DQ) which discretizes weights in different layers to different values. In memristor-based NCS,
this can be realized by programming the resistance states
of each crossbar to different values [12]. Our experiments
on LeNet show that when applying the aforementioned
The error norms
0.3
0.3
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
-3
l1-norm
l2-norm
proposed-norm
-2
-1
0
1
2
3
x
0.3
0.3
Fig. 3. The blue and orange bars denote the original weight
distribution of different layers and the learned discrete weights
after quantization regularization (QR) in LeNet, respectively.
naive method, the test accuracy of 1-level quantization
quickly drops from 99.15% to 90.77%, while our proposed
distribution-aware quantization can still achieve 98.31%
accuracy. Note that without explicit mention, the quantization levels are selected by cross-validation [18].
B. Quantization Regularization
Distribution-aware quantization separates the training
and quantifying processes and therefore it cannot avoid
the accuracy loss once the quantization is completed. To
further improve system performance, we propose quantization regularization (QR) which directly learns a neural
network with discrete weights.
During the training of a network, a regularization term
can be added to the error function to control the distribution of weights and avoid overfitting . For example,
`2 -norm regularization can learn weights with normal distribution and `1 -norm is commonly utilized to learn sparse
networks [17]. The total error function to be minimized
with a generic regularization term can be formulated as
E(W ) = ED (W ) + λ · EW (W ),
(2)
where λ is the coefficient controlling the importance between data-dependent error ED (W ) and regularization
term EW (W ). W is the set of all weights in neural networks. We propose a new quantization regularization as
q
EW
(W ) = sgn (Wk − Q(Wk )) · (Wk − Q(Wk )) ,
(3)
where Wk is the k -th weight, Q(Wk ) is the quantization
value nearest to Wk and sgn(·) is the sign function. After
forwarding and back propagation, the weight updating
with learning rate η can be formulated as:
∂ED (W )
− η · sgn(Wk − Q(Wk )).
Wk ← Wk − η ·
∂Wk
(4)
Through the third term on the right side of (4), our
regularization descents (reduces) the distance between a
weight and its nearest quantization level with a constant
gradient (±1). Compared with the `1 -norm and `2 -norm
regularization, our proposed regularization method can
quantify learning weights to the desired discrete values
more precisely, meanwhile properly control the weight distribution and overfitting.
Fig. 4. Comparison of `1 -norm, `2 -norm and our proposed
regularization.
Fig. 4 demonstrates and compares the three regularization methods. Zero is one of the targeted quantification
values in this work, which is usually realized through `1 norm based neural network sparsification. In addition,
our proposed method include more discrete quantification values. Orange bars in Fig. 3 correspond to the new
weight distribution of LeNet after applying QR, indicating
our method can efficiently learn weights around quantization levels. Compared with the naive 1-level quantization,
including QR only can improve accuracy 6.21%. Combining with DQ, the accuracy drop from the ideal case is
controlled within only 0.20% with 1-level quantization.
More experiments will be discussed in section IV.
C. Bias Tuning
The quantization of weights deviating the information
can be formulated as
yj + ∆yj =
X
(Wji + ∆Wji ) · xi + bj ,
(5)
i
where Wji is the weight connecting the i -th neuron in the
previous
Player to the j -th neuron in this layer. ∆Wji and
∆yj = i ∆Wji · xi are the deviation of weight and input
of activation function, respectively, resulted from quantization. The deviation ∆yj propagates through layers
toward the output classifier neurons and deteriorates the
classification accuracy.
In circuit design of neuron model, the bias usually is an
adjustable parameter, e.g. the fire threshold in TrueNorth
neuron model works as bias. Therefore, to compensate
the deviation, we may adjust the neuron bias from bj to
bj + ∆bj such that
∆bj = −∆yj = −
X
∆Wji · xi .
(6)
i
As such, the neuron activation can remain the original
value before quantization. Unfortunately, the input xi
varies randomly with the input samples (e.g., images) and
a unique bias compensation ∆bj cannot be identified.
We propose bias tuning (BT) which learns the optimal bias compensation to minimize the impact of quantization. Fig. 5 shows the framework of the bias tuning:
first, both weights and biases are trained without quantization; second, weights are quantified and programmed
into NCS; third, weights are frozen and biases are learned
to improve classification accuray; and finally, the tuned
Weight and
bias training
Weight
quantification
Deployment
in NCS
Weight freezing
and bias learning
Memristor
programming
TABLE I
Network and dataset
Bias
programming
Fig. 5. The framework of proposed bias tuning method.
biases are programmed into NCS. Impressively, bias tuning method can achieve 7.89% classification improvement
compared to the naive 1-level quantization baseline on
LeNet. Combining with the above DQ and QR methods,
the total accuracy drop can be reduced to merely 0.19%.
D. Convolution in Memristor Crossbar Array
The memristor crossbar structure can be naturally
mapped to fully connected layers. Here, we extend its
use to convolution layers. A pixel value (y) in a post
feature map is computed by
y=
X
Fk · wk + b,
(7)
k
where wk is the k -th weight in the filter and Fk is the
corresponding input feature. Because the essence of convolution is multiplication-accumulation, we can employ
memristor crossbar array to compute.
Fig. 6 shows an example to compute the convolution
of a 5-by-5 feature map with a 3-by-3 filter. At the time
stamp t0, the green elements are converted to a vector and
sent into a memristor array through word lines. And at
t1, the pink elements are processed similarly to the green
ones. As the filter shifts, the corresponding features in the
previous layer are sent into the crossbar in a time-division
sequence, such that the output features are computed by
the bit line (blue) whose weights belong to the filter. As
shown in the figure, each bitline is mapped to one filter in
the convolutional layer. We note that the proposed DQ,
DR and BT methods also work for weights in CNN.
IV. Experiments
A. Experiment Setup
To evaluate the effectiveness of proposed methods,
we conducted three experiments using multilayer perception (MLP) and CNN neural network structures on two
datasets: MNIST and CIFAR-10 (a 32×32 color image
database). The first two experiments are both conducted
C0 C1C2
…
…
…
t0
t1
…
t1 t0
Fig. 6. Convolution implementation in memristor crossbar array.
Dataset
Input
Conv1
Conv2
Conv3
Ip1
Ip2
Ip3
§ 20×5×5
Network 1
MNIST
28×28
−
−
−
784×500
500×300
300×10
Network 2
MNIST
28×28
20×5×5§
50×5×5
−
800×500
500×10
−
Network 3
CIFAR-10
32×32
32×5×5
32×5×5
64×5×5
1024×10
−
−
means 20 filters with each filter size 5×5.
on MNIST dataset using a MLP and a CNN network, respectively. The third experiment is conducted on CIFAR10 dataset using a CNN network. The adopted deep leaning framework is Caffe developed by the Berkeley Vision
and Learning Center (BVLC) and community contributors [19]. Detailed network parameters and dataset are
summarized in Table I.
B. Function Validation of MLP on MNIST
Network 1 is a MLP network with a size of 784 × 500 ×
300 × 10, which can’t be directly implemented in NCS.
Previously, we presented the hardware implementation of
mapping a large network to small crossbar arrays [13].
Here, 784 corresponds to the 28×28 MNIST image input
pattern; 500 and 300 are the neuron numbers of the first
and second hidden layers, respectively; and 10 is the final
classification outputs.
The baseline is set as the highest accuracy (all the layers
quantified to 0.06) of all naive 1-level quantization situations without applying any proposed method. To explore
the effectiveness of each single method and their combination situations, we conducted 8 separate experiments
with combinations, the experiment results of which are
summarized in Table II.
Compared with the baseline accuracy, there is a large
accuracy increase when applied only one of three accuracy improvement methods (1.52%, 1.26%, 0.4%, respectively). Applying any two of three methods will make the
accuracy further increased. Combining all three methods together can achieve a highest accuracy with only
0.39% accuracy drop compared with the ideal value without any quantization. We note that, in some cases (e.g.
TABLE II
The accuracy measurement for MLP on MNIST dataset
DQ
Ideal§
0 (Baseline)
1
2
3
4
5
6
7
∗√
√∗
QR
BT
√
√
√
√
√
√
√
√
√
√
√
Accuracy
98.39%
95.97%
97.49%
97.23%
96.37%
97.91%
98.00%
97.23%
98.00%
Drop
2.42%
0.90%
1.16%
2.02%
0.48%
0.39%
1.16%
0.39%
§ The ideal accuracy without quantization;
denotes that the corresponding method is utilized.
TABLE III
The accuracy measurement for CNN on MNIST dataset
DQ
Ideal
0 (Baseline)
1
2
3
4
5
6
7
QR
BT
√
√
√
√
√
√
√
√
√
√
√
√
Accuracy
99.15%
90.77%
98.31%
96.98%
98.66%
98.96%
98.68%
98.75%
98.96%
Drop
8.38%
0.84%
2.17%
0.49%
0.19%
0.47%
0.40%
0.19%
DQ+QR+BT vs. DQ+BT), integrating more than one
proposed methods does not improve accuracy much. This
is because MNIST is a relative simpler database so the
effectiveness of these methods on accuracy improvement
quickly approaches to a saturated level. In more challenging CIFAR-10 database, experiments show that more
methods of DQ, QR and BT are harnessed, higher accuracy can always be obtained by a large margin.
C. Function Validation of LeNet
LeNet, which has strong robustness to image geometric
transformations, is a much more popular network. We
utilized it for MNIST and shows the results in Table III.
Compared with the MLP network, 1-level precision LeNet
can achieve an even lower accuracy drop (0.19% compared
with 0.39%) after combining all our methods. Remarkably, although the DQ method separates the training and
quantifying processes, directly quantifying weights in each
layer has accuracy loss less than 1%, without further finetuning. The orthogonality among DQ, QR and BT is also
indicated by the results.
D. Function Validation of CNN on CIFAR-10
We also evaluate the proposed methods in more challenging natural image dataset CIFAR-10 to verify their
generality. The CNN in [20] is adopted without data augmentation. Table IV presents the results of all the interested combinations.
As expected, CNN has a large accuracy drop (64.32%)
when applying the naive 1-level quantization while each
our proposed technique can dramatically hinder the accuracy loss. However, unlike the experiments on MNIST,
a sole method cannot improve the accuracy of CNN to
a satisfactory level. Some combinations of two methods
TABLE IV
The accuracy measurement for CNN on CIFAR-10 dataset
DQ
Ideal
0 (Baseline)
1
2
3
4
5
6
7
QR
BT
√
√
√
√
√
√
√
√
√
√
√
√
Accuracy
82.12%
17.80%
32.92%
66.88%
46.54%
74.43%
57.74%
67.22%
76.59%
Drop
64.32%
49.2%
15.24%
35.58%
7.69%
24.38%
14.90%
5.53%
Fig. 7. The learned floating-point (upper) and quantified (lower)
conv1 filters in LeNet (the gray-scale ones) and CNN on
CIFAR-10 (the color ones). A zero weight is mapped to pixel value
128, and negative (positive) weights are darker (brighter) ones.
perform excellent accuracy improvement. For example,
DQ+RQ makes the accuracy level to 74.43%
BinaryConnect neural network in [7] performs state-ofthe-art accuracy when the last layer utilizes L2-SVM. The
parameters in the L2-SVM layer are floating-point and
critical for accuracy maintaining. However, the SVM is
not good for circuit implementation. Our work quantifies
all weights to one level and controls the accuracy loss
within 5.53% for more efficient circuit (e.g., memristor
crossbar) design.
E. Learned Filters
Fig. 7 presents the learned floating-point and 1-level
precision conv1 filters in LeNet and CNN on CIFAR-10,
respectively. Our methods can efficiently learn the feature extractors similar to the corresponding original ones,
even with 1-level precision. Furthermore, the number of
input channels (RGB) of CIFAR-10 image is 3, such that
each pixel in the filter has 33 possible colors. For filters
with n channels, a 1-level precision filter still has a large
learning space with 3n·k·k possibilities, where k is the filter size. Those explain why our method can maintain the
comparable accuracy.
F. Bias Tuning to Alleviate Crossbar Variation
As aforementioned, the memristive variations caused
by fabrication imperfection can result in deviation of the
programmed weights [4]. Our bias tuning method can
also be extended to overcome memristor variation. After programming weights to memristors under the impact
of variation, we read out the real programmed weights,
then fine-tune the bias with weights frozen, and finally
the tuned biases are reprogrammed to the circuit neuron
models to compensate the impact of weight variation.
Fig. 8 plots the accuracy vs. the variance of programming process. The entry 4 in Table III is taken as the
baseline in this investigation on variation impact. The
figure shows that the bias tuning method successfully hinders the negative impact of variation.
G. Discussion
Our previous research study [5] specifies for spiking neural networks, where the probability distribution can only
be biased to two poles (0 or 1). In this work, we extend the
1
Accuracy
0.95
opinions, findings and conclusions or recommendations
expressed in this material are those of the authors and
do not necessarily reflect the views of grant agencies or
their contractors.
ideal
noisy
bias tuning
0.9
0.85
0.8
References
0.75
0
0.01
0.02
0.03
0.04
!
Fig. 8. The bias tuning in LeNet. The yellow line denotes the
accuracy after applying DQ and QR without noise; The red line is
the baseline with quantization and noise; The green line denotes
the accuracy recovered from the baseline after bias tuning; σ is the
standard deviation of Gaussian noise.
method to memristor-based neural networks adopted by
state-of-the-art research and large-scale applications [21].
The proposed methods can regularize the floating-point
weights to multiple levels with uniform or nonuniform
quantization. For example in our CIFAR-10 experiments,
the quantization points in layer conv1, conv2, conv3 and
ip1 are [−0.12, 0, 0.12], [−0.08, 0, 0.08], [−0.02, 0, 0.02] and
[−0.008, 0, 0.008], respectively. Moreover, we discharge
the reliance on the floating-point layer in [5] and explore
a pure one-level precision solution. Comprehensive experiments and analyses on MLP and CNN using MNIST and
CIFAR-10 datasets are conducted. Our experiments on
MNIST shows negligible accuracy drop (0.19% in CNN),
which is much better than the previous work like [5].
From the aspect of the system implementation, there
are extensive research studies on binary neural networks
deployed in traditional platforms such as CPUs, GPUs
and FPGAs. However, those approaches may not suitable
for the hardware characteristics of brain-inspired systems
like memristor-based systems. For example, BinaryConnect [7] uses L2-SVM layer, which is very costly to be implemented by memristor hardware. In circuit design, bias
has the characteristic of adjustability, which inspires our
bias tuning method in this work. As shown in the paper,
bias tuning can be used to control quantization accuracy
as well as overcome the process variation of memristor
technology.
V. Conclusions
In this work, we analyze the impact on accuracy degradation of low-resolution synapses in neuromorphic hardware implementations theoretically and propose three orthogonal methods to learn synapses with 1-level precision.
We applied these proposed methods and their combinations to MLP on MNIST, CNN on MNIST and CNN on
CIFAR-10 database, comparable state-of-the-art achievements are obtained: only 0.39%, 0.19%, and 5.53% accuracy loss, respectively. Our work will be more suitable for
memristor-based neural networks.
Acknowledgment
This work was supported in part by NSF CCF-1615475,
NSF XPS-1337198 and AFRL FA8750-15-2-0048. Any
[1] A. S. Cassidy, et al., “Cognitive computing building block: A
versatile and efficient digital neuron model for neurosynaptic
cores,” The 2013 International Joint Conference on Neural
Networks (IJCNN), pp. 1–10, 2013.
[2] M. Hu, et al., “Memristor crossbar-based neuromorphic computing system: A case study,” IEEE transactions on neural
networks and learning systems, vol. 25, pp. 1864–1878, 2014.
[3] S. Gaba, et al., “Memristive devices for stochastic computing,”
2014 IEEE International Symposium on Circuits and Systems
(ISCAS), pp. 2592–2595, 2014.
[4] M. Hu, et al., “The stochastic modeling of TiO 2 memristor and
its usage in neuromorphic system design,” 2014 19th Asia and
South Pacific Design Automation Conference (ASP-DAC), pp.
831–836, 2014.
[5] W. Wen, et al., “A new learning method for inference accuracy, core occupation, and performance co-optimization on
TrueNorth chip,” arXiv preprint arXiv:1604.00697, 2016.
[6] M. Rastegari, et al., “XNOR-Net: ImageNet classification using
binary convolutional neural networks,” CoRR, 2016.
[7] M. Courbariaux, et al., “BinaryConnect: Training deep neural
networks with binary weights during propagations,” CoRR, vol.
abs/1511.00363, 2015.
[8] M. Courbariaux, et al., “BinaryNet: Training deep neural networks with weights and activations constrained to +1 or -1,”
CoRR, vol. abs/1602.02830, 2016.
[9] L. Chua, et al., “Memristor-the missing circuit element,” IEEE
Transactions on circuit theory, vol. 18, pp. 507–519, 1971.
[10] D. B. Strukov, et al., “The missing memristor found,” nature,
vol. 453, pp. 80–83, 2008.
[11] W. Yi, et al., “Feedback write scheme for memristive switching
devices,” Applied Physics A, vol. 102, pp. 973–982, 2011.
[12] M. Hu, et al., “Hardware realization of BSB recall function using memristor crossbar arrays,” Proceedings of the 49th Annual
Design Automation Conference, vol. 17, pp. 498–503, 2012.
[13] W. Wen, et al., “An EDA framework for large scale hybrid neuromorphic computing systems,” Proceedings of the 52nd Annual
Design Automation Conference, pp. 12, 2015.
[14] M. Kim, et al., “Bitwise neural networks,” arXiv preprint
arXiv:1601.06071, 2016.
[15] Y. LeCun, et al., “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, pp. 2278–
2324, 1998.
[16] Y. LeCun, et al., “The MNIST database of handwritten digits”,
1998.
[17] X. Glorot, et al., “Understanding the difficulty of training deep
feedforward neural networks,” Aistats, vol. 9, pp. 249–256, 2010.
et al.,
[18] G. H. Golub, et al., “Generalized cross-validation as a method
for choosing a good ridge parameter,” Technometrics, vol. 21,
pp. 215–223, 1979.
[19] Y. Jia, et al., “Caffe: Convolutional architecture for fast feature
embedding,” arXiv preprint arXiv:1408.5093, 2014.
[20] A. Krizhevsky et al., “ImageNet classification with deep convolutional neural networks,” Advances in neural information
processing systems, 2012.
[21] T. Tang et al., “Spiking neural network with RRAM: can we
use it for real-world application?,” Proceedings of the 2015 Design, Automation & Test in Europe Conference & Exhibition,
pp. 860–865, 2015.
| 9cs.NE
|
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
Invitation to Ezhil: A Tamil Programming Language
for Early Computer-Science Education
Muthiah Annamalai, Ph.D.
Boston, USA.
Abstract:
Ezhil is a Tamil programming language with support for
imperative programming, with mixed use of Tamil and English
identifiers and function-names. Ezhil programing system is
targeted toward the K-12 (junior high-school) level Tamil
speaking students, as an early introduction to thinking like a
computer-scientist. We believe this 'numeracy' knowledge is
easily transferred over from a native language (Tamil) to the
pervasive English language programming systems, in Java, dotNet, Ruby or Python. Ezhil is an effort to improve access to
computing in the 21st Century.
Introduction:
Ezhil has exclusively Tamil keywords, and provides built-in
list, string, numeric data-types, mathematical and graphics
functions, by building on the Python programming environment, as
first reported in 2008 [1]. Recent Ezhil development, has added
Turtle-graphics [2], See Fig. 4, Windows installation, and
editor syntax-highlighting support, see Fig. 1, to enable a
smooth development experience. Ezhil is the first freely
available and completely open-source Tamil programming language
with source downloads available via our website [3a,b]. This
paper talks about the state of Ezhil-Language interpreter
development as of current development release, v0.5 [3a].
In this paper we describe the installation and use of Ezhil
language for curriculum development, graphics, and lessonplanning and showcase our system which is continually developed
and expected to reach widespread deployment, trials in near
future.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
Language Overview:
Ezhil keywords
Ezhil Language has the following keywords, see [4], defining the
control structures and imperative programming language
constructs, which we explain with the grammar in the following
sections.
1.- :
"
"
(PRINT)
2.- : "
"
(BREAK)
3.- : "
"
(CONTINUE)
4.- : "
"
5.- : "
ELSEIF-ELSE)
6.- : "
"
", "
7.- : "
8.- : "
(RETURN)
"
(IF-
(END)
"
"
", "
(FUNCTION)
(WHILE)
Data type system
The data type system supported by Ezhil currently includes four
basic types, (floating point, integer) numbers, logical Boolean
value types, string types, and lists. Lists (arrays) are a
heterogeneous constructor. Dictionary (hashtable) support is
planned for future.
Development Environment:
Ezhil programs can be run from the command-line mode in both
Linux, and Windows. It can also be run from the web interface as
we explain below.
Ezhil Interpreter in Command Line Mode
Ezhil Interpreter is invoked by the command ‘ez’. So when you
type,”$ ez” on the command-line you will see, the following
prompt,
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
எழில் 1>
and you can start entering your commands at this prompt and
interactively use the interpreter.
You can try typing the following program line-by-line in the
interpreter to see the output,
"
"
10 + 15, 17/3.0, 17%3, 50 - 6*5, ( 50 - 5*6 )/4
printf("Hello \n World")
"
"
30 - 5, 5^2 , 2^7
"
"
5 * 5
"
"
625 / 5
"
"
25^2
"
"
(((5 + (25^2))/6)^2)
( ( ( 5 ^ 2 ) + 5 ) / 6 ) ^ 2
= 10 #
"
"
" , 2*pi()*
", pi()* ^2
= 6 #
"
", 1/2* ^2*sin(pi()/3)
To use Ezhil interpreter in a batch mode fire it with a file
names, “$./ez ./ezhil_tests/hello.n”
The help for the Ezhil interpreter can be obtained by,
$ ./ez --help
usage: ./ez [-h] [-debug] [-stdin] [files [files ...]]
positional arguments:
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
files
optional arguments:
-h, --help show this help message and exit
-debug
enable debugging information on screen
-stdin
read input from the standard input
Typically user can write a Ezhil language program, a simple one
for example, to get user input and print their age in this case,
# (C)
2013
#
=
_
("
"
age
",
=
( "
@( age >= 18
"+
+ ",
? >> " )
)
"
"
? >> ")
!
age
"
"
"
",(18-age),"
!"
exit(0)
and save the file as ‘age.n’. All Ezhil programs should be
encoded in UTF-8 without BOM, and you should be able to
configure your text editor for this task To execute the program,
type, “./ez age.n”
Syntax Hilighting
A good editor to program Ezhil language is more motivation for
the user and catches some common mistakes like missing
parentheses, brackets, matching keywords in a “begin-end”
fashion.
All modern programming languages have syntax highlighting
capability with popular editors. We try to provide the same for
Ezhil language with three editors so far – emacs, Notepad++ on
Windows, and gedit. In Fig. 1, Ezhil emacs mode, shows a
guessing game program written in Ezhil language, and saved as
“guess.n”.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
The style files are found in the “Ezhil-Lang/ancilla” folder of
our GitHub repository [3a], and need to be installed separately.
Fig. 1, Ezhil emacs mode, showing a guessing
game written in Ezhil language
Library Functions
Currently Ezhil-Language supports over 350+ library functions,
allowing you to do things like file I/O, turtle graphics,
structured programming with user inputs, etc. We only see adding
more library functions in the future based of Python.
Language Support
Ezhil Language will have support for additional constructs for
iterators statements like, for-each, and for loops. A switchcase like construct is being thought about for relevance.
Arrays/lists are already supported for natural expressions, and
you may define heterogeneous arrays and nested arrays. Array
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
indexing with notation ‘[expr]’ and lvalue, rvalue semantics is
missing and will be added in future.
Ezhil Web Interface
Since Tamil input to the computer is still evolving with
different
Input-Methods
(IME)
and
also
to
exploit
the
pervasiveness of the global web access, and remove barriers to
trying out Ezhil, (i.e. without owning a computer, or
Python/Ezhil installation etc.) we have decided on a web
interface
to
Ezhil
on
the
lines
of
tryruby.org
or
interactivepython [5a,b].
The web interface is not yet hosted on ezhillang.org our primary
website. Launch the webserver on your local machine, by typing,
$ ./webserver.sh from the Ezhil-Lang source directory. This
should setup the server.
Now open the web browser, firefox, IE or google-chrome, and
type, http://localhost:8080 in your address bar, which should
open up a page, like in Fig. 2. The browser displays the Ezhil
Language web interface client (to a server running on your local
machine).
Now you may enter your programs in the editor at the center, and
click execute to see the output, like in Fig. 3. Use Keywords
helper for adding keywords in your program by a click. If you
have a program that contains errors, it will show up in red
text. You may navigate back from the results page (Fig. 3.) to
source program page (Fig. 2) by clicking “Go Back” link.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
Fig.2. Ezhil Language Web Interface client to a server running
on localhost. Enter your programs in the editor at the center,
and click execute to see the output, like in Fig.3. Use Keywords
helper for adding keywords in your program by a click.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
Fig.3. Ezhil Language Web Interface server output from a program
like in Fig. 2. Here the program has successfully executed, and
the output is shown in green text. You can click the link “Go
Back” to head back to the program evaluation screen, Fig.2.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
Curriculum Development:
Children
learn
from
immediate,
positive
feedback,
in
a
experiential manner. So we plan lessons in Ezhil around simple
programs, and interactive experiences with user input, graphics,
sounds and terminal/console output.
A simple program to print a square in Ezhil graphics will be,
for the first attempt, straight-line code, which they can
understand the meaning of the constructs, and have fun executing
to see a output from their command.
#
( 90 )
( 90 )
( 90 )
( 90 )
( 90 )
( 90 )
( 90 )
( 90 )
Next lesson would be to improve the program to eliminate
repetition – i.e. provide automation using a while loop – which
is a key lesson here – and they will write the following
program,
#
#
'
'
= 0
@(
< 4 )
( 90 )
( 90 )
=
+ 1
Ezhil language for curriculum development is
thought about and our exercises so far, include
1. hello word program,
2. iteration tests,
being
seriously
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
3.
4.
5.
6.
7.
printing multiplication tables,
developing a quiz
drawing squares with turtle graphics
complex symbols with turtle graphics
Calculation/graphics based on recursion
We have incorporated these ideas into an upcoming book [6], to
be available at a widely accessible cost.
A slightly advanced program to draw, Yin-Yang, Chinese symbols,
as interlocking circles with color filling is shown in Fig. 4.
See our source code examples for details “ezhil_tests/yinyang.n”
[3a]
Fig. 4. Ezhil mode Turtle-graphics to draw Yin-Yang Chinese
energy symbols.
Conclusions:
Ezhil is a Tamil programming language, developed in 2007-08, and
dormant for a few years, and it has been recently revived.
We believe this ‘numeracy’ knowledge is easily transferred
over from a native language (Tamil) to the pervasive English
language programming systems, in Java, dot-Net, Ruby or Python.
Ezhil is a promising effort to improve access to computing
in the 21st Century through ‘numeracy’ education, by being free
and open-source, easily accessibly on multiple platforms, and
being well documented.
Invitation to Ezhil : A Tamil Programming Language for Early Computer-Science Education 07/10/13
References:
1. M. Annamalai, “Ezhil (
ArXiV/0907.4960 (2008).
) : A Tamil Programming Language”,
2. Python 2.7 docs, “turtle - Turtle graphics for Tk”,
http://docs.python.org/2/library/turtle.html
3a. Ezhil-Lang download page on Git-Hub:
https://github.com/arcturusannamalai/EzhilLang/archive/master.zip,
3b. Ezhil-Lang project homepage : http://ezhillang.org
4. Ezhil on Wikipedia : (Tamil) http://ta.wikipedia.org/wiki/
_(
_
),(English) -
http://en.wikipedia.org/wiki/Ezhil_(programming_language)
5a. Try Ruby – 15 minute introduction to Ruby Language,
http://tryruby.org/levels/1/challenges/0 (accessed July,2013)
5b. Interactive Python – programming lessons online,
http://interactivepython.org/ (accessed July, 2013)
6. Ezhil Language Team, “An Introduction to Ezhil – Programming
the Computer in Tamil” (TBD) 2013.
| 6cs.PL
|
arXiv:1705.07382v1 [math.ST] 21 May 2017
GRADIENT FLOWS: APPLICATIONS TO CLASSIFICATION, IMAGE
DENOISING, AND RIEMANNIAN MARKOV CHAIN MONTE CARLO
NICOLÁS GARCÍA TRILLOS
∗ AND
DANIEL SANZ-ALONSO
∗
Abstract. A pressing question in Bayesian statistics and machine learning is to introduce a
unified theoretical framework that brings together some of the many statistical models and algorithmic
methodologies employed by practitioners. In this paper we suggest that the variational formulation of
the Bayesian update, coupled with the theory of gradient flows, provides both an overarching structure
and a powerful tool for the analysis of many such models and algorithms. As particular instances of
our general framework, we provide three variational formulations of the Bayesian update with three
associated gradient flows that converge to the posterior. We highlight the key unifying role of the
concept of geodesic convexity, as it determines —in all three instances— the rate of convergence of
the flows to the posterior. These gradient flows naturally suggest stochastic processes to be used to
build proposals for Markov chain Monte Carlo (MCMC) algorithms. Moreover, by construction the
processes are guaranteed to satisfy certain optimality criteria. A core part of the paper is to explore
three areas of application: the variational formulation of high-dimensional classification and Bayesian
image denoising, and the optimal choice of metric in Riemannian MCMC methods.
Key words. Gradient flows, Wasserstein Space, Geodesic Convexity, Classification, Image Denoising, Riemannian MCMC
1. Introduction In this paper we propose a general framework for the analysis of
several models and sampling schemes used in Bayesian statistics and machine learning.
Our framework relies on the recent advances in the theory of optimal transport and
gradient flows. We demonstrate its usefulness in three application areas: classification
via semi-supervised learning, the Bayesian formulation of total variation denoising for
image processing, and the design and analysis of proposals for Markov chain Monte
Carlo (MCMC) algorithms. In particular, we show how our framework suggests clear
optimality criteria for the choice of metric in Riemannian MCMC methods.
Bayesian statistics consists in the update of a prior probability measure π ∈ P(M)
—that represents our various degrees of belief on the value of a quantity of interest
u ∈ M— into a posterior probability measure µ ∈ P(M), based on observed data y.
In the above, P(M) denotes the space of probability measures on a manifold M (our
motivation for setting up the problem on a manifold will become apparent below). The
Bayesian update is built upon a postulated statistical relationship between the observed
data and the quantity of interest, encoded in a likelihood function exp −φ(u;y) . Under
mild conditions, the posterior can be written in terms of the prior and the likelihood as
a change of measure:
µ(du) =
1
exp −φ(u;y) π(du).
Z
(1.1)
Here and throughout it is tacitly assumed that φ is integrable under π, so that the
normalizing constant Z is positive and finite. Formula (1.1) can be thought of as
generalizing Bayes’ formula P(u|y) ∝ P(u)P(y|u), with P(u|y) representing the posterior,
P(y|u) the likelihood, and P(u) the prior.
Our starting point are three variational formulations of the Bayesian update. The
first two characterize the posterior measure µ as the minimizer of functionals JKL and Jχ2
defined on P(M) —see (3.1) and (3.4). The third one characterizes the posterior density
as the minimizer of the Dirichlet energy Dµ —see (2.4). While the latter characterization
∗ Division
of Applied Mathematics, Brown University.
1
2
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
makes use of the posterior itself, both JKL and Jχ2 are defined in terms of the prior and
the likelihood. As will be detailed in section 1.1, we envision several unexplored benefits
of the old idea of seeing the Bayesian update as an optimization problem (a literature
review will be given in section 1.2). The main benefit investigated in this paper is that
these variational formulations naturally suggest a way to “move” towards the posterior
measure in Wasserstein space (or the posterior density in L2 ) through the theory of
gradient flows. We now describe briefly this idea, which will be later developed.
Recall that a gradient flow in Euclidean space defines a curve whose tangent always
points in the direction of steepest descent of a given function (see (2.5)). In the same
fashion, a gradient flow in a more general metric space can be thought of as a curve on
said space that always points in the direction of steepest descent of a given functional.
In this paper we study (i) the gradient flows defined by JKL and Jχ2 in the space of
probabilities on M with finite second moments endowed with the Wasserstein distance—
definitions are given in (2.1). By construction these flows give “curves” of probabilities
in Wasserstein space that evolve following the direction of steepest descent of JKL and
Jχ2 in said space, converging to µ in the large time limit; and (ii) the gradient flow
defined by the Dirichlet energy Dµ in L2 . By construction, this flow gives a “curve” of
densities in L2 that evolves following the direction of steepest descent of Dµ , converging
to the posterior density in the large time limit.
Our presentation highlights the key unifying role of the notion of geodesic convexity
of functionals. Indeed, geodesic convexity not only guarantees the existence and uniqueness of flows, but also determines their rate of convergence to the posterior in all cases.
In the L2 setting we show that positive geodesic convexity is equivalent to the posterior
satisfying a Poincaré inequality, and to the existence of a spectral gap —see subsection
3.2.2. On the other hand, the geodesic convexity of JKL andJχ2 in Wasserstein space is
determined by the Ricci curvature of the manifold, as well as by the likelihood function
and prior density —see (3.1), (3.2), (3.4), (3.5).
It is known that the flow of JKL in Wasserstein space and the flow of the Dirichlet
energy in L2 produce the same path towards the posterior: their evolution of densities
solve the same Fokker-Planck partial differential equation (PDE), and they result in the
same Langevin diffusion. By construction, the invariant distribution of said diffusion
is the posterior. The MCMC methodology allows to use a discretization of the diffusion —combined with an accept-reject mechanism to remove the discretization bias—
to produce, in the large-time asymptotic, correlated posterior samples. From the above
it follows an optimality criteria satisfied by the Langevin diffusion in Euclidean space
employed within the Metropolis-Adjusted Langevin algorithm (MALA), and by its Riemannian extension when a different metric is used in the parameter space: the densities
associated with the (possibly weighted) Langevin diffusion approach the posterior as
fast as possible in a Kullback-Leibler divergence sense (with the geometry given by the
underlying Wasserstein space, constructed using the appropriate metric) and also in a
Dirichlet energy sense with the geometry given by L2 (M,µ). On the other hand, the
evolution of densities of the gradient flow of Jχ2 in Wasserstein space is given by a porous
medium equation. The Markov process associated with this diffusion is inhomogeneous,
and direct simulation seems unfeasible unless more structure is imposed over the prior
measure and likelihood function.
In short, we investigate the following framework for the study of the Bayesian update
and sampling algorithms:
(i) Characterize the posterior as the minimizer of a functional.
(ii) Study the geodesic convexity of the functional on a given metric space. Define the
N. Garcı́a Trillos, and D. Sanz-Alonso
3
associated gradient flow, and establish convergence rates (functional inequalities)
towards the posterior in terms of the geodesic convexity.
(iii) Determine the evolution PDE governing the flow of densities, and translate the
PDE into a diffusion process to be used within a sampling scheme, notably to
build proposals for MCMC algorithms.
The diffusion built from the above framework will by construction have the optimality property that the associated densities evolve towards the posterior density in the
direction of steepest descent (defined in terms of the metric space where the gradient
flow is defined) of the functional. A further appeal of this line of thinking is that the
convergence rate of the densities towards the posterior is obtained for free from the
sharp theory of functional inequalities of gradient flows. Moreover, changing the metric
on the manifold M results in different flows and diffusions, whose rates of convergence
can then be compared, establishing a sound theoretical basis for the choice of metric in
Riemannian MCMC methods.
We now provide a comparison of the three choices of functionals that we consider
within the framework outlined above.
1. The two gradient flows in Wasserstein space (arising from the functionals JKL
and Jχ2 ) are fundamentally connected with the variational formulation: these
variational formulations can be used to define posterior-type measures via a penalization of deviations from the prior and deviations from the data in situations
where establishing the existence of conditional distributions by desintegration
of measures is technically problematic. On the other hand, the variational formulation for the Dirichlet energy is less natural and requires previous knowledge
of the posterior.
2. The precise level of geodesic convexity of the funcitonals JKL (and Jχ2 ) can
be computed from point evaluation of the Ricci tensor and derivatives of the
densities. In particular, knowledge of the underlying metric suffices to compute
these quantities. In contrast, establishing a sharp Poincaré inequality —the
level of geodesic convexity of the Dirichlet energy in L2 (M,µ)— is in practice
unfeasible, as it effectively requires solving an infinite dimensional optimization problem. It is for this reason —and because of the explicit dependence of
the convexity in Wasserstein space with the geometry induced by the manifold
metric tensor— that our analysis of the choice of metric in Riemannian MCMC
methods is based on the JKL functional (see section 4.3, and in particular Theorem 4.3).
3. On the flip side of point 2, a Poincaré inequality for the posterior with a not
necessarily optimal constant can be established using only tail information. In
particular, even when the functional JKL is not geodesically convex in Wasserstein space, one may still be able to obtain a Poincaré inequality (see subsection
4.1.2 for an example).
4. In contrast to the diffusions arising from the JKL or Dirichlet flows, the stochastic processes arising from the Jχ2 formulation are inhomogeneous, and hence
not computationally appealing.
We conclude this introduction noting that there are other important scenarios —
e.g. the study of conditioned diffusions— where a measure of interest µ is naturally
defined in terms of a reference measure π and an un-normalized density. All the ideas
above extend straightforwardly to those applications, but for concreteness we will focus
on the Bayesian interpretation. Also, in this paper we consider only Kullback-Leibler
and χ2 prior penalizations to define the functionals JKL and Jχ2 , but the ideas extend
4
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
to the family of m-divergences, see [22]. Kullback-Leibler and χ2 prior penalization
correspond to m → 1 and m = 2 within this family.
1.1. Contributions The main contribution of this paper is to demonstrate how
the theory of gradient flows provides an overarching framework for the analysis of
Bayesian models and related algorithms, to highlight the key unifying role of the notion
of geodesic convexity in the general theory, and to illustrate the usefulness of this approach in three concrete applications. In particular, we show how this general framework
suggests optimality criteria for the choice of metric in Riemannian MCMC schemes. We
summarize below some additional outcomes:
• We bring attention to different variational formulations of the Bayesian update.
These formulations have the potential of extending the theory of Bayesian inverse problems in function spaces, in particular in cases with infinite dimensional, non-additive, and non-Gaussian observation noise. Moreover, they suggest numerical approximations to the posterior by restricting the space of allowed measures in the minimization, by discretization of the associated gradient
flows, or by sampling via simulation of the associated diffusion.
• The variational framework considered in this paper provides a new tool for
the study of robustness of Bayesian models, and convergence from discrete to
continuum models by use of variational techniques. Robustness can be studied
at two different levels, 1) posterior distributions are defined as minimizers of an
energy, and hence posterior robustness tantamounts to stability of minimizers
of a sequence of energies, and 2) paths connecting to posterior distributions are
written as gradient flows, and hence path robustness is equivalent to stability
of gradient flows.
• We give new understanding of the ubiquity of Kullback-Leibler penalizations in
sampling methodology. In practice Kullback-Leibler is often used for computational and analytical tractability. The results presented in section 3.3 show
that Kullback-Leibler prior penalization is the only m-divergence prior penalization that leads to a heat-type flow and, therefore, to an easily implementable
diffusion process.
1.2. Literature The idea of approximating a probability measure µ by finding
the minimizer of JKL (or equivalently, of DKL (·kµ) —see equation (3.3)) over certain class
of probability measures is used in Variational Bayes methods [2], [36], [12]. Typically
the optimizer is seeked over the class of product measures. A similar approach is taken
in [25] to study Gaussian approximations of measures defined in Hilbert space.
The textbooks [32], [33] present a modern view of the theory of optimal transport,
and its applications in partial differential equations, geometry, and probability. A more
applied perspective is given in [30]. For the formulation and analysis of gradient flows in
metric spaces we refer to [1]. The interplay between gradient flows and the Wasserstein
metric was made apparent in [17], which established a new variational formulation of
the Fokker-Planck equation. The key concept of geodesic (or displacement) convexity
was introduced by McCann in [21] with the purpose of establishing uniqueness for a
variational problem. A deeper understanding in terms of Riemannian geometry is found
in [23], which studies variational formulations and gradient flows for the porous medium
equation. Our presentation is heavily indebted to [22], that generalizes several aspects
of the theory of Kullback-Leibler gradient flows to the new family of m-divergences. For
a readable account of the theory of Langevin diffusions based on the Dirichlet energy
we refer to [24]. The relationship between concentration, log-Sobolev and Poincaré
inequalities is studied for instance in [19] and [15]. The study of log-Sobolev inequalities
N. Garcı́a Trillos, and D. Sanz-Alonso
5
and Ricci curvature on graphs is an active research area, e.g. [6], [11], [9], and [10]. The
role of spectral gaps in the theory of MCMC stems from the fact that it can be used to
bound the asymptotic variance of empirical expectations [18].
We conclude this literature review by providing some context for the three application areas that we consider. First, the Bayesian formulation of Probit semi-supervised
learning follows the presentation in [3]. Second, total variation denoising was introduced
in [29], and is now known as the ROF model. It has been widely used in image processing. Third, we show the relevance of the general framework in the design of MCMC
proposals. The use of Langevin diffusions to build MCMC methods was first suggested
in [4]. Their exponential ergodicity was analysed in [28], and their optimal scaling in
[27]. The idea of changing the metric on the parameter state in order to accelerate
MCMC algorithms by taking into account the geometric structure of the posterior was
proposed in [14]. Finally, other sampling algorithms have made use of transport ideas,
and the use of transport maps (not necessarily optimal) for posterior sampling is reviewed in [20]. Optimal transport maps have been employed within sequential Monte
Carlo methods [26].
1.3. Outline The rest of the paper is organized as follows. Section 2 contains
some background on the Wasserstein space, geodesic convexity of functionals, and gradient flows in metric spaces. The general framework (i)-(ii)-(iii) outlined above is spelled
out in Section 3. Finally, in section 4 we consider three applications of the general
framework to problems in machine learning, Bayesian image processing, and sampling.
1.4. Notation (M,g) will denote a smooth connected m-dimensional manifold
with metric tensor g. We will denote by d the associated Riemannian distance, and
assume that (M,d) is a complete metric space, so that in particular it is a geodesic
metric space by the Hopf-Rinow theorem —see section (2.1) for a short discussion on
geodesic spaces. We denote by volg the associated volume form. The reader not versed
in Riemannian geometry may focus on the case M = Rm with the usual metric tensor
so that d is the Euclidean distance and dvolg = dx is the Lebesgue measure. However,
in section 4.3 where we discuss applications to Riemannian MCMC, we endow Rm with
a general metric tensor g and hence familiarity with some notions from differential
geometry is desirable.
The space of probability measures on M (with the Borel σ-algebra) is denoted
P(M). We define the prior (denoted π) as a change of measure from volg , and the
posterior as a change of measure from π as follows:
π = e−Ψ volg ,
µ ∝ e−φ π.
(1.2)
2. Preliminaries
In this section we provide some background material. The Wasserstein space, and
the notion of λ-geodesic convexity of functionals are reviewed in subsection 2.1. The
basics of the theory of gradient flows in metric spaces is reviewed in subsection 2.2.
2.1. Geodesic Spaces and Geodesic Convexity of Functionals A geodesic
space (X,dX ) is a metric space with a notion of length of curves that is compatible
with the metric, and where every two points in the space can be connected by a curve
whose length achieves the distance between the points (see [1] for more details). In
this paper we consider three geodesic spaces. First, the base space (M,d), i.e. the
manifold M equipped with its Riemannian distance. Second, the space P2 (M) of square
integrable Borel probability measures defined on M, endowed with the Wasserstein
6
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
R
distance W2 . Third, the space of functions f ∈ L2 (M,µ), with M f dµ = 1, equipped
with the L2 (M,µ) norm. We spell out the definitions of P2 (M) and W2 :
Z
n
o
P2 (M) := ν ∈ P(M) :
d2 (x,x0 )dν(x) < ∞, for some x0 ∈ M ,
M
Z
(2.1)
W22 (ν1 ,ν2 ) := inf
d(x,y)2 dα(x,y), ν1 ,ν2 ∈ P2 (M).
α
M×M
The infimum in the previous display is taken over all transportation maps between ν1
and ν2 , i.e. over α ∈ P(M × M) with marginals ν1 and ν2 on the first and second factors.
The space (P2 (M),W2 ) is indeed a geodesic space: geodesics in (P2 (M),W2 ) are
induced by those in (M,d). All it takes to construct a geodesic connecting ν0 ∈ P2 (M)
and ν1 ∈ P2 (M) is to find an optimal transport plan between ν0 and ν1 to determine
source locations and target locations, and then transport the mass along geodesics in
M (see [32] and [30]).
R
The space of functions f ∈ L2 (M,µ), with M f dµ = 1, equipped with the L2 (M,µ)
norm is also a geodesic space, where a constant speed geodesic connecting f0 and f1 is
given by linear interpolation: t ∈ [0,1] 7→ (1 − t)f0 + tf1 .
We will consider several functionals E : X → R ∪ {∞} throughout the paper. They
will all be defined in one of our three geodesic spaces —that is, X = M, X = P2 (M) or
X = L2 (M,µ). Important examples will be, respectively:
1. Functions Ψ : M → R ∪ {∞}.
2. The Kullback-Leibler and χ2 divergences DKL (·kπ), Dχ2 (·kπ) : P(M) → [0,∞],
where π is a given (prior) measure and, for ν1 ,ν2 ∈ P(M),
(R
dν1
dν1
(u)log
(u)
dν2 (u), ν1 ≪ ν2 ,
dν2
(2.2)
DKL (ν1 ||ν2 ) := M dν2
∞,
otherwise,
Dχ2 (ν1 ||ν2 ) :=
R
dν1
M
∞,
dν2 (u) − 1
2
dν2 (u), ν1 ≪ ν2 ,
(2.3)
otherwise;
and the potential-type functional J : P(M) → R ∪ {∞} given by
Z
h(u)dν(u),
J(ν) :=
M
where h is a given potential function.
3. The Dirichlet energy Dµ : L2 (M,µ) → [0,∞] defined by
(R
k∇f (u)k2 dµ(u), f ∈ L2 (M,µ) ∩ H 1 (M),
.
Dµ (f ) = M
+∞
otherwise.
(2.4)
A crucial unifying concept will be that of λ-geodesic convexity of functionals. We
recall it here:
Definition 2.1. Let (X,dX ) be a geodesic space and let λ ∈ R. A functional E : X →
R ∪ {∞} is called λ-geodesically convex provided that for any x0 ,x1 ∈ X there exists a
constant speed geodesic t ∈ [0,1] 7→ γ(t) ∈ X such that γ(0) = x0 , γ(1) = x1 , and
t(1 − t) 2
dX (x0 ,x1 ),
E γ(t) ≤ (1 − t)E(x0 ) + tE(x1 ) − λ
2
∀t ∈ [0,1].
N. Garcı́a Trillos, and D. Sanz-Alonso
7
The following simple remark characterizes the λ-convexity of functions.
Remark 2.2. Let Ψ ∈ C 2 (M) so that we can define its Hessian at all points in M
(see the proof of Theorem 4.3 in the appendix for the definition). Then the following
conditions are equivalent:
(i) Ψ is λ-geodesically convex.
(ii) HessΨx (v,v) ≥ λ for all x ∈ M and all unit vectors v ∈ Tx M.
If (M,d) is the Euclidean space, (i) and (ii) are also equivalent to:
(iii) Ψ − λ2 |·|2 is a convex function.
This latter condition is known in the optimization literature as strong convexity.
2.2. Gradient Flows in Metric Spaces In this subsection we review the basic
concepts needed to define gradient flows in a metric space (X,dX ). We follow Chapter
8 of [30]; a standard technical reference is [1].
To guide the reader, we first recall the formulation of gradient flows in Euclidean
space. Let E : Rd → R be a differentiable function, and consider the equation
(
ẋ(t) = −∇E x(t) , t ≥ 0,
(2.5)
x(0) = x0 .
Then, the solution to (2.5) is the gradient flow of E in the Euclidean space with initial
condition x0 ; it is a curve whose tangent vector at every point in time is the negative
of the gradient of the function E at that time. In order to generalize the notion of a
gradient flow to functionals defined on more general metric spaces, and in particular
when the metric space has no differential structure, we reformulate (2.5) in the form:
Z
Z
2
1 t
1 t
2
|ẋ(r)| dr +
(2.6)
∇E x(r) dr, t > 0,
E(x0 ) = E x(t) +
2 0
2 0
This identity, known as energy dissipation equality, is equivalent to (2.5) —see Chapter
8 of [30] for further details and other possible formulations. Crucially (2.6) involves
notions that can be defined in an arbitrary metric space (X,dX ): the metric derivative
of a curve t 7→ x(t) ∈ X is given by
dX x(t),x(s)
|ẋ(t)| := lim
,
s→t
|s − t|
and the slope of a functional E : X → R ∪ {∞} is defined as the map |∇E| : {E < ∞} →
R ∪ {∞} given by
+
E(x) − E(y)
|∇E|(x) := limsup
.
dX (x,y)
y→x
The identity (2.6) is the standard way to introduce gradient flows in arbitrary metric
spaces. In this paper we consider gradient flows in L2 and Wasserstein spaces, where
the notion of tangent vector is available. L2 has Hilbert space structure, whereas the
Wasserstein space can be seen as an infinite dimensional manifold (see [1], [30]).
3. Variational Characterizations of the Posterior and Gradient Flows
In this section we carry out the framework (i)-(ii)-(iii) outlined in the introduction.
Subsection 3.1 details the variational formulations of the Bayesian update, defined in
terms of the functionals JKL , Jχ2 and Dµ . Subsection 3.2 studies the geodesic convexity
of JKL and Jχ2 in Wasserstein space and of Dµ in L2 . Finally, subsection 3.3 spells
out the PDEs governing the evolution of densities of the gradient flows, as well as the
corresponding diffusion processes.
8
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
3.1. Variational Formulation of the Bayesian Update The variational formulation of the posterior as the minimizer of JKL and Jχ2 share the same structure, and
are outlined in subsection 3.1.1. The variational formulation in terms of the Dirichlet
energy is given in subsection 3.1.2.
3.1.1. The Functionals JKL and Jχ2 It is well-known in the analysis [16] and
probability [8] communities that the posterior µ —defined by (1.1)— is the minimizer
of the functional
JKL (ν) := DKL (ν||π) + FKL (ν;φ),
ν ∈ P(M),
(3.1)
where
FKL (ν;φ) :=
Z
φ(u)dν(u),
(3.2)
M
and the integral is interpreted as +∞ if φ is not integrable with respect to ν. In
physical terms, the Kullback-Leibler divergence represents an internal energy, whereas
FKL represents a potential energy. The problem of minimizing (3.4) has also a clear
statistical interpretation. The Kullback-Leibler term DKL (·kπ) penalizes deviations from
our prior beliefs, while the term FKL (ν;φ) penalizes deviations from the data. The
normalizing constant Z is known as the partition function in the physical literature,
and as the marginal likelihood in statistics, where it plays an important role in model
selection.
For completeness, we remark that the fact that µ minimizes JKL follows immediately
from the identity
DKL (·kµ) = JKL (·kπ) + logZ.
(3.3)
It is thus equivalent to minimize JKL (·) or DKL (·kµ), but the functional JKL makes
apparent the roles of the prior and the likelihood.
The posterior µ also minimizes the functional
Jχ2 (ν) := Dχ2 (ν||π) + Fχ2 (ν;φ),
ν ∈ P(M),
(3.4)
where
Fχ2 (ν;φ) :=
Z
M
φ̃(u)dν(u),
φ̃ = g(exp(φ(u))),
g(t) := t − 1,
t > 0.
(3.5)
We refer to [22] for details. Note that both JKL and Jχ2 are defined in terms of the two
starting points of the Bayesian update: the prior π and the likelihood φ(u;y). The associated variational formulations suggest a way to define posterior-type measures based
on these two ingredients in scenarios where establishing the existence of conditional
distributions via desintegration of measures is technically demanding. This appealing
feature of the two variational formulations above is not shared by the one described in
the next subsection.
3.1.2. The Dirichlet Energy Dµ Let now the posterior µ be given, and consider
the space L2 (M,µ) of L2 -functions defined on M which are square integrable with
respect to µ; we denote by k·kµ the L2 -norm with respect to µ. Then it is clear that
the measure µ can be characterized as the probability measure with density ρµ ≡ 1 a.s.
with respect to µ. Hence, the posterior density ρµ ≡ 1 isRthe minimizer of the Dirichlet
energy Dµ over probability densities ρ ∈ L2 (M,µ) with M ρdµ = 1.
9
N. Garcı́a Trillos, and D. Sanz-Alonso
3.2. Geodesic Convexity and Functional Inequalities In this section we
study the geodesic convexity of the functionals JKL , Jχ2 , and Dµ . The geodesic convexity
of JKL and Jχ2 in Wasserstein space is considered in subsection 3.2.1, and the geodesic
convexity of Dµ in L2 in subsection 3.2.2, where we show its equivalence to the posterior
satisfying a Poincaré inequality.
3.2.1. Geodesic Convexity of JKL and Jχ2 The next proposition can be found
in [35] and [31]. It shows that the convexity of JKL can be determined by the socalled curvature-dimension condition —a condition that involves the curvature of the
manifold and the Hessian of the combined change of measure Ψ + φ. We recall the
notation π = e−Ψ volg and µ ∝ e−φπ.
Proposition 3.1. Suppose that Ψ,φ ∈ C 2 (M). Then JKL (or DKL (·kµ)) is λ-geodesically
convex if, and only if,
Ric(v) + HessΨ(v,v) + Hessφ(v,v) ≥ λ,
∀x ∈ M,
∀v ∈ Tx M with g(v,v) = 1,
where Ric denotes Ricci curvature.
The following simple example illustrates the geodesic convexity of DKL (·kµ) for
Gaussian µ.
Example 3.2. Let µ = N (θ,Σ) be a Gaussian measure in Rm , with Σ positive definite.
Then DKL (·kµ) is 1/Λmax (Σ)-geodesically convex, where Λmax (Σ) is the largest eigenvalue of Σ. This follows immediately from the above, since here Ψ(x) = 12 hx − θ,Σ−1 (x −
θ)i, and the Euclidean space is flat (its Ricci curvature is identically equal to zero).
Note that the level of convexity of the functional depends only on the largest eigenvalue
of the covariance, but not on the dimension m of the underlying space.
The λ-convexity of JKL guarantees the existence of the gradient flow of JKL in
Wasserstein space. Moreover, it determines the rate of convergence towards the posterior
µ. Precisely, if µ0 is absolutely continuous with respect to µ, then the gradient flow
t ∈ [0,∞) 7→ µt of JKL with respect to the Wasserstein metric starting at µ0 is well defined
and we have:
DKL (µt ||µ) ≤ e−λt DKL (µ0 ||µ),
W2 (µt ,µ)2 ≤ λe−λt DKL (µ0 ||µ),
t ≥ 0,
t ≥ 0.
(3.6)
The second inequality, known as Talagrand inequality [32], establishes a comparison
between Wasserstein geometry and information geometry. It can be established directly
combining the λ-geodesic convexity of JKL (for positive λ) with the first inequality.
We now turn to the geodesic convexity properties of Jχ2 . We recall that m denotes the dimension of the manifold M. The following proposition can be found in [22,
Theorem 4.1].
Proposition 3.3. Jχ2 is λ-geodesically convex if and only if both of the following two
properties are satisfied:
1
1. Ric(v) + HessΨ(v,v) + m+1
h∇Ψ,vi2 ≥ 0, ∀x ∈ M, ∀v ∈ Tx M.
2. φ is λ-geodesically convex as a real valued function defined on M.
There are two main conclusions we can extract from the previous proposition. First,
that condition 1) is only related to the prior distribution π whereas condition 2) is only
related to the likelihood; in particular, the convexity properties of Jχ2 can indeed be
studied by studying separately the prior and the likelihood (notice that the proposition
gives an equivalence). Secondly, notice that condition 1) is a qualitative property and
if it is not met there is no hope that the functional Jχ2 has any level of global convexity
even when the likelihood function is a highly convex function. In addition, if 1) is
10
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
satisfied, the convexity of φ determines completely the level of convexity of Jχ2 . These
features are markedly different from the ones observed in the Kullback-Leibler case.
As for the functional JKL , one can establish the following functional inequalities,
under the assumption of λ-geodesic convexity of Jχ2 :
Jχ2 (µt ) − Jχ2 (µ) ≤ e−λt Jχ2 (µ0 ) − Jχ2 (µ) , t ≥ 0,
W2 (µt ,µ)2 ≤ λe−λt Jχ2 (µ0 ) − Jχ2 (µ) , t ≥ 0.
(3.7)
3.2.2. Geodesic Convexity of Dirichlet Energy Let us start recalling
Poincaré inequality.
Definition 3.4. We say that a Borel probability measure µR on M has a Poincaré
inequality with constant λ if for every f ∈ L2 (M,µ) satisfying M f dµ = 0 we have
1
kf k2µ ≤ Dµ (f ).
λ
We now show that Poincaré inequalities are directly related to the geodesic convexity
of the functional Dµ in the L2 (M,µ) space.
Proposition 3.5. Let λ be a positive real number and let µ be a Borel probability
measure on M. Then, the measure µ has a Poincaré inequality with constant λ if and
only if the Rfunctional Dµ is 2λ-geodesically convex in the space of functions f ∈ L2 (M,µ)
satisfying f dµ = 1.
Proof. First of all we claim that
Dµ (tf0 + (1 − t)f1 ) + t(1 − t)D(f0 − f1 ) = tDµ (f0 ) + (1 − t)Dµ (f1 ),
(3.8)
for all f0 ,f1 ∈ L2 (M,µ) and every t ∈ [0,1]. To see this, it is enough to assume that both
Dµ (f0 ) and Dµ (f1 ) are finite and then notice that equality (3.8) follows from the easily
verifiable fact that for an arbitrary Hilbert space V with induced norm | · | one has
|tv0 + (1 − t)v1 |2 + t(1 − t)|v0 − v1 |2 = t|v0 |2 + (1 − t)|v1 |2 ,
∀v0 ,v1 ∈ V,
∀t ∈ [0,1].
Now, suppose that µ has a Poincaré
inequality
with constant λ and consider two
R
R
functions f0 ,f1 ∈ L2 (M,µ) satisfying M f0 dµ = M f1 dµ = 1. Then, (3.8) combined with
Poincaré inequality (taking f := f0 − f1 ) gives:
(3.9)
Dµ tf0 + (1 − t)f1 + λt(1 − t)kf0 − f1 k2µ ≤ tDµ (f0 ) + (1 − t)Dµ (f1 ),
which is precisely the 2λ-geodesic convexity condition for Dµ .
Conversely, suppose that Dµ is 2λ-geodesic convex in the Rspace of L2 (M,µ) functions that integrate to one. Let f ∈ L2 (M,µ) be such that M f dµ = 0 and without
the loss of generality assume that Dµ (f ) < ∞ and that ||f ||µ 6= 0. Under these conditions,
the positive
negative parts of f , f + and f − , satisfy Dµ (f + ),Dµ (f − ) < ∞ and
R and
R
−
+
M f dµ = r = M f dµ where r > 0. The inequality
1
||f ||2µ ≤ Dµ (f )
λ
is obtained directly from (3.8) and (3.9) applied to
1
f0 := f − ,
r
1
f1 := f + ,
r
t = 1/2.
11
N. Garcı́a Trillos, and D. Sanz-Alonso
Remark 3.6. It is well known that the best Poincaré constant for a measure µ is equal
to the smallest non-trivial eigenvalue of the operator −∆µ defined formally as
−∆µ f := −
1
div (e−φ−ψ ∇f ),
Z
where div and ∇ are the divergence and gradient operators in (M,g). This eigenvalue
can be written variationally as
λ2 :=
Dµ (f )
,
− fµ ||2µ
min
f ∈L2 (M,µ) ||f
where
fµ :=
Z
f dµ.
M
Let us now consider t ∈ (0,∞) 7→ µt the flow of Dµ in L2 (M,µ) with some initial
0
condition dµ
dµ = ρ0 . It is well known that this flow coincides with that of the functional JKL in Wasserstein space. However, taking the Dirichlet-L2 point f view, one can
use Poincaré inequality (i.e. the geodesic convexity of Dµ ) to deduce the exponential
convergence of µt towards µ in the χ2 -sense. Indeed, let
ρt :=
dµt
,
dµ
t ∈ (0,∞).
A standard computation then shows that,
Z
Z
∂ρ
1 d
2
(ρt − 1) dµ = −
|∇(ρt (u) − 1)|2 dµ(u) ≤ −λkρt − 1k2µ .
kρt − 1kµ =
2 dt
∂t
M
M
In the second equality, we have used the fact that
below. Hence by Gronwall’s inequality
∂ρ
∂t
= ∆µ ρ as discussed in section 3.3
kρt − 1kµ ≤ exp(−λt)kρ0 − 1kµ ,
t > 0.
3.3. PDEs and Diffusions Here we describe the PDEs that govern the evolution
of densities of the three gradient flows, and the stochastic processes associated to these
PDEs. We consider first the flows defined with the functionals JKL and Dµ in subsection
3.3.1, and then the flow defined by the functional Jχ2 in subsection 3.3.2.
3.3.1. JKL -Wasserstein and Dµ -L2 (M,µ) It was shown in [17] —in the Euclidean setting and in the unweighted case π = dx— that the gradient flow of the
Kullback-Leibler functional DKL (·kπ) in Wasserestein space produces a solution to the
Fokker-Planck equation. More generally, under the convexity conditions guaranteeing
the existence of the gradient flow t ∈ (0,∞) 7→ µt of DKL (·||µ) (equivalently of JKL ) starting from µ0 ∈ P(M), the densities
ρt :=
dµt
,
dµ
θt :=
dµt
,
dvolg
t ∈ (0,∞)
satisfy (formally) the following Fokker-Planck equations
∂ρ
= ∆µ ρ.
∂t
(3.10)
12
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
∂θ
= ∆θ + div θ(∇φ + ∇Ψ) .
∂t
(3.11)
The latter equation can be identified as the evolution of the densities (w.r.t. dvolg ) of
the diffusion
√
(3.12)
dX(t) = −∇ Ψ X(t) + φ X(t) + 2dB(t),
where B denotes a Brownian motion defined on (M,g). Naturally, the Dµ -L2 flow has
the same associated Fokker-Planck equation (3.10) and diffusion process (3.12).
3.3.2. Jχ2 -Wasserstein The PDE satisfied (formally) by the densities
ρ̃t :=
dµt
,
dπ
of the Jχ2 -Wasserstein flow t ∈ (0,∞) 7→ µt is the (weighted) porous medium equation:
∂ ρ̃
= ∆π ρ̃2 + divπ (ρ̃∇φ),
∂t
(3.13)
where the weighted Laplacian and divergence are defined formally as
∆π f := ∆f − h∇f,∇Ψi,
divπ F := divF − hF,∇Ψi.
(3.14)
Consider now the stochastic process {ut }t≥0 formally defined as
√
1/2
dBt ,
dut = − ρ̃(t,ut )∇Ψ(ut ) + ∇φ(ut ) dt + 2 ρ̃(t,ut )
u 0 ∼ ρ0 ,
(3.15)
where ρ̃ is the solution to (3.13). Let θt be the evolution of the densities (with respect
to dvolg ) of the above diffusion. Then a formal computation shows that θ satisfies the
Fokker-Plank equation:
∂θ
= −div θ −ρ̃∇Ψ − ∇φ + ∆(ρ̃θ).
∂t
If we let β = Z1 exp(−Ψ)θ we see, using (3.14), that
∆π β 2 = eΨ ∆(β 2 e−Ψ ) + div (β 2 e−Ψ )∇Ψ ,
divπ (β∇φ) = eΨ div (βe−Ψ )∇Ψ ,
implying that the distributions of the stochastic process (3.15) are those generated by
the gradient flow of Jχ2 in Wasserstein space.
Remark 3.7. In contrast with the Langevin diffusion (3.12), the process (3.15) is defined in terms of the solution of the equation satisfied by its densities. In particular,
if one wanted to simulate (3.15) one would need to know the solution of (3.13) before
hand.
4. Applications
In this section we show the usefulness of the theory in three settings: Bayesian
semi-supervised learning in subsection 4.1; the Bayesian formulation of the ROF total
variation denoising model in subsection 4.2; and the choice of metric in Riemannian
MCMC in subsection 4.3.
N. Garcı́a Trillos, and D. Sanz-Alonso
13
4.1. Bayesian Formulation of Semi-supervised Learning
In semisupervised learning one is interested in the following task: given a data cloud X =
{x1 ,...,xn } together with (noisy) labels yi ∈ {−1,1} for some of the data points xi ,
i ∈ Z ⊂ {1,...,n}, classify the unlabeled data points by assigning labels ±1 to them.
Here we assume to have access to a weight matrix W quantifying the level of similarity between the points in X. Thus, we focus on the graph-based approach to semisupervised learning, which boils down to propagating the known labels to the whole
cloud, using the geometry of the weighted graph (X,W ). We will investigate the existence and convergence of gradient flows for several Bayesian graph-based classification
models proposed in [3]. In the Bayesian approach, the geometric structure that the
weighted graph imposes on the data cloud is used to build a prior on a latent space,
and the noisy given labels are used to build the likelihood. The Bayesian solution to
the classification problem is a measure on the latent space, that is then push-forwarded
into a measure on the label space {−1,1}n. This latter measure contains information
on the most likely labels, and also provides a principled way to quantify the remaining
uncertainty on the classification process.
Let (X,W ) then be a weighted graph, where X = {x1 ,...,xn } is the set of nodes of
the graph and W is the weight matrix between the points in X. All the entries of W
are non-negative real numbers and we assume that W is symmetric. Let L be the graph
Laplacian matrix defined by
L := D − W,
where D is the degreePmatrix of the weighted graph, i.e., the diagonal matrix with
n
diagonal entries Dii = j=1 Wij . The above corresponds to the unnormalized graph
Laplacian, but different normalizations are possible [34]. The graph-Laplacian will
be used in all the models below to favor prior draws of the latent variables that are
consistent with the geometry of the data cloud.
4.1.1. Probit and Logistic Models Traditionally, the probit approach to semisupervised learning is to classify the unlabeled data points by first optimizing the functional G : Rn → R given by
Z w
X
1
log H(yj uj );γ , H(w;γ) :=
exp(−t2 /2γ 2 )dt
(4.1)
G(u) := hLu,ui −
2
−∞
j∈Z
P
over all u ∈ Rn satisfying ni=1 ui = 0, and then thresholding the optimizer with the sign
function. The minimizer of the functional G can be interpreted as the MAP (maximum
a posteriori estimator) in the Bayesian formulation of probit semi-supervised learning
(see [3]) that we now recall:
P
Prior: Consider the subspace U := {u ∈ Rn : ni=1 ui = 0} and let π be the Gaussian
measure on U defined by
1
dπ
(u) ∝ exp − hLu,ui =: exp −Ψ(u) .
(4.2)
du
2
The measure π is interpreted as a prior distribution on the space of real valued functions
on the point cloud X with average zero.
Likelihood function: For a fixed u ∈ U and for j ∈ Z define
yj = S(uj + ηj ),
14
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
where the ηj are i.i.d. N (0,γ 2 ), and S is the sign function. This specifies the distribution
of observed labels given the underlying latent variable u. We then define, for given data
y, the negative log-density function
X
φ(u;y) := −
log H(yj uj ;γ) , u ∈ U,
(4.3)
j∈Z
where H is given by (4.1).
Posterior distribution: As shown in [3], a simple application of Bayes’ rule gives
the posterior distribution of u given y (denoted by µy ):
dµy
(v) ∝ exp −Ψ(v) − φ(v;y) ,
dv
dµy
(v) = exp −φ(v;y) ;
dπ
where Ψ is given by (4.2), and φ is given by (4.3).
From what has been discussed in the previous sections, the posterior µy can be
characterized as the unique minimizer of the energy
Z
JKL (ν) := DKL (ν||π) +
φ(u;y)dν(u), ν ∈ P(Rd ).
Rn
Moreover, we can study the geodesic convexity of this functional by studying independently the convexity properties of DKL (ν||π) and of φ(·;y). Precisely:
i) Since π is a Gaussian measure with covariance L−1 , Example 3.2 shows that
DKL (ν||π) is Λmin (L)-geodesically convex in Wasserstein space, where Λmin(L) is
the smallest non-trivial eigenvalue of L.
ii) The function
φ(·;y) is convex —see the appendix of [3]. Hence, the functional
R
FKL (ν) = Rn φ(u;y)dν(u) is 0-geodesicaly convex in Wasserstein space.
It then follows from Proposition 3.1 that JKL is Λmin (L)-geodesically convex in Wasserstein space. As a consequence, if we consider t ∈ [0,∞) 7→ µt, the gradient flow of JKL
with respect to the Wasserstein distance starting at µ0 (an absolutely continuous measure with respect to µ), we obtain from (3.6) that
DKL (µt ||µy ) ≤ e−t/Λ DKL (µ0 ||µy ),
∀t ≥ 0,
y 2
y
−tΛ
y 2
y
−t/Λ
W2 (µt ,µ ) ≤ ΛDKL (µt ||µ ) ≤ Λe
dTV (µt ,µ ) ≤ 2DKL (µt ||µ ) ≤ 2e
DKL (µ0 ||µy ),
y
DKL (µ0 ||µ ),
∀t ≥ 0,
∀t ≥ 0.
Remark 4.1. A special case of a weighted graph (X,W ) frequently found in the literature is that in which the points in X are i.i.d. points sampled from some distribution
on a manifold M embedded in Rd , and the similarity matrix W is obtained as
|xi − xj |
Wij = K
.
ε
In the above, K is a compactly supported kernel function, |xi − xj | is the Euclidean
distance between the points xi and xj , and ε > 0 is a parameter controlling data density.
It can be shown (see [5] and [13]) that the smallest non-trivial eigenvalue of a rescaled
version of the resulting graph Laplacian is close to the smallest non-trivial eigenvalue of
a weighted Laplacian on the manifold. This in particular guarantees a robust behavior
of Langevin diffusions in the large n limit.
15
N. Garcı́a Trillos, and D. Sanz-Alonso
Remark 4.2. In the above we have considered a probit model for the likelihood function.
The ideas generalize straightforwardly to other settings, notably the logistic model
X
φ(u;y) := −
log σ(yj uj ;γ) , u ∈ U,
(4.4)
j∈Z
where
σ(t;γ) :=
1
.
1 + e−t/γ
The convexity of φ for the logistic model (4.4) can be established by direct computation
of the second derivative of σ.
4.1.2. Ginzburg-Landau Model In what follows consider the 1d function
Wǫ (t) :=
1 2
(t − 1)2 ,
4ǫ
t ∈ R.
The Ginzburg-Landau model consist of:
Prior:
1
X
dπ
(u) ∝ exp − hu,Lui −
Wǫ uj =: exp −Ψ(u) ,
dv
2
j∈Z
u ∈ U.
(4.5)
Likelihood function: For j ∈ Z,
yj = uj + ηj ,
ηj ∼ N (0,γ 2 ).
This leads to the following negative log-density function:
φ(u;y) =
1 X
|yj − uj |2 .
2γ 2
(4.6)
j∈Z
Posterior distribution: Combining the prior and the likelihood via Bayes’ formula gives the posterior distribution
dµy
(u) = exp −φ(u;y) ;
dπ
dµy
(u) ∝ exp −Ψ(u) − φ(u;y) ,
du
u ∈ U,
where Ψ is given by (4.5), and φ is given by (4.6).
For this model, the negative prior log-density Ψ is not convex, and Wasserstein
λ-geodesic convexity of the functional DKL (·kπ) only holds for negative λ. In particular,
it is not possible to deduce exponential decay taking the Wasserstein flow point of view.
If instead we take a look at the L2 /Dirichlet energy flow point of view we can still
deduce exponential convergence towards the posterior µy . Indeed, because the negative
log-likelihood of µy satisfies:
|∇Ψ(u) + ∇φ(u;y)|2
− ∆Ψ(u) − ∆φ(u;y) → ∞,
2
as |u| → ∞,
there exists some λ > 0 for which µy has a Poincaré inequality with constant λ (see
Chapter 4.5 in [24]). In this example we can say more, and in particular we are able to
find a Poincaré constant that depends explicitly on ε, the smallest non-trivial eigenvalue
of L, and k := |Z|.
16
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
P
Let ψ(u) := j∈Z Wε (uj ) and let ψc be its convex envelope, i.e. let ψc be the largest
convex function that is below ψ. It is straightforward to show that ψc (0) = 0 and that
k
infn {exp(ψc (u) − ψ(u))} = exp −
.
u∈R
4ε
Consider now the probability measure µc with Lebesgue density
1
1
exp − hLu,ui − φ(u;y) − ψc (u) du,
dµc (u) =
Zc
2
and define λ2 and λ2,c as in Remark 3.6 using µy and µc instead of µ. For any given
f ∈ L2 (µ) we then have
R
R
R
|∇f |2 dµ
|∇f |2 dµ
|∇f |2 dµc
k
R
R
R
≥
exp
−
,
≥
ε
|f − fµ |2 dµ
|f − fµc |2 dµ
|f − fµc |2 dµc
R
where the first inequality follows from the fact that fµ = argmina∈R |f − a|2 dµ and the
second inequality follows directly from the fact that 0 ≥ ψc − ψ ≥ − kε . It follows that
k
k
λ2,c ≥ exp −
Λmin (L).
λ2 ≥ exp −
ε
ε
where the last inequality follows from the fact that the negative log-likelihood of µy
satisfies the Bakry-Emery condition with constant Λmin(L) (see Chapter 4.5 in [24]).
Clearly, the Poincaré constant above is very large in the small ǫ limit.
4.2. Application to Image Analysis: a Bayesian Formulation of the ROF
Model Let (X,W ) be a weighted graph as introduced in the previous section, with
X = {x1 ,...,xm }. In imaging, the graph is usually taken to be a regular grid on the unit
square. Instead of working with the Dirichlet energy for a prior, define the graph total
variation for a vector u ∈ Rm as
X
Wi,j |ui − uj |.
GT V (u) :=
i,j
In the ROF model for image denoising (Rudin-Osher-Fatemi model [29]) one seeks for
a solution to the optimization problem:
min GT V (u) +
u
1
|u − y|2 .
2γ
The optimizer of the above energy is seen as the MAP for a posterior distribution from
a model that we now describe.
Prior: We consider the prior π in terms of a change of measure with respect to
Lebesgue measure given by
X
Wi,j |ui − uj |.
(4.7)
dπ(u) ∝ exp −Ψ(u) du, Ψ(u) := GT V (u) =
i,j
The prior represents our prior beliefs of what an image looks like.
Noise Model: Suppose that random noise contaminates a clean image u ∈ Rm
producing a noisy version of it
yi = ui + ηi ,
ηi ∼ N (0,γ 2 ) i.i.d.
17
N. Garcı́a Trillos, and D. Sanz-Alonso
This results in the following negative log-density function:
φ(u;y) :=
1
|u − y|2 .
2γ 2
(4.8)
Posterior: Given our prior beliefs on u and given the noise model just introduced,
the posterior distribution of u given y can be written as
dµy
(v) ∝ exp −Ψ(v) − φ(v;y)
dv
dµy
(v) = exp −φ(v;y) ;
dπ
with Ψ given by (4.7), and φ given by (4.8).
We notice that since Ψ is a convex function in Rm , the functional DKL (·||π) is 0geodesically convex in the Wasserstein space (P2 (Rm ),W2 ). Moreover, from the fact
that the function 2γ1 2 |u − y|2 is γ12 convex, we conclude that the functional DKL (·||µy )
is 1/γ 2-geodesically convex in the Wasserstein space. As a consequence, if we consider
t ∈ [0,∞)] 7→ µt , the gradient flow of JKL with respect to the Wasserstein distance starting
at µ0 (an absolutely continuous measure with respect to µy ), we obtain from (3.6) that:
DKL (µt ||µy ) ≤ e−t/γ DKL (µ0 ||µy ),
∀t ≥ 0,
y 2
y
−tγ
y 2
y
−tγ
W2 (µt ,µ ) ≤ γDKL (µt ||µ ) ≤ γe
dTV (µt ,µ ) ≤ 2DKL (µt ||µ ) ≤ 2e
DKL (µ0 ||µy ),
y
DKL (µ0 ||µ ),
∀t ≥ 0,
∀t ≥ 0.
4.3. Applications to Sampling and Riemannian MCMC So far we have
treated the Riemannian manifold (M,g) as fixed. In this section we take a different
perspective and treat the metric g as a free parameter. We do this motivated by the so
called Riemannian MCMC methods for sampling, where a change of metric in the base
space is introduced in order to produce Langevin-type proposals that are adapted to the
geometric features of the target, thereby exploring regions of interest and accelerating
the convergence of the chain to the posterior. There are different heuristics regarding
the choice of metric (see [14]), but no principled way to compare different metrics and
rank their performance for sampling purposes. With the developments presented in this
paper we propose one such principled criterion as we describe below. Here, as in the rest
of the paper, we focus on the analysis of diffusion processes and not their discretizations.
We restrict our attention to the case M = Rm .
Let g be a Riemannian metric tensor on Rm defined via
gx (u,v) := hG(x)u,vi,
u,v ∈ Tx M,
where for every x ∈ Rm , G(x) is a m × m positive definite matrix. In what follows we
identify g with G and refer to both as ‘the metric’ and we use terms such as g-geodesic, gWassertein distance, etc. to emphasize that the notions considered are being constructed
using the metric g. Let dg be the distance induced by the metric tensor g and let volg
be the associated volume form. Notice that in terms of the Lebesgue measure and the
metric G, we can write
q
dvolg (x) = det G(x) dx.
We use the canonical basis for Rm as global chart for Rm and consider the canonical
vector fields ∂x∂ 1 ,..., ∂x∂m . The Christoffel symbols associated to the Levi-Civita connection of the Riemannian manifold (Rm ,g) can be written in terms of derivatives of the
18
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
metric as
Γlij =
1
2
∂
∂
∂
Gki +
Gkj −
Gij G−1
lk ,
∂xj
∂xi
∂xk
(4.9)
where in the right hand-side —and in what follows— we use Einstein’s summation
convention. The proof of the following result is in the Appendix.
Theorem 4.3. Let F ∈ C 2 (Rm ) and
µ(du) ∝ exp −F (u) du.
The sharp constant λ for which JKL (or DKL (·||µ)) is λ-geodesically convex in the gWasserstein distance is equal to
λG := infm Λmin G−1/2 B + HessF − C G−1/2 ,
x∈R
where HessF is the usual (Euclidean) Hessian matrix of F, B is the matrix with coordinates
Bij :=
∂ l
Γ − Γlij Γlml ,
∂xl ij
(4.10)
and C is the matrix with coordinates
Cij := Γlij
∂F
.
∂xl
(4.11)
Moreover, for any a > 0,
1
λaG = λG .
a
(4.12)
Note that λG is a key quantity in evaluating the quality of a metric G in building
geometry-informed Langevin diffusions for sampling purposes, as it gives the exponential rate at which the evolution of probabilities built using the metric G converges
towards the posterior: larger λG corresponds to faster convergence. However, in order
to establish a fair performance comparison, the metrics need to be scaled appropriately.
Indeed a faster rate can be obtained by scaling up the metric (which can be thought of
as time-rescaling), as it is clearly seen by the scaling property (4.12) of the functional
λG . We suggest that a fair comparison can be established by allowing only for metrics
G with
sup Λmax G(x) ≥ 1.
(4.13)
x∈Rm
Remark 4.4. The functional λG can be used to determine the optimal metric among
a certain subclass of metrics of interest satisfying the condition (4.13). For instance, it
may be of interest to find the optimal constant metric G (see Proposition 4.5 below), or
to find the best metric within a finite family of metrics.
To illustrate the previous remark we show that for a Gaussian target measure the
optimal preconditioner is, unsurprisingly, given by the Fisher information. More precisely we have the following proposition:
N. Garcı́a Trillos, and D. Sanz-Alonso
19
Proposition 4.5. Let µ = N (0,Σ). Then
G∗ := Λmin(Σ)Σ−1
(4.14)
maximizes λG over the class of constant metrics G satisfying Λmax (G) ≥ 1, as in (4.13).
Moreover, the maximum value is
λG∗ =
1
.
Λmin(Σ)
Proof. Note that for constant metric G the geodesics are the same as in Euclidean space,
and that the the Christoffel symbols are identically equal to zero. Therefore using that
HessF = Σ−1 we have that
λG = Λmin(G−1/2 Σ−1 G−1/2 ).
Note also that G∗ given by (4.14) is feasible, since
Λmax (G∗ ) = 1.
(4.15)
Suppose now for the sake of contradiction that there is feasible G such that λG > λG∗ .
This implies that for any nonzero u ∈ Rm
hG−1/2 Σ−1 G−1/2 u,ui
≥ λG > λG∗ .
hu,ui
(4.16)
Let ψ1 ,...,ψm be an orthonormal (in the Euclidean sense) basis of eigenvectors of G
with corresponding eigenvalues g1 ,...,gm . Then from (4.15) and (4.16) we have that, for
any 1 ≤ k ≤ m,
1
Λmin(Σ)
= λG∗ <
hG−1/2 Σ−1 G−1/2 ψk ,ψk i
hψk ,ψk i
= hΣ−1 G−1/2 ψk ,G−1/2 ψk i
−1/2
−1/2
= hΣ−1 gk ψk ,gk ψk i
1
1
1
= hΣ−1 ψk ,ψk i ≤
.
gk
gk Λmin (Σ)
Hence gk < 1 for all k, which contradicts the fact that G was feasible. This completes
the proof.
Acknowledgments. The authors are thankful to Matı́as Delgadino for pointing
out to us the reference [22] while participating in the CNA Ki-net workshop “Dynamics and Geometry from High Dimensional Data” that took place at Carnegie Mellon
University in March 2017.
REFERENCES
[1] L. Ambrosio, N. Gigli, and G. Savaré. Gradient flows: in metric spaces and in the space of
probability measures. Springer Science & Business Media, 2008.
[2] H. Attias. A Variational Bayesian Framework for Graphical Models. In NIPS, volume 12, 1999.
[3] A. L. Bertozzi, X. Luo, A. M. Stuart, and K. C. Zygalakis. Uncertainty quantification in the
classification of high dimensional data.
20
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
[4] J. E. Besag. Comments on “representations of knowledge in complex systems” by u. grenander
and m. i. miller. J. Roy. Statist. Soc. Ser. B, 56:591–592, 1994.
[5] D. Burago, S. Ivanov, and Y. Kurylev. A graph discretization of the laplace-beltrami operator.
arXiv preprint arXiv:1301.2222, 2013.
[6] P. Caputo, P. Dai Pra, and G. Posta. Convex entropy decay via the Bochner–Bakry–Emery
approach. In Annales de l’IHP Probabilités et statistiques, volume 45, pages 734–753, 2009.
[7] M. P. do Carmo Valero. Riemannian Geometry. 1992.
[8] P. Dupuis and R. S. Ellis. A weak convergence approach to the theory of large deviations, volume
902. John Wiley & Sons, 2011.
[9] M. Erbar and J. Maas. Ricci curvature of finite Markov chains via convexity of the entropy.
Archive for Rational Mechanics and Analysis, pages 1–42, 2012.
[10] M. Erbar, J. Maas, and P. Tetali. Discrete Ricci curvature bounds for Bernoulli-Laplace and
random transposition models. arXiv preprint arXiv:1409.8605, 2014.
[11] M. Fathi and J. Maas. Entropic Ricci curvature bounds for discrete interacting systems. The
Annals of Applied Probability, 26(3):1774–1806, 2016.
[12] C. W. Fox and S. J. Roberts. A tutorial on variational Bayesian inference. Artificial intelligence
review, 38(2):85–95, 2012.
[13] N . Garcia Trillos, M . Gerlach, M. Hein, and D Slepcev. Spectral convergence of empirical graph
laplacians. In preparation, 2017.
[14] M. Girolami and B. Calderhead. Riemann manifold langevin and hamiltonian monte carlo methods. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(2):123–
214, 2011.
[15] N. Gozlan, C. Roberto, and P. M. Samson. From concentration to logarithmic sobolev and
poincaré inequalities. Journal of Functional Analysis, 260(5):1491–1522, 2011.
[16] R. Jordan and D. Kinderlehrer. 18, an extended variational principle. Partial differential equations
and applications: collected papers in honor of Carlo Pucci, page 187, 1996.
[17] R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the fokker–planck equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998.
[18] C. Kipnis and S. R. S. Varadhan. Central limit theorem for additive functionals of reversible
Markov processes and applications to simple exclusions. Communications in Mathematical
Physics, 104(1):1–19, 1986.
[19] M. Ledoux. Concentration of measure and logarithmic Sobolev inequalities. In Seminaire de
probabilites XXXIII, pages 120–216. Springer, 1999.
[20] Y. Marzouk, T. Moselhy, M. Parno, and A. Spantini. An introduction to sampling via measure
transport. arXiv preprint arXiv:1602.05023, 2016.
[21] R. J. McCann. A convexity principle for interacting gases. Advances in mathematics, 128(1):153–
179, 1997.
[22] S. Ohta and A. Takatsu. Displacement convexity of generalized relative entropies. Advances in
Mathematics, 228(3):1742–1787, 2011.
[23] F. Otto. The geometry of dissipative evolution equations: the porous medium equation. 2001.
[24] G. A. Pavliotis. Stochastic processes and applications. Texts in Applied Mathematics. Springer,
Berlin, 2014.
[25] F. J. Pinski, G. Simpson, A. M. Stuart, and H. Weber. Kullback–leibler approximation for probability measures on infinite dimensional spaces. SIAM Journal on Mathematical Analysis,
47(6):4091–4122, 2015.
[26] S. Reich. A nonparametric ensemble transform method for Bayesian inference. SIAM Journal on
Scientific Computing, 35(4):A2013–A2024, 2013.
[27] G. O. Roberts and J. S. Rosenthal. Optimal scaling of discrete approximations to Langevin diffusions. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 60(1):255–
268, 1998.
[28] G. O. Roberts and R. L. Tweedie. Exponential convergence of Langevin distributions and their
discrete approximations. Bernoulli, pages 341–363, 1996.
[29] L. I. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise removal algorithms.
Physica D: Nonlinear Phenomena, 60(1-4):259–268, 1992.
[30] F. Santambrogio. Optimal transport for applied mathematicians. Birkäuser, NY, 2015.
[31] K. T. Sturm. On the geometry of metric measure spaces. Acta mathematica, 196(1):65–131, 2006.
[32] C. Villani. Topics in optimal transportation. Number 58. American Mathematical Soc., 2003.
[33] C. Villani. Optimal transport: old and new, volume 338. Springer Science & Business Media,
2008.
[34] U. Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007.
[35] M. K. von Renesse and K. T. Sturm. Transport inequalities, gradient estimates, entropy and
Ricci curvature. Communications on pure and applied mathematics, 58(7):923–940, 2005.
N. Garcı́a Trillos, and D. Sanz-Alonso
21
[36] M. J. Wainwright and M. I. Jordan. Graphical models, exponential families, and variational
inference. Foundations and Trends R in Machine Learning, 1(1–2):1–305, 2008.
Appendix A. Proof of Theorem 4.3.
Notice that the measure µ can be rewritten as
p
dµ(x) ∝ exp −F (x) − log
det(G(x))
dvolg (x) = exp(−Fg (x))dvolg (x),
where
Fg (x) := F (x) + log
p
det(G) ,
x ∈ Rm .
From Proposition 3.1 the sharp constant λ for which DKL (·||µ) is λ-geodesically convex
in the g-Wasserstein space is given by
inf
min
x∈Rm v:g(v,v)=1
Ricg (v) + Hessg Fg (v,v),
where Ricg and Hessg stand for the Ricci curvature and Hessian in the gmetric. To establish the proposition, it suffices to show that for any given x ∈
Rm , minv:g(v,v)=1 Ricg (v) + Hessg Fg is equal to the smallest eigenvalue of the matrix
B + HessF − C.
Let us start by recalling that the g-Hessian of a C 2 function I, denoted by Hessg I,
is the symmetric (2,0)-tensor satisfying
Hessg I(v,v) =
d2
I γ(t)
dt2
t=0
,
for every v ∈ Rm and every constant speed g-geodesic curve with γ(0) = x and γ̇(0) = v.
It is convenient to rewrite Hessg I(v,v) in terms of the Christoffel symbols of the metric
g, the Euclidean inner product, and the regular (Euclidean) gradient and Hessian of the
function I. Let γ : (−ε,ε) → Rm be a constant speed g-geodesic with γ(0) = x, γ̇(0) = v.
We can then write:
d
d2
∇I(γ(t)),
γ̇(t)
I
γ(t)
=
dt2
dt
(1.1)
D
E D
E
= HessI γ(t) γ̇(t), γ̇(t) + ∇I(γ(t)), γ̈(t) ,
where ∇I and HessI are the usual gradient and Hessian matrix of I, respectively. The
acceleration of the curve γ can be written in terms of the Christoffel symbols. Namely,
if we write γ in coordinates as
γ(t) = γ1 (t),...,γm (t) , t ∈ (−ε,ε),
the following system of second order ODEs holds:
X
Γlij γ̇i (t)γ̇j (t), l = 1,...,m.
γ̈l (t) = −
ij
Plugging the geodesic equations back into (1.1), and setting t = 0, it follows that
D
E
d2
I γ(t) |t=0 = HessI − A(I) v,v ,
2
dt
22
Gradient Flows: Applications to Classification, Image Denoising, and Riemannian MCMC
where A(I) is the matrix with coordinates
A(I)ij = Γlij
∂I
.
∂xl
Hence,
Hessg I(v,v) =
D
E
HessI − A(I) v,v ,
∀v ∈ Rm .
(1.2)
Taking I := Fg we can write the coordinate ij of the matrix HessI − A(I) as
p
p
∂2F
∂2
∂
∂F
+
log
− Γlij
log
det(G) − Γlij
det(G) .
∂xi xj ∂xi xj
∂xl
∂xl
Using now the fact that the g-divergence of the vector field
of the Christoffel symbols as
we deduce that
∂
∂xl
can be written in terms
p
∂
det(G) = Γklk ,
log
∂xl
∂2F
∂Γl
∂F
HessI − A(I) ij =
+ il − Γlij
− Γlij Γklk
∂xi xj ∂xj
∂xl
On the other hand, the Ricci curvature Ricg (v) can be written in terms of the Christoffel
symbols and its derivatives (alternatively in terms of the metric and its first and second
order derivatives) as
Ricg (v) = hRv,vi,
v ∈ Rm ,
(1.3)
where R is the (symmetric) matrix with entries
Rij =
∂Γlij ∂Γlil
−
+ Γlij Γkkl − Γkil Γljk ,
∂xl
∂xj
see [7] for details. After some cancellations using the symmetry of the symbols, we
obtain that
and so
∂Γlij
∂2F
∂F
− Γkil Γljk +
− Γlij
,
R + HessI − A(I) ij =
∂xl
∂xi ∂xj
∂xl
R + HessI − A(I) = B + HessF − C.
Using (1.2) and (1.3) we deduce that
Ricg (v) + Hessg Fg (v,v) = (B + HessF − C)v,v ,
Therefore the variational problem (for every fixed x ∈ Rm )
min
v:g(v,v)=1
Ricg (v) + Hessg F (v,v)
∀v ∈ Rm .
N. Garcı́a Trillos, and D. Sanz-Alonso
23
can be rewritten, applying the change of variables w := G1/2 v, as
E
D
G−1/2 B + HessF − C G−1/2 w,w .
min
v:hw,wi=1
In turn this coincides with
Λmin G−1/2 B + HessF − C G−1/2 ,
i.e., the smallest eigenvalue of the matrix
G−1/2 B + HessF − C G−1/2 .
This concludes the proof of the first part of the theorem.
e = aG. By definition
Now we show the scaling property (4.12). Let G
−1/2
e + HessF − C̃ G
e
e−1/2 B
,
λGe = infm Λmin G
x∈R
e and C
e are defined as in (4.10) and (4.11) but in terms of the metric G̃. From
where B
the expression (4.9) for the Christoffel symbols, it follows that they are invariant under
e and C, Ce depend on the metric only through the
rescaling of the metric and, since B, B
e
e
symbols, we deduce that B = B, C = C. Therefore,
1
infm Λmin G−1/2 B + Hess(F ) − C G−1/2
a x∈R
1
= λG .
a
λGe =
| 10math.ST
|
arXiv:1712.06957v3 [physics.med-ph] 7 Jan 2018
MURA Dataset: Towards Radiologist-Level Abnormality Detection
in Musculoskeletal Radiographs
Pranav Rajpurkar * 1 Jeremy Irvin * 1 Aarti Bagul 1 Daisy Ding 1 Tony Duan 1
Hershel Mehta 1 Brandon Yang 1 Kaylie Zhu 1 Dillon Laird 1 Robyn L. Ball 2
Curtis Langlotz 3 Katie Shpanskaya 3 Matthew P. Lungren 3 Andrew Ng 1
Abstract
We introduce MURA, a large dataset of musculoskeletal radiographs containing 40,895
images from 14,982 studies, where each study
is manually labeled by radiologists as either
normal or abnormal. On this dataset, we
train a 169-layer densely connected convolutional network to detect and localize abnormalities. To evaluate our model robustly and
to get an estimate of radiologist performance,
we collect additional labels from boardcertified Stanford radiologists on the test set,
consisting of 209 musculoskeletal studies. We
compared our model and radiologists on the
Cohen’s kappa statistic, which expresses the
agreement of our model and of each radiologist with the gold standard, defined as
the majority vote of a disjoint group of radiologists. We find that our model achieves
performance comparable to that of radiologists. Model performance is higher than the
best radiologist performance in detecting abnormalities on finger studies and equivalent
on wrist studies. However, model performance is lower than best radiologist performance in detecting abnormalities on elbow,
forearm, hand, humerus, and shoulder studies, indicating that the task is a good challenge for future research. To encourage advances, we have made our dataset freely available at https://stanfordmlgroup.github.
io/projects/mura.
*
1
Equal contribution
Stanford University Department of Computer Science 2 Stanford University Department of Medicine 3 Stanford University Department of Radiology.
Correspondence to:
Pranav
Rajpurkar <[email protected]>, Jeremy Irvin
<[email protected]>.
Study (Elbow)
Study (Humerus)
Normal
Study (Finger)
Abnormal
Study (Forearm)
Abnormal
Abnormal
Figure 1. The MURA dataset contains 14,982 musculoskeletal studies of the upper extremity, where each study
contains one or more views and is manually labeled by
radiologists as either normal or abnormal. These examples show a normal elbow study (top left), an abnormal
humerus study with a fracture (top right), an abnormal
finger study with degenerative changes (bottom left), and
an abnormal forearm study (bottom right) demonstrating
operative plate and screw fixation of radial and ulnar fractures.
1. Introduction
Large, high-quality datasets have played a critical role
in driving progress of fields with deep learning methods
(Deng et al., 2009). To this end, we introduce MURA,
a large dataset of radiographs, containing 14,982 musculoskeletal studies of the upper extremity. Each study
contains one or more views (images) and is manually
labeled by radiologists as either normal or abnormal.
The abnormality detection task, or in other words determining whether a radiographic study is normal or
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
Study
Normal
Abnormal
Total
Elbow
Finger
Forearm
Hand
Humerus
Shoulder
Wrist
1,203
1,389
677
1,613
411
1,479
2,295
768
753
380
602
367
1,594
1,451
1,971
2,142
1,057
2,215
778
3,073
3,746
Total
9,067
5,915
14,982
Table 1. MURA contains 9,067 normal and 5,915 abnormal
musculoskeletal radiographic studies of the upper extremity including the shoulder, humerus, elbow, forearm, wrist,
hand, and finger. MURA is one of the largest public radiographic image datasets.
Study (Elbow)
View 1
View 2
169-layer
Convolutional Neural Network
Arithmetic mean of output probabilities
Probability of Abnormality
abnormal, is a critical radiological task: a study interpreted as normal rules out disease and can eliminate the need for patients to undergo further diagnostic procedures or interventions. Musculoskeletal conditions affect more than 1.7 billion people worldwide
(BMU, 2017), and are the most common cause of severe, long-term pain and disability (Woolf & Pfleger,
2003), with 30 million emergency department visits annually and increasing. Our dataset, MURA, contains
9,067 normal and 5,915 abnormal musculoskeletal radiographic studies of the upper extremity including the
shoulder, humerus, elbow, forearm, wrist, hand, and
finger. MURA is one of the largest public radiographic
image datasets.
On MURA, we developed a competitive abnormality
detection model. The model takes as input one or more
views for a study of an upper extremity. On each view,
a 169-layer convolutional neural network predicts the
probability of abnormality; the per-view probabilities
are then averaged to output the probability of abnormality for the study.
To evaluate models robustly and to get an estimate
of radiologist performance, we collected six additional
labels from board-certified radiologists on a holdout
test set of 209 studies. We compare the performance
of our model and radiologists, and find that our model
achieves performance comparable to that of radiologists. Model performance is higher than the best radiologist’s performance in detecting abnormalities on
finger studies, and equivalent in wrist studies. However, model performance is lower than best radiologist’s performance in detecting abnormalities on elbow, forearm, hand, humerus, and shoulder studies.
We have made our dataset freely available to encourage advances in medical imaging models.
Figure 2. The model takes as input one or more views for
a study. On each view, our 169-layer convolutional neural
network predicts the probability of abnormality; the perview probabilities are then averaged to output the probability of abnormality for the study.
2. MURA
The MURA abnormality detection task is a binary classification task, where the input is an upper
exremity radiograph study — with each study containing one or more views (images) — and the expected
output is a binary label y ∈ {0, 1} indicating whether
the study is normal or abnormal, respectively.
2.1. Data Collection
Our institutional review board approved study collected de-identified, HIPAA-compliant images from
the Picture Archive and Communication System
(PACS) of Stanford Hospital. We assembled a dataset
of musculoskeletal radiographs consisting of 14,982
studies from 12,251 patients, with a total of 40,895
multi-view radiographic images. Each belongs to one
of seven standard upper extremity radiographic study
types: elbow, finger, forearm, hand, humerus, shoulder, and wrist. Table 1 summarizes the distribution of
normal and abnormal studies.
Each study was manually labeled as normal or abnormal by board-certified radiologists from the Stanford
Hospital at the time of clinical radiographic interpretation in the diagnostic radiology environment between
2001 and 2012. The labeling was performed during interpretation on DICOM images presented on at least
3 megapixel PACS medical grade display with max
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
Radiologist 1
(0.770,
(0.249,
(0.772,
(0.623,
(0.850,
(0.847,
(0.766,
Radiologist 2
Elbow
Finger
Forearm
Hand
Humerus
Shoulder
Wrist
0.795
0.304
0.796
0.661
0.867
0.864
0.791
0.819)
0.358)
0.821)
0.698)
0.883)
0.881)
0.817)
Overall
0.724 (0.719, 0.728)
0.733
0.403
0.802
0.927
0.733
0.791
0.931
(0.703,
(0.339,
(0.779,
(0.917,
(0.703,
(0.765,
(0.922,
Radiologist 3
0.764)
0.467)
0.825)
0.937)
0.764)
0.816)
0.940)
0.766 (0.762, 0.770)
0.738
0.410
0.798
0.789
0.933
0.864
0.931
(0.709,
(0.358,
(0.774,
(0.762,
(0.925,
(0.847,
(0.922,
Model
0.767)
0.463)
0.822)
0.815)
0.942)
0.881)
0.940)
0.780 (0.777, 0.784)
0.733
0.525
0.670
0.780
0.800
0.795
0.931
(0.703,
(0.476,
(0.634,
(0.752,
(0.776,
(0.770,
(0.922,
0.764)
0.573)
0.705)
0.809)
0.824)
0.819)
0.940)
0.749 (0.744, 0.753)
Table 2. We compare radiologists and our model on the Cohen’s kappa statistic, which expresses the agreement of each
radiologist/model with the gold standard, defined as the majority vote of a disjoint group of radiologists. We highlight
the best (green) and worst (red) performances on each of the study types and in aggregate. On finger studies, model
performance is significantly higher than the best radiologist performance. On wrist studies, model performance is equivalent to the best radiologist performance. On hand, humerus, and shoulder studies, model performance is higher than the
worst radiologist performance but lower than the best radiologist performance. On elbow and shoulder studies, model
performance is comparable to the worst radiologist performance. On forearm studies, model performance is significantly
worse than the worst radiologist performance.
luminance 400 cd/m2 and min luminance 1 cd/m2
with pixel size of 0.2 and native resolution of 1500
x 2000 pixels. The clinical images vary in resolution
and in aspect ratios. We split the dataset into training (11,255 patients, 13,565 studies, 37,111 images),
validation (788 patients, 1,208 studies, 3,225 images),
and test (208 patients, 209 studies, 559 images) sets.
There is no overlap in patients between any of the sets.
2.2. Test Set Collection
To evaluate models and get a robust estimate of radiologist performance, we collected additional labels from
board-certified Stanford radiologists on the test set,
consisting of 209 musculoskeletal studies. The radiologists individually retrospectively reviewed and labeled each study in the test set as a DICOM file as
normal or abnormal in the clinical reading room environment using the PACS system. The radiologists
have 8.83 years of experience on average ranging from
2 to 25 years. The radiologists did not have access to
any clinical information. Labels were entered into a
standardized data entry program.
2.3. Abnormality Analysis
To investigate the types of abnormalities present in the
dataset, we reviewed the radiologist reports to manually label 100 abnormal studies with the abnormality
finding: 53 studies were labeled with fractures, 48 with
hardware, 35 with degenerative joint diseases, and 29
with other miscellaneous abnormalities, including lesions and subluxations.
3. Model
The model takes as input one or more views for a study
of an upper extremity. On each view, our 169-layer
convolutional neural network predicts the probability
of abnormality. We compute the overall probability
of abnormality for the study by taking the arithmetic
mean of the abnormality probabilities output by the
network for each image. The model makes the binary
prediction of abnormal if the probability of abnormality for the study is greater than 0.5. Figure 2 illustrates
the model’s prediction pipeline.
3.1. Network Architecture and Training
We used a 169-layer convolutional neural network to
predict the probability of abnormality for each image
in a study. The network uses a Dense Convolutional
Network architecture – detailed in Huang et al. (2016)
– which connects each layer to every other layer in a
feed-forward fashion to make the optimization of deep
networks tractable. We replaced the final fully connected layer with one that has a single output, after
which we applied a sigmoid nonlinearity.
For each image X of study type T in the training set,
we optimized the weighted binary cross entropy loss
L(X, y)
= −wT,1 · y log p(Y = 1|X)
−wT,0 · (1 − y) log p(Y = 0|X),
where y is the label of the study, p(Y = i|X) is the
probability that the network assigns to the label i,
wT,1 = |NT |/(|AT | + |NT |), and wT,0 = |AT |/(|AT | +
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
Dataset
Study Type
Label
Images
MURA
Musculoskeletal (Upper Extremity)
Abnormality
41,299
Musculoskeletal (Hand)
Bone Age
14,236
Musculoskeletal (Knee)
K&L Grade
8,892
Musculoskeletal (Left Hand)
Bone Age
1,390
Chest
Multiple Pathologies
112,120
Chest
Multiple Pathologies
7,470
Chest
Abnormality
138
Chest
Tuberculosis
662
Chest
Pulmonary Nodule
247
Mammogram
Breast Cancer
Pediatric Bone Age
(AIMI)
0.E.1
(OAI)
Digital Hand Atlas
(Gertych et al., 2007)
ChestX-ray14
(Wang et al., 2017)
OpenI
(Demner-Fushman et al., 2015)
MC
(Jaeger et al., 2014)
Shenzhen
(Jaeger et al., 2014)
JSRT
(Shiraishi et al., 2000)
DDSM
(Heath et al., 2000)
10,239
Table 3. Overview of publicly available medical radiographic image datasets.
|NT |) where |AT | and |NT | are the number of abnormal images and normal images of study type T in the
training set, respectively.
Before feeding images into the network, we normalized
each image to have the same mean and standard deviation of images in the ImageNet training set. We then
scaled the variable-sized images to 224 × 224. We augmented the data during training by applying random
lateral inversions and rotations.
The weights of the network were initialized with
weights from a model pretrained on ImageNet (Deng
et al., 2009). The network was trained end-to-end
using Adam with default parameters β1 = 0.9 and
β2 = 0.999 (Kingma & Ba, 2014). We trained the
model using minibatches of size 8. We used an initial
learning rate of 0.0001 that is decayed by a factor of 10
each time the validation loss plateaus after an epoch,
and chose the model with the lowest validation loss.
4. Radiologist vs. Model Performance
We assessed the performance of both radiologists and
our model on the test set. Recall that for each study in
the test set, we collected additional normal/abnormal
labels from 6 board-certified radiologists. We randomly chose 3 of these radiologists to create a gold
standard, defined as the majority vote of labels of the
radiologists. We used the other 3 radiologists to get
estimates of radiologist performance on the task.
We compared radiologists and our model on the Cohen’s kappa statistic (κ), which expresses the agreement of each radiologist/model with the gold standard.
We also reported the 95% confidence interval using the
standard error of kappa (McHugh, 2012). Table 2 summarizes the performance of both radiologists and the
model on the different study types and in aggregate.
The radiologists achieved their highest performance on
either wrist studies (radiologist 2) or humerus studies
(radiologists 1 and 3), and their lowest performance on
finger studies. The model also achieved its highest performance on wrist studies and its lowest performance
on finger studies.
We compared the best radiologist performance on each
of the study types against model performance. On
finger studies, the model performance of 0.525 (95%
CI 0.476, 0.573) was significantly higher than the best
radiologist performance of 0.410 (95% CI 0.358, 0.463).
On wrist studies, the model performance of 0.931 (95%
CI 0.922, 0.940) was equivalent to the best radiologist’s
performance. On all other study types, and overall,
model performance was significantly lower than best
radiologist’s performance.
We also compared the worst radiologist performance
on each of the study types against model performance.
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
(a) Three views of the right wrist in a patient with distal radius and ulna
fractures after casting. The model predicted the probability of abnormality in
each view to be greater than 99%, correctly predicting the study as abnormal.
(b) Two radiographic views of the finger
in a patient with a prior fifth proximal
phalynx fracture status post screw and
plate fixation. The model predicted the
probability of abnormality in both views
to be greater than 99.9%, correctly predicting the study as abnormal.
Figure 3. Our model localizes abnormalities it identifies using Class Activation Maps, which highlight the areas of the
radiograph that are most important for making the prediction of abnormality.
On forearm studies, the model performance of 0.670
(95% CI 0.634, 0.705) was significantly lower than
the worst radiologist’s performance of 0.796 (95% CI
0.772, 0.821). On elbow studies, the model performance of 0.733 (95% CI 0.703, 0.764) was not significantly different than the worst radiologist’s performance of 0.733 (95% CI 0.703, 0.764). On all
other study types, and overall, model performance was
higher than the worst radiologist performance, and significantly so except on shoulder studies.
5. Model Interpretation
We visualize the parts of the radiograph which contribute most to the model’s prediction of abnormality by using class activation mappings (CAMs) (Zhou
et al., 2016). We input the radiograph X into the
fully trained network to obtain the feature maps output by the final convolutional layer. To compute the
CAM M (X), we take a weighted average of the feature maps using the weights of the final fully connected
layer. Denote the kth feature map output by the network on image X by fk (X) and the kth fully connected
weight by wk . Formally,
M (X) =
X
k
wk fk (X).
To highlight the salient features in the original radiograph which contribute the most to the network predictions, we upscale the CAM M (X) to the dimensions
of the image and overlay the image. We provide several
examples of CAMs along with radiologist commentary
in Figure 3.
6. Related Work
Large datasets have led to deep learning algorithms
achieving or approaching human-level performance on
tasks such as image recognition (Deng et al., 2009),
speech recognition (Hannun et al., 2014), and question answering (Rajpurkar et al., 2016). Large medical
datasets have led to expert-level performance on detection of diabetic retinopathy (Gulshan et al., 2016),
skin cancer (Esteva et al., 2017), heart arrhythmias
(Rajpurkar et al., 2017a), brain hemorrhage (Grewal
et al., 2017), pneumonia (Rajpurkar et al., 2017b), and
hip fractures (Gale et al., 2017).
There has been a growing effort to make repositories of
medical radiographs openly available. Table 3 provides
a summary of the publicly available datasets of medical radiographic images. Previous datasets are smaller
than MURA in size, with the exception of the recently
released ChestX-ray14 (Wang et al., 2017), which contains 112,120 frontal-view chest radiographs with up
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
to 14 thoracic pathology labels. However, their labels
were not provided directly from a radiologist, but instead automatically generated from radiologists’ text
reports.
There are few openly available musculoskeletal radiograph databases. The Stanford Program for Artificial
Intelligence in Medicine and Imaging hosts a dataset
containing pediatric hand radiographs annotated with
skeletal age (AIMI). The Digital Hand Atlas consists
of left hand radiographs from children of various ages
labeled with radiologist readings of bone age (Gertych
et al., 2007). The OsteoArthritis Initiative hosts the
0.E.1 dataset which contains knee radiographs labeled
with the K&L grade of osteoarthritis (OAI). Each of
these datasets contain less than 15,000 images.
portion of the image that is recognized as abnormal
by the model, drawing the attention of the clinician.
If effective, this could lead to more efficient interpretation of the imaging examination, reduce errors, and
help standardize quality. More studies are necessary
to evaluate the optimal integration of this model and
other deep learning models in the clinical setting.
8. Acknowledgements
We would like to acknowledge the Stanford Program for Artificial Intelligence in Medicine and
Imaging for clinical dataset infrastructure support
(AIMI.stanford.edu).
References
7. Discussion
Abnormality detection in musculoskeletal radiographs
has important clinical applications. First, an abnormality detection model could be utilized for worklist
prioritization. In this scenario, the studies detected as
abnormal could be moved ahead in the image interpretation workflow, allowing the sickest patients to receive
quicker diagnoses and treatment. Furthermore, the
examinations identified as normal could be automatically assigned a preliminary reading of ”normal”; this
could mean (1) normal examinations can be properly
triaged as lower priority on a worklist (2) more rapid
results can be conveyed to the ordering provider (and
patient) which could improve disposition in other areas of the healthcare system (i.e., discharged from the
ED more quickly) (3) a radiology report template for
the normal study could be served to the interpreting
radiologist for more rapid review and approval.
Second, automated abnormality localization could
help combat radiologist fatigue. Radiologists all over
the world are reading an increasing number of cases
with more images per case. Physician shortages exacerbate the problem, especially for radiologists in medically underserved areas (Nakajima et al., 2008). While
physician fatigue is a common problem that affects
all healthcare professionals, radiologists are particularly susceptible, and there is evidence that workloads
are so demanding that fatigue may impact diagnostic accuracy. (Bhargavan & Sunshine, 2005; Lu et al.,
2008; Berlin, 2000; Fitzgerald, 2001). A study examining radiologist fatigue in the interpretation of musculoskeletal radiographs found a statistically significant
decrease in fracture detection at the end of the work
day compared to beginning of work day (Krupinski
et al., 2010). Thus, a model which can perform automatic abnormality localization could highlight the
2017. URL http://www.boneandjointburden.org/
2014-report.
AIMI.
Artificial intelligence in medicine
&
imaging:
Available
labeled
medical
datasets.
https://aimi.stanford.edu/
available-labeled-medical-datasets.
[Online; accessed 2-December-2017].
Berlin, Leonard. Liability of interpreting too many
radiographs. American Journal of Roentgenology,
175(1):17–22, 2000.
Bhargavan, Mythreyi and Sunshine, Jonathan H. Utilization of radiology services in the united states:
levels and trends in modalities, regions, and populations. Radiology, 234(3):824–832, 2005.
Demner-Fushman, Dina, Kohli, Marc D, Rosenman,
Marc B, Shooshan, Sonya E, Rodriguez, Laritza,
Antani, Sameer, Thoma, George R, and McDonald,
Clement J. Preparing a collection of radiology examinations for distribution and retrieval. Journal of
the American Medical Informatics Association, 23
(2):304–310, 2015.
Deng, Jia, Dong, Wei, Socher, Richard, Li, Li-Jia, Li,
Kai, and Fei-Fei, Li. Imagenet: A large-scale hierarchical image database. In Computer Vision and
Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.
Esteva, Andre, Kuprel, Brett, Novoa, Roberto A,
Ko, Justin, Swetter, Susan M, Blau, Helen M, and
Thrun, Sebastian. Dermatologist-level classification
of skin cancer with deep neural networks. Nature,
542(7639):115–118, 2017.
Fitzgerald, Richard. Error in radiology. Clinical radiology, 56(12):938–946, 2001.
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
Gale, W., Oakden-Rayner, L., Carneiro, G., Bradley,
A. P., and Palmer, L. J. Detecting hip fractures
with radiologist-level performance using deep neural
networks. ArXiv e-prints, November 2017.
Gertych, Arkadiusz, Zhang, Aifeng, Sayre, James,
Pospiech-Kurkowska, Sylwia, and Huang, HK. Bone
age assessment of children using a digital hand atlas. Computerized Medical Imaging and Graphics,
31(4):322–331, 2007.
Grewal, Monika, Srivastava, Muktabh Mayank, Kumar, Pulkit, and Varadarajan, Srikrishna. Radnet:
Radiologist level accuracy using deep learning for
hemorrhage detection in ct scans. arXiv preprint
arXiv:1710.04934, 2017.
Gulshan, Varun, Peng, Lily, Coram, Marc, Stumpe,
Martin C, Wu, Derek, Narayanaswamy, Arunachalam, Venugopalan, Subhashini, Widner, Kasumi,
Madams, Tom, Cuadros, Jorge, et al. Development
and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. Jama, 316(22):2402–2410, 2016.
Hannun, Awni, Case, Carl, Casper, Jared, Catanzaro,
Bryan, Diamos, Greg, Elsen, Erich, Prenger, Ryan,
Satheesh, Sanjeev, Sengupta, Shubho, Coates,
Adam, et al. Deep speech: Scaling up end-to-end
speech recognition. arXiv preprint arXiv:1412.5567,
2014.
Heath, Michael, Bowyer, Kevin, Kopans, Daniel,
Moore, Richard, and Kegelmeyer, W Philip. The
digital database for screening mammography. In
Proceedings of the 5th international workshop on
digital mammography, pp. 212–218. Medical Physics
Publishing, 2000.
Huang, Gao, Liu, Zhuang, Weinberger, Kilian Q, and
van der Maaten, Laurens. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993,
2016.
Jaeger, Stefan, Candemir, Sema, Antani, Sameer,
Wang, Yi-Xiang J, Lu, Pu-Xuan, and Thoma,
George.
Two public chest x-ray datasets for
computer-aided screening of pulmonary diseases.
Quantitative imaging in medicine and surgery, 4(6):
475, 2014.
Kingma, Diederik and Ba, Jimmy.
Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Krupinski, Elizabeth A, Berbaum, Kevin S, Caldwell,
Robert T, Schartz, Kevin M, and Kim, John. Long
radiology workdays reduce detection and accommodation accuracy. Journal of the American College of
Radiology, 7(9):698–704, 2010.
Lu, Ying, Zhao, Shoujun, Chu, Philip W, and Arenson, Ronald L. An update survey of academic radiologists’ clinical productivity. Journal of the American College of Radiology, 5(7):817–826, 2008.
McHugh, Mary L. Interrater reliability: the kappa
statistic. Biochemia Medica, 22(3):276–282, October
2012. ISSN 1330-0962. URL https://www.ncbi.
nlm.nih.gov/pmc/articles/PMC3900052/.
Nakajima, Yasuo, Yamada, Kei, Imamura, Keiko, and
Kobayashi, Kazuko. Radiologist supply and workload: international comparison. Radiation medicine,
26(8):455–465, 2008.
OAI. Osteoarthritis initiative: a multi-center observational study of men and women. https://oai.
epi-ucsf.org/datarelease/. [Online; accessed 2December-2017].
Rajpurkar, Pranav, Zhang, Jian, Lopyrev, Konstantin,
and Liang, Percy. Squad: 100,000+ questions for
machine comprehension of text. arXiv preprint
arXiv:1606.05250, 2016.
Rajpurkar, Pranav, Hannun, Awni Y, Haghpanahi,
Masoumeh, Bourn, Codie, and Ng, Andrew Y.
Cardiologist-level arrhythmia detection with convolutional neural networks.
arXiv preprint
arXiv:1707.01836, 2017a.
Rajpurkar, Pranav, Irvin, Jeremy, Zhu, Kaylie, Yang,
Brandon, Mehta, Hershel, Duan, Tony, Ding, Daisy,
Bagul, Aarti, Langlotz, Curtis, Shpanskaya, Katie,
et al. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv
preprint arXiv:1711.05225, 2017b.
Shiraishi, Junji, Katsuragawa, Shigehiko, Ikezoe, Junpei, Matsumoto, Tsuneo, Kobayashi, Takeshi, Komatsu, Ken-ichi, Matsui, Mitate, Fujita, Hiroshi,
Kodera, Yoshie, and Doi, Kunio. Development of
a digital image database for chest radiographs with
and without a lung nodule: receiver operating characteristic analysis of radiologists’ detection of pulmonary nodules. American Journal of Roentgenology, 174(1):71–74, 2000.
Wang, Xiaosong, Peng, Yifan, Lu, Le, Lu, Zhiyong,
Bagheri, Mohammadhadi, and Summers, Ronald M.
Chestx-ray8: Hospital-scale chest x-ray database
and benchmarks on weakly-supervised classification
and localization of common thorax diseases. arXiv
preprint arXiv:1705.02315, 2017.
MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs
Woolf, Anthony D and Pfleger, Bruce. Burden of major musculoskeletal conditions. Bulletin of the World
Health Organization, 81(9):646–656, 2003.
Zhou, Bolei, Khosla, Aditya, Lapedriza, Agata, Oliva,
Aude, and Torralba, Antonio. Learning deep features for discriminative localization. In Proceedings
of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 2921–2929, 2016.
| 2cs.AI
|
An integral quadratic constraint framework for real-time steady-state
optimization of linear time-invariant systems*
arXiv:1710.10204v1 [math.OC] 27 Oct 2017
Zachary E. Nelson and Enrique Mallada
Abstract— Achieving optimal steady-state performance in
real-time is an increasingly necessary requirement of many
critical infrastructure systems. In pursuit of this goal, this paper
builds a systematic design framework of feedback controllers
for Linear Time-Invariant (LTI) systems that continuously
track the optimal solution of some predefined optimization
problem. The proposed solution can be logically divided into
three components. The first component estimates the system
state from the output measurements. The second component
uses the estimated state and computes a drift direction based
on an optimization algorithm. The third component computes
an input to the LTI system that aims to drive the system toward
the optimal steady-state.
We analyze the equilibrium characteristics of the closed-loop
system and provide conditions for optimality and stability. Our
analysis shows that the proposed solution guarantees optimal
steady-state performance, even in the presence of constant
disturbances. Furthermore, by leveraging recent results on the
analysis of optimization algorithms using integral quadratic
constraints (IQCs), the proposed framework is able to translate
input-output properties of our optimization component into
sufficient conditions, based on linear matrix inequalities (LMIs),
for global exponential asymptotic stability of the closed loop
system. We illustrate the versatility of our framework using
several examples.
I. I NTRODUCTION
Infrastructure systems are the foundation of our modern
society. The Internet, power grids, and transportation networks are just some examples of the several critical systems
that our current lifestyle relies on. Due to their large scale,
the operational and fault-associated costs that these systems
incur are both in the range of hundreds of millions of
dollars to several billion dollars [1]. Therefore, operators are
continuously faced with the conflicting tasks of operating
these systems as efficiently as possible and guaranteeing
certain levels of security or robustness.
Traditionally, this balancing between efficiency and security is achieved by separating tasks across different timescales. Efficiency goals are achieved using optimization algorithms running at a slow time-scale, and stability/robustness
goals are achieved using fast time-scale controllers. For example, in power systems, generators are optimally scheduled
by solving an (economic dispatch) optimization problem at a
slow time-scale (every 5/15 minutes, hour, or day) [2], but at
the fast time-scale the scheduling uses controllers based on
*This work was supported by the Army Research Office contract
W911NF-17-1-0092, NSF grants (CNS 1544771, EPCN 1711188), and
Johns Hopkins WSE startup funds.
Zachary E. Nelson and Enrique Mallada are with the Department of Electrical and Computer Engineering, The Johns Hopkins
University, 3400 N. Charles Street, Baltimore, MD 21218, emails:
{znelson2,mallada}@jhu.edu
frequency measurements [3] that are focused on preserving
the system stability [4], not efficiency.
Unfortunately, the state of flux that these infrastructure
systems currently experience due to the growing population,
deployment of sensing and communication technologies, and
sustainability trends, is pushing its operation towards their
limits and, in this way, rendering this approach obsolete.
Operating at maximum capacity does not leave room for the
inefficiencies incurred by the timescale separation. Moreover,
the limited coordination capabilities that today’s controllers
provide, when compared with optimization algorithms, does
not allow the system to quickly react to unprescribed events.
Motivated by this problem, this paper aims to remove the
time-scale separation by building controllers that can simultaneously achieve steady-state optimality while preserving
the system stability.
More precisely, this paper proposes a systematic design
framework for feedback controllers that, given a LTI system
and an unconstrained optimization problem, generates a
family of nonlinear controllers that seek to drive the system
towards the optimal solution of the optimization problem.
Our equilibrium analysis connects notions of controllability and observability of the LTI system with a criterion
for steady-state optimality of the closed loop equilibrium.
Furthermore, we leverage recent analyses of optimization
algorithms using Integral Quadratic Constraints (IQCs) [5],
[6], [7] to provide sufficient conditions based on Linear
Matrix Inequalities [8] that guarantee global exponential
asymptotic stability. The derived LMIs provide an explicit
bound on the rate of convergence and allow us to design an
algorithm that computes the maximum rate of convergence.
Our controllers have two main distinctive features. Firstly,
they can be functionally separated into three components/modules: (i) an Estimator, that aims to estimate the
system’s state from the output; (ii) an Optimizer, that uses
the estimated state to compute the drift direction necessary to
achieve optimality or outputs zero when the estimated state
is optimal; and (iii) a Driver (PI controller) that generates the
necessary input to drive the system toward the optimal solution. Secondly, the Optimizer module can be implemented
using one of many optimization algorithms, leading to a
family of optimization-based nonlinear controllers. The only
required conditions are that (i) in steady state the output of
the optimizer is zero if and only if its input (the estimated
state) is the optimal solution of the optimization problem, and
(ii) there exists an Integral Quadratic Constraint that captures
the input-output relationship of the optimizer.
Related Work: Optimization-based control design for achieving optimal steady-state performance has a been a popular
subject of research for more than three decades. It has been
used in communication networks to reverse engineer TCP/IP
congestion control protocols [9], [10] and provide a design
framework for novel congestion control algorithms [11],
distributed multi-path routing [12], [13], and admission control [14], and access control in wireless networks [15]. In
the context of power systems and micro-grids, optimizationbased control design has been used for the design of distributed controllers that can achieve efficient supply-demand
balance [16], frequency restoration [17], [18], congestion
management [19], and economic steady-state optimality [20],
[21], [22], [23]. Some of these approaches have been further
extended for more general settings such as [24] and [25].
In general, these solutions either require that the dynamical
system to be optimized has a specific structure, such as being
passive [19], [20], having primal-dual dynamics [18], [23],
[25], or having direct access to (a subset of) the system
state [24].
More recently, real-time optimization algorithms have
been proposed as a mean to mitigate the large fluctuations
that renewable energy introduce in power networks. The
solutions fall within two categories depending on whether
the system dynamics are considered as perturbations of the
optimization algorithms [26], [27], or the system is modeled
as a set of nonlinear algebraic constraints with slowly time
varying parameters [28], [29]. Our work distinguishes from
these works by explicitly modeling the system dynamics
and simultaneously guaranteeing stability of the dynamical
system and convergence to the optimal solution. Notably,
while our framework today does not include optimization
constraints or nonlinearities in the system dynamics, extending our framework to incorporate these features is the subject
of our current research.
Paper Organization: The organization of the paper is as
follows. Section II gives the reader preliminary tools that
are necessary for the later analysis. Section III sets up the
problem and discusses some of the challenges. Section IV
proposes a design framework of controllers that addresses
the challenges. Section V shows the systematic procedure
for analyzing steady-state optimality and stability. Section
VI considers multiple numerical examples to illustrate the
practicality of this approach. Lastly, Section VII summarizes
the major points of the paper and suggests future work.
II. P RELIMINARIES
A. Notation
The following notation will be used throughout the remainder of the paper. The n × n identity matrix is denoted
as In . The m × n zero matrix is denoted as 0m×n . The
zero vector with length n is denoted as 0n . The subscripts
are removed when the dimensions are implied by context.
A positive (semi) definite matrix P ∈ Rn×n is denoted as
P ≻ 0 ( 0). All norms || · || : Rn → R are the standard
ℓ2-norm. The Kronecker product of two matrices is denoted
by the symbol ⊗.
B. Integral Quadratic Constraints
Given a nonlinear mapping φ : p 7→ q, with p, q ∈ Rn ,
and an input-output reference (p∗ , φ(p∗ )) ∈ Rn × Rn , we
consider the following class of IQCs:
Definition 1 (Pointwise IQC): The mapping φ is said to
satisfy the pointwise IQC defined by (Qφ , p∗ , φ(p∗ )) if
T
p − p∗
p − p∗
≥0
Qφ
φ(p) − φ(p∗ )
φ(p) − φ(p∗ )
holds for all (p, p∗ ) ∈ Rn × Rn , where QTφ = Qφ ∈ R2n×2n
is an indefinite matrix.
Next, we discuss two particular choices of the nonlinear
map φ that are commonly used in optimization algorithms.
Gradient Mapping:
One source of nonlinearity that commonly arises in optimization algorithms is the gradient ∇f (p) of a function
f : Rn → R. In particular, characterizing the input-output
properties of the gradient of a strongly convex function with
a Lipschitz continuous gradient is of interest.
Definition 2: The gradient mapping ∇f : Rn → Rn is
Lipschitz continuous with parameter L if
||∇f (p) − ∇f (p∗ )|| ≤ L||p − p∗ ||
holds for all (p, p∗ ) ∈ Rn × Rn , where L ≥ 0 is a real
constant.
Definition 3: The function f : Rn → Rn is said to be
strongly convex if
(∇f (p) − ∇f (p∗ ))T (p − p∗ ) ≥ m||p − p∗ ||2
holds for all (p, p∗ ) ∈ Rn × Rn , where m > 0 is a real
constant.
Using these two properties, it is possible to show that ∇f
satisfies the pointwise IQC (Qf , p∗ , ∇f (p∗ )) defined by the
matrix
−2mL L + m
⊗ In .
(1)
Qf :=
L+m
−2
We refer the reader to [5] or [30] for a proof of this statement.
Proximal Mapping:
The second type of nonlinearity that will be used in this
paper arises from the proximal mapping of a function.
Definition 4: The proximal mapping Πρf : Rn → Rn of
the function f : Rn → R with real parameter ρ > 0 is
defined as
1
Πρf (p) := arg min f (v) + ||v − p||2 .
(2)
2ρ
v∈Rn
The optimality condition of the optimization problem
associated with (2) is:
1
0 = ∇f (Πρf (p)) + (Πρf (p) − p).
(3)
ρ
From (3), the proximal mapping can be viewed as the
composition of the gradient mapping with an affine operator,
followed by an inversion operation:
Πρf (p) = (I + ρ∇f )−1 (p).
A known result is that Πρf satisfies the pointwise IQC
(QΠρf , p∗ , Πρf (p∗ )) defined by the matrix
0
1
0 ρ−1
.
⊗
I
Q
⊗
I
QΠρf :=
n
f
n
ρ−1 −ρ−1
1 −ρ−1
This result can be derived by using Lemma 1 followed by
an IQC for inversion operations [6].
Âξ∗ + B̂q∗ = 0, the third property can be expressed as
Affine Composition of IQCs:
The following lemma, whose proof can be found in [6],
shows how to derive IQCs when a nonlinearity φ is composed with an affine map.
Lemma 1: (IQC for Affine Operations) Consider the nonlinear mapping φ that satisfies the pointwise IQC defined by
(Qφ , p∗ , φ(p∗ )). Define the affine mapping ψ : Rn → Rn to
be ψ(p) := S2 p + S1 φ(S0 p) where S0 , S1 , S2 ∈ Rn×n and
S1 is invertible. Then, ψ satisfies the pointwise IQC defined
by (Qψ , p∗ , ψ(p∗ )), where
T
S0
0
S
−(S1−1 S2 )T
.
Q
Qψ := 0
φ
−S1−1 S2 S1−1
0
(S1−1 )T
= 2(δξ)T P ((Âξ+ B̂q)−(Âξ∗ + B̂q∗ ))+α(δξ)T P δξ
Stability Analysis Using IQCs:
The following lemma is useful when deriving stability
conditions in terms of a LMI. See [8] and [31] for details.
Lemma 2: (Lossless S-Lemma) Let AT = A ∈ Rn×n and
T
B = B ∈ Rn×n . Then, A σB holds for some σ ≥ 0 if
and only if xT Bx ≥ 0 =⇒ xT Ax ≥ 0 for all x ∈ Rn .
We will now show how the input-output properties of an
IQC can be used to generate a sufficient stability condition
for the feedback interconnection of a LTI system and nonlinearity φ.
Proposition 1: Consider a LTI system defined by the matrices  ∈ Rn×n , B̂ ∈ Rn×m , Ĉ ∈ Rm×n , and D̂ ∈ Rm×m
with state ξ ∈ Rn , input q ∈ Rm , and output p ∈ Rm :
˙ = Âξ(t) + B̂q(t)
ξ(t)
p(t) = Ĉξ(t) + D̂q(t).
Suppose the LTI system has the nonlinearity φ : Rm → Rm
as feedback so that q = φ(Ĉξ + D̂q). Assume φ satisfies the
pointwise IQC (Qφ , p∗ , φ(p∗ )) and the feedback interconnection is well-posed.1 Then, the closed-loop equilibrium point
ξ∗ ∈ Rn has global exponential asymptotic stability of at
least rate α if the LMI
T
T
Ĉ
0
 P + P  + αP P B̂
Ĉ D̂
0
Q
+σ
φ
0 In
D̂T In
B̂ T P
0
(4)
is feasible for some σ ≥ 0, α > 0, and P ≻ 0.
Proof: Assume that (4) is feasible. Let δξ := ξ − ξ∗ and
δq := q − q∗ , where q∗ is the input that achieves equilibrium.
Consider the quadratic function V (δξ) = (δξ)T P δξ, where
P ∈ Rn×n , P ≻ 0. Lyapunov theory states that if V satisfies:
•
•
•
V (0) = 0 and V (δξ) > 0 for all δξ ∈ Rn \ {0},
if ||δξ|| → ∞, then V (δξ) → ∞ (radially unbounded),
V̇ (δξ) ≤ −αV (δξ) for all δξ ∈ Rn \ {0} and α > 0,
then the equilibrium point has global exponential asymptotic
stability of at least rate α [32].
Clearly, V (0n ) = 0Tn P 0n = 0. The property V (δξ) > 0
holds ∀δξ 6= 0 because P ≻ 0. The radial unboundedness
property similarly follows from P ≻ 0. Using the fact that
1 The
definition of well-posedness is given in Section III.
V̇ (δξ)+αV (δξ) = 2(δξ)T P ξ̇+α(δξ)T P δξ
= 2(δξ)T P (Âδξ+ B̂δq)+α(δξ)T P δξ
= (δξ)T P (Âδξ+ B̂δq)+(Âδξ+ B̂δq)T P δξ+α(δξ)T P δξ
T T
 P +P Â+αP P B̂ δξ
δξ
≤ 0.
=
δq
δq
B̂ T P
0
(5)
Finally, since the pointwise IQC (Qφ , p∗ , φ(p∗ ) is satisfied,
T T
Ĉ
0
δξ
Ĉ D̂ δξ
≥ 0.
(6)
Qφ
δq
0 In δq
D̂T In
Since (4) is feasible and (6) holds, it directly follows from
Lemma 2 that (5) holds. Hence, the equilibrium ξ∗ has global
exponential asymptotic stability of at least rate α.
III. P ROBLEM S ETUP
The problem setup is illustrated in Fig. 1, where we
consider a LTI system represented by a state-space model
where x ∈ Rn is the state, u ∈ Rm is the input, and y ∈ Rp
is the output:
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t).
(7)
The input u(t) is the sum of a control signal r(t) ∈ Rm
and an unknown constant disturbance w(t) = w ∈ Rm , i.e.
u(t) = w(t)+r(t) = w+r(t). Finally, the feedback operator
Ψ(·) denotes the (possibly nonlinear) feedback control to be
designed.
w(t) +
u(t)
+
!
B
D
A
C
"
y(t)
r(t)
Ψ
Fig. 1. LTI system interconnected with a nonlinear mapping and constant
disturbance signal.
Our goal is to, given the measurement y, design a control
input r = Ψ(y) that drives the system (7) to a steady-state
x∗ that is an optimal solution of a predefined optimization
problem
min f (x),
x∈Rn
(8)
where f : Rn → R is a given cost function.
Therefore, given the measurement y(t), the feedback Ψ(·)
must produce a control r = Ψ(y) such that x(t) → X ∗ ,
where X ∗ is the set of optimal solutions to (8), i.e.,
X ∗ = {x ∈ Rn : ∇f (x) = 0}.
Throughout this paper we make the following assumption.
Assumption 1: The cost function f (x) of the optimization
problem (8) is continuously differentiable, strongly convex,
and has a Lipschitz continuous gradient. This implies that
the set X ∗ is a singleton.2
Finally, we provide a concrete model for Ψ(·). As the
optimality conditions for optimization problem (8) are in
general nonlinear, the feedback controllers to be designed
will be necessarily of the same type. Thus, we consider the
nonlinear feedback Ψ using the nonlinear dynamics
Ψ:
η̇(t) = F (η(t), y(t))
r(t) = H(η(t), y(t)),
(9)
where η ∈ Rd is the state of the feedback dynamics, r ∈ Rm
is the output of the feedback dynamics, and the mappings
F : Rd × Rp → Rd and H : Rd × Rp → Rm are possibly
nonlinear.
Remark 1 (Well-Posedness): From the feedthrough terms
present in (7) and (9), it is possible a priori that the feedback
interconnection is not well-posed.3 A sufficient condition that
prevents this problem is by enforcing that whenever D 6= 0,
the map H depends only on η, i.e., r(t) = H(η(t)). We will
further discuss this condition in Section IV.
A. Design Challenges
There are several challenges associated to designing (9)
such that in steady state x∗ ∈ X ∗ .
• Lack of direct access to x(t): The system output matrix
C is not necessarily invertible. Thus, recovering x(t)
from y(t) is not straightforward.
∗
• Finding the solution x∗ ∈ X : Finding the optimal solution to the optimization problem is usually challenging
or the cost function may change, giving not enough time
to recompute x∗ .
∗
• Driving x(t) to x∗ ∈ X : Even if one has access to the
optimal solution x∗ , one then needs to design the right
r(t) that ensures that x(t) converges to it.
Interestingly, some of these challenges can be easily
handled using tools from control theory, such as recovering
x(t) from y(t) or driving x(t) to x∗ . On the other hand,
finding an optimal solution x∗ is the major goal within
optimization theory. Therefore, when the timescale of the
control and optimization tasks do not intersect, our problem
can be easily solved using standard tools from control and
optimization. However, when the timescale separation is no
longer present, the problem becomes more challenging as
there is no standard tools to address it. In particular, it is
usually hard to assess the stability of such an interconnected
system. This problem is systematically addressed in the next
section.
IV. O PTIMIZATION -BASED C ONTROL D ESIGN
In this section we describe the proposed optimizationbased controllers, that combine tools from control and optimization, and leverage the IQC framework described in the
preliminaries. The crux of our solution is a modularized
architecture that breaks down the feedback dynamics (9)
into three serial components that systematically addresses
2 Relaxing
this assumption is desired and is a subject of future work.
feedback interconnection of (7) and (9) is well-posed if u(t) and
y(t) are uniquely defined for every choice of states x(t) and η(t).
3 The
the challenges described in the previous section and allow a
straightforward application of Proposition 1 to certify global
exponential asymptotic stability.
r(t)
e(t)
D
ϕ
z(t)
E
y(t)
Ψ
Fig. 2.
Optimization-based control feedback breakdown.
The proposed architecture is described in Fig. 2. The first
component E : y(t) 7→ z(t) is a state estimator that takes
the output of the LTI system and produces a state estimate
z(t). The second component ϕ : z(t) 7→ e(t), referred to
as the optimizer, takes the state estimate and produces a
measurement of the optimality error or direction of desired
drift e(t), which is required to be zero if and only if the input
is in the set X ∗ . The optimizer can be thought of as the part
of optimization algorithm that dictates the direction of the
next step. The third component D : e(t) 7→ r(t), the driver,
takes the optimality error and produces the input to the LTI
system that ensures that the equilibrium satisfies e∗ = 0.
Remark 2: One of the advantages of the proposed architecture is the role independence of each component. This
allows for a subsystem to be skipped if the functionality is
not required. For example, in cases where y(t) = x(t) or
the optimization problem uniquely depends on y(t), then the
estimator block can be avoided.
In the remainder of this section, we describe the design
requirements of each proposed component/subsystem and
give some examples on how to implement them.
A. Design of the Estimator E
The estimator component E : y(t) 7→ z(t) is perhaps the
simplest to design. Its goal to build an estimate of the state,
z(t), from y(t). The design requirement of E is:
• A.1: If the system is in equilibrium, z(t) = z∗ = x∗ .
Therefore, an obvious choice for E is an observer/state
estimator. The dynamics of E are therefore given by
E:
x̂˙ = (A − LC)x̂ + (B − LD)u + Ly
z = x̂,
where L ∈ Rn×p is a constant matrix to be designed.
A standard argument for observers shows that the evolution of the error δx(t) := x(t) − z(t) is given by
˙
δx(t)
= (A − LC)δx(t).
Moreover, if (7) is observable, L can be chosen to satisfy
rank(A − LC) = n.
(10)
B. Design of the Optimizer ϕ
The optimizer ϕ has two design requirements:
• B.1: The optimizer must take the estimated state z(t)
as an input and then produce a measure of optimality
error or direction of drift e(t) such that e(t) = 0 if and
only if z(t) = x∗ ∈ X ∗ .
B.2: The input-output characteristics of ϕ must be
captured by an IQC (Qϕ , z∗ , ϕ(z∗ )).
For the purpose of this paper, we consider two possible
solutions.
•
ϕ1 : Gradient Descent. The first solution considered is the
standard gradient descent mapping, i.e.,
ϕ1 := −∇f.
(11)
It is straightforward to verify that e(t) = −∇f (z(t)) = 0 if
and only if z(t) = x∗ ∈ X ∗ . The following lemma explicitly
computes the IQC for ϕ1 .
Lemma 3: Assume the pointwise IQC (Qf , z∗ , ∇f (z∗ )) is
satisfied. Then, the pointwise IQC (Qϕ1 , z∗ , ϕ1 (z∗ )) defined
by the matrix
1 0
1 0
⊗ In Qf
⊗ In
Qϕ1 :=
0 −1
0 −1
is satisfied.
Proof: The IQC immediately follows from Lemma 1
with φ = ∇f , S0 = In , S1 = −In , and S2 = 0n×n .
ϕ2 : Proximal Tracking. Our second option for the optimizer
block is inspired by the proximal mapping (2). It essentially
computes the error between the input z(t) and the solution
given by the proximal operator Πρf (z(t)), that is,
ϕ2 := Πρf − In .
1
0 = ∇f (Πρf (z∗ )) + (Πρf (z∗ ) − z∗ ) ⇐⇒ 0 = ∇f (z∗ ).
ρ
By the definition of X ∗ , z∗ ∈ X ∗ .
Conversely, assume that z∗ ∈ X ∗ . Since
v∈Rn
v∈Rn
we have arg min f (v) +
v∈Rn
D:
ėI = e, r = KI eI + KP e
where KI , KP ∈ Rm×n are constant matrices to be designed.
It is straightforward to show that ėI = 0 if and only if
e(t) = 0, which satisfies our design requirements. In fact,
this also shows that we only need an integrator to satisfy the
design requirement. However, a PI controller provides better
dynamic properties than a pure integrator and therefore we
choose to add the proportional term.
D. Integrated System
This resulting interconnected system is shown in Fig. 3.
w(t) +
!
u(t)
+
r(t)
!
0n×n
KI
In
KP
"
PI Controller
B
D
A
C
e(t)
ϕ
"
z(t)
y(t)
!
A − LC
In
B − LD
0n×m
L
0n×p
"
Estimator
(12)
The following proposition shows that (12) satisfies the first
design requirement.
Proposition 2: The mapping ϕ2 satisfies the property that
e(t) = Πρf (z(t)) − z(t) = 0 if and only if z(t) = z∗ ∈ X ∗ .
Proof: Let z(t) = z∗ and assume that ϕ2 (z∗ ) = 0.
Then from (12), Πρf (z∗ ) = z∗ . It follows from (3) that
arg min f (v) = z∗ and arg min
towards the optimal solution of (8). The design requirement
for D is:
• C.1: If the system is in equilibrium, then e(t) = e∗ = 0.
Thus, one possible choice would be to use a ProportionalIntegral (PI) controller defined by the dynamics
1
||v − z∗ ||2 = z∗ ,
2ρ
1
||v − z∗ ||2 = z∗ .
2ρ
This is equivalent to Πρf (z∗ ) = z∗ . Thus, ϕ2 (z∗ ) = 0.
Finally, the next lemma computes the IQC that characterizes ϕ2 .
Lemma 4: Assume the pointwise IQC (Qf , z∗ , ∇f (z∗ )) is
satisfied. Then, the pointwise IQC (Qϕ2 , z∗ , ϕ2 (z∗ )) defined
by the matrix
1 0
1 1
⊗ In .
⊗ In QΠρf
Qϕ2 :=
1 1
0 1
is satisfied.
Proof: The IQC immediately follows from Lemma 1
with φ = Πρf , S0 = In , S1 = In , and S2 = −In .
C. Design of the Driver D
The last component of the proposed solution is in charge
of generating the control signal r(t) that drives the system
Fig. 3. LTI system in feedback with a state estimator E, optimizer ϕ, and
PI controller as the driver D.
In terms of the dynamics defined by (9), the feedback
design is given by
L
A− LC +(B − LD)KP ϕ (B − LD)KI
y
η+
η̇ =
0
ϕ
0
T
∈ R2n . (13)
r = KP ϕ KI η, where η := x̂T eTI
Here we have used the notation ϕx̂ := ϕ(x̂).
Remark 3: Whenever the estimator subsystem E is included in the interconnection, the feedback interconnection
will be well-posed because r(t) only depends on η(t).
However, well-posedness is not guaranteed when D 6= 0,
KP 6= 0, and there is no estimator subsystem because
r(t) = KP ϕ(Cx(t) + D(w + r(t))) + KI eI (t)
depends on itself. One simple solution to overcome this issue
is add a module E such that z(t) = y(t) − Du(t).
The next section shows that indeed the integrated system
is able to guarantee steady-state optimality under mild conditions and illustrates how the IQC framework can be leveraged
to guarantee global exponential asymptotic stability.
V. O PTIMALITY AND C ONVERGENCE
A. Optimality Analysis
The optimality analysis requires the following assumption.
Assumption 2: The system is steady-state controllable.
That is, given any steady-state x∗ , there exists an input u∗
such that Ax∗ + Bu∗ = 0.
Assumption 2 is in some sense necessary to ensure that
the system can achieve an arbitrary steady-state. Although
this assumption is stronger than the standard controllability
assumption, we point out that while controllability is sufficient to drive x(t) towards any state x∗ in finite time, it does
not requires that x(t) remains equal to x∗ .
Theorem 1: Consider the interconnection of the LTI system (7) and nonlinear feedback (13), where design requirements A.1, B.1, B.2, and C.1 are satisfied. Suppose (7) is
a minimal realization and Assumption 2 is satisfied. Then,
(x∗ , η∗ ) is an equilibrium point of the interconnected system
for some point η∗ and matrix L that satisfies (10) if and only
if x∗ ∈ X ∗ .
Proof: Assume (x∗ , η∗ ) is an equilibrium point of the
T
interconnected system, where η∗ = z∗T eTI∗ . The LTI
system must then be in equilibrium, meaning that
0 = Ax∗ + Bu∗ , y∗ = Cx∗ + Du∗ , u∗ = r∗ + w. (14)
Additionally, the dynamics of Ψ must be in equilibrium,
meaning that components E and D are in equilibrium. It
follows from D being in equilibrium and B.1 that
ėI = e = ϕ(z∗ ) = 0 ⇐⇒ z∗ ∈ X ∗ .
(15)
point considered in Theorem 1. Assume the pointwise IQC
(Q, z∗ , ϕ(z∗ )) is satisfied and the assumptions of Theorem
1 hold. Then, the equilibrium point (x∗ , η∗ ) has global
exponential asymptotic stability of at least rate α if the LMI
T
 P + P  + αP P B̂e
Ĉ 0
Ĉ T 0
0
Q
+σ
0 In
0 In
B̂eT P
0
(16)
is feasible for some σ ≥ 0, α > 0, and P ≻ 0.
Proof: This is essentially an application of Proposition
1, where the Lyapunov function is given by
V (δξ) = (δξ)T P δξ > 0, P ∈ R3n×3n , P ≻ 0.
Using the fact that Âz∗ + B̂e e∗ + B̂w w = 0,
V̇ (δξ) + αV (δξ)
= 2(δξ)T P [Âξ + B̂e e + B̂w w] + α(δξ)T P δξ
= 2(δξ)T P [Âδξ + B̂e δe] + α(δξ)T P δξ
T T
δξ
 P + P  + αP P B̂e δξ
=
≤ 0.
δe
δe
B̂eT P
0
The rest of the proof follows from Proposition 1.
It follows from E being in equilibrium that
x̂˙ = (A − LC)z∗ + (B − LD)u∗ + L(Cx∗ + Du∗ )
= (A − LC)z∗ + Bu∗ + LCx∗ = 0.
Adding Ax∗ to both sides and using (14) results in
(A − LC)(z∗ − x∗ ) = 0.
Since the LTI system is observable, we can choose L such
that A − LC is Hurwitz and therefore x∗ = z∗ . From (15),
x∗ ∈ X ∗ .
Conversely, assume that x∗ ∈ X ∗ . Consider e∗ = 0 and
an eI∗ such that Ax∗ + B(KI eI∗ + w) = 0, which exists
because of Assumption 2. Then, ẋ = 0 and ėI = 0 directly
follow. Next, consider z∗ = x∗ . It is then straightforward
to show that x̂˙ = 0. The optimality property of ϕ gives
that ϕ(z∗ ) = e∗ = 0, which is consistent with the previous
definition of e∗ . Therefore, (x∗ , η∗ ) is an equilibrium point.
C. Convergence Rate
Finally, we show how the LMI condition derived in
Theorem 2 can be leveraged to compute the maximum
convergence rate that the system can achieve. Our goal here
is to solve the optimization problem:
maximize α subject to (16).
σ≥0,α>0,P ≻0
(17)
The main challenge is that because α multiplies P in (16),
the optimization problem is non-convex. However, for a fixed
α, finding whether (16) is feasible can be done efficiently.
Therefore, it is possible to implement a line search in α that
finds the maximum value αmax that satisfies (16).
VI. N UMERICAL E XAMPLES
A. Scalar System
Consider the scalar system
ẋ = −5x + u, y = x,
B. Stability Analysis
This section will derive a sufficient condition for the
global exponential asymptotic stability of the equilibrium
point considered in the optimality analysis. For this analysis,
it is useful to group the linear dynamics of E and D into
the LTI system to essentially create a larger dimension LTI
system. The resulting system is
A
0
BKI
B
BKP
ξ̇ = LC A − LC BKI ξ + BKP e + B w
0
0
0
0
In
|
{z
}
| {z }
| {z }
Â
z = 0 In
{z
|
Ĉ
B̂e
T
0 ξ, where ξ := x
}
T
x̂
B̂w
T
eTI .
Theorem 2: Consider the interconnection of the LTI system (7) and nonlinear feedback (13) with the equilibrium
with x, u, y ∈ R. Since the estimator module E is not needed,
the dimension of the LMI can be reduced. The feedback
interconnection is still well-posed because D = 0. Let the
cost function be of the form
1
f (x) = qx2 + cx + v,
2
where q, c, v ∈ R are constants. For this case, the Lipschitz
constant and strong convexity constant are m = L = q. Let
the control parameters be given by ki = 1 and kp = 1.
Fig. 4 shows the solution of (17) as a function of ρ. Several
curves representing different steepness levels of f are plotted.
The plots demonstrate that larger values of ρ lead to a larger
αmax , with no marginal improvement after a certain point.
Additionally, it is interesting to note that there are cases
when the ϕ2 optimizer achieved a larger αmax than the ϕ1
optimizer.
Fig. 4. Maximum feasible α versus ρ for several different scalar cost
functions when using (a): ϕ2 (proximal optimizer) and (b): ϕ1 (gradient
optimizer).
After choosing a sufficiently large ρ, the solution of
(17) was plotted as a function of q as shown in Fig. 5.
As expected, larger values of q, corresponding to steeper
quadratic functions, resulted in a larger αmax . There was also
no marginal improvement past a certain threshold of q. This
threshold was a very large q for ϕ1 and a very small q for
ϕ2 . For q < 1, ϕ2 achieved a larger αmax and for q > 1, ϕ1
achieved a larger αmax .
Fig. 5.
Maximum feasible α versus q when using ϕ1 and ϕ2 optimizers.
The system’s state as a function of time, when the cost
function was f (x) = (x − 10)2 , is given in Fig. 6. Several
trajectories, corresponding to different values of ρ, were
plotted. The disturbance signal was initially set as w = 2
and at t = 50s was changed to w = −10. For all cases,
the state was able to recover from the change in disturbance
and continue tracking the optimal solution. The trajectories
illustrate that the performance of the ϕ2 optimizer is very
much related to the choice of ρ and if chosen correctly can
outperform the ϕ1 optimizer.
module is obviously needed. Let the cost function be of the
form
1
f (x) = xT Qx + cT x,
2
2×2
where Q ∈ R , Q ≻ 0, and c ∈ R2 . For this case, the
Lipschitz constant L is the larger eigenvalue of Q and the
strong convexity constant m is the smaller eigenvalue of Q.
Let the feedback parameters be given by
0
1
and L = 1 1 .
KI = KP =
1/4 −1/4
Fig. 7 shows the solution of (17) as a function of ρ. Several
curves corresponding to different eigenvalue choices of Q are
plotted. With a sufficiently large ρ, ϕ2 was able to achieve
a larger αmax than ϕ1 when m = 0.75, but was not able to
when m = 1.25. For both optimizer types, L = 1.25 resulted
in a larger αmax than L = 1.5.
Fig. 7. Maximum feasible α versus ρ for several different multivariable
cost functions when using (a): ϕ2 (proximal optimizer) and (b): ϕ1 (gradient
optimizer).
After choosing a sufficiently large ρ, the solution of (17)
was plotted as a function of q as shown in Fig. 8. The
Lipschitz constant was chosen as different multiples of m.
The ϕ1 optimizer resulted in a larger αmax when L was
chosen closer to m. Conversely, the ϕ2 optimizer resulted
in a larger αmax when the multiple was chosen farther from
m.
Fig. 8. Maximum feasible α versus m for several different choices of L
when using ϕ1 and ϕ2 optimizers.
Fig. 6. System state versus time when using ϕ1 and ϕ2 optimizers. Several
choices of ρ are shown for the ϕ2 optimizer.
B. MIMO System with State Estimator
Consider the MIMO system defined by
1 4
0
1
, C = 1 0 , D = 01×2 ,
, B=
A=
1 0
−10 −5
with x, u ∈ R2 and y ∈ R. Since the output of the LTI
system only has information about the first state, an estimator
The system’s state as a function of time, when the cost
function was defined by
1 1/6
, cT = −17/3 −4/3 ,
Q=
1/6 2/3
is given in Fig. 9. In this case, m≈ 0.5976, L ≈ 1.0690,
T
and the optimal solution is x∗ ≈ 5.5652 0.6087 . The
disturbance was initially
set
T to zero, but at t = 75s was
changed to w = 1 1 . Similar to the scalar case,
the trajectories were able to recover from the change in
disturbance and there were cases when ϕ2 outperformed ϕ1 .
It is particularly interesting that state 2 has the ability to
reach the optimal solution despite the fact that it was not
being measured.
Fig. 9. System (a) state 1 and (b) state 2 versus time when using ϕ1 and
ϕ2 optimizers. Several choices of ρ are shown for the ϕ2 optimizer.
VII. C ONCLUSIONS
This paper introduces a framework of nonlinear controllers
whose purpose is to drive a given LTI system to the optimal
solution of some predefined optimization problem. The controllers are composed of an estimator, optimizer, and driver.
We give specific design requirements and possible design
choices for each of these modules. Our analysis shows that
under mild assumptions, optimal steady-state performance
can be guaranteed. Moreover, we give a sufficient condition,
in terms of a LMI, such that global exponential asymptotic
stability of the optimal steady-state can be guaranteed. Lastly,
we present numerical illustrations that demonstrate how the
design choices relate to the rate of exponential convergence.
The main focus of future work includes further generalizing
the proposed framework. In particular, we will consider
driving a LTI system to the optimal solution of a constrained
optimization problem and when there are multiple LTI systems that occur in a distributed setting.
R EFERENCES
[1] R. J. Campbell, “Weather-related power outages and electric system
resiliency.” Congressional Research Service, Library of Congress
Washington, DC, 2012.
[2] A. K. Bejestani, A. Annaswamy, and T. Samad, “A hierarchical
transactive control architecture for renewables integration in smart
grids: Analytical modeling and stability,” IEEE Transactions on Smart
Grid, vol. 5, no. 4, pp. 2054–2065, July 2014.
[3] A. J. Wood and B. F. Wollenberg, Power generation, operation, and
control, 3rd ed. John Wiley & Sons, 2013.
[4] Ibraheem and D. P. Kothari, “Recent philosophies of automatic generation control strategies in power systems,” IEEE Transactions on
Power Systems, vol. 20, no. 1, pp. 346–357, Feb 2005.
[5] L. Lessard, B. Recht, and A. Packard, “Analysis and design of optimization algorithms via integral quadratic constraints,” SIAM Journal
on Optimization, vol. 26, no. 1, pp. 57–95, 2016.
[6] M. Fazlyab, A. Ribeiro, M. Morari, and V. M. Preciado, “Analysis of
optimization algorithms via integral quadratic constraints: Nonstrongly
convex problems,” arXiv preprint arXiv:1705.03615, 2017.
[7] B. Hu and L. Lessard, “Control interpretations for first-order optimization methods,” in 2017 American Control Conference (ACC), May
2017, pp. 3114–3119.
[8] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix
Inequalities in System and Control Theory, ser. Studies in Applied
Mathematics. Philadelphia, PA: SIAM, Jun. 1994, vol. 15.
[9] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan, “Rate control for
communication networks: Shadow prices, proportional fairness and
stability,” The Journal of the Operational Research Society, vol. 49,
no. 3, pp. 237–252, 1998.
[10] 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, Dec 1999.
[11] D. X. Wei, C. Jin, S. H. Low, and S. Hegde, “FAST TCP: Motivation,
Architecture, Algorithms, Performance,” IEEE/ACM Transactions on
Networking, vol. 14, no. 6, pp. 1246–1259, 2006.
[12] M. Q. Rieck, S. Pai, and S. Dhar, “Distributed routing algorithms for
multi-hop ad hoc networks using d-hop connected d-dominating sets,”
Computer Networks, vol. 47, no. 6, pp. 785–799, 2005.
[13] E. Mallada and F. Paganini, “Stability of node-based multipath routing
and dual congestion control,” in 2008 47th IEEE Conference on
Decision and Control, Dec 2008, pp. 1398–1403.
[14] A. Ferragut and F. Paganini, “Achieving network stability and user
fairness through admission control of TCP connections,” in 2008 42nd
Annual Conference on Information Sciences and Systems, March 2008,
pp. 1195–1200.
[15] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Cross-layer
congestion control, routing and scheduling design in ad hoc wireless networks,” in 25th IEEE International Conference on Computer
Comunications, 2006, pp. 1–13.
[16] C. Zhao, U. Topcu, N. Li, and S. Low, “Design and stability of loadside primary frequency control in power systems,” IEEE Transactions
on Automatic Control, vol. 59, no. 5, pp. 1177–1189, 2014.
[17] E. Mallada and S. H. Low, “Distributed frequency-preserving optimal
load control,” in IFAC World Congress, 08 2014, pp. 5411–5418.
[18] E. Mallada, C. Zhao, and S. H. Low, “Optimal load-side control for
frequency regulation in smart grids,” IEEE Transactions on Automatic
Control, 6 2017.
[19] C. Zhao, E. Mallada, S. H. Low, and J. W. Bialek, “A unified
framework for frequency control and congestion management,” in
Power Systems Computation Conference, 06 2016, pp. 1–7.
[20] F. Dörfler, J. Simpson-Porco, and F. Bullo, “Breaking the Hierarchy:
Distributed Control & Economic Optimality in Microgrids,” IEEE
Transactions on Control of Network Systems, vol. 3, no. 3, pp. 241–
253, 2016.
[21] A. Cherukuri and J. Cortés, “Distributed generator coordination for
initialization and anytime optimization in economic dispatch,” IEEE
Transactions on Control of Network Systems, vol. 2, no. 3, pp. 226–
237, Sept 2015.
[22] C. Zhao, E. Mallada, and F. Dörfler, “Distributed frequency control
for stability and economic dispatch in power networks,” in American
Control Conference, 07 2015, pp. 2359–2364.
[23] N. Li, C. Zhao, and L. Chen, “Connecting automatic generation control
and economic dispatch from an optimization view,” IEEE Transactions
on Control of Network Systems, vol. 3, no. 3, pp. 254–264, Sept 2016.
[24] A. Jokic, M. Lazar, and P. P. J. van den Bosch, “On constrained steadystate regulation: Dynamic KKT controllers,” IEEE Transactions on
Automatic Control, vol. 54, no. 9, pp. 2250–2254, Sept 2009.
[25] X. Zhang, A. Papachristodoulou, and N. Li, “Distributed optimal
steady-state control using reverse- and forward-engineering,” in 2015
54th IEEE Conference on Decision and Control (CDC), Dec 2015,
pp. 5257–5264.
[26] E. Dall’Anese and A. Simonetto, “Optimal power flow pursuit,” IEEE
Transactions on Smart Grid, 2016, in press.
[27] E. Dall’Anese, S. V. Dhople, and G. B. Giannakis, “Photovoltaic
inverter controllers seeking ac optimal power flow solutions,” IEEE
Transactions on Power Systems, vol. 31, no. 4, pp. 2809–2823, July
2016.
[28] F. Dörfler and S. Grammatico, “Amidst centralized and distributed
frequency control in power systems,” in 2016 American Control
Conference (ACC), July 2016, pp. 5909–5914.
[29] Y. Tang, K. Dvijotham, and S. Low, “Real-time optimal power flow,”
IEEE Transactions on Smart Grid, 2017, in press.
[30] Y. Nesterov, Introductory Lectures on Convex Optimization: A Basic
Course, 1st ed. Springer Publishing Company, Incorporated, 2014.
[31] K. Derinkuyu and M. Ç. Pınar, “On the s-procedure and some
variants,” Mathematical Methods of Operations Research, vol. 64,
no. 1, pp. 55–77, 2006.
[32] H. K. Khalil, Nonlinear systems; 3rd ed. Upper Saddle River, NJ:
Prentice-Hall, 2002.
| 3cs.SY
|
Learning Structural Weight Uncertainty for
Sequential Decision-Making
Ruiyi Zhang1
Chunyuan Li1 Changyou Chen2 Lawrence Carin1
Duke University 2 University at Buffalo
[email protected], [email protected], [email protected], [email protected]
arXiv:1801.00085v1 [stat.ML] 30 Dec 2017
1
Abstract
Learning probability distributions on the
weights of neural networks (NNs) has recently proven beneficial in many applications. Bayesian methods, such as Stein
variational gradient descent (SVGD), offer
an elegant framework to reason about NN
model uncertainty. However, by assuming
independent Gaussian priors for the individual NN weights (as often applied), SVGD
does not impose prior knowledge that there
is often structural information (dependence)
among weights. We propose efficient posterior learning of structural weight uncertainty,
within an SVGD framework, by employing
matrix variate Gaussian priors on NN parameters. We further investigate the learned
structural uncertainty in sequential decisionmaking problems, including contextual bandits and reinforcement learning. Experiments on several synthetic and real datasets
indicate the superiority of our model, compared with state-of-the-art methods.
1
Introduction
Deep learning has achieved state-of-the-art performance on a wide range of tasks, including image classification [Krizhevsky et al., 2012], language modeling [Sutskever et al., 2014], and game playing [Silver
et al., 2016]. One challenge in training deep neural networks (NNs) is that such models may overfit
to the observed data, yielding over-confident decisions
in learning tasks. This is partially because most NN
learning only seeks a point estimate for the model paProceedings of the 21st International Conference on Artificial Intelligence and Statistics (AISTATS) 2018, Lanzarote,
Spain. JMLR: W&CP volume 7X. Copyright 2018 by the
author(s).
rameters, failing to quantify parameter uncertainty. A
natural way to ameliorate these problems is to adopt a
Bayesian neural network (BNN) formulation. By imposing priors on the weights, a BNN utilizes available
data to infer an approximate posterior distribution on
NN parameters [MacKay, 1992]. When making subsequent predictions (at test time), one performs model
averaging over such learned uncertainty, effectively
yielding a mixture of NN models [Li et al., 2016a, Gal
and Ghahramani, 2016, Liu and Wang, 2016, Li et al.,
2016b]. BNNs have shown improved performance on
modern achitectures, including convolutional and recurrent networks [Li et al., 2016c, Gan et al., 2017,
Fortunato et al., 2017].
For computational convenience, traditional BNN
learning typically makes two assumptions on the
weight distributions: independent isotropic Gaussian
distributions as priors, and fully factorized Gaussian
proposals as posterior approximation when adopting
variational inference [Hernández-Lobato and Adams,
2015, Blundell et al., 2015, Liu and Wang, 2016]. By
examining this procedure, we note two limitations: (i)
the independent Gaussian priors can ignore the anticipated structural information between the weights,
and (ii) the factorized Gaussian posteriors can lead to
unreasonable approximation errors and underestimate
model uncertainty (underestimate variances).
Recent attempts have been made to overcome these
two issues. For example, [Louizos and Welling, 2016,
Sun et al., 2017] introduced structural priors with the
matrix variate Gaussian (MVG) distribution [Gupta
and Nagar, 1999] to impose dependency between
weights within each layer of a BNN. Further, nonparametric variational inference methods, e.g., Stein
variational gradient descent (SVGD) [Liu and Wang,
2016], iteratively transport a set of particles to approximate the target posterior distribution (without making explicit assumptions about the form of the posterior, avoiding the aforementioned factorization assumption). SVGD represents the posterior approximately in terms of a set of particles (samples), and is
Learning Structural Weight Uncertainty for Sequential Decision-Making
endowed with guarantees on the approximation accuracy when the number of particles is exactly infinity
[Liu, 2017]. However, since the updates within SVGD
learning involve kernel computation in the parameter
space of interest, the algorithm can be computationally
expensive in a high-dimensional space. This becomes
even worse in the case of structural priors, where a
large amount of additional parameters are introduced,
rendering SVGD inefficient when directly applied for
posterior inference.
We propose an efficient learning scheme for accurate
posterior approximation of NN weights, adopting the
MVG structural prior. We provide a new perspective to unify previous structural weight uncertainty
methods [Louizos and Welling, 2016, Sun et al., 2017]
via the Householder flow [Tomczak and Welling, 2016].
This perspective allows SVGD to approximate a target
structural distribution in a lower-dimensional space,
and thus is more efficient in inference. We call the
proposed algorithm Structural Stein Variational Gradient Descent (S2 VGD).
We investigate the use of our structural-weightuncertainty framework for learning policies in sequential decision problems, including contextual bandits
and reinforcement learning. In these models, uncertainty is particularly important because greater uncertainty on the weights typically introduces more variability into a decision made by a policy network [Kolter
and Ng, 2009], naturally leading the policy to explore.
As more data are observed, the uncertainty decreases,
allowing the decisions made by a policy network to
become more deterministic as the environment is better understood (exploitation when the policy becomes
more confident). In all these models, structural weight
uncertainty is inferred by our proposed S2 VGD.
We conduct several experiments, first demonstrating
that S2 VGD yields effective performance on classic
classification/regression tasks. We then focus our experiments on the motivating applications, sequential
decision problems, for which accounting for NN weight
uncertainty is believed to be particularly beneficial. In
these applications the proposed method demonstrates
particular empirical value, while also being computationally practical. The results show that structural
weight uncertainty gives better expressive power to describe uncertainty driving better exploration.
2
2.1
Preliminaries
Matrix variate Gaussian distributions
The matrix variate Gaussian (MVG) distribution
[Gupta and Nagar, 1999] has three parameters, describing the probability of a random matrix W ∈
R`1 ×`2 :
p(W) , MN (W; M, U, V)
1
−1
> −1
(1)
tr[V (W − M) U (W − M)]
exp
2
=
(2π)`1 `2 /2 |V|`1 /2 |U|`2 /2
where M ∈ R`1 ×`2 is the mean of the distribution.
U ∈ R`1 ×`1 and V ∈ R`2 ×`2 encode covariance information for the rows and columns of W respectively.
The MVG is closely related to the multivariate Gaussian distribution.
Lemma 1 (Golub and Van Loan [2012]). Assume W
follows the MVG distribution in (1), then
vec(W) ∼ N (vec(M), V ⊗ U)
(2)
where vec(M) is the vectorization of M by stacking the
columns of M, and ⊗ denotes the standard Kronecker
product [Golub and Van Loan, 2012].
Furthermore, a linear transformation of an MVG distribution is still an MVG distribution.
Lemma 2 (Golub and Van Loan [2012]). Assume W
follows the MVG distribution in (1), A ∈ R`2 ×`1 , C ∈
R`2 ×`1 , then,
B , AW ∼ MN (B; AM, AUA> , V)
B , WC ∼ MN (B; MC, U, C> VC)
(3)
MVG priors for BNNs For classification and regression tasks on data D = {d1 , · · · , dN }, where
di = {xi , yi }, with input xi and output yi , an L-layer
NN parameterizes the mapping {g` }L
`=1 , defining the
prediction of yi for xi as:
ŷi = f (xi ) = gL ◦ gL−1 ◦ · · · ◦ g0 (xi ), ∀i.
(4)
where ◦ represents function composition, i.e., A ◦ B
means A is evaluated on the output of B. Each
layer g` represents a nonlinear transformation. For
example, with the Rectified Linear Unit (ReLU) activation function [Nair and Hinton, 2010], g` (xi ) =
ReLU(W`> xi + b` ), where ReLU(x) , max (0, x), W`
is the weight matrix for the `th-layer, and b` the corresponding bias term.
The MVG can be adopted as a prior for the weight
matrix in each layer, to impose the prior belief that
there are intra-layer weight correlations,
W` ∼ MN (W; 0, U` , V` ),
(5)
where the covariances U` , V` have components drawn
independently from Inv-Gamma(a0 , b0 ). The parameters are θ , {W` , log U` , log V` }, and the above
distributions represent the prior p(θ). Our goal
with BNNs is to learn layer-wise structured weight
uncertainty, described by the posterior distribution
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
p(θ|D) ∝ p(θ)p(D|θ), represented below as p for simplicity. When U = σI and V = σI, we reduce to BNNs
with independent isotropic Gaussian priors [Blundell
et al., 2015].
2.2
Stein Variational Gradient Descent
SVGD considers a set of particles {θi }M
i=1 drawn from
distribution q, and transforms them to better match
the target distribution p, by update:
θi ← θi + φ(θi ),
∂
KL(q[φ] ||p) ,
φ = arg max
φ∈F
∂
(6)
where q[φ] is the updated empirical distribution, with
as the step size, and φ as a function perturbation direction chosen to minimize the KL divergence between
q and p. SVGD considers F as the unit ball of a vectorvalued reproducing kernel Hilbert space (RKHS) H
associated with a kernel κ(θ, θ 0 ). The RBF kernel is
usually used as default. It has been shown [Liu and
Wang, 2016] that:
−
∂
KL(q[φ] kp)|=0 = Eθ∼q [Γp φ(θ)],
(7)
∂
>
with Γp φ(θ) , ∇θ log p(θ|D) φ(θ) + ∇θ · φ(θ),
where ∇θ log p(θ) denotes the derivative of the logdensity of p; Γp is the Stein operator. Assuming that
the update function φ(θ) is in a RKHS with kernel
κ(·, ·), it has been shown in [Liu and Wang, 2016] that
(7) is maximized with:
φ(θ) = Eθ∼q [κ(θ, θ 0 )∇θ log p(θ|D) + ∇θ κ(θ, θ 0 )].
(8)
The expectation Eθ∼q [·] can be approximated by an
empirical averaging of particles {θi }M
i=1 , resulting in a
practical SVGD procedure as:
θi ← θi +
M
Xh
κ(θj , θi )∇θi log p(θj |D)
M j=1
i
+∇θj κ(θj , θi ) .
(9)
The first term to the right of the summation in (9)
drives the particles θi towards the high probability regions of p, with information sharing across similar particles. The second term repels the particles away from
each other, encouraging coverage of the entire distribution. SVGD applies the updates in (9) repeatedly, and
the samples move closer to the target distribution p in
each iteration. When using state-of-the-art stochastic gradient-based algorithms, e.g., RMSProp [Hinton
et al., 2012] or Adam [Kingma and Ba, 2015], SVGD
becomes a highly efficient and scalable Bayesian inference method.
Computational Issues Applying SVGD with
structured distributions for NNs has many challenges.
For a weight matrix W of size `1 × `2 , the number
of parameters in U and V are `21 and `22 , respectively.
Hence, the total number of parameters θ needed to
describe the distribution is `1 `2 + `21 + `22 , compared
to `1 `2 + 1 in traditional BNNs that employ isotropic
Gaussian priors and factorization.
The increase of parameter dimension by `21 + `22 − 1
leads to significant computational overhead. The problem becomes even more severe in two aspects in calculating the kernels: (i) The computation increases
quadratically by M (M − 1)(`21 + `22 − 1)/2, (ii) the approximation to the repelling term in (9) using limited
particles can be inaccurate in high dimensions. Therefore, it is desirable to transform the MVG distribution
to a lower-dimensional representation.
3
SVGD & Imposition of Structure
3.1 Reparameterization of the MVG
Since the covariance matrices U and V are positive definite, we can decompose them as U =
PΛ1 Λ1 P> , V = QΛ2 Λ2 Q> , where P and Q are the
corresponding orthogonal matrices, Λ1 and Λ2 are diagonal matrices with positive diagonal elements. According to Lemma 2, we show the following reparameterization of MVG:
Proposition 3. For a random matrix C following an
independent Gaussian distribution:
−1
vec(C) ∼ N (·, P> Λ−1
1 MΛ2 Q, I),
(10)
the corresponding full-covariance MVG W in (1) can
be reparameterized as W = PΛ1 CΛ2 Q> .
The proof is in Section A of the Supplementary Material. Therefore, W drawn from MVG can be decomposed as the linear product of five random matrices:
• C as a standard weight matrix with an independent Gaussian distribution.
• Λ1 and Λ2 as the diagonal matrices, encoding the
structural information within each row and column, respectively.
• P and Q as orthogonal matrices, which characterize the structural information of weights between
rows and columns, respectively.
3.2
MVG as Householder Flows
Based on Proposition 3, we propose a layer decomposition: the one-layer weight matrix W with an MVG
prior can be decomposed into a linear product of five
matrices, as illustrated in Figure 1(a). Our layer decomposition provides an interesting interpretation for
the original MVG layer: it is equivalent to several
Learning Structural Weight Uncertainty for Sequential Decision-Making
W
⇤1
P
⇥
⇥
Q
⇤2
C
⇥
⇥
(a) Layer Decomposition
H1
P or Q
H2
HK
1
⇥· · · ⇥
⇥
HK
⇥
(b) Householder Flow
Figure 1: Illustration of the two proposed techniques to
reduce parameter size in learning the distribution of W:
(a) decomposition of W as a linear product of five matrices,
and (b) approximation of P or Q as a linear product of K
Householder matrices. Note that each rectangle indicates
a matrix, “×” indicates the matrix product, and H in (b)
is constructed using (12).
within-layer transformations. The representations imposed in the standard weight matrix C are rotated by
P and Q, and re-weighted by Λ1 and Λ2 .
Note that the layer decomposition maintains similar
computational complexity as the original MVG layer.
To reduce the computation bottleneck in the layer decomposition, we further propose to represent P and Q
using Householder flows [Tomczak and Welling, 2016].
Formally, a Householder transformation is a linear
transformation that describes a reflection about a hyperplane containing the origin. Householder flow is a
series of Householder transformations.
Lemma 4 ([Sun and Bischof, 1995]). Any orthogonal
matrix M of degree K can be expressed as a Householder flow, i.e., a product of exactly K nontrivial
Householder matrices., i.e.,
M = HK HK−1 · · · H1 .
(11)
Importantly, each Householder matrix H is constructed from a Householder vector v (which is orthogonal to the hyperplane):
H=I−2
>
vv
.
kvk2
(12)
According to Lemma 4, P and Q can be represented as
(p)
(q)
a product of Householder matrices {Hk } and {Hk }:
(p)
(p)
(p)
(q)
(q)
(q)
(p)
Hk
=I−
(q)
(p) (p)>
2vk vk /
(q) (q)>
Hk = I − 2vk vk
(p)
(q)
Since Householder flows allow one to represent P or
Q as K Householder vectors, the parameter sizes reduce from `21 and `22 to K`1 and K`2 . Overall, we
can model W with structured weight priors using
only (K + 1)(`1 + `2 ) + `1 `2 parameters. Therefore,
we can efficiently capture the structure information
with only a slight increase of computational cost (i.e.,
(K + 1)(`1 + `2 )).
Interestingly, our method provides a unifying perspective of previous methods on learning structured weight
uncertainty. In terms of prior distributions, when
K = 0 (i.e., P = Q = I), our reparameterization
reduces to [Louizos and Welling, 2016]. When K = 1,
and Λ1 = Λ2 = I, our reparameterization reduces to
[Sun et al., 2017]. In terms of posterior learning methods, when P = Q = Λ1 = Λ2 = I and M > 1, S2 VGD
reduces to SVGD; when M = 1, it reduces to learning
an MAP solution.
3.3 Structural BNNs Revisited
We can leverage the layer decomposition and Householder flow above to construct an equivalent BNN by
approximating the `th MVG layer in (5) with standard
Gaussian weight matrices:
p(C|λ) = N (C` ; 0, λ) ,
(p)
(p)
p(vk` |φ) = N vk` ; 0, φI ,
(q)
(q)
p(vk` |ψ) = N vk` ; 0, ψI ,
(1)
p(Λ(1) |φ) = N Λ` ; 0, φI
(2)
p(Λ(2) |ψ) = N Λ` ; 0, ψI ,
(p)> (p)
vk vk
(13)
(q)> (q)
/ vk vk ,
where vk and vk are the kth Householder vector for
P and Q, respectively. This is illustrated in Figure 1
(14)
log λ, log φ, log ψ ∼ Inv-Gamma(·; a` , b` ).
The forms of the likelihood for the last layer are defined
according to the specific applications. For regression
problems on real-valued response y:
y|x, WL ∼ N (y; f (x, WL ), γI)
(15)
log γ ∼ Inv-Gamma(·; aL , bL ) ,
with f (·) a neural network defined in (4). For classification problems on discrete labels y:
y|x, WL ∼ Categorical(y; Softmax(f (x, WL )).
P = HK HK−1 · · · H1
Q = HK HK−1 · · · H1
(b). Note that the degree K ≤ min{`1 , `2 }, with proof
in Section B of Supplementary Material. In practice,
K is a trade-off hyperparameter, balancing the approximation accuracy and computation trade-off.
(16)
Note that WL can follow the same proposed techniques to reduce parameter size. Therefore, standard
SVGD algorithms can be applied to sample from the
posterior distribution of each model parameter. Intuitively, in SVGD the kernel function governs the interactions between particles, which employs this information to accelerate convergence and provide better
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
performance. Similarly, the Householder flow, encoding structural information, controls the interactions
between weights in each particle.
4
Sequential Decision Problems
A principal motivation for the proposed S2 VGD framework is sequential decision problems, including contextual multi-arm bandits (CMABs) and Markov decision processes (MDPs). A challenge in sequential
decision problems in the face of uncertainty is the exploration/exploitation trade-off: the trade-off between
either taking actions that are most rewarding according to the current knowledge, or taking exploratory actions, which may be less immediately rewarding, but
may lead to better-informed decisions in the future. In
a Bayesian setting, the exploration/exploration tradeoff is naturally addressed by imposing uncertainty into
the parameters of a policy model.
4.1
new procedure Stein Thompson sampling, summarized
in Algorithm 1.
Note our Stein Thompson sampling is a general scheme
for exploration/exploitation balance in CMABs. The
techniques in [Russo et al., 2017, Kawale et al., 2015]
can be adapted in this framework; we leave this for
future work.
Algorithm 1 Stein Thompson Sampling
Require: D = ∅; initialize particles Θ0 = {θi }M
i=1 ;
1: for t = 0, 1, 2, . . . , T do
2:
Receive context st ∼ Ps ;
3:
Draw a particle θ̂ t from Θt ;
4:
Select at = arg maxa Er∼Pr (·|a,st ;θ̂t ) rt ;
5:
Observe reward rt ∼ Pr , by performing at ;
6:
Collect observation: Dt+1 = Dt ∪ (st , at , rt );
7:
Update Θt+1 , according to SVGD in (9);
8: end for
CMABs and Stein Thompson Sampling
CMABs model stochastic, discrete-time and finite
action-state space control problems. A CMAB is formally defined as a tuple C = hS, A, Ps , Pr , ri, where
S is the state/context space, A the action/arm space,
r ∈ R is the reward, Ps and Pr are the unknown environment distributions to draw the context and reward, respectively. At each time step t, the agent (i)
first observes a context st ∈ S, drawn i.i.d. over time
from Ps ; then (ii) chooses an action at at ∈ A and
observes a stochastic reward rt (at , st ), which is drawn
i.i.d. over time from Pr (·|at , st ), conditioned on the
current context and action. The agent makes decisions via a policy π(a|s) that maps each context to
a distribution over actions, yielding the probability of
choosing action a in state s. The goal in CMABs is to
learn a policy to maximize the expected
PT total reward
in T interactions: J(π) = EPs ,π,Pr t=1 rt .
We represent the policy using a θ-parameterized neural network πθ (a|s), where MVG priors p(θ) are employed on the weights. At each time t, given the past
observations Dt , {d}tj=1 , where dj = (sj , aj , rj ), the
posterior distribution of θt is updated as p(θt |Dt ) ∝
Q
t
j=1 p(dj |θ)p(θ).
Thompson sampling [Thompson, 1933] is a popular
method to solve CMABs [Li et al., 2011]. It approximates the posterior p(θ|Dt ) in an online manner. At
each step, Thompson sampling (i) first draws a set
of parameter samples, then (ii) picks the action by
maximizing the expected reward over current step,
i.e., at = arg maxa Er∼Pr (·|a,st ;θt ) rt , (iii) collects data
samples after observing the reward rt , and (iv) updates posterior of the policy. We apply the proposed
S2 VGD for the updates in the final step, and call the
4.2
MDPs and Stein Policy Gradient
An MDP is a sequential decision-making procedure
in a Markovian dynamical system. It can be seen
as an extension of the CMAB, by replacing the context with the notion of a system state, that may dynamically change according to the performed actions
and previous state. Formally, an MDP defines a tuple
M = hS, A, Ps , Pr , r, γi, which is similar to a CMAB C
except that (i) the next state st+1 is now conditioned
on state st and action at , i.e., st+1 ∼ Ps (·|st , at );
and (ii) a discount factor 0 < γ < 1 for the reward is considered. The goal is to find a policy
π(a|s) to maximize the discounted expected reward:
PT
J(π) = EPs ,π,Pr t=1 γ t rt .
Policy gradient [Sutton and Barto, 1998] is a family
of reinforcement learning methods that solves MDPs
by iteratively updating the parameters θ of the policy
to maximize J(θ) , J(πθ (a|s)). Instead of searching for a single policy parameterized by θ, we consider
adopting an MVG prior for p(θ), and learning its variational posterior distribution q(θ) using S2 VGD. Following [Liu et al., 2017], the objective function is modified as:
max{Eq(θ) [J(θ)] − αKL(qkp)},
q
(17)
where α ∈ [0, +∞) is the temperature hyperparameter to balance exploitation and exploration in
the policy. The optimal distribution is shown to have
a simple closed form [Liu et al., 2017]:
q(θ) ∝ exp
1
J(θ) p(θ).
α
(18)
Learning Structural Weight Uncertainty for Sequential Decision-Making
We iteratively approximate the target distribution as:
M
X
1
4θi =
[∇θj
J(θj ) + log p(θj ) κ(θi , θj )
M j=1
α
+ ∇θj κ(θj , θi )],
(19)
where J(θ) can be approximated with REINFORCE [Williams, 1992] or advantage actor
critic [Schulman et al., 2016].
We note two advantages of S2 VGD in sequential
decision-making: (i ) the structural priors can characterize the flexible weight uncertainty, thus providing
better exploration-exploitation when learning the policies; (ii ) the efficient approximation scheme provides
accurate representation of the true posterior while
maintaining similar online-processing speed.
5
Experiments
To demonstrate the effectiveness of our S2 VGD, we
first conduct experiments on the standard regression
and classification tasks, with real datasets (two synthetic experiments on classification and regression are
given in the Supplementary Material). The superiority
of S2 VGD is further demonstrated in the experiments
on contextual bandits and reinforcement learning.
We compare S2 VGD with related Bayesian learning algorithms, including VMG [Louizos and Welling, 2016],
PBP MV [Sun et al., 2017], and SVGD [Liu and Wang,
2016]. The RMSprop optimizer is employed if there is
no specific declaration. For SVGD-based methods, we
use a RBF kernel κ(θ, θ 0 ) = exp(−kθ − θ 0 k22 /h), with
the bandwidth set to h = med2 / log M . [Oates et al.,
2016, Gorham and Mackey, 2017] Here med is the median of the pairwise distance between particles. The
hyper-parameters a` = 1, b` = 0.1. All experiments
are conducted on a single TITAN X GPU.
We first study the role of hyperparameters in S2 VGD:
the number of Householder transformations K and the
number of particles M . This is investigated by a classification task from [Liu and Wang, 2016] on the Covertype dataset with 581,012 data points and 54 features.
0.76
Accuracy
0.74
0.72
M=3
M=5
M=10
M=20
M=30
M=50
0.70
0.68
0.66
1
3
5
K
10
16
Figure 2: Impact of K and M .
We perform 5 runs for each setting and report the
mean of testing accuracy in Figure 2. As expected,
increasing M or K improves the performance, as they
lead to a more accurate approximation. Interestingly,
when M is small, increasing K gives significant improvement. Furthermore, when M is large, the change
of K yields similar performance. Therefore, we set
K = 1 and M = 20 unless otherwise specified.
5.1
Regression
We use a single-layer BNN for regression tasks. Following [Li et al., 2015], 10 UCI public datasets are considered: 100 hidden units for 2 large datasets (Protein
and YearPredict), and 50 hidden units for the other 8
small datasets. We repeat the experiments 20 times for
all datasets except for Protein and YearPredict, which
we repeat 5 times and once, respectively, for computation considerations [Sun et al., 2017]. The batch size
for the two large datasets is set to 1000, while it is 100
for the small datasets. The datasets are randomly split
into 90% training and 10% testing. We adopt the root
mean squared error (RMSE) and test log-likelihood as
the evaluation criteria.
The experimental results are shown in Table 1, from
which we observe that i) weight structure information is useful (SVGD is the only method without
structure, and it yields inferior performance); ii) algorithms with non-parametric assumptions, i.e., the
Stein-based methods, perform better; and iii) when
combined with structure information, our method
achieves state-of-the-art results.
5.2
Classification
We perform the classification tasks on the standard
MNIST dataset, which consists of handwritten digits
of size 28 × 28, with 50,000 images for training and
10,000 for testing. A two-layer model 784-X-X-10 with
ReLU activation function is used, and X is the number
of hidden units for each layer. The training epoch is
set to 100. The test errors for network (X-X) sizes 400400 and 800-800 are reported in Table 2. We observe
that the Bayesian methods generally perform better
than their optimization counterparts. The proposed
S2 VGD improves SVGD by a significant margin. Increasing K also improves performance, demonstrating
the advantages of incorporating structured weight uncertainty into the model. See [Li et al., 2016a, Louizos
and Welling, 2016, Blundell et al., 2015] for details on
the other methods with which we compare.
We wish to verify that the performance gain of S2 VGD
is due to the special structural design of the network
architecture, rather than the increasing number of
model parameters. This is demonstrated by training
a NN with 415-415 hidden units using SVGD, which
yields test error 1.49%. It has slightly more parameters
than our 400-400 network trained by S2 VGD (K=10),
but worse performance.
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
Table 1: Averaged predictions with standard deviations in terms of RMSE and log-likelihood on test sets.
Dataset
Boston
Energy
Concrete
Kin8nm
Naval
CCPP
Winequality
Yacht
Protein
YearPredict
VMG
2.70 ± 0.13
0.54 ± 0.02
4.89 ± 0.12
0.08 ± 0.00
0.00 ± 0.00
4.04 ± 0.04
0.61 ± 0.04
0.48 ± 0.18
4.13 ± 0.02
8.78 ± NA
Test RMSE
PBP MV
SVGD
2.76 ± 0.43
2.96±0.10
0.48 ± 0.04
1.37±0.05
4.66 ± 0.44
5.32±0.10
0.08 ± 0.00
0.09±0.00
0.00 ± 0.00
0.00±0.00
3.91 ± 0.09
4.03±0.03
0.61 ± 0.02
0.61±0.01
0.53 ± 0.14
0.86±0.05
4.38 ± 0.01
4.61±0.01
8.84 ± NA 8.68 ± NA
VMG
-2.46 ± 0.09
-1.06 ± 0.03
-3.01 ± 0.03
1.10 ± 0.01
2.46 ± 0.00
-2.82 ± 0.01
-0.95 ± 0.01
-1.30 ± 0.02
−2.84 ± 0.00
-3.59 ± NA
Test Log likelihood
PBP MV
SVGD
-3.01 ± 0.26 -2.50±0.03
-2.37 ± 0.03 -1.77±0.02
-3.22 ± 0.05 -3.08±0.02
0.78 ± 0.02
0.98±0.01
4.37 ± 0.17
4.09±0.01
-2.81 ± 0.02 -2.82±0.01
-0.99 ± 0.07 -0.93±0.01
-1.67 ± 0.24 -1.23±0.04
-2.91 ± 0.03 -2.95±0.00
-3.58±NA
-3.62 ± NA
S2VGD
SVGD
VMG
2000
( = 0%)greedy
( = 1%)greedy
( = 3%)greedy
0
0
10000 20000 30000 40000 50000
Trial #
(a) Simulation results on Mushroom
1.2
1.0
S2VGD
SVGD
VMG
LinUCB
0.8
0.6
0
1
2
3
Visits Number
4
1e7
(b) New Article Recommendation
S2 VGD
−2.43 ± 0.10
−0.55 ± 0.04
−2.90 ± 0.07
1.15 ± 0.01
4.79 ± 0.05
−2.77 ± 0.02
−0.90 ± 0.03
−0.81 ± 0.14
-2.84 ± 0.01
−3.57 ± NA
S2VGD
SVGD
1.32
Click Through Rate
4000
Click Through Rate
1.4
6000
Cumulative Regret
S2 VGD
2.56 ± 0.33
0.38 ± 0.02
4.25 ± 0.37
0.07 ± 0.00
0.00 ± 0.00
3.84 ± 0.08
0.59 ± 0.02
0.47 ± 0.11
4.15 ± 0.04
8.73 ± NA
1.30
1.28
1.26
1.24
1
3
5
10 20
Particle Size
30
50
(c) Particle Size on Yahoo!Today
Figure 3: Experimental results of Contextual Bandits
Table 2: Classification error of FNN on MNIST.
Method
S2 VGD (K=10)
S2 VGD (K=1)
SVGD
SGLD
RMSprop
RMSspectral
SGD
VMG, variational dropout
BPB, Gaussian
BPB, scale mixture
SGD, dropout
5.3
Test Error
400-400 800-800
1.36%
1.30%
1.43%
1.39%
1.53%
1.47%
1.64%
1.41%
1.59%
1.43%
1.65%
1.56%
1.72%
1.47%
1.15%
1.82%
1.99%
1.32%
1.34%
1.51%
1.33%
Contextual Bandits
Simulation We first simulate a contextual-bandit
problem with the UCI mushrooms dataset. Following
[Blundell et al., 2015], the provided features of each
mushroom are regarded as the context. A reward of 5
is given when an agent eats an edible mushroom. Otherwise, if a mushroom is poisonous and an agent eats it,
a reward of -10 or 5 will be received, both with probability 0.5; if the agent decides not to eat the mushroom,
it receives a reward of 0. We use a two-layer BNN with
ReLU and 50 hidden units to represent the policy of
an agent. We compared our method with a standard
baseline, ε-greedy policy with ε = 0% (pure greedy),
1%, 3%, respectively [Sutton and Barto, 1998].
We evaluate the performance of different algorithm by
cumulative regret [Sutton and Barto, 1998], a measure
of the loss caused by playing suboptimal bandit arms.
The results are plotted in Figure 3(a). Thompson sampling with 3 different strategies to update policy are
considered: S2 VGD, SVGD and VMG. S2 VGD shows
lower regret at the beginning of learning than SVGD,
and the lowest final cumulative regret among all methods. We hypothesize that our method captures the
internal weight correlation, and the structural uncertainty can effectively help the agent learn to make less
mistakes in exploration with less observations.
News Article Recommendation We consider
personalized news article recommendation on Yahoo!
[Li et al., 2010], where each time a user visits the portal, a news article from a dynamic pool of candidates
is recommended based on the user’s profile (context).
The dataset contains 45,811,883 user visits to the Today Module in a 10-day period in May 2009. For each
visit, both the user and each of the 20 candidate articles are associated with a feature vector of 6 dimensions [Li et al., 2010].
The goal is to recommend an article to a user based on
her behavior, or, formally, maximize the total number
of clicks on the recommended articles. The procedure
is regraded as a CMAB problem, where articles are
treated as arms. The reward is defined to be 1 if the
article is clicked on and 0 otherwise. A one-layer NN
with ReLU and 50 hidden units is used as a policy net-
Learning Structural Weight Uncertainty for Sequential Decision-Making
5.4
50
0
50
100
0
Figure 4 plots the mean (dark curves) and standard
derivation (light areas) of discounted rewards over 5
runs. In all tasks and value-estimation setups, S2 VGD
converges faster than SVGD and finally converges to
higher average rewards. The results are even comparable to [Houthooft et al., 2016], in which a subtle
reward mechanism is incorporated to encourage exploration. It demonstrates that simply adding structural
information on policy networks using S2 VGD improves
the agent’s exploration ability. We also add a baseline
method called SVGD* that applies SVGD to train a
network of similar size (25-16 hidden units) with the
one reparameterized by S2 VGD (K=4). The fact that
200
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
400 600 800 1000
Episodes
Average Reward
Average Reward
100
100
0
200
Double Pendulum (A2C)
Double Pendulum (REINFORCE)
400
Average Reward
600
800
1000
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
1200
0
200
400 600
Episodes
800
600
800
1000
1200
1000
0
200
4000
4000
3000
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
1000
0
20
40
60
Episodes
80
100
Average Reward
5000
2000
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
400 600 800 1000
Episodes
Cartpole (A2C)
Cartpole (REINFORCE)
5000
0
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
400 600 800 1000
Episodes
0
100
400
Reinforcement Learning
We apply our S2 VGD to policy gradient learning. All
experiments are conducted with the OpenAI rllab
toolkit [Duan et al., 2016]. Three classical continuous
control tasks are considered: Cartpole Swing-Up, Double Pendulum, and Cartpole. Following the settings in
[Liu et al., 2017, Houthooft et al., 2016], the policy is
parameterized as a two-layer (25-10 hidden units) neural network with tanh as the activation function. The
maximal length of horizon is set to 500. SVGD and
S2 VGD use a sample size of 10000 for policy gradient
estimation, and M = 16, α = 10. For the easy task,
Cartpole, all agents are trained for 100 episodes. For
the two complex tasks, Cartpole Swing-Up and Double
Pendulum, all agents are trained up to 1000 episodes.
We consider two different methods to estimate the
state values: REINFORCE [Williams, 1992] and advantage actor critic (A2C) [Schulman et al., 2016]
200
150
Average Reward
To further verify the influence of the particle size M
on the sequential decision problem, we vary the M
from 1 to 50 on the-first-day data. All algorithms are
repeated 10 times and their mean performances are
plotted in Figure 3(c). We observe that the CTR keeps
increasing when M becomes larger. S2 VGD dominates
the performance of SVGD with much higher CTRs,
the gap becomes larger as M increases. Since larger M
typically leads to more accurate posterior estimation of
the policy, indicating again that the accurately learned
structural uncertainty are beneficial for CMABs.
Cartpole Swing Up (A2C)
Cartpole Swing Up (REINFORCE)
200
Average Reward
work. The classic LinUCB [Li et al., 2010] is also compared as the baseline. The performance is evaluated
by an unbiased offline evaluation protocol: the average
normalized accumulated click-through-rate (CTR) in
every 20000 observations [Li et al., 2010, 2011]. The
normalized CTRs are plotted in Figure 3(b). It is clear
that S2 VGD consistently outperforms other methods.
The fact that S2 VGD and VMG perform better than
SVGD and the baseline LinUCB indicates that structural information helps algorithms to better balance
exploration and exploitation.
3000
2000
1000
0
0
20
SVGD
SVGD*
S2VGD (K=1)
S2VGD (K=4)
40
60
80
100
Episodes
Figure 4: Learning curves by S2 VGD and SVGD with
REINFORCE (left) and A2C (right).
S2 VGD (K=4) converges better than SVGD* demonstrates that our structural uncertainty is key to excellent performance.
6
Conclusions
We have proposed S2 VGD, an efficient Bayesian posterior learning scheme for the weights of BNNs with
structural MVG priors. To achieve this, we derive a
new reparametrization for the MVG to unify previous structural priors, and adopt the SVGD algorithm
for accurate posterior learning. By transforming the
MVG into a lower-dimensional representation, S2 VGD
avoids computation of related kernel matrices in highdimensional space. The effectiveness of our framework
is tested on several real-world tasks, including regression, classification, contextual bandits and reinforcement learning. Extensive experimental results demonstrate its superiority relative to related algorithms.
Our empirical results on sequential decision problems
suggest the benefits of including inter-weight structure within the model, when computing policy uncertainty for online decision-making in an uncertain environments. More sophisticated methods for leveraging
uncertainty for exploration/exploration balance may
be a promising direction for future work. For example, explicitly encouraging exploration using learned
structural uncertainty [Houthooft et al., 2016].
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
Acknowledgements We acknowledge Qiang Liu
and Yang Liu for making their code public. This research was supported in part by ARO, DARPA, DOE,
NGA, ONR and NSF.
J Zico Kolter and Andrew Y Ng. Near-bayesian exploration in polynomial time. In ICML, 2009.
References
Charles Blundell,
Julien Cornebise,
Koray
Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. In ICML, 2015.
Chunyuan Li, Changyou Chen, David E Carlson, and
Lawrence Carin. Preconditioned stochastic gradient langevin dynamics for deep neural networks. In
AAAI, 2016a.
Yan Duan, Xi Chen, Rein Houthooft, John Schulman,
and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. In ICML,
2016.
Chunyuan Li, Changyou Chen, Kai Fan, and Lawrence
Carin. High-order stochastic gradient thermostats
for bayesian learning of deep models. In AAAI,
2016b.
Meire Fortunato, Charles Blundell, and Oriol Vinyals.
Bayesian recurrent neural networks. arXiv preprint
arXiv:1704.02798, 2017.
Chunyuan Li, Andrew Stevens, Changyou Chen,
Yunchen Pu, Zhe Gan, and Lawrence Carin. Learning weight uncertainty with stochastic gradient
mcmc for shape classification. In CVPR, 2016c.
Yarin Gal and Zoubin Ghahramani. Dropout as a
bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016.
Zhe Gan, Chunyuan Li, Changyou Chen, Yunchen Pu,
Qinliang Su, and Lawrence Carin. Scalable bayesian
learning of recurrent neural networks for language
modeling. ACL, 2017.
Soumya Ghosh, Francesco Maria Delle Fave, and
Jonathan Yedidia. Assumed density filtering methods for learning bayesian neural networks. In AAAI,
2016.
Gene H Golub and Charles F Van Loan. Matrix Computations. 2012.
Jackson Gorham and Lester Mackey. Measuring sample quality with kernels. arXiv:1703.01717, 2017.
Arjun K Gupta and Daya K Nagar. Matrix Variate
Distributions. 1999.
José Miguel Hernández-Lobato and Ryan Adams.
Probabilistic backpropagation for scalable learning
of bayesian neural networks. In ICML, 2015.
Geoffrey E Hinton, Nitish Srivastava, and Kevin Swersky. Rmsprop: Divide the gradient by a running average of its recent magnitude. Neural Networks for
Machine Learning, Coursera, 2012.
Rein Houthooft, Xi Chen, Yan Duan, John Schulman, Filip De Turck, and Pieter Abbeel. Vime:
Variational information maximizing exploration. In
NIPS, 2016.
Jaya Kawale, Hung H Bui, Branislav Kveton, Long
Tran-Thanh, and Sanjay Chawla. Efficient thompson sampling for online matrix-factorization recommendation. In NIPS, 2015.
Diederik Kingma and Jimmy Ba. Adam: A method
for stochastic optimization. In ICLR, 2015.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS, 2012.
Lihong Li, Wei Chu, John Langford, and Robert E
Schapire. A contextual-bandit approach to personalized news article recommendation. In WWW, 2010.
Lihong Li, Wei Chu, John Langford, and Xuanhui
Wang. Unbiased offline evaluation of contextualbandit-based news article recommendation algorithms. In WSDM, 2011.
Yingzhen Li, José Miguel Hernández-Lobato, and
Richard E Turner. Stochastic expectation propagation. In NIPS, 2015.
Qiang Liu. Stein variational gradient descent as gradient flow. In NIPS, 2017.
Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference
algorithm. In NIPS, 2016.
Yang Liu, Prajit Ramachandran, Qiang Liu, and Jian
Peng. Stein variational policy gradient. In UAI,
2017.
Christos Louizos and Max Welling. Structured and
efficient variational deep learning with matrix gaussian posteriors. In NIPS, 2016.
David JC MacKay. A practical bayesian framework
for backpropagation networks. Neural Computation,
1992.
Vinod Nair and Geoffrey E Hinton. Rectified linear
units improve restricted boltzmann machines. In
ICML, 2010.
Chris J Oates, Jon Cockayne, François-Xavier Briol,
and Mark Girolami.
Convergence rates for
a class of estimators based on stein’s identity.
arXiv:1603.03220, 2016.
Daniel Russo, David Tse, and Benjamin Van Roy.
Time-sensitive bandit learning and satisficing
thompson sampling. arXiv:1704.09028, 2017.
Learning Structural Weight Uncertainty for Sequential Decision-Making
John Schulman, Philipp Moritz, Sergey Levine,
Michael Jordan, and Pieter Abbeel.
Highdimensional continuous control using generalized
advantage estimation. In ICLR, 2016.
David Silver, Aja Huang, et al. Mastering the game
of go with deep neural networks and tree search.
Nature, 2016.
Shengyang Sun, Changyou Chen, and Lawrence Carin.
Learning structured weight uncertainty in bayesian
neural networks. In AISTATS, 2017.
Xiaobai Sun and Christian Bischof. A basis-kernel representation of orthogonal matrices. SIAM Journal
on Matrix Analysis and Applications, 1995.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks.
In NIPS, 2014.
Richard S Sutton and Andrew G Barto. Reinforcement
learning: An introduction. 1998.
William R Thompson. On the likelihood that one unknown probability exceeds another in view of the
evidence of two samples. Biometrika, 1933.
Jakub M Tomczak and Max Welling. Improving
variational auto-encoders using householder flow.
arXiv:1611.09630, 2016.
Ronald J Williams.
Simple statistical gradientfollowing algorithms for connectionist reinforcement
learning. Machine Learning, 1992.
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
Learning Structural Weight Uncertainty for
Sequential Decision-Making: Supplementary Material
Ruiyi Zhang1
Chunyuan Li1 Changyou Chen2 Lawrence Carin1
Duke University 2 University at Buffalo
[email protected], [email protected], [email protected], [email protected]
1
A
B
Proof of Proposition 3
Proof. Let a MVG distributed matrix W be
W ∼ MN (W; M, U, V),
B.1
(20)
Since the covariance matrices U and V are positive
definite, we can decompose them as
U = PΛ1 Λ1 P> ,
(21)
>
(22)
V = QΛ2 Λ2 Q ,
where P and Q are the corresponding orthogonal matrices, i.e., PP> = I, QQ> = I.
According to Lemma 2, we have,
P> W ∼ MN (P> W; M, P> UP, V),
(23)
>
Since U = PΛ1 Λ1 P , and we have:
P> W ∼ MN (P> W; M, P> PΛ1 Λ1 P> P, V). (24)
Then,
P> W ∼ MN (P> W; M, Λ1 Λ1 , V),
(25)
Similarly, we have,
P> WQ ∼ MN (P> WQ; M, Λ1 Λ1 , Λ2 Λ2 ).
(26)
Further,
>
−1
−1 >
−1
Λ−1
1 P WQΛ2 ∼ MN (Λ1 P WQΛ2 ; 0, I, I), (27)
−1
>
Define C = Λ−1
1 P WQΛ2 , then C follows an independent Gaussian distribution:
−1
C ∼ MN (C; P> Λ−1
1 MΛ2 Q, I, I),
(28)
(28) can also be expressed as:
−1
vec(C) ∼ N (vec(C); P> Λ−1
1 MΛ2 Q, I),
showing that vectorized elements form of C follows
an isotropic Gaussian distribution. Finally, since
Λ1 CΛ2 =P> WQ , we have:
W = PΛ1 CΛ2 Q> .
(30)
The Upperbound of Orthogonal Degree
Lemma 5 ([Sun and Bischof, 1995] The Basis-Kernel
Representation of Orthogonal Matrices). For any m ×
m orthogonal matrix Q, there exist a full-rank m × k
matrix Y and a nonsingular k × k matrix S, k ≤ m,
such that:
Q , Q(Y, S) = I − YSY>
(31)
Definition 6 ([Sun and Bischof, 1995] Active Subspace). Orthogonal matrix Q acts on the space R(Y)⊥
as the identity and changes every nonzero vector in
R(Y), and R(Y)⊥ is the active space of Q
In basis-kernel representation, S is the kernel, and Y
is the basis. The degree of an orthogonal matrix is defined as the dimension of its active subspace. Specifically, Householder matrix is an orthogonal matrix of
degree 1. With the introduced definitions and Lemma
5, the degree of an orthogonal matrix is bounded by
Lemma 7.
Lemma 7 ([Sun and Bischof, 1995]). Let A and B be
two m-by-k matrices, k < m. If B = QA for some
orthogonal matrix Q, then Q is either of degree no
greater than k or can be replaced by an orthogonal factor of its own with degree no greater than k.
Since the degree of orthogonal matrix is bounded by
size of the matrix (k), the number of Householder
transformations needed for Householder flow is also
bounded.
B.2
(29)
Properties of Householder Flow
Intuitive Explanation
In our setting, the orthogonal matrix works as a rotation matrix. The Householder transformation reflects
the weights by a hyperplane orthogonal to its corresponding Householder vector. Hence, Householder
flows applies a series of reflections to the original
weights. In another word, it rotates the weight matrix, equivalent to the effects of the rotation matrix.
Learning Structural Weight Uncertainty for Sequential Decision-Making
C
C.1
Running Time Analysis
models, because it is more similar to the ground truth
and has a balanced posterior density. PBP MV employed the structure information but is a little unbalanced. The SVGD is more unbalanced compared
with PBP MV. For non-linear regression, we follow
Time Comparison
Table 3: Time Comparison between SVGD and S2 VGD.
Time (sec)
Dataset
S2 VGD SVGD
Boston Housing
35
32
52
39
Energy
Concrete
48
40
50
39
Kin8nm
Naval
63
53
44
33
CCPP
Winequality
54
44
Yacht
43
36
671
495
Protein
YearPredict
8209
6306
We use a single-layer NN, and compare their running
time on Intel Xeon E5-2640 v4 2.4GHz (CPU). On all
datasets, S2 VGD takes slightly more time to converge
than SVGD.
C.2
Computational Trick
Figure 5: Binary classification on the synthetic dataset,
Top left is SVGD, top right is the ground truth, bottom
left is PBP MV and bottom right is ours. The dots indicate both training and testing data points, different colors
illustrates the prediction density, and the higher confidence
the model has, the deeper the color will be.
Assume v is a Householder vector, and its correspondvv >
ing Householder matrix H , I − 2
. For a feature
kvk2
vector x, defining v̂ , v/kvk, then it is easy to show:
Hx = x − 2
vv > x
kvk2
v
kvk2
= x − 2hx, v̂iv̂ .
= x − 2v > x
(32)
(33)
(34)
To efficiently compute the Householder transformation, we do not need to revert the Householder matrix
H and apply Hx to complete one Householder transformation. (32) can be used to drastically reduce the
computational cost and thus make the Householder
flows more efficient.
D
Experimental Results
We follow the toy example in [Ghosh et al., 2016]
to consider the binary classification task, in which
we uniformly sample 10 data points in seperate
distributions:[−3, −1] × [−3, −1] and [1, 3] × [1, 3]. A
one-layer BNNs with 30 hidden units is employed,
we use at most 200 epochs for PBP MV, SVGD and
S2 VGD. The posterior prediction density is plotted
in Figure D. The S2 VGD performs better than other
Figure 6: Regression on synthetic dataset. From the left
to the right are PBP MV, SVGD, S2 VGD. The blue line
is the ground truth. The light grey area shows the ±3
standard derivation confidence intervals; The green line is
the mean of predictions.
the experiment setup of [Louizos and Welling, 2016,
Sun et al., 2017]. We randomly generated 20 points
as input xn , in which 12 points are sampled from
Uniform(0,0.6), and 8 points are sampled from Uniform(0.8,1). The output yn is corresponding to xn ,
and yn = xn + n + sin(4(xn + n )) + sin(13(xn +
n )), where n ∼ N (n , 0.0009). We fit a one-layer
ReLU neural network with 100 hidden units. We run
PBP MV, SVGD and S2 VGD for at most 1500 epochs.
Compared with other methods, S2 VGD captures uncertainty on two-sides using its variance, but other
methods only capture part of the uncertainty.
Ruiyi Zhang, Chunyuan Li, Changyou Chen, Lawrence Carin
Table 4: Hyper-parameter settings for policy gradient experiments.
Datasets
CartpoleSwingup DoublePendulum Cartpole
Batch Size
5000
5000
5000
Step Size
5×10−3
5×10−3
5×10−3
1000
1000
100
#Episodes
Discount
0.99
0.99
0.99
Network (hidden layers)
[25, 10]
[25, 10]
[25, 10]
Variance in prior
0.01
0.01
0.01
E
Experimental setting
We discuss the hyper-parameter settings for S2 VGD,
and then provide their values for our experiments.
E.1
resented by a series of Householder transformations,
when K is the degree. See details in Supplementary
D. In experiments, the number of Householder transformations balanced orthogonal matrix approximation
accuracy and computational cost.
Discussion of Hyper-parameters
Step Size The step size for SVGD and S2 VGD
corresponds to the optimization counterparts. A block
decay strategy is used on several datasets, it decreases
by the stepsize by half every L epochs.
Mini-batch Size The gradient at step t is evaluated on a batch of data St . For small datasets,
the batch size can be set to the training sample size
|St | = N , giving the true gradient for each step. For
large datasets, a stochastic gradient evaluated from a
mini-batch of size |St | < N is used to
Variance of Gaussian Prior The prior distributions on the parameterized weights of DNNs are Gaussian, with mean 0 and variance σ 2 . The variance of
this Gaussian distribution determines the prior belief
of how strongly these weights should concentrate on 0.
This setting depends on user perception of the amount
variability existing in the data. A larger variance in
the prior leads to a wider range of weight choices, thus
higher uncertainty. The weight decay value of `2 regularization in stochastic optimization is related to the
prior variance in SVGD and S2 VGD.
Number of Particles The number of particles M ,
is employed to approximate the posterior. SVGD
and S2 VGD represents the posterior approximately in
terms of a set of particles (samples), and is endowed
with guarantees on the approximation accuracy when
the number of particles is exactly infinity [Liu, 2017].
The number of particles balanced posterior approximation accuracy and computational cost.
Number of Householder Transformations Instead of maintaining a full orthogonal matrix, we
approximate it employing householder flows containing K Householder transformations. According to
Lemma 4, the orthogonal matrix can be exactly rep-
Table 5: Hyper-parameter settings for contextual bandits.
Datasets
Batch Size
Step Size
#Trial
Network (hidden layers)
Variance in prior
Mushroom
64
10−3
5×104
[50]
1
Yahoo!Today
100
10−3
4.5×107
[50]
1
Table 6: Hyper-parameter settings for MNIST.
Datasets
MNIST
Batch Size
100
100
Step Size
5× 10−4
5× 10−4
#Epoch
150
150
RMSProp
0.99
0.99
Network (hidden layers) [400, 400] [800, 800]
Variance in prior
1
1
E.2
Settings in Our Experiments
The hyper-parameter settings of SVGD and S2 VGD on
each dataset is specified in Table 6 for MNIST, Table 5
for contextual bandits and Table 4 for reinforcement
learning.
| 2cs.AI
|
arXiv:1603.04337v1 [math.CO] 14 Mar 2016
On the structure of the power graph and
the enhanced power graph of a group ∗
Ghodratollah Aalipour†
Saieed Akbari‡
Peter J. Cameron§
Reza Nikandish¶
Farzad Shaveisik
Abstract
Let G be a group. The power graph of G is a graph with the vertex set G,
having an edge between two elements whenever one is a power of the other. We
characterize nilpotent groups whose power graphs have finite independence number. For a bounded exponent group, we prove its power graph is a perfect graph
and we determine its clique/chromatic number. Furthermore, it is proved that for
every group G, the clique number of the power graph of G is at most countably
infinite. We also measure how close the power graph is to the commuting graph
by introducing a new graph which lies in between. We call this new graph as the
enhanced power graph. For an arbitrary pair of these three graphs we characterize
finite groups for which this pair of graphs are equal.
1
Introduction
We begin with some standard definitions from graph theory and group theory.
Let G be a graph with vertex set V (G). If x ∈ V (G), then the number of vertices
adjacent to x is called the degree of x, and denoted by deg(x). The distance between
two vertices in a graph is the number of edges in a shortest path connecting them.
The diameter of a connected graph G, denoted by diam(G), is the maximum distance
between any pair of vertices of G. If G is disconnected, then diam(G) is defined to be
∗
Key Words: Power graph, Clique number, Chromatic number, Independence number, Group.
2010 Mathematics Subject Classification: 05C25, 05C69, 20D60.
†
Department of Mathematics and Computer Sciences, Kharazmi University, 50 Taleghani Avenue,
Tehran, Iran and and Department of Mathematical and Statistical Sciences, University of Colorado
Denver, CO 80217, USA ([email protected])
‡
Department of Mathematical Sciences, Sharif University of Technology (s [email protected])
§
School of Mathematics and Statistics, University of St Andrews and School of Mathematical Sciences, Queen Mary, University of London ([email protected])
¶
Department of Basic Sciences, Jundi-Shapur University of Technology, Dezful, Iran. P.O. BOX
64615-334 ([email protected])
k
Department of Mathematics, Faculty of Sciences, Razi University, Kermanshah, Iran
([email protected])
1
infinite. A star is a graph in which there is a vertex adjacent to all other vertices, with
no further edges. The center of a star is a vertex that is adjacent to all other vertices.
Let U ⊆ V (G). The induced subgraph on U is denoted by hU i. An independent set is
a set of vertices in a graph, no two of which are adjacent; that is, a set whose induced
subgraph is null. The independence number of a graph G is the cardinality of the largest
independent set and is denoted by α(G). A subset S of the vertex set of G is called a
dominating set if for every vertex v of G, either v ∈ S or v is adjacent to a vertex in S.
The minimum size of dominating sets of G, denoted by γ(G), is called the domination
number of G. A clique in a graph is a set of pairwise adjacent vertices. The supremum
of the sizes of cliques in G, denoted by ω(G), is called the clique number of G. By
χ(G), we mean the chromatic number of G, i.e., the minimum number of colours which
can be assigned to the vertices of G in such a way that every two adjacent vertices have
different colours.
The cyclic group of order n is denoted by Cn . A group G is called periodic if every
element of G has finite order. For every element g ∈ G, the order of g is denoted by
o(g). If there exists an integer n such that for all g ∈ G, gn = e, where e is the identity
element of G, then G is said to be of bounded exponent. If G is of bounded exponent,
then the exponent of G is the least common multiple of the orders of its elements; that
is, the least n for which gn = e for all g ∈ G. A group G is said to be torsion-free if
apart from the identity every element of G has infinite order. Let p be a prime number.
The p-quasicyclic group (known also as the Prüfer group) is the p-primary component
of Q/Z, that is, the unique maximal p-subgroup of Q/Z. It is denoted by Cp∞ . The
center of a group G, denoted by Z(G), is the set of elements that commute with every
element of G. A group G is called locally finite if every finitely generated subgroup of
G is finite. A group is locally cyclic if any finitely generated subgroup is cyclic.
Other concepts will be defined when needed.
Now, we define the object of interest to us in this paper.
Let G be a group. The power graph of G, denoted by G(G), is the graph whose
vertex set is G, two elements being adjacent if one is a power of the other. This
graph was first introduced for semigroups in [8] and then was studied in [6] and [7] for
groups. It was shown that, for a finite group, the undirected power graph determines
the directed power graph up to isomorphism. As a consequence, two finite groups which
have isomorphic undirected power graphs have the same number of elements of each
order. The authors in [7] have shown that the only finite group whose automorphism
group is the same as that of its power graph is the Klein group of order 4.
Our results about the power graph fall into four classes.
• In Section 2.1, we consider the independence number α(G(G)). We show that if
the independence number is finite then G is a locally finite group whose centre has
finite index. Using this we are able to give precise characterizations of nilpotent
2
groups G for which α(G(G)) is finite – such a group (if infinite) is the direct
product of a p-quasicyclic group and a nilpotent p′ -group.
• In Section 2.2, we show that the power graph of every group has clique number at
most countable. A group with finite clique number must be of bounded exponent.
Hence we obtain a structure theorem for abelian groups with this property, as
well as showing that it passes to subgroups and supergroups of finite index.
• We do not know whether the chromatic number of every group is at most countable; in Section 2.2.1, we prove this for periodic groups and for free groups. We
show that, if G has bounded exponent, then G(G) is perfect.
• Finally, in Section 2.3 there are some miscellaneous results. A group is periodic
if and only if its power graph is connected, and in this case its diameter must be
at most 2. Also we show that, if all vertex degrees in G(G) are finite, then G is
finite.
In the recent paper [10], the authors prove that the power graph of every finite group
is perfect. We acknowledge that our result on the perfectness along with all results in
the Section 2 were proved independently in 2011.
Another well-studied graph associated to a group G is the commuting graph of G.
This graph appears to be first studied by Brauer and Fowler in 1955 in [5] as a part of
classification of finite simple groups. As the elements of the centre are adjacent to all
other vertices, usually the vertices are assumed to be non-central. For more information
on the commuting graph, see [2, 14, 21] and the references therein.
In Section 3 we relate the power graph to the commuting graph and characterize
when they are equal for finite groups. A new graph pops up while considering these
graphs, a graph whose vertex set consists of all group elements, in which two vertices x
and y are adjacent if they generate a cyclic group. We call this graph as the enhanced
power graph of G and we denote it by Ge (G). The enhanced power graph contains
the power graph and is a subgraph of the commuting graph. We further study some
properties of this graph in the Section 3.
We characterize the finite groups for which equality holds for any two of these three
graphs, and the solvable groups for which the power graph is equal to the commuting
graph. Other results are as follows:
• If the power graphs of G and H are isomorphic, then their enhanced power graphs
are isomorphic.
• A maximal clique in the enhanced power graph is either a cyclic or a locally cyclic
subgroup.
• Ge (G) has finite clique number if and only if G has finite exponent; if this holds,
then the clique number of Ge (G) is equal to the largest order of an element of G.
Also, for any group G, the clique number of Ge (G) is at most countable.
3
2
Power graphs of groups
2.1
Independent sets in power graphs
In this section we provide some results on the finiteness of the independence number of
the power graphs. In the proof of the our first theorem, we need the following definition.
Let G be a group and associate with G a graph Γ(G) as follows: the vertices of Γ(G)
are the elements of G and two vertices g and h of Γ(G) are joined by an edge if and
only if g and h do not commute, see [1] and [16] for more details. Now, we have the
following result.
Theorem 1. Let G be a group and α(G(G)) < ∞. Then
(i) [G : Z(G)] < ∞.
(ii) G is locally finite.
Proof. (i) First we note that if x and y are adjacent in Γ(G), then x and y are not
adjacent in G(G). Thus ω(Γ(G)) ≤ α(G(G)) < ∞. Hence [16, Theorem 6] implies that
[G : Z(G)] < ∞.
(ii) Let H be a finitely generated subgroup of G. Then by (i) and [17, 1.6.11], Z(H) is
finitely generated, too. So by the fundamental theorem for finitely generated abelian
groups we find that Z(H) ∼
= Zn × Cq1 × · · · × Cqk , where n and k are non-negative
integers and every qi , 1 ≤ i ≤ k, is a power of a prime number. Since α(G(Z)) = ∞,
we deduce that H is a finite group and so the proof is complete.
Now, we characterize those abelian groups whose power graphs have finite independence number. First we need the following theorem.
Theorem 2.([17, 4.3.11]) If G is an abelian group which is not torsion-free, then it has
a non-trivial direct summand which is either cyclic or quasicyclic.
Theorem 3. Let G be an abelian group such that α(G(G)) < ∞. Then either G is
finite or G ∼
= Cp∞ × H, where H is a finite group and p ∤ |H|.
Proof. If G is torsion-free, then G contains Z and so α(G(G)) ≥ α(G(Z)) = ∞,
a contradiction. Thus by Theorem 2, G = G1 ⊕ H1 , where G1 is either cyclic or
quasicyclic. If H1 is trivial, then we are done. Otherwise, α(G(H1 )) < ∞ implies that
H1 = G2 ⊕ H2 , where G2 is either cyclic or quasicyclic. So G = G1 ⊕ G2 ⊕ H2 . By
repeating this procedure and using α(G(G)) < ∞, we deduce that there exists a positive
Ln
integer n such that G ∼
= i=1 Gi , where every Gi is either cyclic or quasicyclic. We
show that at most one Gi is quasicyclic. By the contrary, suppose that G contains
4
the group Cp∞ × Cq∞ . It is not hard to see that for every positive integer n, In =
{(1/pi + Z, 1/q n−i+1 + Z) : 1 ≤ i ≤ n} is an independent set of size n, a contradiction.
Qn
Qn
So either G ∼
= Cp∞ × i=1 Cpαi or G ∼
= i=1 Cpαi , where p and pi are prime numbers.
i
i
Now, suppose that the first case occurs. To complete the proof, we show that p 6= pi ,
for every i, 1 ≤ i ≤ n. By contrary, suppose that p = pi , for some i. Then Cp∞ × Cp
is a subgroup of G. Since Cp∞ × {1} is an independent set in G(Cp∞ × Cp ), we get a
contradiction. So, the proof is complete.
Theorem 4. Let p be a prime number and G be a p-group such that α(G(G)) < ∞.
Then either G is finite or G ∼
= Cp ∞ .
Proof. Since α(G(G)) < ∞, we deduce that α(G(Z(G))) < ∞. Thus by Theorem 3,
either Z(G) is finite or Z(G) ∼
= Cp∞ , for some prime number p. If Z(G) is finite, then
by Theorem 1, G is finite. Now, suppose that Z(G) ∼
= Cp∞ . To complete the proof, we
show that G is abelian. To the contrary, suppose that there exists a ∈ G \ Z(G). Let
H = hZ(G) ∪ {a}i. Clearly, H is an abelian p-subgroup of G and α(G(H)) < ∞. So,
by Theorem 3, H ∼
= Cp ∞ ∼
= Z(G). Since every proper subgroup of Cp∞ is finite, we get
a contradiction. Hence G is abelian and G = Z(G) ∼
= Cp ∞ .
Now, we exploit Theorem 4 to extend Theorem 3 to nilpotent groups.
Remark 5. Let H and K be two subgroups of G. If H ∩ K = {e}, G = HK and
H ⊆ Z(G), then G ∼
= H × K.
Theorem 6. Let G be an infinite nilpotent group. Then α(G(G)) < ∞ if and only if
G∼
= Cp∞ × H, for some prime number p, where H is a finite group and p ∤ |H|.
Proof. First suppose that G ∼
= Cp∞ × H, where H is a finite group and p ∤ |H|.
Suppose to the contrary, {(sn /pαn + Z, gn ) : n ≥ 1, sn ∈ Z, p ∤ sn and gn ∈ H} is an
infinite independent set of G(G). Since H is a finite group, there exists g ∈ H such
that the infinite set
{(sn /pαn + Z, g) : n ≥ 1, sn ∈ Z and p ∤ sn }
forms an independent set. Since o(g) < ∞, there exist αi and αj such that pαi ≡
pαj (mod o(g)) and αi > αj . On the other hand, we know that gcd(si , p) = 1. So, let
ti be the multiplicative inverse of si in Cpαj . Thus by Chinese Reminder Theorem,
there exists a positive integer x such that x ≡ ti sj (mod pαj ) and x ≡ pαj −αi (mod o(g)).
Therefore, we have
pαi −αj x
si x − sj
sj
si
∈ Z,
− αj =
pαi
p
pαj
gxp
αi −αj
= g.
Thus, (si /pαi + Z, g) and (sj /pαj + Z, g) are adjacent, a contradiction.
5
Conversely, suppose that α(G(G)) < ∞. Then by Theorem 1, [G : Z(G)] < ∞
and so G = Z(G)H, where H is a finitely generated subgroup of G. Now, Theorem 1
implies that H is finite. By Theorem 3, Z(G) = AB, where A ∼
= Cp∞ and B is a finite
group such that p ∤ |B|. Also, since H is nilpotent, we have H ∼
= Hp Hp1 · · · Hpt , where
Hp and Hpi (1 ≤ i ≤ t) are sylow p-subgroup and sylow pi -subgroup of H, respectively.
We show that Hp ⊆ A. To the contrary, suppose that x ∈ Hp \ A. Then hA, xi is a pgroup and so by Theorem 4, hA, xi ∼
= Cp ∞ ∼
= hAi. Since every proper subgroup of Cp∞
is finite, we get a contradiction. Thus Hp ⊆ Z(G) and so G = ABHp1 · · · Hpt . Since
G is nilpotent, BHp1 · · · Hpt is a finite subgroup of G and p ∤ |BHp1 · · · Hpt |. Hence by
Remark 5, G ∼
= A × BHp1 · · · Hpt , as desired.
2.2
The colouring of power graphs
Let G be a group. In this section, we first show that the chromatic number of the
power graph of G is finite if and only if the clique number of the power graph of G is
finite and this statement is also equivalent to that the exponent of G is finite. Then it
is proved that the clique number of the power graph of G is at most countable. Finally,
it is shown that the power graph of every bounded exponent group is perfect.
Lemma 7. Let G be a group. If ω(G(G)) is finite, then G is of bounded exponent.
Proof. By the contrary, suppose that G is not of bounded exponent. Then for every
positive integer k, there is an element gk ∈ G such that o(gk ) > 2k . So one can
i
easily show that {gk2 | 0 ≤ i ≤ k} is a clique of size k + 1 in G(G). This implies that
ω(G(G)) = ∞, a contradiction. The proof is complete.
Remark 8. The proof uses the Axiom of Choice for families of finite sets.
Corollary 9. Let G be an abelian group and ω(G(G)) < ∞. Then there are some
positive integers r and ni and sets Ii , 1 ≤ i ≤ r such that
G∼
=
r Y
Y
Cn i .
i=1 Ii
Proof. Since ω(G(G)) < ∞, by Lemma 7, G is bounded exponent. So the assertion
follows from Prüfer-Baer Theorem (see [17, 4.3.5]).
Theorem 10. The clique number of the power graph of any group is at most countably
infinite.
Proof. Let C be a clique in the power graph of G, and take x ∈ C. Then the
remaining vertices y of C are of two types:
6
• y = xn for some n;
• x = y n for some n.
Clearly, there are at most countably many of the first type. We denote the set of
vertices of the second type by C(n). We show that C(n) is at most countably infinite.
If there is only one y in C(n), then there is nothing to prove; so suppose there are
at least two elements in C(n). We claim that every element in C(n) has finite order.
Choose y, y ′ ∈ C(n). With no loss of generality, one can assume that y ′ = y k , for
some positive integer k. So y (k−1)n = 1. This implies that the orders of both y and
y ′ are finite. Thus the claim is proved. Now, for every positive integer k, define
S
C(n, k) = {y ∈ C(n) | o(y) = k}. By the claim, C(n) = k≥1 C(n, k). It is not hard to
show that for every a, b ∈ C(n, k), hai = hbi and so C(n, k) is finite. Therefore, C(n) is
at most countably infinite.
We wonder if the same result holds for the chromatic number: Does the power graph
of every group have a countable chromatic number? A group is called a pcc-group if its
power graph has at most countable chromatic number. Free groups have this property
by the next theorem.
Theorem 11. Every free group is a pcc-group.
Proof. By [19, Corollary, p.51], in a free group, every non-identity element lies in a
unique maximal cyclic subgroup, generated by an element which is not a proper power
(when written as a reduced word). So, the power graph of a free group consists of
many copies of the power graph of an infinite cyclic group, with the identity in all these
copies identified.
Next, we show that every abelian group is a pcc-group. First, we need the following
result.
Lemma 12. Every periodic group is a pcc-group.
Proof. Suppose that G is a periodic group. For every positive integer n, let Gn be the
set of all elements of G of order n. If g, h ∈ Gn and g and h are adjacent in the power
graph, then g and h generate the same cyclic group. Hence, the induced subgraph on
Gn is a disjoint union of cliques of size φ(n). So one can colour the induced subgraph
S
on Gn with φ(n) colours. Clearly, G = n≥1 Gn and so the chromatic number of G(G)
is at most countable.
Now, we prove that the class of pcc-groups contains the class of abelian groups.
7
Theorem 13. Every abelian group is a pcc-group.
Proof. Let G be abelian. Then G can be embedded in a divisible abelian group ([17,
Theorem 4.1.6, p.98]). It is known that every divisible abelian group is of the form
H × K, where H is a periodic group and K is a direct sum of many copies of Q, see
[17, Theorem 4.1.5, p.97]. By Lemma 12, H is a pcc-group. Now, the next two claims
prove the assertion of the theorem.
Claim 1.
If M and N are two pcc-groups, then M × N is a pcc-group.
Proof of Claim 1: Let f and g be the proper countable colouring G(M ) and G(N ),
respectively. Then the new map φ : M × N −→ Image(f ) × Image(g) defined by
φ(x, y) = (f (x), g(y)) is a proper countable colouring for G(M × N ). This completes
the proof of Claim 1.
Claim 2.
Every torsion-free abelian group is a pcc-group.
Proof of Claim 2: Let A be a torsion-free abelian group. We show that the connected
components of G(A) are at most countable. First, we show that the degree of any vertex
is at most countable. Now, the identity element 0 is an isolated vertex. Suppose that
x 6= 0. There are only countably many multiples of x. Also, for each natural number n,
there is at most one element y such that ny = x. For, if ny = nz = x, then n(y −z) = 0,
so y − z = 0 since A is torsion-free. So there are at most countably many elements of
which x is a multiple. Hence, the neighbourhood of x is countable. This implies that
the set of vertices of distance 2 from x is countable as well. By induction, we conclude
that the set of vertices of distance k from x is countable. So, the component of G(A)
which contains x is countable. Now, we can colour each connected component with
countably many colours. The proof of Claim 2 is complete.
Other classes that could be looked at would include solvable groups. One could also
ask whether the class of pcc-groups is extension-closed.
2.2.1
Perfectness of the power graph
A graph G is called perfect if for every finite induced subgraph H of G, χ(H) = ω(H).
The Strong Perfect Graph Theorem states that a finite graph G is perfect if and only
if neither G nor G (the complement of G) contains an induced odd cycle of length at
least 5, see [4, Theorem 14.18]. However, this is a deep theorem, and we do not need
it to prove our results.
Utilizing Lemma 7 to colour the power graph with a finite set of colours we require
the group to be bounded exponent. Here we show that for such groups the resulting
8
power graph is always perfect and can be finitely coloured. To prove this result we
facilitate the concepts of comparability graph.
Let ≤ be a binary relation on the elements of a set P . If ≤ is reflexive and transitive,
then (P, ≤) is called a pre-ordered set. All partially ordered sets are pre-ordered. The
comparability graph of a pre-ordered set (P, ≤) is the simple graph Υ(P ) with the vertex
set P and two distinct vertices x and y are adjacent if and only if either x ≤ y or y ≤ x
(or both).
Theorem 14. Let m be a positive integer and P be a pre-ordered set (not necessarily
finite) whose maximum chain size is m. Then the comparability graph Υ(P ) is perfect
and
ω(Υ(P )) = χ(Υ(P )) = m.
Proof. The result is well known for comparability graphs of partial orders; our proof
is a slight extension of this. Since the class of comparability graphs is closed under
taking induced subgraphs, it is enough to prove that the comparability graph of P has
equal clique number and chromatic number. Clearly, a clique in a Υ(P ) is a chain in
P , while a colouring is a partition into antichains.
First we show that ω(Υ(P )) = m. Let C be a clique in Υ(P ). Then C is a chain in
P , and so |C| ≤ m. Thus ω(Υ(P )) = m.
Now, we show that χ(Υ(P )) ≤ m. We form a directed graph by putting an arc
from x to y whenever x ≤ y but y 6≤ x; and, if C is an equivalence class of the relation
≡ defined by x ≡ y if x ≤ y and y ≤ x, then take an arbitrary directed path on the
elements of C. Clearly, the longest directed path contains m vertices. Let Pi be the
set of elements x for which the longest directed path ending at x contains i vertices. It
is easy to see that Pi is an independent set; these sets partition P into m classes.
Now, we show that the power graph of a group is the comparability graph of a preordered set. First, we define some notations. Let n be a positive integer and D(n) be
the set of all divisors of n in N. Define a relation on D(n) by r s if and only if r | s.
Clearly, (D(n), ) is a partially ordered set. Denote the set of all chains of (D(n), )
by C(n). Using this convention we are able to determine the clique/chromatic number
of the power graph of a group of bounded exponent (see Lemma 7).
Theorem 15. Let G be a group of exponent n. Then G(G) is a perfect graph and
)
(
X
φ(d) : C ∈ C(n) ≤ n,
χ(G(G)) = ω(G(G)) = max
d∈C
Gd 6=∅
where Gd is the set of elements of G of order d, for some divisor d of n.
9
Proof. First we consider two following facts:
Fact 1. Suppose that d is a divisor of n and Gd 6= ∅. If g, h ∈ Gd and g and h are
adjacent in the power graph, then g and h generate the same cyclic group. So hGd i is a
disjoint union of cliques of size φ(d). Therefore; if x is an element of a clique H of hGd i
adjacent to an element y of a clique K of hGd′ i, then every element of H is adjacent to
every element of K and moreover, d | d′ or d′ | d.
d
Fact 2. Note that if z is an element of order d, then for each divisor d′ of d, z d′
is of order d′ . So for each clique T of hGd i, every element of T is adjacent to every
element of a clique S of hGd′ i.
Since {Gd : G has an element of order d} forms a partition for G, Fact 1 implies
that every maximal clique of G(G) has the form Cl1 ∪ · · · ∪ Clm , where Cli is a clique of
hGdi i of size φ(di ) and {d1 , . . . , dm } is a chain of length m belonging to C(n). Moreover;
by Fact 2, we deduce that for every chain {d1 , . . . , dm } in C(n), there exists a clique for
G(G) of this form. Now, by |Cli | = φ(di ), we conclude that
)
(
X
X
φ(d) : C ∈ C(n) ≤
φ(d) = n.
ω(G(G)) = max
d∈C
Gd 6=∅
d|n
Define a pre-ordering ≤ on G by x ≤ y if and only if x is a power of y. Clearly, the
power graph of G is the comparability graph of ≤ and so by Theorem 14, the power
graph of G is perfect. Thus χ(G(G)) = ω(G(G)) and the proof is complete.
The next two corollaries are direct consequences of Lemma 7 and Theorem 15.
Corollary 16. For every group G, the following statements are equivalent:
(i) χ(G(G)) < ∞;
(ii) ω(G(G)) < ∞;
(iii) G is bounded exponent.
Moreover, the chromatic number of G(G) does not exceed the exponent of G.
Corollary 17. Let G be an abelian group of exponent n. Then
(
)
X
χ(G(G)) = ω(G(G)) = max
φ(d) : C ∈ C(n) .
d∈C
10
Corollary 18. Let H be a subgroup of G and [G : H] < ∞. Then ω(G(H)) < ∞ if
and only if ω(G(G)) < ∞.
The following example shows that a similar assertion does not hold for the independence number.
Example 19. Let G = C2 × C2∞ and H = {0} × C2∞ . Thus [G : H] = 2. Since
G(H) is a complete graph, α(H) = 1. Clearly, the set {1} × C2∞ is independent and
so α(G) = ∞.
2.3
Miscellaneous properties
We conclude this section with three miscellaneous properties of the power graph of a
group.
Theorem 20. If G(G) is a triangle-free graph, then G is isomorphic to a direct product
of C2 and G(G) is a star.
Proof. First we show that the order of every element of G is at most 2. Let a ∈ G.
If o(a) ≥ 3, then {e, a, a2 } is a triangle, a contradiction. So G is an elementary abelian
2-group. Therefore, by Prüfer-Baer Theorem, G is isomorphic to a direct product of
C2 and so G(G) is a star with the center e.
The following theorem characterizes those groups whose power graphs are connected.
Theorem 21. Let G be a group. The following statements are equivalent.
(i) G(G) is connected;
(ii) G is periodic;
(iii) γ(G(G)) = 1;
(iv) diam(G(G)) ≤ 2.
Proof. (i)=⇒(ii) Let x (x 6= e) be a vertex of G(G). We show that x is of finite order
in G. Since G(G) is connected, there is a path from x to e. Let y be the adjacent vertex
to e in this path. So the order of y is finite. Now, suppose that t is the adjacent vertex
to y in this path. Then the order of t is finite, too. By repeating this procedure, we
deduce that the order of x is finite. So G is periodic.
(ii)=⇒(iii) Since every element in G has a finite order, {e} is a dominating set.
The parts (iii)=⇒(iv) and (iv)=⇒(i) are clear.
11
Theorem 22. If deg(g) < ∞, for every g ∈ G, then G is a finite group.
Proof. Let g ∈ G. Since deg(g) < ∞, g has a finite order in G. Thus G is a periodic
group and so e is adjacent to every other vertices of G(G). Since deg(e) < ∞, we
deduce that G is finite.
Q
Remark 23. Let G ∼
= i≥1 C2 . Then G(G) is an infinite star with the center 0. This
shows that in the previous theorem the condition deg(g) < ∞, for every g ∈ G is
necessary.
3
Power graph and commuting graph
Let G be a group. If the vertices x and y are joined in the power graph of G, then they
are joined in the commuting graph; so the power graph is a spanning subgraph of the
commuting graph.
Question 24. For which groups is it the case that the power graph is equal to the
commuting graph?
The identity is joined to all others in the commuting graph; so if the two graphs
are equal, then G is a periodic group.
Theorem 25. Let G be a finite group with power graph equal to commuting graph.
Then one of the following holds:
• G is a cyclic p-group;
• G is a semidirect product of Cpa by Cqb , where p and q are primes with a, b > 0,
q b | p − 1 and Cqb acts faithfully on Cpa ;
• G is a generalized quaternion group.
Proof. Let G have power graph equal to commuting graph; that is, if two elements
commute, then one is a power of the other. Then G contains no subgroup isomorphic
to Cp × Cq , where p and q are primes, since this group fails the condition.
A theorem of Burnside [15, Theorem 12.5.2] says that a p-group containing no
Cp × Cp subgroup is cyclic or generalized quaternion. So all Sylow subgroups of G are
of one of these two types.
Suppose that all Sylow subgroups are cyclic. Then G is metacyclic [15, Theorem
9.4.3]. The cyclic normal subgroup of G has order divisible by one prime only, say p.
12
Its centraliser in G is a Sylow p-subgroup P of G, since it contains no elements of order
coprime to p. Hence G is a semidirect product of P and a cyclic group Q of order
coprime to p, necessarily a cyclic q-group for some prime q. If |Q| = q b , then we have
q b | p − 1.
So we may suppose that G has generalized quaternion Sylow 2-subgroups. By
Glauberman’s Z*-Theorem [11], G/O(G) has a central involution, where O(G) is the
maximal normal subgroup of G of odd order. This involution must act fixed-pointfreely on O(G), so O(G) is abelian, and hence cyclic of prime power order. But a
generalized quaternion group cannot act faithfully on such a group. So O(G) = 1. Then
the involution in G is central, so G is a 2-group, necessarily a generalized quaternion
group.
Remark 26. In the infinite case, there are other examples, such as Tarski monsters,
which are infinite groups whose non-trivial proper subgroups are all cyclic of prime
order p. Probably no classification is possible.
In the next theorem, we extend Theorem 25 to solvable groups.
Theorem 27. Let G be a solvable group with power graph equal to commuting graph.
Then one of the following holds:
• G is a cyclic p-group;
• G is a semidirect product of Cpa by Cqb , where p and q are primes with a, b > 0,
q b | p − 1 and Cqb acts faithfully on Cpa ;
• G is a generalized quaternion group;
• G is the p-quasicyclic group Cp∞ ;
• G is a semidirect product of p-quasicyclic group Cp∞ and a finite cyclic group.
Proof. We use this fact that every finitely generated periodic solvable group is finite.
We know that G is periodic. We show that there are no three elements whose order
are distinct primes. Suppose that o(a) = p, o(b) = q and o(c) = r, where p, q and r
are distinct primes. Let H be the subgroup generated by a, b and c. Then H is finite.
Clearly, the power graph and the commuting graph of H are equal. This contradicts
Theorem 25. Thus the order of every finite subgroup of G is pα q β , for some nonnegative integers α and β. By the second part of Theorem 25, we may assume that
β is bounded, because q β |p − 1. Also, by Theorem 25, the order of every element of
G is a p-power or a q-power, because every cyclic subgroup is a p-group or a q-group.
If α, β > 0, then by the second part of Theorem 25, ha, bi is semidirect product of hai
and hbi. So, hai and hbi are both cyclic (even in the case q = 2). Let N be the set
13
of all elements of G whose orders are p-power. We show that N is an abelian normal
subgroup of G. To see this first we show that if x and y are two elements of N , then
xy = yx. Let S be the subgroup generated by x and y. Then S is a finite group of
order pα q β . If β = 0, then by Theorem 25, S is a cyclic p-group and we are done. So
assume that β > 0. Let N1 and Q be Sylow p-subgroup and Sylow q-subgroup of S,
respectively. Then by Theorem 25 both are cyclic. Now, by [18, Theorem 6.2.11], Q
has a normal complement. So, N1 ⊳ S. This implies that x, y ∈ N1 and so xy = yx.
Thus we conclude that N is an abelian p-subgroup of G. Now, by the definition of N ,
N ⊳ G.
Now, let Q be a q-subgroup of G which has maximum size. We prove that G = N Q.
Let a ∈ G be an element of G whose order is q-power. Let M be the subgroup
generated by a and Q. Then M = P1 Q1 , where P1 and Q1 are Sylow p-subgroup and
Sylow q-subgroup of M , respectively and P1 ⊳ M . But Q1 is a conjugate of Q and so
M = P1 Q. Since a ∈ M , we have a = bc, where b ∈ P1 and c ∈ Q. But P1 ⊆ P and
this implies that a ∈ P Q, as desired. So G = N Q.
Since N is an abelian group, the commuting graph of N and so the power graph of
N is a complete graph. Now, Theorem 4 yields that N is finite or N = Zp∞ .
4
The enhanced power graph
4.1
Definition and properties
In the Section 2 we investigated some properties of power graphs of groups. In Theorem 25, we characterized finite groups for which the power graph is the same as the
commuting graph. Now, it is natural to ask if they are not equal, how close these
graphs are. To tackle this problem we introduce an intermediate graph. This graph
can be regarded as a measurement for this difference. Given a group G, the enhanced
power graph of G denoted by Ge (G) is the graph with vertex set G, in which x and y
are joined if and only if there exists an element z such that both x and y are powers of
z.
The power graph and commuting graph behave well on restriction to a subgroup
(that is, if H ≤ G, then the power graph of H is the induced subgraph of the power
graph of G on the set H, and similarly for the commuting graph). Because of the
existential quantifier in the definition, it is not obvious that the same holds for the
enhanced graph. That this is so is a consequence of the fact that x and y are joined in
the enhanced power graph if and only if hx, yi is cyclic. Note that
• the power graph is a spanning subgraph of the enhanced power graph;
• the enhanced power graph is a spanning subgraph of the commuting graph.
14
In the next remark, we use the concept of graph squares. For a graph H, the square
of H is a graph with the same vertex set as H in which two vertices are adjacent if
their distance in H is at most two.
Remark 28. If we assume that the (undirected or directed) power graph has a loop
at each vertex, then the enhanced power graph lies between the power graph and its
square. We already saw that it contains the power graph (as a spanning subgraph).
Now, let x and y be two vertices joined by a path (x, z, y) of length 2 in the power
~
graph. There are four cases in the directed power graph D = G(G):
• (x, z), (z, y) ∈ E(D). Then x is a power of z, and z a power of y; so x is a power
of y, and (x, y) ∈ E(D).
• (z, x), (y, z) ∈ E(D). Dual to the first case.
• (x, z), (y, z) ∈ E(D). Then x and y are powers of z, so they are joined in the
square of the power graph.
• (z, x), (z, y) ∈ E(D). In this case there is nothing we can say.
Also the following holds:
Theorem 29. Let G and H be finite groups. If the power graphs of G and H are
isomorphic, then their enhanced power graphs are also isomorphic.
Proof. Note that x and y are joined in the enhanced power graph if and only if there
is a vertex z which dominates both in the directed power graph. So the theorem follows
from the main theorem of [6].
4.2
Comparing to the power graph and commuting graph
Question 30. For which (finite) groups is the power graph equal to the enhanced
power graph?
This question connects with another graph associated with a finite group, the prime
graph, defined by Gruenberg and Kegel [13]: the vertices of the prime graph of G are the
prime divisors of |G|, and vertices p and q are joined if and only G contains an element
of order pq. To state the next result we need a definition. The group G is a 2-Frobenius
group if it has normal subgroups F1 and F2 such that F1 < F2 , F2 is a Frobenius group
with Frobenius kernel F1 , and G/F1 is a Frobenius group with Frobenius kernel F2 /F1 .
In the statement of the following theorem, p and q denote distinct primes.
Theorem 31. For a finite group G, the following conditions are equivalent:
15
(a) the power graph of G is equal to the enhanced power graph;
(b) every cyclic subgroup of G has prime power order;
(c) the prime graph of G is a null graph.
A group G with these properties is one of the following: a p-group; a Frobenius group
whose kernel is a p-group and complement a q-group; a 2-Frobenius group where F1 and
G/F2 are p-groups and F2 /F1 is a q-group; or G has a normal 2-subgroup with quotient
group H, where S ≤ H ≤ Aut(S) and S ∼
= A5 or A6 .
All these types of group exist. Examples include S3 and A4 (Frobenius groups); S4
(a 2-Frobenius group); A5 , A6 and 24 : A5 .
Proof. Let p and q be distinct primes. The cyclic group of order pq does not have
property (a); so a group satisfying (a) must also satisfy (b). Conversely, suppose that
(b) holds. If x and y are adjacent in the enhanced power graph, then hx, yi is cyclic,
necessarily of prime power order; so it must be generated by one of x and y, and so x
and y are adjacent in the power graph.
Clearly, (b) and (c) are equivalent.
Now, let G be a group satisfying these conditions. Either G is a p-group for some
prime p, or the prime graph of G is disconnected. Now, we use the result of Gruenberg
and Kegel [13] (stated and proved in Williams [20]), asserting that a finite group with
disconnected prime graph is Frobenius or 2-Frobenius, simple, π1 by simple, simple by
π1 -solvable, or π1 by simple by π1 . Here π1 is the set of primes in the connected component of the prime graph containing 2, assuming that |G| is even; and a 2-Frobenius
group is a group G with normal subgroups F1 < F2 such that F2 is a Frobenius group
with kernel F1 , and G/F1 is a Frobenius group with kernel F2 /F1 .
It follows from the work of Frobenius that a Frobenius complement either has all
Sylow subgroups cyclic (and so is metacyclic) or has SL(2, 3) or SL(2, 5) as a normal
subgroup. These last two cases cannot occur, since the central involution commutes
with elements of order 3. In the first case, the results of Gruenberg and Kegel (see
the first corollary in Williams [20]) show that the Frobenius complement has only one
prime divisor.
In the case of a 2-Frobenius group, an element of the Frobenius complement in the
top group centralises some element of F1 ; so F1 and G/F2 must be p-groups for the
same prime p.
In the remaining cases, it can be read off from the tables in Williams [20] that the
simple group can only be A5 or A6 , and the conclusions of the theorem follow since
π1 = {2}.
16
Question 32. For which (finite) groups is the enhanced power graph equal to the
commuting graph?
Again, we have a lot of information about such a group.
Theorem 33. For a finite group G, the following conditions are equivalent:
(a) the enhanced power graph of G is equal to its commuting graph;
(b) G has no subgroup Cp × Cp for p prime;
(c) the Sylow subgroups of G are cyclic or (for p = 2) generalized quaternion.
A group satisfying these conditions is either a cyclic p-group for some prime p, or
satisfies the following: if O(G) denotes the largest normal subgroup of G of odd order,
then O(G) is metacyclic, H = G/O(G) is a group with a unique involution z, and
H/hzi is a cyclic or dihedral 2-group, a subgroup of PΓL(2, q) containing PSL(2, q) for
q an odd prime power, or A7 .
An example of a group for the second case is the direct product of the Frobenius
group of order 253 by SL(2, 5).
Proof. The group Cp × Cp has commuting graph not equal to its enhanced power
graph, so cannot be a subgroup of a group satisfying (a); thus (a) implies (b). Conversely, suppose that (b) holds. Let x and y be elements of G which are adjacent in
the commuting graph. Then hx, yi is abelian, and hence is the direct product of two
cyclic groups, say Cr × Cs . Under hypothesis (b), we must have gcd(r, s) = 1, and so
hx, yi ∼
= Crs ; thus x and y are joined in the enhanced power graph.
Conditions (b) and (c) are equivalent by a theorem of Burnside [15, Theorem 12.5.2].
Suppose that a Sylow 2-subgroup P of G is cyclic or generalized quaternion. If
P is cyclic, then by Burnside’s transfer theorem [15, Section 14.3], G has a normal
2-complement: that is, if O(G) is the largest normal subgroup of G of odd order, then
G/O(G) ∼
= P . If P is generalized quaternion, then by Glauberman’s Z ∗ -Theorem,
H = G/O(G) has a unique central involution z. Put Z = hzi. Then the Sylow
2-subgroups Q of H/Z are dihedral; the Gorenstein–Walter theorem [12] shows that
H/Z is isomorphic to a subgroup of PΓL(2, q) containing PSL(2, q) (for odd q), or to
the alternating group A7 , or to Q. For any such group H ∗ = H/Z, an argument of
Glauberman (which can be found in [3]) shows that there is a unique double cover H
with a single involution.
Question 34. What can be said about the difference of the enhanced power graph and
the power graph, or the difference of the commuting graph and the enhanced power
graph? In particular, for which groups is either of these graphs connected?
17
4.3
Maximal cliques in the enhanced power graph
We will now look at maximal cliques in the enhanced power graph. This requires a
lemma which looks trivial, but we couldn’t find an easier proof of it than the one below.
Lemma 35. Let x, y, z be elements of a group G, and suppose that hx, yi, hx, zi and
hy, zi are cyclic. Then hx, y, zi is cyclic.
Proof. The result clearly holds if one of x, y, z is the identity; so suppose not. Now,
a cyclic group cannot contain elements of both finite and infinite order, so either all
three elements have finite order, or all three have infinite order.
Case 1:
x, y, z have finite order. Then they generate a finite abelian group A.
We first note that it suffices to do the case where the orders of x, y, z are powers
of a prime p. For A is the direct sum of p-groups for various primes p; each p-group is
generated by certain powers of x, y, z; and if each p-group is cyclic, then so is A. With
this assumption, suppose that A is not cyclic. Since hx, yi is cyclic, A is the sum of
two cyclic groups, and so contains a subgroup Q ∼
= Cp × Cp . Each of hxi, hyi and hzi
intersects Q in a subgroup of order p; let these subgroups be X, Y, Z. Since hx, yi is
cyclic, it meets Q in a subgroup of order p; so X = Y . Similarly X = Z. So hx, y, zi
meets Q in a subgroup of order p, contradicting the assumption that Q ≤ hx, y, zi.
Case 2:
x, y, z have infinite order.
Then they generate a free abelian group; since hx, yi is cyclic, we see that A =
hx, y, zi has rank at most 2. Consider the Q-vector space A ⊗Z Q, which has dimension
at most 2. Since hx, yi is cyclic, the 1-dimensional subspaces hxi ⊗Z Q and hyi ⊗Z Q
have non-empty intersection, and so are equal. Similarly for hzi ⊗Z Q. Thus A ⊗Z Q is
1-dimensional, so A is cyclic.
Now, we have the following characterization of the maximal cliques in the enhanced
power graph.
Lemma 36. A maximal clique in the enhanced power graph is either a cyclic subgroup
or a locally cyclic subgroup.
Proof. Clearly, a cyclic or locally cyclic subgroup is a clique. Suppose that C is a
maximal clique. If x, y ∈ C, then by Lemma 35, every element of hx, yi is joined to
every element z ∈ C; so C ∪ hx, yi is a clique. By maximality of C, we have hx, yi ⊆ C;
so C is a subgroup. Now, a simple induction shows that any finite subset of C generates
a cyclic group, so that C is locally cyclic.
18
Remark 37. Locally cyclic groups include the additive group of Q (or the subgroup
consisting of rationals whose denominators only involve primes from a prescribed set),
and direct sums of copies of the p-quasicyclic groups (the Prüfer groups) Cp∞ for distinct
primes p.
Now, we have two immediate corollaries:
Corollary 38. Let G be a group. Then ω(Ge (G)) < ∞ if and only if G is a group of
finite exponent. If these conditions hold, then
ω(Ge (G)) = max{o(g) : g ∈ G}.
Remark 39. Note that this may be smaller than the exponent of G.
Proof. Clearly, if G is not a bounded exponent group, then G(G) as a subgraph of
Ge (G) has infinite clique number by Lemma 7. Now, let G be a periodic group. Then
the subsets Gn = {g ∈ G : o(g) = n}, for n ∈ N, partition G into at most countably
many subsets. On each of these subsets the power graph and the enhanced power graph
coincide. In particular, if G is bounded exponent, then there are only finitely many
classes. It is clear that, if x and y have the same order and generate a cyclic group,
then each is a power of the other.
Corollary 40. A clique in the enhanced power graph of a group is at most countable.
Proof. For a locally cyclic group is isomorphic to a subgroup of Q or Q/Z and hence
countable, see [17, Exercise 5, p.105].
Open problems
This paper concerns several graph theoretical parameters of the power graph of a group.
In Section 2.1 we studied groups whose power graph has a finite independence number.
In Theorem 6, we proved that if G is a nilpotent group and α(G) < ∞, then either G
is a finite group or G ∼
= Cp∞ × H, for some prime number p, where H is a finite group
and p ∤ |H|. This result motivates us to pose the following question.
Question 41. Let G be an infinite group. Is it true that α(G(G)) < ∞ if and only if
G∼
= Cp∞ × H, where H is a finite group and p ∤ |H|.
In Section 2.2, we showed that the chromatic number of the power graph of G
is finite if and only if the clique number of the power graph of G is finite and this
statement is also equivalent to the finiteness of the exponent of G. We proved that the
clique number of the power graph of G is at most countable. We also introduced the
concept of pcc groups and we posed the following question.
19
Question 42. Is it true that the chromatic number of the power graph of any group
is at most countably infinite?
It might be interesting to ask how much of Lemma 7 can be proved without the
Axiom of Choice. Is there any way of showing that the chromatic number of a group of
finite exponent is finite? A good test case for this question would be an abelian group
of exponent 3. Colouring the non-identity elements with two colours requires choosing
one of each pair {x, x−1 }, which requires AC (as Bertrand Russell famously pointed
out).
In the study of the commuting graph, it is normal to delete vertices which lie in
the centre of the group, since they would be joined to all other vertices. Similarly, in
the study of the generating graph of a 2-generator group, the identity is an isolated
vertex and is usually excluded. This convention is not used for the power graph. So
any problem we raise will have two different forms, depending on which convention we
use. For the power graph, the question of whether to include or exclude the identity
is more interesting. Some of the results will be completely different in the two cases
especially those dealing with connectedness. For example, if the identity is excluded,
Theorem 21 fails, since indeed the power graph of the infinite cyclic group is connected
when the identity is discarded. The next question seems interesting.
Question 43. Which groups do have the property that the power graph is connected
when the identity is removed?
Or more generally:
Question 44. Which groups do have the property that the power graph is connected
when the set of vertices which dominate the graph is removed?
The following question is the second version for Question 24.
Question 45. For which groups, are the induced subgraphs of the power graph and
the commuting graph on G \ {e} are equal. Note that free groups have this property.
Question 46. Consider the difference of the power graph and commuting graph, the
graph in which x and y are joined if they commute but neither is a power of the other.
What can be said about this difference graph? In particular, for which groups is it
connected? Again this question can be asked with or without the identity. Note that
in a periodic group the identity is isolated in the difference graph, but this is not true
for arbitrary infinite groups.
20
References
[1] A. Abdollahi, S. Akbari, H.R. Maimani, Non-commuting graph of a group, J. Algebra 298
(2) (2006), 468–492.
[2] J. Araújo, W. Bentz, J. Konieczny, The commuting graph of the symmetric inverse semigroup, Israel Journal of Mathematics 207 (1) (2015), 103-149.
[3] L. Babai, P.J. Cameron, Automorphisms and enumeration of switching classes of tournaments, Electronic J. Combinatorics 7 (1) (2000), article #R38.
[4] J.A. Bondy, U.S.R. Murty, Graph Theory, Springer-Verlag, 2008.
[5] R. Brauer, K.A. Fowler, On groups of even order, The Annals of Mathematics 62 (3) (1955),
567–583.
[6] P.J. Cameron, The power graph of a finite group II, J. Group Theory 13 (2010), 779–783.
[7] P.J. Cameron, S. Ghosh, The power graph of a finite group, Discrete Math. 311 (13) (2011),
1220–1222.
[8] I. Chakrabarty, S. Ghosh, M.K. Sen, Undirected power graphs of semigroups, Semigroup
Forum 78 (2009), 410–426.
[9] N.G. De Bruijn, P. Erdös, A colour problem for infinite graphs and a problem in the theory
of relations, Indagationes Math. 13 (1951), 369–373.
[10] M. Feng, X. Ma, K. Wang, The structure and metric dimension of the power graph of a
finite group, European Journal of Combinatorics 43 (2015), 82–97.
[11] G. Glauberman, Central elements in core-free groups, J. Algebra 4 (1966), 403–420.
[12] D. Gorenstein, J.H. Walter, The characterization of groups with dihedral Sylow 2subgroups, J. Algebra 2 (1965), 85–151, 218–270, 354–393.
[13] K.W. Gruenberg, O.H. Kegel, unpublished manuscript, 1975.
[14] M. Giudici, A. Pope, On bounding the diameter of the commuting graph of a group,
Journal of Group Theory 17 (1) (2014), 131–149.
[15] M. Hall, Jr., The Theory of Groups, Macmillan, New York, 1959.
[16] B.H. Neumann, A problem of Paul Erdös on groups, J. Austral. Math. Soc. 21 (4) (1976),
467–472.
[17] D.J.S. Robinson, A Course in the Theory of Groups, Second edition, Springer-Verlag, New
York, 1995.
[18] W.R. Scott, Group Theory, Dover Publ., New York, 1987.
[19] M. Takashi, On partitions of free products of groups, Osaka Math. J. 1 (1) (1949), 49–51.
[20] J.S. Williams, Prime graph components of finite groups, J. Algebra 69 (1981),
487–513.
21
[21] T.J. Woodcock, Commuting Graphs of Finite Groups. PhD thesis, University of
Virginia, 2010.
22
| 4math.GR
|
arXiv:1802.06698v1 [cs.AI] 19 Feb 2018
Analysis of Cause-Effect Inference via
Regression Errors
Patrick Blöbaum1 , Dominik Janzing2 , Takashi Washio1 , Shohei
Shimizu1,3 , and Bernhard Schölkopf2
1 Osaka
2 MPI
University, Japan
for Intelligent Systems, Tübingen, Germany
3 Shiga University, Japan
Abstract
We address the problem of inferring the causal relation between two variables by comparing the least-squares errors of the predictions in both possible causal directions. Under the assumption of an independence between
the function relating cause and effect, the conditional noise distribution, and
the distribution of the cause, we show that the errors are smaller in causal
direction if both variables are equally scaled and the causal relation is close
to deterministic. Based on this, we provide an easily applicable algorithm
that only requires a regression in both possible causal directions and a comparison of the errors. The performance of the algorithm is compared with
different related causal inference methods in various artificial and real-world
data sets.
1
Introduction
Causal inference [2, 3] is becoming an increasingly popular topic in machine
learning, since the results are often not only from interest in predicting the result of
This is an extended version of the AISTATS 2018 paper [1]. Compared to [1], this work
mainly differs by providing detailed mathematical proofs and further extensive experiments.
1
potential interventions, but also in general statistical and machine learning applications [4]. In particular, the identification of the causal relation between only two
observed variables is a challenging task [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16].
As regards the present work, we address this bivariate setting, where one variable
is the cause and the other variable is the effect. That is, given observed data X,Y
that are drawn from a joint distribution pX,Y , we are interested in inferring whether
X caused Y or Y caused X. For instance, is a certain change of the phyiscal state
of a patient a symptom or a cause of a certain disease? If we are not able to observe the effect of an intervention on one of the variables, the identification of
the correct causal relation generally relies on the exploitation of statistical asymmetries between cause and effect [2, 3, 10, 14, 7]. Conventional approaches to
causal inference rely on conditional independences and therefore require at least
three observed variables. Given the observed pattern of conditional dependences
and independences, one infers a class of directed acyclic graphs (DAGs) that are
compatible with the respective pattern (subject to Markov condition and faithfulness assumption [2, 3]). Whenever there are causal arrows that are common to
all DAGs in the class, conditional (in)dependences yield definite statements about
causal directions.
In case of bivariate data, however, we rely on those types of asymmetries between cause and effect that are already apparent in the bivariate distribution alone.
One kind of asymmetry is given by restricting the structural equations relating
cause and effect to a certain function class: For linear relations with non-Gaussian
independent noise, the linear non-Gaussian acyclic model (LiNGAM) [7] guarantees to identify the correct causal direction. For nonlinear relations, the additive
noise model (ANM) [10] and its generalization to post-nonlinear models (PNL)
[9] identify the causal direction by assuming an independence between cause and
noise, where, apart from some exceptions such as bivariate Gaussian, a model can
only be fit in the correct causal direction such that the input is independent of the
residual.
Further recent approaches for the bivariate scenario are based on an informal
independence assumption stating that the distribution of the cause, denoted by pC
contains no information about the conditional distribution of the effect, given the
cause, denoted by pE|C . Here, the formalization of ‘no information’ is a challenging task. For the purpose of foundational insights (rather than for practical
purposes), [17] and [18] formalize the idea via algorithmic information and postulate that knowing pC does not enable a shorter description of pE|C and vice
versa. The information-geometric approach for causal inference (IGCI) [14] is
inspired by the independence assumption and is able to infer the causal direction
2
in deterministic nonlinear relationships subject to a certain independence condition between the slope of the function and the distribution of the cause. A related
but different independence assumption is also used by a technique called unsupervised inverse regression (CURE) [15], where the idea is to estimate a prediction
model of both possible causal directions in an unsupervised manner, i.e. only the
input data is used for the training of the prediction models. According to the above
independence assumption, the effect data may contain information about the relation between cause and effect that can be employed for predicting the cause from
the effect, but the cause data alone does not contain any information that helps the
prediction of the effect from the cause (as hypothesized in [19]). Accordingly, the
unsupervised regression model in the true causal direction should be less accurate
than the prediction model in the wrong causal direction.
For our approach, we address the causal inference problem by exploiting an
asymmetry in the mean-squared error (MSE) of predicting the cause from the
effect and the effect from the cause, respectively, and show, that under appropriate
assumptions and in the regime of almost deterministic relations, the prediction
error is smaller in causal direction. A preliminary version of this idea can be
found in [20, 21] but in these works the analysis is based on a simple heuristic
assuming that the regression of Y on X and the regression of X on Y yield functions
that are inverse to each other, which holds approximately in the limit of small
noise. Moreover, the analysis is also based on the assumption of an additive noise
model in causal direction and on having prior knowledge about the functional
relation between X and Y , which makes it impractical for generic causal inference
problems.
In this work, we aim to generalize and extend the two aforementioned works
in several ways: 1) We explicitly allow a dependency between cause and noise. 2)
We give a proper mathematical proof of the theory that justifies the method subject
to clear formal assumptions. 3) We perform a more extensive evaluation for the
application in causal inference and compare it with various related approaches.
The theorem stated in this work might also be of interest for general statistical
purposes. A briefer version of this work with less extensive experiments and
without detailed proofs can be found in [1].
This paper is structured as follows: In Section 2, we introduce the used notations and assumptions, which are necessary for the main theorem of this work
stated in Section 3. An algorithm that utilizes this theorem is proposed in Section
4 and evaluated in various artificial and real-world data sets in Section 5.
3
2
Preliminaries
In the following, we introduce the preliminary notations and assumptions.
2.1
Notation and problem setting
The goal of this paper is to correctly identify cause and effect variable of given
observations X and Y . Throughout this paper, a capital letter denotes a random
variable and a lowercase letter denotes values attained by the random variable.
Variables X and Y are assumed to be real-valued and to have a joint probability
density (with respect to the Lebesgue measure), denoted by pX,Y . By slightly
abusing terminology, we will not further distinguish between a distribution and its
density since the Lebesgue measure as a reference is implicitly understood. The
notations pX , pY , and pY |X are used for the corresponding marginal and conditional densities, respectively. The derivative of a function f is denoted by f 0 .
2.2
General idea
As mentioned before, the general idea of our approach is to simply compare the
MSE of regressing Y on X and the MSE of regressing X on Y . If we denote cause
and effect by C, E ∈ {X,Y }, respectively, our approach explicitly reads as follows.
Let φ denote the function that minimizes the expected least squares error when
predicting E from C, which implies that φ is given by the conditional expectation
φ (c) = E[E|c]. Likewise, let ψ be the minimizer of the least squares error for
predicting C from E, that is, ψ(e) = E[C|e]. Then we will postulate assumptions
that imply
E[(E − φ (C))2 ] ≤ E[(C − ψ(E))2 ],
(1)
in the regime of almost deterministic relations. This conclusion certainly relies on
some kind of scaling convention. For our theoretical results we will assume that
both X and Y attain values between 0 and 1. However, in some applications, we
will also scale X and Y to unit variance to deal with unbounded variables. (1) can
be rewritten in terms of conditional variance as
E[Var[E|C]] ≤ E[Var[C|E]].
4
2.3
Assumptions
First, recall that we assume throughout the paper that either X is the cause of Y
or vise versa in an unconfounded sense, i.e. there is no common cause. To study
the limit of an almost deterministic relation in a mathematically precise way, we
consider a family of effect variables Eα by
Eα := φ (C) + αN,
(2)
where α ∈ R+ is a parameter controlling the noise level and N is a noise variable
that has some (upper bounded) joint density pN,C with C. Note that N here does
not need to be statistically independent of C (in contrast to ANMs). Here, φ is a
function that is further specified below. Therefore, (2) does not, a priori, restrict
the set of possible causal relations, because for any pair (C, E) one can define the
noise N as the residual N := E − φ (C) and thus obtain E1 = E for any arbitrary
function φ .
For this work, we make use of the following assumptions:
1. Invertible function: φ is a strictly monotonically increasing two times differentiable function φ : [0, 1] → [0, 1]. For simplicity, we assume that φ is
monotonically increasing with φ (0) = 0 and φ (1) = 1 (similar results for
monotonically decreasing functions follow by reflection E → 1 − E). We
0
also assume that φ −1 is bounded.
2. Compact supports: The distribution of C has compact support. Without
loss of generality, we assume that 0 and 1 are the smallest and the largest
values, respectively, attained by C. We further assume that the distribution
of N has compact support and that there exist values n+ > 0 > n− such
that for each c, [n− , n+ ] is the smallest interval containing the support of
pN|c . This ensures that we know [αn− , 1 + αn+ ] is the smallest interval
containing the support of pEα . Then the shifted and rescaled variable
Ẽα :=
1
1 + αn+ − αn−
(Eα − αn− )
(3)
attains 0 and 1 as minimum and maximum values and thus is equally scaled
as C.
3. Unbiased noise: We use the convention E[N|c] = 0 for all values c of C
without loss of generality (this can easily be achieved by modifying φ ).
Then φ is just the conditional expectation, that is, φ (c) = E[E|c].
5
4. Unit noise variance: The expected conditional noise variance is E[Var[N|C]] = 1,
which is also not a proper restriction, because we can scale α accordingly
since we are only interested in the limit α → 0.
5. Independence postulate: While the above assumptions are just technical,
we now state the essential assumption that generates the asymmetry between cause and effect. To this end, we consider the unit interval [0, 1]
as probability space with uniform distribution as probability measure. The
functions c 7→ φ 0 (c) and c 7→ Var[N|c]pC (c) define random variables on this
space, which we postulate to be uncorrelated, formally stated as
Cov[φ 0 , Var[N|c]pC ] = 0.
(4)
More explicitly, (4) reads:
Z 1
0
0
φ (c) Var[N|c]pC (c)dc −
Z 1
0
Z 1
φ (c)dc
0
0
Var[N|c]pC (c)dc = 0.
(5)
The justification of (4) is not obvious at all. For the special case where the conditional variance Var[N|c] is a constant in c (e.g. for ANMs), (4) reduces to
Cov[φ 0 , pC ] = 0,
(6)
which is an independence condition for deterministic relations stated in [19]. Conditions of similar type as (6) have been discussed and justified in [14]. They are
based on the idea that φ contains no information about pC . This, in turn, relies on
the idea that the conditional pE|C contains no information about pC .
To discuss the justification of (5), observe first that it cannot be justified as
stating some kind of ‘independence’ between pC and pE|C . To see this, note
first that (5) states an uncorrelatedness of the two functions c 7→ φ 0 (c) and c 7→
Var[N|c]pC (c). Since the latter function contains the map c 7→ Var[N|c], which
is a porperty of the conditional pE|C and not of the marginal pC , it thus contains
components from both pC and pE|C . Nevertheless, to justify (5) we assume that
the function φ represents a law of nature that persists when pC and N change due
to changing background conditions. From this perspective, it becomes unlikely
that they are related to the background condition at hand. This idea follows the
general spirit of ‘modularity and autonomy’ in structural equation modelling, that
some structural equations may remain unchanged when other parts of a system
change (see Chapter 2 in [4] for a literature review).1
1 Note,
however, that the assignment E = φ (C) + N is not a structural equation in a strict sense,
because then C and N would need to be statistically independent.
6
A simple implication of (5) reads
Z 1
0
φ 0 (c) Var[N|c]pC (c)dc = 1,
(7)
due to 01 φ 0 (c)dc = 1 and 01 Var[N|c]pC (c)dc = Var[N|C] = 1.
In the following, the term independence postulate is referred to the aforementioned postulate and the term independence to a statistical independence, which
should generally become clear from the context.
R
R
3
Theory
As introduced in Section 2.2, we aim to exploit an inequality of the expected prediction errors in terms of E[Var[E|C]] ≤ E[Var[C|E]] to infer the causal direction.
Under the aforementioned assumptions, this can be stated if the noise variance
is sufficiently small. In order to conclude this inequality and, thus, to justify an
application to causal inference, we must restrict our analysis to the case where the
noise variance is sufficiently small, since a more general statement is not possible
under the aforementioned assumptions. The analysis can be formalized by the
ratio of the expectations of the conditional variances in the limit α → 0.
We will then show
E[Var[C|Ẽα ]]
≥ 1.
lim
α→0 E[Var[Ẽα |C]]
3.1
Error asymmetry theorem
For our main theorem, we first need an important lemma:
Lemma 1 (Limit of variance ratio) Let the assumptions 1-4 in Section 2.3 hold.
Then the following limit holds:
E[Var[C|Ẽα ]]
=
lim
α→0 E[Var[Ẽα |C]]
Z 1
1
0
φ 0 (c)2
Var[N|c]pC (c)dc
(8)
The formal proof is a bit technical and is provided in Appendix A, but the idea is
quite simple: just think of the scatter plot of an almost deterministic relation as a
thick line. Then Var[Eα |c] and Var[C|Eα = φ (c)] are roughly the squared widths
of the line at some point (c, φ (c)) measured in vertical and horizontal direction,
respectively. The quotient of the widths in vertical and horizontal direction is then
7
given by the slope. This intuition yields the following approximate identity for
small α:
Var[C|Ẽα = φ (c)] ≈
1
(φ 0 (c))2
Var[Ẽα |C = c] = α 2
1
(φ 0 (c))2
Var[N|c].
(9)
Taking the expectation of (9) over C and recalling that Assumption 4 implies
E[Var[Ẽα |C] = α 2 E[Var[N|C]] = α 2 already yields (8).
With the help of Lemma 1, we can now formulate the core theorem of this
paper, which is shown in Appendix B:
Theorem 1 (Error Asymmetry) Let the assumptions 1-5 in Section 2.3 hold.
Then the following limit always holds
E[Var[C|Ẽα ]]
≥ 1,
α→0 E[Var[Ẽα |C]]
lim
with equality only if the functional relation in Assumption 3 is linear.
Note that if φ is non-invertible, there is an information loss in anticausal direction, since multiple possible values can be assigned to the same input. Therefore,
we can expect that the error difference becomes even higher in these cases, which
is supported by the experiments in Section 5.2.2.
3.2
Remark
Theorem 1 states that the inequality holds for all values of α smaller than a certain
finite threshold. Whether this threshold is small or whether the asymmetry with
respect to regression errors already occurs for large noise cannot be concluded
from the above theoretical insights. Presumably, this depends on features of φ ,
pC , pN|C in a complicated way. However, the experiments in Section 5 suggest
that the asymmetry often appears even for realistic noise levels.
4
Algorithm
A causal inference algorithm that exploits Theorem 1 can be formulated in a
straightforward way. Given observations X,Y sampled from a joint distribution
pX,Y , the key idea is to fit regression models in both possible directions and compare the MSE. We call this approach Regression Error based Causal Inference
(RECI) and summarize the algorithm in Algorithm 1.
8
Algorithm 1 The proposed causal inference algorithm.
function RECI(X, Y )
. X and Y are the observed data.
(X,Y ) ← RescaleData(X,Y )
f ← FitModel(X,Y )
. Fit regression model f : X → Y
g ← FitModel(Y, X)
. Fit regression model g : Y → X
MSEY |X ← MeanSquaredError( f , X,Y )
MSEX|Y ← MeanSquaredError(g,Y, X)
if MSEY |X < MSEX|Y then
return X is the cause and Y is the effect.
else if MSEX|Y < MSEY |X then
return Y is the cause and X is the effect.
else
return No decision
end if
end function
Although estimating the conditional expectations E[Y |X] and E[X|Y ] by regression is a standard task in machine learning, we should emphasize that the
usual issues of over- and underfitting are critical for our purpose (like for methods
based on ANMs or PNLs) because they under- or overestimate the noise levels. It
may, however, happen that the method even benefits from underfitting: if there is a
simple regression model in causal direction that fits the data quite well, but in anticausal relation the conditional expectation becomes more complex, a regression
model with underfitting increases the error even more for the anticausal direction
than for the causal direction. This speculative remark is somehow supported by
our experiments, where we observed that simple models performed better than
complex models, even though they probably did not represent the true conditional
expectation.
Also, an accurate estimation of the MSE with respect to the regression model
and appropriate preprocessing of the data like removing isolated points in lowdensity regions, might improve the performance.
5
Experiments
For the experiments, we compared our algorithm with three different related causal
inference methods in various artificial and real-world data sets. In each evaluation,
9
observations of two variables were given and the goal was to correctly identify
cause and effect variables.
5.1
Causal inference methods for comparison
In the following, we briefly discuss and compare the causal inference methods
which we used for the evaluations.
LiNGAM The model assumptions of LiNGAM are
E = βC + N,
where β ∈ R, C ⊥
⊥ N and N is non-Gaussian. While in real-world data we may
often have non-Gaussian noise, the functional relationships are mostly nonlinear.
The computational cost is relatively low.
For the experiments, we used a state-of-the-art implementation of LiNGAM
that utilizes an entropy based method for calculating the likelihood ratios of the
possible causal directions [22]. For this, Eq. (3) in [22] is used in order to estimate
the likelihood ratio Eq. (2) in [22].
ANM The ANM approach assumes that
E = f (C) + N,
where f is nonlinear and C ⊥
⊥ N. This method requires fitting a regression function
and performing an additional evaluation of the relation between input and residual,
which lead to a high computational cost. Important to note is that the choice of
the evaluation method is crucial for the performance.
We used an implementation provided by [16], which uses a Gaussian process
regression for the prediction and provides different methods for the evaluation of
the causal direction. For the experiments, we chose different evaluation methods;
HSIC for statistical independence tests, an entropy estimator for estimating the
mutual information between input and residuals (denoted as ENT) and a Bayesian
model comparison that assumes Gaussianity (denoted as FN). Implementation
details and parameters can be found in Table 2 of [16].
10
IGCI The IGCI approach is able to determine the causal relationship in a deterministic setting
E = f (C),
under the ‘independence assumption’ Cov[log f 0 , pC ] = 0. The corresponding algorithm has been applied to noisy causal relations with partial success (and some
heuristic justifications [14]), but generalizations of IGCI for non-deterministic relations are actually not known and we consider Assumption 5 in Section 2.3 as
first step towards a possibly more general formulation. The computational cost
depends on the utilized method for estimating the information criterion, but is
generally low. Therefore, IGCI is the fastest of the methods.
For the experiments, we also used an implementation provided by [16], where
we always tested all possible combinations of reference measures and information estimators. These combinations are denoted as IGCIi j , where the subscripts
indicate:
• i = U: Uniform reference measure (normalizing X and Y )
• i = G: Gaussian reference measure (standardizing X and Y )
• j = 1: Entropy estimator using Eq. (12) in [13]
• j = 2: Integral approximation of Eq. (13) in [13]
• j = 3: Integral approximation of Eq. (22) in [16]
CURE CURE is based on the idea that the distribution pC does not help in better regressing E on C, while the distribution pE may help in better regressing C
on E. CURE implements this idea in a Bayesian way via a modified Gaussian
process regression. However, since CURE requires the generation of MarkovChain-Monte-Carlo (MCMC) samples, the biggest drawback is a very high computational cost.
An implementation of CURE by the authors has been provided for our experiments. Here, we used similar settings as described in Section 6.2 of [15], where
200 data samples were used and 10000 MCMC samples were generated. The
number of internal repetitions depends on the experimental setting.
11
RECI Our approach addresses non-deterministic nonlinear relations and, in particular, allows a dependency between cause and noise. Since we only require the
fitting of a least-squares solution in both possible causal directions, RECI can be
easily implemented, does not rely on any independence tests and, depending on
the regression model and implementation details, can be performed quickly.
In the experiments, we have always used the same class of regression function for the causal and anticausal direction to compare the errors, but performed
multiple experiments with different function classes. For each evaluation, we randomly split the data into training and test data, where we tried different ratios. The
utilized regression models were:
• a logistic function (LOG) of the form a + (b − a)/(1 + exp(c · (d − x)))
• shifted monomial functions (MON) of the form axn + b with n ∈ [2, 9]
• polynomial functions (POLY) of the form ∑ki=0 ai xi with k ∈ [1, 9]
• support vector regression (SVR) with a linear kernel
• neural networks (NN) with different numbers of hidden neurons: 2, 5, 10,
20, 2-4, 4-8, where ’-’ indicates two hidden layers
The logistic and monomial functions cover rather simple regression models, which
are probably not able to capture the true function f in most cases. On the other
hand, support vector regression and neural networks should be complex enough
to capture f . The polynomial functions are rather simple too, but more flexible
than the logistic and monomial functions.
We used the standard Matlab implementation of these methods and have always chosen the default parameters, where the parameters of LOG, MON and
POLY were fitted by minimizing the least-squared error.
During the experiments, we observed that the MSE varied a lot in many data
sets due to relatively small sample sizes and the random selection of training
and test data. Therefore, we averaged the MSE over all performed runs first before comparing them, seeing that this should give more accurate estimations of
E[Var[Y |X]] and E[Var[X|Y ]] with respect to the class of the regression function.
Since the performances of averaging the MSE over multiple runs were better than
only a single run and since we did not optimize the choice of regression functions
for each setting, we only summarize the results of the best performing MON,
POLY and NN setups of each setting in the following. A detailed overview of all
results, including the performance of estimating the MSE in single runs, can be
found in Appendix D.
12
General Remark Each evaluation was performed in the original data sets and in
preprocessed versions where isolated points (low-density points) were removed.
For the latter, we used the implementation and parameters from [15], where a
kernel density estimator with a Gaussian kernel is utilized. Note that CURE per
default uses this preprocessing step, also in the original data. In all evaluations,
we forced a decision by the algorithms, where in case of ANM the direction with
the highest score of the independence test was taken.
Except for CURE, we averaged the performances of each method over 100
runs, where we uniformly sampled 500 data points for ANM and SVR if the data
set contains more than 500 data points. For CURE, we only performed four internal repetitions in the artificial and eight internal repetitions in the real-world data
sets due to the high computational cost. Since we used all data points for IGCI
and LiNGAM, these methods have a consistent performance over all runs.
5.2
Artificial data
For experiments with artificial data, we performed evaluations with simulated
cause-effect pairs generated for a benchmark comparison in [16]. Further, we
generated additional pairs with linear, nonlinear invertible and nonlinear noninvertible functions where input and noise are strongly dependent.
5.2.1
Simulated benchmark cause-effect pairs
The work of [16] provides simulated cause-effect pairs with randomly generated
distributions and functional relationships under different conditions. As pointed
out by [16], the scatter plots of these simulated data look similar to those of realworld data. We took the same data sets as used in [16] and extended the reported
results with an evaluation with CURE, LiNGAM, RECI and further provide results in the preprocessed data.
The data sets are categorized into four different categories:
• SIM: Pairs without confounders. The results are shown in Figure 1(a)-1(b)
• SIM-c: A similar scenario as SIM, but with one additional confounder. The
results are shown in Figure 1(c)-1(d)
• SIM-ln: Pairs with low noise level without confounder. The results are
shown in Figure 2(a)-2(b)
13
• SIM-G: Pairs where the distributions of C and N are almost Gaussian without
confounder. The results are shown in Figure 2(c)-2(d)
The general form of the data generation process without confounder but with measurement noise2 is
C0 ∼ pC , N ∼ pN
NC ∼ N (0, σC ), NE ∼ N (0, σE )
C = C0 + NC
E = fE (C0 , N) + NE
and with confounder
C0 ∼ pC , N ∼ pN , Z ∼ pZ
C00 = fC (C0 , Z)
NC ∼ N (0, σC ), NE ∼ N (0, σE )
C = C00 + NC
E = fE (C00 , Z, N) + NE ,
where NC , NE represent independent observational Gaussian noise and the variances σC and σE are chosen randomly with respect to the setting.3 More details
can be found in Appendix C of [16].
Generally, ANM performs the best in all data sets. However, the difference between ANM and RECI, depending on the regression model, becomes smaller in
the preprocessed data where isolated points were removed. According to the observed performances, removing these points seems to often improve the accuracy
of RECI, but decrease the accuracy of ANM.
In all data sets, except for SIM-G, RECI always outperforms IGCI, CURE and
LiNGAM if a simple logistic or polynomial function is utilized for the regression.
However, in the SIM-G data set, our approach performs comparably poor, which
could be explained by the violation of the assumption of a compact support. In this
nearly Gaussian data, IGCI performs the best with a Gaussian reference measure.
2 Note,
however, that adding noise to the cause (as it is done here) can also be considered as a
kind of confounding. Actually, C0 is the cause of E in the below generating model, while the noisy
version C is not the cause of E. Accordingly, C0 is the hidden common cause of C and E. Here we
refer to the scenario as an unconfounded case with measurement noise as in [16].
3 Note that only N is Gaussian, while the regression residual is non-Gaussian due to the nonE
linerity of fE and non-Gaussianity of N, Z.
14
0.8
0.8
0.7
0.7
0.6
0.6
0.4
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
A
M
A
A
N
M
N
A
-H
S
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.5
N
Accuracy
1
0.9
-H
Accuracy
1
0.9
(b) Preprocessed SIM
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
M
N
A
A
M
N
A
-H
S
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.4
(c) Original SIM-c
N
0.5
A
Accuracy
1
0.9
-H
Accuracy
(a) Original SIM
(d) Preprocessed SIM-c
Figure 1: Evaluation result of all methods in the SIM and SIM-c data sets. The
figures on the left side show the results of the evaluations in the original data and
on the right side the results in the preprocessed versions where low-density points
were removed.
However, we also evaluated RECI with standardized data in the SIM-G data sets,
which is equivalent to a Gaussian reference measure for IGCI. A summary of the
results can be found in Figures 7(c)-7(d) in Appendix C and more detailed results
in Table 4 and Table 5 in Appendix D. These results are significantly better than
15
0.8
0.8
0.7
0.7
0.6
0.6
0.4
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
A
M
A
A
N
M
N
A
-H
S
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.5
N
Accuracy
1
0.9
-H
Accuracy
1
0.9
(b) Preprocessed SIM-ln
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
M
N
A
A
M
N
A
-H
S
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.4
(c) Original SIM-G
N
0.5
A
Accuracy
1
0.9
-H
Accuracy
(a) Original SIM-ln
(d) Preprocessed SIM-G
Figure 2: Evaluation result of all methods in the SIM-ln and SIM-G data sets. The
figures on the left side show the result of the evaluations in the original data and
on the right side the results in the preprocessed versions where low-density points
were removed.
normalizing the data in this case. However, although our theorem only justifies a
normalization, a different scaling, such as standardization, might be a reasonable
alternative.
Even though Theorem 1 does not exclude cases of a high noise level, it makes
16
a clear statement about low level noise. Therefore, as expected, RECI performs
the best in SIM-ln, where the noise level is low. In all cases, LiNGAM performs
very poorly due to the violations of its core assumptions.
ANM and RECI require a least-squares regression, but ANM additionally depends on an independence test, which can have a high computational cost and a
big influence on the performance. Therefore, even though RECI does not outperform ANM, it represents a competitive alternative with a lower computational
cost, depending on the regression model and MSE estimation. Also, seeing that
RECI explicitly allows both cases, a dependency and an independency between C
and N while ANM only allows the latter, it can be expected that RECI performs
significantly better than ANM in cases where the dependency between C and N
is strong. This is evaluated in Section 5.2.2. In comparison with IGCI, LiNGAM
and CURE, RECI outperforms in almost all data sets. Note that [16] performed
more extensive experiments and show more comparisons with ANM and IGCI in
these data sets, where additional parameter configurations were tested. However,
they report no results for the preprocessed data.
5.2.2
Simulated cause-effect pairs with strong dependent noise
Since the data sets of the evaluations in Section 5.2.1 are generated by structural
equations with independence noise variables, we additionally performed evaluations with artificial data sets where the input distribution and the noise distribution
are strongly dependent. For this, we considered a similar data generation process
as described in [13]. We generated data with various different cause and noise
distributions, different functions and varying values for α ∈ [0, 1]. In order to ensure a dependency between C and N, we additionally introduced two unobserved
source variables S1 and S2 that are randomly sampled from different distributions.
Variables C and N then consist of a randomly weighted linear combination of S1
and S2 . The general causal structure of these data sets is illustrated in Figure 3.
All used distributions for the source variables and all used functions can be found
in Table 1, where GMµ,σ denotes a Gaussian mixture distribution with density
pGMµ,σ (c) = 12 (ϕ(c|µ1 , σ1 ) + ϕ(c|µ2 , σ2 )) and Gaussian pdf ϕ(c|µ, σ ).
Apart from rather simple functions for φ , [13] proposed to generate more general functions in the form of convex combinations of mixtures of cumulative Gaussian distribution functions ψ(C|µi , σi ):
n
sn (C) = ∑ βi ψ(C|µi , σi ),
i=1
17
Table 1: All distributions pS , functions φ and functions f that were used for the
generation of the Linear, Non-invertible and Invertible data sets. In case
of the functions for Non-invertible, rescale(X, −n, n) denotes a rescaling of the
input data X on [−n, n].
pS
U(0, 1)
N (0, σ 2 )
N (0.5, σ 2 )
N (1, σ 2 )
GM[0.3,0.7]T ,[0.1,0.1]T
Data set
Linear
Invertible
Non-invertible
φ (C)
C
s5 (C)
rescale(C, −2, 2)2
rescale(C, −2, 2)4
sin(rescale(C, −2 · π, 2 · π))
f (X)
X
exp(X)
s5 (X)
Figure 3: The general structure of the data generation process where C and N are
dependent. In order to achieve this, cause and noise consist of a mixture of two
sources S1 and S2 .
where βi , µi ∈ [0, 1] and σi ∈ [0, 0.1]. For the experiments, we set n = 5 and chose
the parameters of s5 (C) randomly according to the uniform distribution. Note
that ψ(C|µi , σi ) is always monotonically increasing and thus s5 (C) can have an
arbitrary random shape while being monotonically increasing.
18
Cause-effect pairs were then generated in the following way:
w1 , w2 ∼ U(0, 1)
S1 , S2 ∼ pS
S1 = S1 − E[S1 ]
S2 = S2 − E[S2 ]
C0 = w1 · f1 (S1 ) + (1 − w1 ) · f2 (S2 )
N 0 = w2 · f3 (S1 ) + (1 − w2 ) · f4 (S2 )
C = normalize(C0 )
N = α · normalizeVariance(N 0 )
E = φ (C) + N,
where the distributions of S1 and S2 and the functions f1 , f2 , f3 , f4 were chosen
randomly from pS and f in Table 1, respectively. Note that S1 and S2 can follow different distributions. The choice of φ depends on the data set, where we
differentiated between three data sets:
• Linear: Only the identity function φ (C) = C
• Invertible: Arbitrary invertible functions φ (C) = s5 (C)
• Non-invertible: Functions that are not invertible on the respective domain
In total, we generated 100 data sets for each value of parameter α, which controls the amount of noise in the data. In each generated data set, we randomly
chose different distributions and functions. For Linear and Invertible the step
size of α is 0.1 and for Non-invertible 0.025. Here, we only performed one
repetition on each data set for all algorithms. Figure 4(a)-4(c) summarize all results and Table 7 in Appendix D shows the best performing functions and parameters of the different causal inference methods. Note that we omitted experiments
with CURE in these data sets due to the high computational cost.
Linear: As expected, ANM, IGCI and RECI perform very poorly, since they
require nonlinear data. In case of RECI, Theorem 1 states an equality of the
MSE if the functional relation is linear and, thus, the causal direction can not be
inferred. While LiNGAM performs well for α = 0.1, and probably for smaller
values of α too, the performance drops if α increases. The poor performances of
19
LiNGAM and ANM can also be explained by the violation of its core assumption
of an independence between cause and input.
Invertible: In this data set, IGCI performs quite well for small α, since all
assumptions approximately hold, but the performance decreases when the noise
becomes stronger and violates the assumption of a deterministic relation. In case
of RECI, we made a similar observation, but it performs much better than IGCI
if α < 0.5. Aside from the assumption of linear data for LiNGAM, the expected
poor performance of LiNGAM and ANM can be explained by the violation of the
independence assumption between C and N.
Non-invertible: These results seem very interesting, since it supports the
argument that the error asymmetry becomes even more clearer if the function
is not invertible due to an information loss of regressing in anticausal direction.
On the other hand, IGCI performs reasonably well, while ANM and LiNGAM
perform even worse than a baseline of just guessing. The constant results of each
method can be explained by the rather simple and similar choice of data generating
functions.
While the cause and noise also have a dependency in the SIM-c data sets,
the performance gap between ANM and RECI is vastly greater in Invertible
and Non-invertible than in SIM-c due to a strong violation of the independent
noise assumption. Therefore, RECI might perform better than ANM in cases with
a strong dependency between cause and noise.
5.3
Real-world data
In real-world data, the true causal relationship generally requires expert knowledge and can still lead to rather philosophical discussions in unclear cases. For our
evaluations, we considered the commonly used cause-effect pairs (CEP) benchmark data sets. These benchmark data sets provided, at the time of these evaluations, 106 data sets with given cause and effect variables.4 However, since we
only consider a two variable problem, we omit the multivariate data sets, which
leaves 100 data sets for the evaluations. Each data set comes with a corresponding weight. This is because several data sets are too similar to consider them as
independent examples, hence they get lower weights. Therefore, the accuracy is
4 The data set can be found on https://webdav.tuebingen.mpg.de/cause-effect/.
More details and
discussion about the causal relationship of the first 100 pairs can be found in [16].
20
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Accuracy
Accuracy
1
0.5
0.4
0.3
0.5
0.4
0.3
0.2
0.2
RECI
ANM
IGCI
LiNGAM
0.1
RECI
ANM
IGCI
LiNGAM
0.1
0
0
0
0.2
0.4
0.6
0.8
0
1
0.2
0.4
,
0.6
0.8
1
,
(a) Linear
(b) Invertible
1
0.9
0.8
Accuracy
0.7
0.6
0.5
0.4
0.3
0.2
RECI
ANM
IGCI
LiNGAM
0.1
0
0
0.2
0.4
0.6
0.8
1
,
(c) Non-Invertible
Figure 4: Evaluation result of all methods in the Linear, Invertible and
Non-Invertible data sets. The parameter α controls the amount of noise in
the data.
calculated in the following way
accuracy =
∑M
m=1 wm δdˆm ,dm
∑M
m=1 wm
21
,
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Accuracy
0.5
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
A
N
M
-H
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
G
A
M
A
-H
M
N
A
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0
Y R N
G
O ON L V I-N
I-L I-M I-PO CI-S EC
EC C C E R
R RE RE R
A
Accuracy
1
0.9
e
(a) Original CEP
(b) Preprocessed CEP
Figure 5: Evaluation result of all methods in the real-world CEP data sets. The
figure on the left side shows the result of the evaluations in the original data and
on the right side the results in the preprocessed versions where low-density points
were removed.
where M is the number of data sets, wm the weight of data set m, dm the correct
causal direction and dˆm the inferred causal direction of the corresponding method.
The experimental setup is the same as for the artificial data sets, but we doubled
the number of internal repetition of CURE to eight times in order to provide the
same experimental conditions as in [15].
Figure 5(a) and 5(b) shows the results of the evaluations in the original and
preprocessed data, respectively. In all cases, RECI performs better than all other
methods.5 Further, the performance difference gap even increases in the preprocessed data with removed low-density points. Surprisingly, as Table 2 in Appendix
D indicates, the very simple monomial functions ax3 + c and ax2 + c perform the
best, even though it is very unlikely that these functions are able to capture the
true function φ . We obtained similar observations In the artificial data sets.
In order to show that RECI still performs reasonably well under a different
scaling, we also evaluated RECI in the real-world data set with standardized data.
5 The
work of [16] provides further evaluations of ANM and IGCI in the original CEP data
set with parameter configurations that reached slightly higher accuracies than the presented results in this work. Regarding CURE, we had to use a simplified implementation due to the high
computational cost, which did not perform as well as the results reported in [15].
22
Algorithm 2 Causal inference algorithm that uses (10) as rejection criterion.
function RECI(X, Y , t)
. X and Y are the observed data and t ∈ [0, 1] the
threshold for rejecting a decision.
(X,Y ) ← RescaleData(X,Y )
f ← FitModel(X,Y )
. Fit regression model f : X → Y
g ← FitModel(Y, X)
. Fit regression model g : Y → X
MSEY |X ← MeanSquaredError( f , X,Y )
MSEX|Y ← MeanSquaredError(g,Y, X)
ξ←
min(MSEX|Y ,MSEY |X )
max(MSEX|Y ,MSEY |X )
if ξ ≤ t and MSEY |X 6= MSEX|Y then
if MSEY |X < MSEX|Y then
return X is the cause and Y is the effect.
else
return Y is the cause and X is the effect.
end if
else
return No decision
end if
end function
These results can be found summarized in Figures 7(a)-7(b) in Appendix C and
more detailed in Table 4 and Table 5 in Appendix D. While standardizing the data
improve the performance in the SIM-G data, it slightly decreases the performance
in the real-world data as compared to a normalization of the data, but still performs
resonably well. This shows some robustness with respect to a different scaling.
5.4
Error ratio as rejection criterion
It is not clear how to define a confidence measure for the decision of RECI. However, since Theorem 1 states that the correct causal direction has a smaller error,
we evaluated the idea of using the error ratio
min(E[Var[X|Y ]], E[Var[Y |X]])
max(E[Var[X|Y ]], E[Var[Y |X]])
(10)
as a rejection criterion for a decision. The idea is that the smaller the error ratio
the higher the confidence of the decision due to the large error difference. Note
23
0
0.2 0.4 0.6 0.8
Error ratio
(a) CEP-LOG
50
0
0.2 0.4 0.6 0.8
100
Accuracy (%)
50
100
Accuracy (%)
Accuracy (%)
Accuracy (%)
100
50
0
0.2 0.4 0.6 0.8
Error ratio
Error ratio
(b) SIM-LOG
(c) SIM-c-NN
100
50
0
0.2 0.4 0.6 0.8
Error ratio
(d) SIM-ln-POLY
Figure 6: The exemplary performance of RECI with respect to the error ratio (10)
as rejection criterion. A small ratio indicates a high error difference.
that we formulated the ratio inverse to Theorem 1 in order to get a value on [0, 1].
Algorithm 1 can be modified in a straight forward manner to utilize this error ratio.
The modification is summarized in Algorithm 2.6
We re-evaluated the experimental results by considering only data sets where
Algorithm 2 returns a decision. Figures 6(a)-6(d) show some examples of the
performance of RECI if we use the error ratio as rejection criterion. In the figures,
an error ratio of 0.2, for instance, can be seen as a surrogate indicator of RECI’s
accuracy if we only consider data sets where the error ratio was ≤ 0.2 and exclude
data sets where the error ratio was > 0.2. In this sense, we can get an idea of
how useful the error ratio is as rejection criterion. While Figures 6(a), 6(c) and
6(d) support the intuition that the smaller the error ratio, the higher the decision
confidence, Figure 6(b) has a contradictive behavior. Generally, it seems that
a small error ratio (big error difference) is rather an indicator for an uncertain
decision, while an error ratio > 0.3 might be a more reliable indicator. Note that
in the most cases, depending on the regression model and scenario, the majority
of data sets have an error ratio between 0.3 and 0.9. All plots for the original data
can be found in Figure 8, all plots for the preprocessed data in Figure 9 and all
plots for the standardized data in Figure 10 in Appendix E.
5.5
Discussion
Due to the greatly varying behavior and the choice of various optimization parameters, a clear rule of which regression function is the best choice for RECI
remains an unclear and difficult problem. Overall, it seems that simple functions
are better in capturing the error asymmetries than complex models. However, a
6 Algorithm
1 and Algorithm 2 are equivalent if t = 1.
24
clear explanation for this is still lacking. A possible reason for this might be that
simple functions in causal direction already achieve a small error, while in anticausal direction, more complex models are required to achieve a small error.7
Regarding the removal of low-density points, the performance of methods that are
based on the Gaussiantiy assumption, such as FN and IGCI with Gaussian reference measure, seems not to be influenced by the removal. On the other hand,
the performance of HSIC, ENT, and IGCI with uniform measure is negatively affected, while the performance of LiNGAM and RECI increases. In case of RECI,
this can be explained by a better estimation of the true MSE with respect to the
regression function class.
Regarding the computational cost, we want to emphasize again that RECI,
depending on the implementation details, can have a significantly lower computational cost than ANM and CURE, while providing comparable or even better
results. Further, it can be easily implemented and applied.
6
Conclusion
We presented an approach for causal inference based on an asymmetry in the
prediction error. Under the assumption of an independence among the data generating function, the noise, and the distribution of the cause, we proved (in the limit
of small noise) that the conditional variance of predicting the cause by the effect
is greater than the conditional variance of predicting the effect by the cause. Here,
the additive noise is not assumed to be independent of the cause (in contrast to
so-called additive noise models). The stated theorem might also be interesting for
other statistical applications.
We proposed an easily implementable and applicable algorithm, which we call
RECI, that exploits this asymmetry for causal inference. The evaluations show
supporting results and leave room for further improvements. By construction, the
performance of RECI depends on the regression method. According to our limited
7 To
justify speculative remarks of this kind raises deep questions about the foundations of
causal inference. Using algorithmic information theory, one can, for instance, show that the algorithmic independence of pC and pE|C implies
K(pC ) + K(pE|C ) ≤ K(pE ) + K(pC|E ),
if K denotes the description length of a distribution in the sense of Kolmogorov complexity, for
details see Section 4.1.9 in [4]. In this sense, appropriate independence assumptions between pC
and pE|C imply that pE,C has a simpler description in causal direction than in anticausal direction.
25
experience so far, regression with simple model classes (that tend to underfit the
data) performs reasonably well. To clarify whether this happens because the conditional distributions tend to be simpler – in a certain sense – in causal direction
than in anticausal direction has to be left for the future.
Appendices
A
Proof of Lemma 1
We first give some reminders of the definition of the conditional variance and
some properties. For two random variables Z, Q the conditional variance of Z,
given q is defined by
Var[Z|q] := E[(Z − E[Z|q])2 |q],
while Var[Z|Q] is the random variable attaining the value Var[Z|q] when Q attains
the value q. Its expectation reads
Z
E[Var[Z|Q]] :=
Var[Z|q]pQ (q)dq.
For any a ∈ R, we have
Var[Z|q]
Z
q =
.
Var
a
a2
For any function h we have
Var[h(Q) + Z|q] = Var[Z|q],
which implies Var[h(Q)|q] = 0. Moreover, we have
Var[Z|h(q)] = Var[Z|q],
if h is invertible.
To begin the main part of the proof, we first observe
E[Var[Eα |C]] = E[Var[φ (C) + αN|C]] = α 2 E[Var[N|C]] = α 2 .
|
{z
}
= 1 (Assumpt. 4))
26
(11)
Moreover, one easily verifies that
E[Var[C|Ẽα ]]
E[Var[C|Eα ]]
,
= lim
α→0 E[Var[Ẽα |C]]
α→0 E[Var[Eα |C]]
lim
(12)
due to (3) provided that these limits exist. Combining (11) and (12) yields
E[Var[C|Ẽα ]]
E[Var[C|Eα ]]
C
lim
= lim
= lim E Var
E .
α→0 E[Var[Ẽα |C]]
α→0
α→0
α2
α
(13)
Now, we can rewrite (13) as
−1
C
φ (Eα − αN)
lim E Var
Eα = lim E Var
Eα
α→0
α→0
α
α
−1
Z φ (1)+αn+
φ (e − αN)
= lim
e pEα (e)de
Var
α→0 φ (0)+αn−
α
−1
Z φ (1)
φ (e − αN)
e pEα (e)de,
= lim
Var
(14)
α→0 φ (0)
α
where in the latter step we exploit that the function
e 7→ Var φ −1 (e − αN)/α e pEα (e)
is uniformly bounded in α. This is because, first, φ −1 attains only values in [0, 1],
and hence the variance is bounded by 1. Second, pEα (e) is uniformly bounded
due to
Z n+
pEα (e) =
pφ (C),N (e − αn, n)dn =
n−
−10
≤ kφ
Z n+
n−
0
pC,N (φ −1 (e − αn), n)φ −1 (e − αn)dn
k∞ kpC,N k∞ (n+ − n− ).
Hence, the bounded convergence theorem states
−1
−1
Z φ (1)
Z φ (1)
φ (e − αN)
φ (e − αN)
e pEα (e)de =
lim Var
e pEα (e) de.
lim
Var
α→0 φ (0)
α
α
φ (0) α→0
To compute the limit of
φ −1 (e − αN)
Var
e ,
α
27
we use Taylor’s theorem to obtain
φ −1 (e − αn) = φ −1 (e) − αnφ −10 (e) −
α 2 n2 φ −100 (E2 (n, e))
,
2
(15)
where E2 (n, e) is a real number in the interval (e − αn, e). Since (15) holds for
−1 are bijections of [0, 1], thus e + αn lies
every n ∈ [− αe , 1−e
α ] (note that φ and φ
in [0, 1]) it also holds for the random variable N if E2 (n, e) is replaced with the
random variable E2 (N, e) (here, we have implicitly assumed that the map n 7→
e2 (n, e) is measurable). Therefore, we see that
−1
αN 2 φ −100 (E2 (N, e))
φ (e − αN)
−10
lim Var
e = lim Var −Nφ (e) −
e
α→0
α→0
α
2
0
= φ −1 (e)2 Var[N|e].
(16)
lim pEα (e) = pE0 (e).
(17)
Moreover, we have
α→0
Inserting (17) and (16) into (14) yields
Z φ (1)
0
C
E0
=
φ −1 (e)2 Var[N|e]pE0 (e)de
lim E Var
α→0
α
φ (0)
Z 1
=
0
0
φ −1 (φ (c))2 Var[N|φ (c)]pC (c)dc
Z 1
1
0
φ 0 (c)2
=
Var[N|φ (c)]pC (c)dc,
where the second equality is a variable substituition using the deterministic relation E0 = φ (C) (which implies pE0 (φ (c)) = pC (c)/φ 0 (c) or, equivalently, the
simple symbolic equation pE0 (e)de = pC (c)dc). This completes the proof due to
(13).
B
Proof of Theorem 1
We first recall that Lemma 1 states
E[Var[C|Ẽα ]]
lim
=
α→0 E[Var[Ẽα |C]]
Z 1
1
0
φ 0 (c)2
28
Var[N|c]pC (c)dc.
We then have
Z 1
1
0
φ 0 (c)2
Z 1
=
0
Var[N|c]pC (c)dc
1
1
Var[N|c]p
(c)dc
·
Var[N|c]pC (c)dc
C
φ 0 (c)2
0
|
{z
}
Z
= 1 (Assumpt. 4)
2
s
Z 1
2
Z 1p
2
1
Var[N|c] pC (c)dc
=
Var[N|c] pC (c)dc ·
0
φ (c)
0
0
s
2
2
Z 1
p
1
≥
Var[N|c] Var[N|c]pC (c)dc
0
φ (c)
0
2
Z 1
1
Var[N|c]pC (c)dc ,
=
0
0 φ (c)
(18)
where the inequalityR is just the Cauchy Schwarz inequality applied to the bilinear
form f , g 7→ f (c)g(c)pC (c)dc for the space of functions f for which
R 2
f (c)pc (c)dc exists. We can make a statement about (18) in a similar way by
using (7) implied by the independence postulate and using Cauchy Schwarz:
Z 1
1
0
φ 0 (c)
Z 1
=
0
Var[N|c]pC (c)dc
1
1
Var[N|c]p
(c)dc
·
φ 0 (c) Var[N|c]pC (c)dc
C
φ 0 (c)
|0
{z
}
Z
= 1 (7)
Z 1
s
2
Z 1p
2
1
0 (c) Var[N|c] p (c)dc
Var[N|c]
p
(c)dc
·
φ
C
C
φ 0 (c)
0
0
s
!2
Z 1
p
1
≥
Var[N|c] φ 0 (c) Var[N|c]pC (c)dc
φ 0 (c)
0
2
=
Z 1
= Var[N|c]pC (c)dc = 1.
| 0
{z
}
= 1 (Assumpt. 4)
Combining (18) and (19) with Lemma 1 completes the proof.
29
(19)
30
Results of standardized data
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
M
N
M
N
A
A
A
(b) Prep. CEP with standardized variables
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Accuracy
1
0.9
0.5
0.4
0.5
0.4
0.2
0.1
0.1
0
0
N
M
A HS
N
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
0.2
N SIC
M
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
C G3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.3
A
A
M
-H
Accuracy
(a) Org. CEP with standardized variables
A
N
-H
S
M IC
A ENT
N
M
-F
IG N
C
I
IG U1
C
I
IG U2
C
I
IG U3
C
I
IG G1
C
I
IG G2
C
I
G
C 3
U
Li RE
N
R G
EC A
M
R I-L
EC O
R I-M G
EC O
I- N
R PO
EC L
I- Y
R SV
EC R
I-N
N
0.4
N
0.5
A
Accuracy
1
0.9
-H
Accuracy
C
(c) Org. SIM-G with standardized variables
(d) Prep. SIM-G with standardized variables
Figure 7: Evaluation result of all methods in the real-world CEP and SIM-G data
sets. Here, a standardization instead of a normalization was used for the scaling
in RECI. The figure on the left side shows the results of the evaluations in the
original data and on the right side the results in the preprocessed versions where
low-density points were removed.
31
D
Tables
Table 7: The best performing parameters for each causal inference method in the
artificial data sets Linear, Invertible and Non-invertible.
Data set
Linear
Invertible
Non-invertible
ANM estimator
FN
HSIC
HSIC
IGCI configuration
IGCIU1
IGCIU2
IGCIU1
32
RECI regression model
MON: ax9 + c
NN 4-8
POLY: ∑4i=0 ai xi
33
LOG
ax2 + c
ax3 + c
ax4 + c
ax5 + c
ax6 + c
ax7 + c
ax8 + c
ax9 + c
∑1i=0 ai xi
∑2i=0 ai xi
∑3i=0 ai xi
∑4i=0 ai xi
∑5i=0 ai xi
∑6i=0 ai xi
∑7i=0 ai xi
∑8i=0 ai xi
∑9i=0 ai xi
SVR
NN 2
NN 5
NN 10
NN 20
NN 2-4
NN 4-8
CEP
68.93%
72.13%
72.27%
71.58%
69.92%
68.84%
67.75%
67.61%
66.56%
71.58%
70.09%
68.61%
66.63%
69.32%
67.52%
66.36%
68.87%
69.33%
68.84%
65.87%
68.48%
64.82%
67.81%
64.28%
67.47%
Prep. CEP
70.02%
75.41%
72.67%
72.67%
71.99%
70.16%
68.93%
67.42%
69.93%
66.51%
64.26%
64.49%
67%
66.54%
67%
67.55%
66.74%
66.74%
66.51%
67%
65.85%
66.63%
70.53%
64.12%
65.85%
SIM
61%
43%
45%
50%
51%
52%
52%
51%
50%
46%
53%
57%
57%
59%
60%
60%
60%
60%
48%
57%
56%
57%
62%
55%
58%
Prep. SIM
71%
52%
55%
53%
54%
55%
54%
55%
58%
54%
58%
63%
61%
63%
65%
65%
64%
64%
54%
61%
59%
62%
63%
63%
62%
SIM-c
70%
57%
58%
56%
60%
61%
62%
62%
63%
51%
65%
67%
68%
68%
68%
68%
68%
68%
52%
68%
67%
68%
66%
68%
68%
Prep. SIM-c
74%
63%
60%
58%
62%
63%
64%
62%
62%
58%
68%
71%
72%
71%
70%
72%
69%
69%
59%
72%
70%
69%
70%
71%
68%
SIM-ln
72%
44%
45%
45%
44%
42%
42%
41%
42%
45%
69%
78%
80%
81%
82%
81%
82%
81%
45%
81%
79%
80%
81%
78%
82%
Prep. SIM-ln
80%
47%
47%
47%
47%
46%
45%
45%
44%
43%
75%
76%
77%
75%
76%
74%
75%
74%
43%
79%
76%
74%
72%
76%
76%
SIM-G
72%
44%
43%
43%
45%
45%
47%
44%
45%
41%
54%
57%
60%
60%
64%
64%
65%
64%
40%
60%
61%
62%
60%
60%
61%
Prep. SIM-G
66%
33%
34%
37%
38%
39%
38%
38%
36%
26%
49%
52%
55%
53%
59%
58%
57%
55%
26%
56%
56%
56%
57%
54%
55%
Table 2: The performances of all used classes of regression functions for RECI in the artificial and real-world
benchmark data sets when averaging the MSE over all runs.
34
LOG
ax2 + c
ax3 + c
ax4 + c
ax5 + c
ax6 + c
ax7 + c
ax8 + c
ax9 + c
∑1i=0 ai xi
∑2i=0 ai xi
∑3i=0 ai xi
∑4i=0 ai xi
∑5i=0 ai xi
∑6i=0 ai xi
∑7i=0 ai xi
∑8i=0 ai xi
∑9i=0 ai xi
SVR
NN 2
NN 5
NN 10
NN 20
NN 2-4
NN 4-8
CEP
63.66% ± 3.48
70.73% ± 1.55
70.73% ± 1.73
69.47% ± 1.48
69.20% ± 1.82
68.84% ± 0.00
67.35% ± 2.54
67.27% ± 2.31
66.78% ± 2.25
68.84% ± 0.00
67.7% ± 0.00
67.24% ± 0.00
63.87% ± 2.78
64.85% ± 0.00
64.8% ± 2.91
65.08% ± 0.00
65.39% ± 0.00
65.39% ± 0.00
66% ± 0.38
63.64% ± 1.94
63.92% ± 2.06
64.83% ± 2.44
64.9% ± 3.02
63.26% ± 1.87
64.33% ± 2.23
Prep. CEP
64.34% ± 3.57
72.90% ± 0.00
72.67% ± 0.00
72.67% ± 0.00
70.5% ± 2.18
69.63% ± 2.47
68.01% ± 2.61
67.27% ± 2.44
66.67% ± 2.77
66.28% ± 0.00
66% ± 2.64
65.4% ± 1.97
66.77% ± 0.00
66.54% ± 0.00
67% ± 0.00
67.55% ± 0.00
66.74% ± 0.00
66.74% ± 0.00
64.67% ± 2.73
64.84% ± 1.6
65.69% ± 1.52
66.07% ± 1.46
66.87% ± 2.19
65.33% ± 1.61
66.05% ± 1.44
SIM
57.62% ± 3.57
43.84% ± 1.17
45.47% ± 1.78
50% ± 0.00
51% ± 0.00
52% ± 0.00
50% ± 0.00
48.14% ± 2.24
48.39% ± 2.5
45.77% ± 1.54
53.53% ± 1.86
57.67% ± 2.4
57.47% ± 2.37
57.3% ± 2.72
57.21% ± 2.39
59% ± 0.00
60% ± 0.00
58.01% ± 2.82
46.12% ± 1.79
55.58% ± 2.32
56.69% ± 2.56
56.06% ± 2.63
58.69% ± 2.62
56.72% ± 2.5
56.33% ± 2.53
Prep. SIM
60.83% ± 3.83
52% ± 0.00
54% ± 0.00
53% ± 0.00
54% ± 0.00
55% ± 0.00
53.92% ± 1.73
54.89% ± 1.8
55.74% ± 1.9
54.37% ± 1.47
59.54% ± 2.11
61.43% ± 2.58
60.72% ± 2.59
61.79% ± 2.12
62.68% ± 2.92
62.5% ± 2.92
60% ± 0.00
62.15% ± 2.82
54.33% ± 2.06
59.83% ± 2.39
58.66% ± 2.51
59.21% ± 2.72
59.72% ± 3.23
59.35% ± 2.71
58.84% ± 2.71
SIM-c
65.02% ± 3.82
57% ± 0.00
58% ± 0.00
56.42% ± 1.44
58.11% ± 1.78
59% ± 0.00
60% ± 0.00
60.27% ± 1.58
61% ± 0.00
51% ± 0.00
65% ± 0.00
67% ± 0.00
68% ± 0.00
68% ± 0.00
68% ± 0.00
68% ± 0.00
62.06% ± 2.86
68% ± 0.00
52% ± 0.00
66.54% ± 1.78
66.49% ± 0.97
66.82% ± 1.1
66.36% ± 2.78
66.52% ± 1.02
66.48% ± 1.77
Prep. SIM-c
67.52% ± 3.3
63% ± 0.00
59% ± 0.00
58.52% ± 1.87
61% ± 0.00
62.99% ± 0.1
64% ± 0.00
62% ± 0.00
62% ± 0.00
58.24% ± 1.62
68% ± 0.00
70% ± 0.00
70% ± 0.00
70% ± 0.00
69% ± 0.00
69% ± 0.00
68% ± 0.00
69% ± 0.00
59% ± 0.00
68.36% ± 2
68.26% ± 2.28
68.71% ± 2.26
69.46% ± 2.42
68.3% ± 2.29
68.5% ± 2.59
SIM-ln
71.74% ± 4.02
44.45% ± 1.29
45.08% ± 1.28
45% ± 0.00
43.76% ± 1.66
42.69% ± 1.93
41.76% ± 2.31
41% ± 0.00
41% ± 0.00
45% ± 0.00
69% ± 0.00
78% ± 0.00
80% ± 0.00
81% ± 0.00
82% ± 0.00
81% ± 0.00
82% ± 0.00
81% ± 0.00
41.37% ± 2.1
80.68% ± 1.48
79.72% ± 1.74
76.99% ± 2.41
76.36% ± 2.84
80.25% ± 1.81
78.71% ± 2.01
Prep. SIM-ln
72.71% ± 3.49
47% ± 0.00
47.63% ± 1.13
47% ± 0.00
47% ± 0.00
46% ± 0.00
45% ± 0.00
45% ± 0.00
44% ± 0.00
43% ± 0.00
73% ± 0.00
75.54% ± 1.22
75.63% ± 1.45
75.49% ± 1.56
76% ± 0.00
74% ± 0.00
75% ± 0.00
74% ± 0.00
43% ± 0.00
76.97% ± 2.31
74.58% ± 3.17
74.67% ± 3.17
73.10% ± 2.83
76.05% ± 2.76
73.31% ± 2.6
SIM-G
67.37% ± 4.15
43.74% ± 0.91
42.77% ± 1.72
44.11% ± 1.62
44.59% ± 1.99
44% ± 0.00
44% ± 0.00
43.27% ± 2.26
43% ± 0.00
41% ± 0.00
54.08% ± 1.68
57.26% ± 1.95
59.35% ± 2.11
59% ± 0.00
60.14% ± 2.27
60.04% ± 2.29
61% ± 0.00
60% ± 0.00
40% ± 0.00
59.65% ± 2.43
60.26% ± 2.53
59.48% ± 2.38
58.98% ± 2.9
59.98% ± 2.46
59.76% ± 2.51
Table 3: The performances of all used classes of regression functions for RECI in the artificial and real-world
benchmark data sets when not averaging the MSE.
Prep. SIM-G
61.95% ± 4.35
33.56% ± 1.78
34.80% ± 1.69
37.14% ± 1.8
38.59% ± 2.07
38.3% ± 2.35
38% ± 0.00
35% ± 1.98
34.3% ± 2.83
26% ± 0.00
49% ± 0.00
52.01% ± 2.31
54% ± 0.00
53.06% ± 2.37
54.82% ± 2.41
55.15% ± 2.89
55.84% ± 3.06
55% ± 0.00
25.69% ± 1.88
54.93% ± 2.89
54.67% ± 3.11
54.49% ± 2.85
54.83% ± 3.21
54.26% ± 2.83
55.44% ± 2.61
35
LOG
ax2 + c
ax3 + c
ax4 + c
ax5 + c
ax6 + c
ax7 + c
ax8 + c
ax9 + c
∑1i=0 ai xi
∑2i=0 ai xi
∑3i=0 ai xi
∑4i=0 ai xi
∑5i=0 ai xi
∑6i=0 ai xi
∑7i=0 ai xi
∑8i=0 ai xi
∑9i=0 ai xi
SVR
NN 2
NN 5
NN 10
NN 20
NN 2-4
NN 4-8
Standardized CEP
69.29%
63.57%
61.99%
65.44%
64.61%
65.65%
64.49%
66.72%
65.86%
60.83%
74.07%
67.47%
69.27%
69.54%
72.94%
67.30%
67.87%
68.89%
55.30%
55.19%
56.85%
60.57%
56.75%
60.09%
59.13%
Standardized Prep. CEP
66.87%
61.59%
64.71%
66.79%
62.71%
65.72%
64.26%
65.76%
65.84%
61.25%
63.99%
66.01%
66.07%
66.95%
71.85%
71.19%
65.70%
68.31%
61.58%
64.82%
61.35%
58.90%
60.89%
57.78%
59.21%
Standardized SIM-G
80%
38%
60%
40%
63%
56%
63%
61%
64%
50%
83%
83%
85%
85%
83%
83%
83%
81%
42%
80%
81%
80%
79%
72%
80%
Standardized Prep. SIM-G
78%
34%
47%
23%
48%
43%
47%
54%
51%
49%
83%
83%
83%
82%
81%
82%
82%
81%
47%
78%
76%
80%
79%
77%
80%
Table 4: The performances of all used classes of regression functions for RECI in the standardized CEP and SIM-G
data sets when averaging the MSE over all runs.
36
LOG
ax2 + c
ax3 + c
ax4 + c
ax5 + c
ax6 + c
ax7 + c
ax8 + c
ax9 + c
∑1i=0 ai xi
∑2i=0 ai xi
∑3i=0 ai xi
∑4i=0 ai xi
∑5i=0 ai xi
∑6i=0 ai xi
∑7i=0 ai xi
∑8i=0 ai xi
∑9i=0 ai xi
SVR
NN 2
NN 5
NN 10
NN 20
NN 2-4
NN 4-8
Standardized CEP
53.86% ± 5.33
62.41% ± 0.00
61.99% ± 0.00
65.44% ± 0.00
56.80% ± 0.00
65.65% ± 0.00
60.53% ± 0.00
65.73% ± 0.63
63.72% ± 0.92
56.05% ± 5.73
60.03% ± 5.66
60.52% ± 0.00
61.04% ± 4.80
62.03% ± 0.00
61.45% ± 4.99
61.45% ± 4.02
62.37% ± 5.10
63.70% ± 4.53
55.12% ± 5.50
55.89% ± 5.30
55.61% ± 5.87
56.63% ± 5.37
55.93% ± 6.55
56.26% ± 6.47
56.57% ± 3.70
Standardized Prep. CEP
54.18% ± 6.12
61.59% ± 0.00
60.81% ± 0.05
66.79% ± 0.00
58.79% ± 0.02
64.18% ± 0.00
57.39% ± 0.00
62.91% ± 0.00
58.79% ± 0.00
57.78% ± 6.10
59.82% ± 0.00
59.14% ± 4.47
61.14% ± 4.65
62.07% ± 4.54
62.28% ± 4.87
63.36% ± 4.56
63.35% ± 4.20
63.27% ± 4.16
57.15% ± 5.49
57.12% ± 5.70
57.31% ± 6.08
57.08% ± 3.52
58.07% ± 5.76
56.27% ± 6.03
57.65% ± 3.47
Standardized SIM-G
76.67% ± 3.26
40.83% ± 4.77
54.42% ± 3.53
37.10% ± 3.34
57% ± 0.00
45% ± 0.00
53% ± 0.00
50.37% ± 3.76
53% ± 0.00
54% ± 0.00
83% ± 0.00
83% ± 0.00
83% ± 0.00
81% ± 0.00
80% ± 0.00
83% ± 0.00
83% ± 0.00
81% ± 0.00
46.69% ± 4.65
80.07% ± 2.38
78.10% ± 2.23
76.68% ± 2.41
75.45% ± 2.76
78.71% ± 2.69
77.04% ± 2.19
Standardized Prep. SIM-G
76.21% ± 3.13
39.79% ± 3.67
47.91% ± 4.12
29.85% ± 3.90
45.33% ± 4.81
35.60% ± 3.80
44.65% ± 4.10
41.36% ± 5.09
50% ± 0.00
50.01% ± 4.17
83% ± 0.00
83% ± 0.00
83% ± 0.00
81% ± 0.00
80% ± 0.00
82% ± 0.00
82% ± 0.00
81% ± 0.00
47% ± 0.00
80.04% ± 2.74
77.29% ± 2.35
77.10% ± 2.50
75.78% ± 2.94
78.26% ± 2.62
77.60% ± 2.40
Table 5: The performances of all used classes of regression functions for RECI in the standardized CEP and SIM-G
data sets when not averaging the MSE.
37
ANM-HSIC
ANM-ENT
ANM-FN
IGCIU,1
IGCIU,2
IGCIU,3
IGCIG,1
IGCIG,2
IGCIG,3
CURE
LiNGAM
CEP
62.89% ± 1.56
49.07% ± 1.88
50.80% ± 2.80
60.66% ± 0.00
63.97% ± 0.00
60.67% ± 0.00
56.62% ± 0.00
55.15% ± 0.00
45.45% ± 0.00
59.67% ± 2.86
57.76% ± 0.00
Prep. CEP
65.01% ± 1.74
43.57% ± 2.31
53.50% ± 2.18
62.13% ± 0.00
61.40% ± 0.00
61.99% ± 0.00
56.26% ± 0.00
54.79% ± 0.00
41.26% ± 0.00
59.67% ± 2.86
57.06% ± 0.00
SIM
75.54% ± 2.14
72.91% ± 3.09
57.50% ± 1.60
36% ± 0.00
41% ± 0.00
41% ± 0.00
36% ± 0.00
37% ± 0.00
37% ± 0.00
59.62% ± 3.30
43% ± 0.00
Prep. SIM
73.33% ± 2.08
71.51% ± 3.35
51.94% ± 1.92
39% ± 0.00
42% ± 0.00
42% ± 0.00
36% ± 0.00
38% ± 0.00
38% ± 0.00
59.62% ± 3.30
49% ± 0.00
SIM-c
80.54% ± 2.14
76.36% ± 3.14
61.15% ± 1.99
46% ± 0.00
50% ± 0.00
50% ± 0.00
42% ± 0.00
46% ± 0.00
46% ± 0.00
68.60% ± 3.68
47% ± 0.00
Prep. SIM-c
77.93% ± 2.02
74.95% ± 3.57
59.57% ± 1.93
47% ± 0.00
52% ± 0.00
52% ± 0.00
41% ± 0.00
45% ± 0.00
45% ± 0.00
68.60% ± 3.68
52% ± 0.00
SIM-ln
76.60% ± 2.53
78.88% ± 2.63
85.31% ± 1.36
51% ± 0.00
52% ± 0.00
52% ± 0.00
59% ± 0.00
61% ± 0.00
61% ± 0.00
55.58% ± 1.82
24% ± 0.00
Table 6: All performances of ANM, IGCI, CURE and LiNGAM.
Prep. SIM-ln
70.98% ± 2.40
73.93% ± 2.88
80.03% ± 1.31
57% ± 0.00
57% ± 0.00
57% ± 0.00
57% ± 0.00
60% ± 0.00
60% ± 0.00
55.58% ± 1.82
25% ± 0.00
SIM-G
74.49% ± 2.57
70.32% ± 3.77
77.72% ± 1.90
54% ± 0.00
55% ± 0.00
55% ± 0.00
87% ± 0.00
86% ± 0.00
86% ± 0.00
47.46% ± 0.96
28% ± 0.00
Prep. SIM-G
73.78% ± 2.44
69.78% ± 3.32
78% ± 2.23
48% ± 0.00
48% ± 0.00
48% ± 0.00
87% ± 0.00
86% ± 0.00
86% ± 0.00
47.46% ± 0.96
25% ± 0.00
Plots
0
0.2 0.4 0.6 0.8
100
50
0
Error ratio
100
50
0
0.2 0.4 0.6 0.8
100
50
0
Error ratio
SIM-ln-LOG
0.2 0.4 0.6 0.8
SIM-c-MON
100
50
0
50
0
0.2 0.4 0.6 0.8
0
50
0
Error ratio
SIM-ln-MON
50
0
100
50
0
Error ratio
50
0
SIM-ln-
0.2 0.4 0.6 0.8
Error ratio
SIM-c-SVR
100
0.2 0.4 0.6 0.8
SIM-NN
0.2 0.4 0.6 0.8
POLY
0.2 0.4 0.6 0.8
100
Error ratio
SIM-SVR
100
Error ratio
0.2 0.4 0.6 0.8
100
0.2 0.4 0.6 0.8
CEP-NN
Accuracy (%)
100
Error ratio
50
0
Error ratio
0.2 0.4 0.6 0.8
SIM-c-POLY
Accuracy (%)
Accuracy (%)
0.2 0.4 0.6 0.8
SIM-POLY
Error ratio
Accuracy (%)
Accuracy (%)
SIM-c-LOG
0
50
Error ratio
CEP-SVR
Error ratio
SIM-MON
Accuracy (%)
Accuracy (%)
50
0.2 0.4 0.6 0.8
50
Error ratio
SIM-LOG
100
CEP-POLY
100
0.2 0.4 0.6 0.8
Accuracy (%)
0
Error ratio
0
100
Error ratio
Accuracy (%)
50
50
SIM-c-NN
Accuracy (%)
0.2 0.4 0.6 0.8
100
0.2 0.4 0.6 0.8
Accuracy (%)
0
0
100
Error ratio
CEP-MON
Accuracy (%)
Accuracy (%)
50
0.2 0.4 0.6 0.8
50
Error ratio
CEP-LOG
100
Accuracy (%)
0
Error ratio
100
Accuracy (%)
0.2 0.4 0.6 0.8
Accuracy (%)
0
50
Accuracy (%)
50
100
Accuracy (%)
100
Accuracy (%)
Accuracy (%)
E
100
50
0
0.2 0.4 0.6 0.8
Error ratio
0.2 0.4 0.6 0.8
Error ratio
SIM-ln-NN
0.4
0.6
0.8
Error ratio
SIM-G-LOG
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-MON
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-POLY
100
Accuracy (%)
0
50
100
Accuracy (%)
50
100
Accuracy (%)
100
Accuracy (%)
Accuracy (%)
SIM-ln-SVR
50
0
0.4
0.6
0.8
Error ratio
SIM-G-SVR
100
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-NN
Figure 8: The performance of RECI in the original data sets with respect to the
error ratio (10) as rejection criterion.
38
0
0.4
0.6
0.8
100
50
0
Error ratio
100
50
0
0.6
0.8
SIM-c-MON
0.2 0.4 0.6 0.8
100
50
0
Error ratio
SIM-ln-LOG
0.4
Accuracy (%)
Accuracy (%)
0.2 0.4 0.6 0.8
100
50
0
Accuracy (%)
50
0
0.4
50
0
0.2 0.4 0.6 0.8
0
0.2 0.4 0.6 0.8
Error ratio
SIM-ln-MON
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-NN
100
50
0
100
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-c-SVR
POLY
50
100
Error ratio
100
Error ratio
0.8
100
Error ratio
0.2 0.4 0.6 0.8
0.6
0.2 0.4 0.6 0.8
CEP-NN
SIM-SVR
0.2 0.4 0.6 0.8
SIM-c-POLY
0
Error ratio
SIM-POLY
Error ratio
Accuracy (%)
Accuracy (%)
SIM-c-LOG
0
50
Error ratio
Accuracy (%)
50
50
100
Error ratio
SIM-MON
Accuracy (%)
Accuracy (%)
100
0.2 0.4 0.6 0.8
100
Error ratio
SIM-LOG
0.2 0.4 0.6 0.8
CEP-SVR
Accuracy (%)
0
Error ratio
0
SIM-c-NN
Accuracy (%)
0.2 0.4 0.6 0.8
50
50
100
Error ratio
CEP-POLY
Accuracy (%)
0
100
0.5
100
Error ratio
CEP-MON
Accuracy (%)
Accuracy (%)
50
0
0
Error ratio
CEP-LOG
100
0.2 0.4 0.6 0.8
50
Accuracy (%)
0
Error ratio
100
Accuracy (%)
0.2 0.4 0.6 0.8
Accuracy (%)
0
50
Accuracy (%)
50
100
Accuracy (%)
Accuracy (%)
Accuracy (%)
100
100
50
0
0.2 0.4 0.6 0.8
Error ratio
0.2 0.4 0.6 0.8
SIM-ln-
Error ratio
SIM-ln-NN
0.4
0.6
0.8
Error ratio
SIM-G-LOG
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-MON
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-POLY
100
Accuracy (%)
0
50
100
Accuracy (%)
50
100
Accuracy (%)
100
Accuracy (%)
Accuracy (%)
SIM-ln-SVR
50
0
0.4
0.6
0.8
Error ratio
SIM-G-SVR
100
50
0
0.2 0.4 0.6 0.8
Error ratio
SIM-G-NN
Figure 9: The performance of RECI in the preprocessed data sets with respect to
the error ratio (10) as rejection criterion.
39
0
Org. SIM-G-MON
0.2 0.4 0.6 0.8
100
50
0
0.6
Error ratio
Accuracy (%)
Accuracy (%)
50
0
0.2
0.4
0.6
0.8
Error ratio
Prep. SIM-G-LOG
50
0
0
100
100
0.2 0.4 0.6 0.8
Error ratio
Prep. SIM-G-MON
50
0
0.2 0.4 0.6 0.8
Error ratio
Prep. SIM-G-POLY
Accuracy (%)
0
0
0.85
0.9
50
0
0.2 0.4 0.6 0.8
50
0
0
0.85
0.9
0.95
Error ratio
Prep. SIM-G-SVR
0.2 0.4 0.6 0.8
Error ratio
Org. SIM-G-NN
100
50
0
0.2 0.4 0.6 0.8
Error ratio
Prep. CEP-SVR
50
0.8
100
Error ratio
100
0.6
Org. CEP-NN
0.95
100
0.4
Error ratio
Org. SIM-G-SVR
0.5
Prep. CEP-POLY
0
50
Error ratio
Prep. CEP-MON
50
Error ratio
Org. SIM-G-POLY
0.8
50
100
Error ratio
Error ratio
Prep. CEP-LOG
100
0.2 0.4 0.6 0.8
100
0.9
Org. CEP-SVR
Accuracy (%)
50
0
Error ratio
Accuracy (%)
Accuracy (%)
100
0.2 0.4 0.6 0.8
50
0.8
Accuracy (%)
0
Error ratio
100
0.7
Accuracy (%)
50
0
100
Error ratio
Accuracy (%)
100
50
Prep. CEP-NN
Accuracy (%)
0.2 0.4 0.6 0.8
Org. SIM-G-LOG
0.2 0.4 0.6 0.8
Org. CEP-POLY
Accuracy (%)
0
0
100
Error ratio
Org. CEP-MON
Accuracy (%)
Accuracy (%)
50
0.8
50
Error ratio
Org. CEP-LOG
100
0.6
Accuracy (%)
0
0.4
Error ratio
100
Accuracy (%)
0.2 0.4 0.6 0.8
Accuracy (%)
0
50
Accuracy (%)
50
100
Accuracy (%)
Accuracy (%)
Accuracy (%)
100
100
50
0
0.2 0.4 0.6 0.8
Error ratio
Prep. SIM-G-NN
Figure 10: The performance of RECI in the standardized data sets with respect to
the error ratio (10) as rejection criterion.
40
References
[1] P. Blöbaum, D. Janzing, T. Washio, S. Shimizu, and B. Schölkopf. Causeeffect inference by comparing regression errors. In Proceedings of the 21st
International Conference on Artificial Intelligence and Statistics (AISTATS
2018), 2018.
[2] J. Pearl. Causality: Models, Reasoning and Inference. Cambridge University
Press, New York, NY, USA, 2nd edition, 2009.
[3] P. Spirtes, C. Glymour, and R. Scheines. Causation, prediction, and search.
MIT press, 2000.
[4] J. Peters, D. Janzing, and B. Schölkopf. Elements of Causal Inference –
Foundations and Learning Algorithms. MIT Press, 2017. http://www.
math.ku.dk/~peters/.
[5] Y. Kano and S. Shimizu. Causal inference using nonnormality. In Proceedings of the International Symposium on Science of Modeling, the 30th Anniversary of the Information Criterion, pages 261–270, Tokyo, Japan, 2003.
[6] J. W. Comley and D. L. Dowe. General Bayesian networks and asymmetric
languages. In Proceedings of the Second Hawaii International Conference
on Statistics and Related fields, June 2003.
[7] S. Shimizu, P. Hoyer, A. Hyvärinen, and A. Kerminen. A linear nonGaussian acyclic model for causal discovery. Journal of Machine Learning
Research, 7:2003–2030, 2006.
[8] X. Sun, D. Janzing, and B. Schölkopf. Causal inference by choosing graphs
with most plausible Markov kernels. In Proceedings of the 9th International Symposium on Artificial Intelligence and Mathematics, pages 1–11,
Fort Lauderdale, FL, 2006.
[9] K. Zhang and A. Hyvärinen. On the identifiability of the post-nonlinear
causal model. In Proceedings of the 25th Conference on Uncertainty in
Artificial Intelligence, pages 647–655, Arlington, Virginia, United States,
June 2009. AUAI Press.
[10] P. Hoyer, D. Janzing, J. Mooij, J. Peters, and B. Schölkopf. Nonlinear causal
discovery with additive noise models. In Advances in Neural Information
41
Processing Systems 21, pages 689–696, Red Hook, NY, USA, June 2009.
Curran Associates, Inc.
[11] D. Janzing, X. Sun, and B. Schölkopf. Distinguishing cause and effect via
second order exponential models. eprint http://arxiv.org/abs/0910.
5561, 2009.
[12] J. Peters, D. Janzing, and B. Schölkopf. Causal inference on discrete data
using additive noise models. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 33(12):2436–2450, 2011.
[13] P. Daniušis, D. Janzing, J. Mooij, J. Zscheischler, B. Steudel, K. Zhang, and
B. Schölkopf. Inferring deterministic causal relations. In Proceedings of the
26th Conference on Uncertainty in Artificial Intelligence, pages 143–150,
Corvallis, OR, USA, July 2010. AUAI Press.
[14] D. Janzing, J. Mooij, K. Zhang, J. Lemeire, J. Zscheischler, P. Daniušis,
B. Steudel, and B. Schölkopf. Information-geometric approach to inferring
causal directions. Artificial Intelligence, 182:1–31, May 2012.
[15] E. Sgouritsa, D. Janzing, P. Hennig, and B. Schölkopf. Inference of cause
and effect with unsupervised inverse regression. In Artificial Intelligence
and Statistics, pages 847–855, 2015.
[16] J. Mooij, Peters J., D. Janzing, J. Zscheischler, and B. Schölkopf. Distinguishing cause from effect using observational data: Methods and benchmarks. Journal of Machine Learning Research, 17(32):1–102, January 2016.
[17] D. Janzing and B. Schölkopf. Causal inference using the algorithmic Markov
condition. IEEE Transactions on Information Theory, 56(10):5168–5194,
2010.
[18] J. Lemeire and D. Janzing. Replacing causal faithfulness with algorithmic
independence of conditionals. Minds and Machines, 23(2):227–249, 7 2012.
[19] B. Schölkopf, D. Janzing, J. Peters, E. Sgouritsa, K. Zhang, and J. Mooij.
Semi-supervised learning in causal and anticausal settings, chapter 13,
pages 129–141. Festschrift in Honor of Vladimir Vapnik. Springer-Verlag,
2013.
42
[20] P. Blöbaum, T. Washio, and S. Shimizu. Error asymmetry in causal and
anticausal regression. Behaviormetrika, pages 1–22, 2017.
[21] P. Blöbaum, S. Shimizu, and T. Washio. A novel principle for causal inference in data with small error variance. In ESANN, 2017.
[22] A. Hyvärinen and S. Smith. Pairwise likelihood ratios for estimation of nongaussian structural equation models. Journal of Machine Learning Research,
14(Jan):111–152, 2013.
43
| 2cs.AI
|
Cost-aware Pre-training for Multiclass Cost-sensitive Deep Learning
arXiv:1511.09337v3 [cs.LG] 24 May 2016
Yu-An Chung
Department of CSIE
National Taiwan University
[email protected]
Hsuan-Tien Lin
Department of CSIE
National Taiwan University
[email protected]
Abstract
Deep learning has been one of the most prominent machine learning techniques nowadays, being the state-of-the-art on a broad range of applications where automatic feature extraction is needed.
Many such applications also demand varying costs
for different types of mis-classification errors, but
it is not clear whether or how such cost information
can be incorporated into deep learning to improve
performance. In this work, we first design a novel
loss function that embeds the cost information for
the training stage of cost-sensitive deep learning.
We then show that the loss function can also be integrated into the pre-training stage to conduct costaware feature extraction more effectively. Extensive experimental results justify the validity of the
novel loss function for making existing deep learning models cost-sensitive, and demonstrate that our
proposed model with cost-aware pre-training and
training outperforms non-deep models and other
deep models that digest the cost information in
other stages.
1
Introduction
In many real-world machine learning applications [Tan, 1993;
Chan and Stolfo, 1998; Fan et al., 2000; Zhang and Zhou,
2010; Jan et al., 2011], classification errors may come with
different costs; namely, some types of mis-classification errors may be (much) worse than others. For instance, when
classifying bacteria [Jan et al., 2011], the cost of classifying a Gram-positive species as a Gram-negative one should
be higher than the cost of classifying the species as another
Gram-positive one because of the consequence on treatment
effectiveness. Different costs are also useful for building a
realistic face recognition system, where a government staff
being mis-recognized as an impostor causes only little inconvenience, but an imposer mis-recognized as a staff can result
in serious damage [Zhang and Zhou, 2010]. It is thus important to take into account the de facto cost of every type of
error rather than only measuring the error rate and penalizing
all types of errors equally.
The classification problem that mandates the learning algorithm to consider the cost information is called cost-sensitive
Shao-Wen Yang
Intel Labs
Intel Corporation
[email protected]
classification. Amongst cost-sensitive classification algorithms, the binary classification ones [Elkan, 2001; Zadrozny
et al., 2003] are somewhat mature with re-weighting the
training examples [Zadrozny et al., 2003] being one major approach, while the multiclass classification ones are
continuing to attract research attention [Domingos, 1999;
Margineantu, 2001; Abe et al., 2004; Tu and Lin, 2010].
This work focuses on multiclass cost-sensitive classification, whose algorithms can be grouped into three categories
[Abe et al., 2004]. The first category makes the prediction procedure cost-sensitive [Kukar and Kononenko, 1998;
Domingos, 1999; Zadrozny and Elkan, 2001], generally done
by equipping probabilistic classifiers with Bayes decision theory. The major drawback is that probability estimates can often be inaccurate, which in term makes cost-sensitive performance unsatisfactory. The second category makes the training
procedure cost-sensitive, which is often done by transforming
the training examples according to the cost information [Chan
and Stolfo, 1998; Domingos, 1999; Zadrozny et al., 2003;
Beygelzimer et al., 2005; Langford and Beygelzimer, 2005].
However, the transformation step cannot take the particularities of the underlying classification model into account and
thus sometimes has room for improvement. The third category specifically extends one particular classification model
to be cost-sensitive, such as support vector machine [Tu and
Lin, 2010] or neural network [Kukar and Kononenko, 1998;
Zhou and Liu, 2006]. Given that deep learning stands as an
important class of models with its special properties to be discussed below, we aim to design cost-sensitive deep learning
algorithms within the third category while borrowing ideas
from other categories.
Deep learning models, or neural networks with deep architectures, are gaining increasing research attention in recent
years. Training a deep neural network efficiently and effectively, however, comes with many challenges, and different
models deal with the challenges differently. For instance, conventional fully-connected deep neural networks (DNN) generally initialize the network with an unsupervised pre-training
stage before the actual training stage to avoid being trapped
in a bad local minimal, and the unsupervised pre-training
stage has been successfully carried out by stacked autoencoders [Vincent et al., 2010; Krizhevsky and Hinton, 2011;
Baldi, 2012]. Deep belief networks [Hinton et al., 2006;
Le Roux and Bengio, 2008] shape the network as a gener-
ative model and commonly take restricted Boltzmann machines [Le Roux and Bengio, 2008] for pre-training. Convolutional neural networks (CNN) [LeCun et al., 1998] mimic
the visual perception process of human based on special network structures that result in less need for pre-training, and
are considered the most effective deep learning models in
tasks like image or speech recognition [Ciresan et al., 2011;
Krizhevsky et al., 2012; Abdel-Hamid et al., 2014].
While some existing works have studied cost-sensitive
neural networks [Kukar and Kononenko, 1998; Zhou and
Liu, 2006], none of them have focused on cost-sensitive deep
learning to the best of our knowledge. That is, we are the first
to present cost-sensitive deep learning algorithms, with the
hope of making deep learning more realistic for applications
like bacteria classification and face recognition. In Section 2,
we first formalize the cost-sensitive classification problem
and review related deep learning works. Then, in Section 3,
we start with a baseline algorithm that makes the prediction procedure cost-sensitive (first category). The features
extracted from the training procedure of such an algorithm,
however, are cost-blind. We then initiate a pioneering study
on how the cost information can be digested in the training
procedure (second category) of DNN and CNN. We design
a novel loss function that matches the needs of neural network training while embedding the cost information. Furthermore, we argue that for DNN pre-trained with stacked autoencoders, the cost information should not only be used for the
training stage, but also the pre-training stage. We then propose a novel pre-training approach for DNN (third category)
that mixes unsupervised pre-training with a cost-aware loss
function. Experimental results on deep learning benchmarks
and standard cost-sensitive classification settings in Section 4
verified that the proposed algorithm based on cost-sensitive
training and cost-aware pre-training indeed yields the best
performance, outperforming non-deep models as well as a
broad spectrum of deep models that are either cost-insensitive
or cost-sensitive in other stages. Finally, we conclude in Section 5.
2
Background
We will formalize the multiclass cost-sensitive classification
problem before introducing deep learning and related works.
2.1
Multiclass Cost-sensitive Classification
We first introduce the multiclass classification problem and
then extend it to the cost-sensitive setting. The K-class classification problem comes with a size-N training set S =
{(xn , yn )}N
n=1 , where each input vector xn is within an input space X , and each label yn is within a label space Y =
{1, 2, ..., K}. The goal of multiclass classification is to train a
classifier g : X → Y such that the expected error Jy 6= g(x)K
on test examples (x, y) is small.1
Multiclass cost-sensitive classification extends multiclass
classification by penalizing each type of mis-classification error differently based on some given costs. Specifically, consider a K by K cost matrix C, where each entry C(y, k) ∈
[0, ∞) denotes the cost for predicting a class-y example as
1
J·K is 1 when the inner condition is true, and 0 otherwise.
class k and naturally C(y, y) = 0. The goal of cost-sensitive
classification is to train a classifier g such that the expected
cost C(y, g(x)) on test examples is small.
The cost-matrix setting is also called cost-sensitive classification with class-dependent costs. Another popular setting is
to consider example-dependent costs, which means coupling
an additional cost vector c ∈ [0, ∞)K with each example
(x, y), where the k-th component c[k] denotes the cost for
classifying x as class k. During training, each cn that accompanies (xn , yn ) is also fed to the learning algorithm to train a
classifier g such that the expected cost c[g(x)] is small with
respect to the distribution that generates (x, y, c) tuples. The
cost-matrix setting can be cast as a special case of the costvector setting by defining the cost vector in (x, y, c) as row y
of the cost matrix C. In this work, we will eventually propose
a cost-sensitive deep learning algorithm that works under the
more general cost-vector setting.
2.2
Neural Network and Deep Learning
There are many deep learning models that are successful for
different applications nowadays [Lee et al., 2009; Krizhevsky
and Hinton, 2011; Ciresan et al., 2011; Krizhevsky et al.,
2012; Simonyan and Zisserman, 2014]. In this work, we
first study the fully-connected deep neural network (DNN)
for multiclass classification as a starting point of making
deep learning cost-sensitive. The DNN consists of H hidden layer and parameterizes each layer i ∈ {1, 2, ..., H} by
θi = {Wi , bi }, where Wi is a fully-connected weight matrix and bi is a bias vector that enter the neurons. That is, the
weight matrix and bias vector applied on the input are stored
within θ1 = {W1 , b1 }. For an input feature vector x, the H
hidden layers of the DNN describe a complex feature transform function by computing φ(x) = s(WH · s(· · · s(W2 ·
1
s(W1 · x + b1 ) + b2 ) · · · ) + bH ), where s(z) = 1+exp(−z)
is
the component-wise logistic function. Then, to perform multiclass classification, an extra softmax layer, parameterized
by θsm = {Wsm , bsm }, is placed after the H-th hidden layer.
There are K neurons in the softmax layer, where the j-th neu(j)
(j)
ron comes with weights Wsm and bias bsm and is responsible
for estimating the probability of class j given x:
(j)
(j)
exp(φ(x)T Wsm + bsm )
P (y = j|x) = PK
.
(k)
(k)
T
k=1 exp(φ(x) Wsm + bsm )
(1)
Based on the probability estimates, the classifier trained from
the DNN is naturally g(x) = argmax1≤k≤K P (y = k|x).
Traditionally, the parameters {{θi }H
i=1 , θsm } of the DNN
are optimized by the back-propagation algorithm, which is
essentially gradient descent, with respect to the negative loglikelihood loss function over the training set S:
LNLL (S) =
N
X
− ln(P (y = yn |xn )).
(2)
n=1
The strength of the DNN, through multiple layers of nonlinear transforms, is to extract sophisticated features automatically and implement complex functions. However, the training of the DNN is non-trivial because of non-convex optimization and gradient diffusion problems, which degrade the
test performance of the DNN when adding too many layers.
[Hinton et al., 2006] first proposed a greedy layer-wise pretraining approach to solve the problem. The layer-wise pretraining approach performs a series of feature extraction steps
from the bottom (input layer) to the top (last hidden layer) to
capture higher level representations of original features along
the network propagation.
In this work, we shall improve a classical yet effective
unsupervised pre-training strategy, stacked denoising autoencoders [Vincent et al., 2010], for the DNN. Denoising autoencoder (DAE) is an extension of regular auto-encoder. An
auto-encoder is essentially a (shallow) neural network with
one hidden layer, and consists of two parameter sets: {W, b}
for mapping the (normalized) input vector x ∈ [0, 1]d to the
d0 -dimensional latent representation h by h = s(W·x+b) ∈
0
[0, 1]d ; {W0 , b0 } for reconstructing an input vector x̃ from h
by x̃ = s(W0 · h + b0 ). The auto-encoder is trained by minimizing the total cross-entropy loss LCE (S) over S, defined
as
N X
d
X
−
xn [j] ln x̃n [j] + (1 − xn [j]) ln(1 − x̃n [j]) , (3)
classification neural network, and then make the prediction
stage of the trained network cost-sensitive by including the
costs in the prediction formula; the third approach modifies
the learning rate of the training algorithm base on the costs;
the fourth approach, called MIN (minimization of the misclassification costs), modifies the loss function of neural network training directly. Among the four proposed algorithms,
MIN consistently achieves the lowest test cost [Kukar and
Kononenko, 1998] and will be taken as one of our competitors. Nevertheless, none of the existing works, to the best
of our knowledge, have conducted careful study on costsensitive algorithms for deep neural networks.
3
Cost-sensitive Deep Learning
Before we start describing our proposed algorithm, we highlight a naı̈ve algorithm. For the DNN/SDAE/CNN that estimate the probability with (1), when given the full picture of
the cost matrix, a cost-sensitive prediction can be obtained
using Bayes optimal decision, which computes the expected
cost of classifying an input vector x to each class and predicts
the label that reaches the lowest expected cost:
n=1 j=1
where xn [j] denotes the j-th component of xn and x̃n [j] is
the corresponding reconstructed value.
The DAE extends the regular auto-encoder by randomly
adding noise to inputs xn before mapping to the latent representation, such as randomly setting some components of xn
to 0. Several DAEs can then be stacked to form a deep network, where each layer receives its input from the latent representation of the previous layer. For the DNN, initializing
with stacked DAEs is known to perform better than initializing with stacked regular auto-encoders [Vincent et al., 2010]
or initializing randomly. Below we will refer the DNN initialized with stacked DAEs and trained (fine-tuned) by backpropagation with (2) as the SDAE, while restricting the DNN
to mean the model that is initialized randomly and trained
with (2).
In this work, we will also extend another popular deep
learning model, the convolutional neural network (CNN), for
cost-sensitive classification. The CNN is based on a locallyconnected network structure that mimics the visual perception process [LeCun et al., 1998]. We will consider a standard
CNN structure specified in Caffe2 [Jia et al., 2014], which
generally does not rely on a pre-training stage. Similar to the
DNN, we consider the CNN with a softmax layer for multiclass classification.
2.3
Cost-sensitive Neural Network
Few existing works have studied cost-sensitive classification using neural networks [Kukar and Kononenko, 1998;
Zhou and Liu, 2006]. [Zhou and Liu, 2006] focused on studying the effect of sampling and threshold-moving to tackle
the class imbalance problem using neural network as a core
classifier rather than proposing general cost-sensitive neural network algorithms. [Kukar and Kononenko, 1998] proposed four approaches of modifying neural networks for costsensitivity. The first two approaches train a usual multiclass
2
https://github.com/BVLC/caffe/tree/master/examples/cifar10
g(x) = argmin
K
X
P (y|x)C(y, k).
(4)
16k6K y=1
We will denote these algorithms as DNNBayes , SDAEBayes
and CNNBayes , respectively. These algorithms do not include
the costs in the pre-training nor training stages. Also, those
algorithms require knowing the full cost matrix, and cannot
work under the cost-vector setting.
3.1
Cost-sensitive Training
The DNN essentially decomposes the multiclass classification problem to per-class probability estimation problems
via the well-known one-versus-all (OVA) decomposition. [Tu
and Lin, 2010] proposed the one-sided regression algorithm
that extends OVA for support vector machine (SVM) to a
cost-sensitive SVM by considering per-class regression problems. In particular, if regressors rk (x) ≈ c[k] can be learned
properly, a reasonable prediction can be made by
gr (x) ≡ argmin rk (x).
(5)
16k6K
[Tu and Lin, 2010] further argued that the loss function of the
regressor rk with respect to c[k] should be one-sided. That
is, rk (x) is allowed to underestimate the smallest cost c[y]
and to overestimate other costs. Define zn,k = 2Jcn [k] =
cn [yn ]K − 1 for indicating whether cn [k] is the smallest
within cn . The cost-sensitive SVM [Tu and Lin, 2010] minimizes a regularized version of the total one-sided loss ξn,k =
max(zn,k ·(rk (xn )−cn [k]), 0), where rk are formed by (kernelized) linear models. With such a design, the cost-sensitive
SVM enjoys the following property [Tu and Lin, 2010]:
cn [gr (xn )] 6
K
X
ξn,k .
(6)
k=1
PK
That is, an upper bound k=1 ξn,k of the total cost paid by gr
on xn is minimized within the cost-sensitive SVM.
If we replace the softmax layer of the DNN or the CNN
with regression outputs (using the identity function instead of
the logistic one for outputting), we can follow [Tu and Lin,
2010] to make DNN and CNN cost-sensitive by letting each
output neuron estimate c[k] as rk and predicting with (5).
The training of the cost-sensitive DNN and CNN can also be
done by minimizing the total one-sided loss. Nevertheless, the
one-sided loss is not differentiable at some points, and backpropagation (gradient descent) cannot be directly applied. We
thus derive a smooth approximation of ξn,k instead. Note that
the new loss function should not only approximate ξn,k but
also be an upper bound of ξn,k to keep enjoying the bounding property of (6). [Lee and Mangasarian, 2001] has shown a
smooth approximation u+ α1 ·ln(1+exp(−αu)) ≈ max(u, 0)
when deriving the smooth SVM. Taking α = 1 leads to
LHS = ln(1 + exp(u)), which is trivially an upper bound of
max(u, 0) because ln(1+exp(u)) > u, and ln(1+exp(u)) >
ln(1) = 0. Based on the approximation, we define
δn,k ≡ ln(1 + exp(zn,k · (rk (xn ) − cn [k]))).
(7)
δn,k is not only a smooth approximation of ξn,k that enjoys
the differentiable property, but also an upper bound of ξn,k to
keep the bounding property of (6) held. That is, we can still
ensure a small total cost by minimizing the newly defined
smooth one-sided regression (SOSR) loss over all examples:
LSOSR (S) =
N X
K
X
δn,k .
LCAE (S) = (1 − β) · LCE (S) + β · LSOSR (S)
(9)
The mixture step is a widely-used technique for multi-criteria
optimization [Hillermeier, 2001], where β controls the balance between reconstructing the original input x and the cost
vector c. A positive β makes CAE cost-aware during its feature extraction, while a zero β makes CAE degenerate to
DAE. Similar to DAEs, CAEs can then be stacked to initialize a deep neural network before the weights are fine-tuned by
back-propagation with (8). The resulting algorithm is named
CSDNN, which is cost-sensitive in both the pre-training stage
(by CAE) and the training stage (by (8)), and can work under
the general cost-vector setting. The full algorithm is listed in
Algorithm 1.
Algorithm 1 CSDNN
Input: Cost-sensitive training set S = {(xn , yn , cn )}N
n=1
1: for each hidden layer θi = {Wi , bi } do
2:
Learn a CAE by minimizing (9).
3:
Take {Wi , bi } of CAE as θi .
4: end for
5: Fine-tune the network parameters {{θi }H
i=1 , θSOSR } by
minimizing (8) using back-propagation.
Output: The fine-tuned deep neural network with (5) as gr .
(8)
n=1 k=1
We will refer to these algorithms, which replace the softmax layer of the DNN/SDAE/CNN with a regression layer
parameterized by θSOSR = {WSOSR , bSOSR } and minimize (8) with back-propagation, as DNNSOSR , SDAESOSR
and CNNSOSR . These algorithms work with the cost-vector
setting. They include costs in the training stage, but not the
pre-training stage.
3.2
balancing coefficient β ∈ [0, 1], yielding the following loss
function for CAE over S:
Cost-aware Pre-training
For multiclass classification, the pre-training stage, either in a
totally unsupervised or partially supervised manner [Bengio
et al., 2007], has been shown to improve the performance of
the DNN and several other deep models [Bengio et al., 2007;
Hinton et al., 2006; Erhan et al., 2010]. The reason is that pretraining usually helps initialize a neural network with better
weights that prevent the network from getting stuck in poor
local minima. In this section, we propose a cost-aware pretraining approach that leads to a novel cost-sensitive deep
neural network (CSDNN) algorithm.
CSDNN is designed as an extension of SDAESOSR . Instead of pre-training with SDAE, CSDNN takes stacked costsensitive auto-encoders (CAE) for pre-training instead. For
a given cost-sensitive example (x, y, c), CAE tries not only
to denoise and reconstruct the original input x like DAE, but
also to digest the cost information by reconstructing the cost
vector c. That is, in addition to {W, b} and {W0 , b0 } for
DAE, CAE introduces an extra parameter set {W00 , b00 } fed
to regression neurons from the hidden representation. Then,
we can mix the two loss functions LCE and LSOSR with a
CSDNN is essentially SDAESOSR with DAEs replaced by
CAEs with the hope of more effective cost-aware feature extraction. We can also consider SCAEBayes which does the
same for SDAEBayes . The CNN, due to its special network
structure, generally does not rely on stacked DAEs for pretraining, and hence cannot be extended by stacked CAEs.
As discussed, DAE is a degenerate case of CAE. Another
possible degeneration is to consider CAE with less complete
cost information. For instance, a naı̈ve cost vector defined by
ĉn [k] = Jyn 6= kK encodes the label information (whether
the prediction is erroneous with respect to the demanded label) but not the complete cost information. To study whether
it is necessary to take the complete cost information into
account in CAE, we design two variant algorithms that replace the cost vectors in CAEs with ĉn [k], which effectively
makes those CAEs error-aware. Then, SCAEBayes becomes
SEAEBayes (with E standing for error); CSDNN becomes
SEAESOSR .
4
Experiments
In the previous section, we have derived many cost-sensitive
deep learning algorithms, each with its own specialty. They
can be grouped into two series: those minimizing with (2) and
predicting with (4) are Bayes-series algorithms (DNNBayes ,
SDAEBayes , SEAEBayes , SCAEBayes , and CNNBayes );
those minimizing with (8) and predicting with (5) are SOSRseries algorithms (DNNSOSR , SDAESOSR , SEAESOSR ,
CSDNN ≡ SCAESOSR , CNNSOSR ). Note that the Bayesseries can only be applied to the cost-matrix setting while the
SOSR-series can deal with the cost-vector setting. The two
series help understand whether it is beneficial to consider the
cost information in the training stage.
Within each series, CNN is based on a locally-connected
structure, while DNN, SDAE, SEAE and SCAE are fullyconnected and differ by how pre-training is conducted, ranging from none, unsupervised, error-aware, to cost-aware. The
wide range helps understand the effectiveness of digesting the
cost information in the pre-training stage.
Next, the two series will be compared with the blind-series
algorithms (DNNblind , SDAEblind , and CNNblind ), which
are the existing algorithms that do not incorporate the cost information at all, to understand the importance of taking the
cost information into account. The two series will also be
compared against two baseline algorithms: CSOSR [Tu and
Lin, 2010], a non-deep algorithm that our proposed SOSRseries originates from; MIN [Kukar and Kononenko, 1998], a
neural-network algorithm that is cost-sensitive in the training
stage like the SOSR-series but with a different loss function.
The algorithms along with highlights on where the cost information is digested are summarized in Table 1.
4.1
Setup
We conducted experiments on MNIST, bg-img-rot (the hardest variant of MNIST provided in [Larochelle et al., 2007]),
SVHN [Netzer et al., 2011], and CIFAR-10 [Krizhevsky and
Hinton, 2009]. The first three datasets belong to handwritten
digit recognition and aim to classify each image into a digit
of 0 to 9 correctly; CIFAR-10 is a well-known image recognition dataset which contains 10 classes such as car, ship and
animal. For all four datasets, the training, validation, and testing split follows the source websites; the input vectors in the
training set are linearly scaled to [0, 1], and the input vectors
in the validation and testing sets are scaled accordingly.
The four datasets are originally collected for multiclass
classification and contain no cost information. We adopt the
most frequently-used benchmark in cost-sensitive learning,
the randomized proportional setup [Abe et al., 2004], to generate the costs. The setup is for the cost-matrix setting. It
first generates a K × K matrix C, and sets the diagonal entries C(y, y) to 0 while sampling the non-diagonal entries
n =k}|
C(y, k) uniformly from [0, 10 |{n:y
|{n:yn =y}| ]. The randomized
proportional setup generates the cost information that takes
the class distribution of the dataset into account, charging
a higher cost (in expectation) for mis-classifying a minority
class, and can thus be used to deal with imbalanced classification problems. Note that we take this benchmark cost-matrix
setting to give prediction-stage cost-sensitive algorithms like
the Bayes-series a fair chance of comparison. We find that the
range of the costs can affect the numerical stability of the algorithms, and hence scale all the costs by the maximum value
within C during training in our implementation. The reported
test results are based on the unscaled C.
Arguably one of the most important use of cost-sensitive
classification is to deal with imbalanced datasets. Nevertheless, the four datasets above are somewhat balanced, and the
randomized proportional setup may generate similar cost for
each type of mis-classification error. To better meet the realworld usage scenario, we further conducted experiments to
evaluate the algorithms with imbalanced datasets. In particular, for each dataset, we construct a variant dataset by randomly picking four classes and removing 70% of the examples that belong to those four classes. We will name these imbalanced variants as MNISTimb , bg-img-rotimb , SVHNimb ,
and CIFAR-10imb , respectively.
All experiments were conducted using Theano. For algorithms related to DNN and SDAE, we selected the hyperparameters by following [Vincent et al., 2010]. The β in (9),
needed by SEAE and SCAE algorithms, was selected among
{0, 0.05, 0.1, 0.25, 0.4, 0.75, 1}. As mentioned, for CNN, we
considered a standard structure in Caffe [Jia et al., 2014].
4.2
Experimental Results
The average test cost of each algorithm along with the standard error is shown in Table 2. The best result3 per dataset
among all algorithms is highlighted in bold.
Is it necessary to consider costs? DNNblind and SDAEblind
performed the worst on almost all the datasets. While
CNNblind was slightly better than those two, it never reached
the best performance for any dataset. The results indicate the
necessity of taking the cost information into account.
Is it necessary to go deep? The two existing cost-sensitive
baselines, CSOSR and MIN, outperformed the cost-blind
algorithms often, but were usually not competitive to costsensitive deep learning algorithms. The results validate the
importance of studying cost-sensitive deep learning.
Is it necessary to incorporate costs during training?
SOSR-series models, especially under the imbalanced scenario, generally outperformed their Bayes counterparts. The
results demonstrate the usefulness of the proposed (7) and (8)
and the importance of incorporating the cost information during the training stage.
Is it necessary to incorporate costs during pre-training?
CSDNN outperformed both SEAESOSR and SDAESOSR ,
and SDAESOSR further outperformed DNNSOSR . The results show that for the fully-connected structure where pretraining is needed, our newly proposed cost-aware pretraining with CAE is indeed helpful in making deep learning
cost-sensitive.
Which is better, CNNSOSR or CSDNN? The last two
columns in Table 2 show that CSDNN is competitive to
CNNSOSR , with both algorithms usually achieving the best
performance. CSDNN is slightly better on two datasets. Note
that CNNs are known to be powerful for image recognition
tasks, which match the datasets that we have used. Hence,
it is not surprising that CNN can reach promising performance with our proposed SOSR loss (8). Our efforts not only
make CNN cost-sensitive, but also result in the CSDNN algorithm that makes the full-connected deep neural network costsensitive with the help of cost-aware pre-training via CAE.
Is the mixture loss necessary? To have more insights on
CAE, we also conducted experiments to evaluate the perfor3
The selected CSDNN that achieved the test cost listed in Table 2 is composed of 3 hidden layers, and each hidden layer consists
of 3000 neurons.
Table 1: cost-awareness of algorithms (O: cost-aware; E: error-aware; X: cost-blind)
aa
Algorithm
aa
DNNblind
aa
Stage
aa
pre-training
training
prediction
SDAEblind
CNNblind
CSOSR
MIN
DNNBayes
SDAEBayes
SEAEBayes
SCAEBayes
CNNBayes
DNNSOSR
SDAESOSR
SEAESOSR
CSDNN
CNNSOSR
X
X
X
none
X
X
none
O
X
none
O
X
none
X
O
X
X
O
E
X
O
O
X
O
none
X
O
none
O
X
X
O
X
E
O
X
O
O
X
none
O
X
none
X
X
Table 2: Average test cost
aa
Algorithm
aa
aa
Dataset
aa
MNIST
bg-img-rot
SVHN
CIFAR-10
MNISTimb
bg-img-rotimb
SVHNimb
CIFAR-10imb
DNNblind
SDAEblind
CNNblind
CSOSR
MIN
DNNBayes
SDAEBayes
SEAEBayes
SCAEBayes
CNNBayes
DNNSOSR
SDAESOSR
SEAESOSR
CSDNN
CNNSOSR
0.11 ± 0.00
3.33 ± 0.06
1.58 ± 0.03
3.46 ± 0.04
0.32 ± 0.01
15.9 ± 0.70
1.79 ± 0.01
19.1 ± 0.09
0.10 ± 0.00
3.28 ± 0.07
1.40 ± 0.03
3.26 ± 0.05
0.31 ± 0.01
13.8 ± 0.70
1.60 ± 0.01
17.7 ± 0.09
0.10 ± 0.00
3.05 ± 0.07
0.91 ± 0.03
2.51 ± 0.04
0.19 ± 0.01
5.04 ± 0.67
0.31 ± 0.01
7.29 ± 0.08
0.10 ± 0.00
3.25 ± 0.06
1.17 ± 0.03
3.30 ± 0.04
0.26 ± 0.01
8.55 ± 0.70
1.05 ± 0.01
10.1 ± 0.09
0.10 ± 0.003
3.02 ± 0.06
1.19 ± 0.03
3.19 ± 0.05
0.27 ± 0.01
8.40 ± 0.69
0.99 ± 0.01
11.2 ± 0.09
0.10 ± 0.00
2.95 ± 0.07
1.07 ± 0.03
2.80 ± 0.05
0.23 ± 0.01
7.19 ± 0.69
0.53 ± 0.01
8.16 ± 0.09
0.09 ± 0.00
2.66 ± 0.07
0.93 ± 0.03
2.52 ± 0.05
0.20 ± 0.01
5.10 ± 0.70
0.33 ± 0.01
7.48 ± 0.09
0.09 ± 0.00
2.85 ± 0.07
0.94 ± 0.03
2.68 ± 0.05
0.20 ± 0.01
4.95 ± 0.70
0.34 ± 0.01
7.25 ± 0.08
0.09 ± 0.00
2.54 ± 0.07
0.88 ± 0.03
2.38 ± 0.04
0.18 ± 0.01
4.73 ± 0.70
0.29 ± 0.01
6.97 ± 0.09
0.09 ± 0.00
2.40 ± 0.07
0.85 ± 0.03
2.34 ± 0.05
0.18 ± 0.01
4.49 ± 0.68
0.28 ± 0.01
6.81 ± 0.09
0.10 ± 0.00
3.21 ± 0.07
1.02 ± 0.03
2.74 ± 0.05
0.22 ± 0.01
6.89 ± 0.70
0.51 ± 0.01
7.86 ± 0.08
0.09 ± 0.00
2.99 ± 0.07
0.92 ± 0.03
2.48 ± 0.04
0.20 ± 0.01
4.99 ± 0.69
0.31 ± 0.01
7.44 ± 0.09
0.09 ± 0.00
3.00 ± 0.07
0.99 ± 0.03
2.52 ± 0.05
0.19 ± 0.01
4.86 ± 0.69
0.31 ± 0.01
7.14 ± 0.09
0.09 ± 0.00
2.34 ± 0.07
0.83 ± 0.03
2.24 ± 0.05
0.17 ± 0.01
4.16 ± 0.68
0.26 ± 0.01
6.48 ± 0.09
0.08 ± 0.00
2.29 ± 0.07
0.82 ± 0.03
2.25 ± 0.04
0.17 ± 0.01
4.39 ± 0.69
0.28 ± 0.01
6.63 ± 0.08
MNIST
bg−img−rot
0.11
3.4
0.105
3.2
0.1
3
0.095
2.8
0.09
2.6
0.085
0.08
2.4
0
0.2
0.4
0.6
0.8
1
2.2
0
0.2
0.4
0.6
0.8
1
0.8
1
CIFAR−10
SVHN
2.6
1
2.5
0.95
2.4
0.9
2.3
0.85
series, the Bayes-series and the SOSR-series were extensively compared with CSDNN carefully to validate the importance of each component of CSDNN. The experimental results demonstrate that incorporating the cost information into
both the pre-training and the training stages leads to promising performance of CSDNN, outperforming those baselines
and alternatives. One key component of CSDNN, namely the
SOSR loss for cost-sensitivity in the training stage, is shown
to be helpful in improving the performance of CNN. The results justify the importance of the proposed SOSR loss for
training and the CAE approach for pre-training.
2.2
0.8
0
0.2
0.4
0.6
0.8
0
1
MNISTimb
0.2
0.4
0.6
bg−img−rotimb
0.24
5.2
5
0.22
4.8
0.2
4.6
4.4
0.18
4.2
0.16
0
0.2
0.4
0.6
0.8
1
4
0
0.4
0.6
0.8
1
7.6
0.34
7.2
0.3
7
0.28
6.8
0.26
6.6
We thank the anonymous reviewers for valuable suggestions.
This material is based upon work supported by the Air Force
Office of Scientific Research, Asian Office of Aerospace
Research and Development (AOARD) under award number
FA2386-15-1-4012, and by the Ministry of Science and Technology of Taiwan under number MOST 103-2221-E-002149-MY3.
References
6.4
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
Figure 1: Relation between β and test cost (note that
SDAESOSR is the data point with β = 0).
mance of CSDNN for β ∈ [0, 1]. When β = 0, CSDNN degenerates to SDAESOSR ; when β = 1, each CAE of CSDNN
performs fully cost-aware pre-training to fit the cost vectors.
The results are displayed in Figure 1, showing a roughly Ushaped curve. The curve implies that some β ∈ [0, 1] that best
balances the tradeoff between denoising and cost-awareness
can be helpful. The results validate the usefulness of the proposed mixture loss (9) for pre-training.
5
Acknowledgement
7.4
0.32
0.24
0.2
CIFAR−10imb
SVHNimb
6
Conclusion
We proposed a novel deep learning algorithm CSDNN for
multiclass cost-sensitive classification with deep learning.
Existing baselines and other alternatives within the blind-
[Abdel-Hamid et al., 2014] Ossama Abdel-Hamid, Abdelrahman Mohamed, Hui Jiang, Li Deng, Gerald Penn, and
Dong Yu. Convolutional neural networks for speech
recognition. Audio, Speech, and Language Processing,
IEEE/ACM Transactions on, 22(10):1533–1545, 2014.
[Abe et al., 2004] Naoki Abe, Bianca Zadrozny, and John
Langford. An iterative method for multi-class costsensitive learning. In KDD, 2004.
[Baldi, 2012] Pierre Baldi. Autoencoders, unsupervised
learning, and deep architectures. Unsupervised and Transfer Learning Challenges in Machine Learning, 2012.
[Bengio et al., 2007] Yoshua Bengio, Pascal Lamblin, Dan
Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. In NIPS, 2007.
[Beygelzimer et al., 2005] Alina Beygelzimer, Varsha Dani,
Tom Hayes, John Langford, and Bianca Zadrozny. Error
limiting reductions between classification tasks. In ICML,
2005.
[Chan and Stolfo, 1998] Philip K. Chan and Salvatore J.
Stolfo. Toward scalable learning with non-uniform class
and cost distributions: A case study in credit card fraud
detection. In KDD, 1998.
[Ciresan et al., 2011] Dan C Ciresan, Ueli Meier, Jonathan
Masci, and Jürgen Schmidhuber. Flexible, high performance convolutional neural networks for image classification. In IJCAI, 2011.
[Domingos, 1999] Pedro Domingos. Metacost: A general
method for making classifiers cost-sensitive. In KDD,
1999.
[Elkan, 2001] Charles Elkan. The foundations of costsensitive learning. In IJCAI, 2001.
[Erhan et al., 2010] Dumitru Erhan, Yoshua Bengio, Aaron
Courville, Pierre-Antoine Manzagol, Pascal Vincent, and
Samy Bengio. Why does unsupervised pre-training help
deep learning? JMLR, 11:625–660, 2010.
[Fan et al., 2000] Wei Fan, Wenke Lee, Salvatore J. Stolfo,
and Matthew Miller. A multiple model cost-sensitive approach for intrusion detection. In ECML, 2000.
[Hillermeier, 2001] Claus Hillermeier. Nonlinear multiobjective optimization. Birkhauser, 2001.
[Hinton et al., 2006] Geoffrey E. Hinton, Simon Osindero,
and Yee-Whye Teh. A fast learning algorithm for deep
belief nets. Neural Computation, 18:1527–1554, 2006.
[Jan et al., 2011] Te-Kang Jan, Hsuan-Tien Lin, Hsin-Pai
Chen, Tsung-Chen Chern, Chung-Yueh Huang, BingCheng Wen, Chia-Wen Chung, Yung-Jui Li, Ya-Ching
Chuang, Li-Li Li, Yu-Jiun Chan, Juen-Kai Wang, Yuh-Lin
Wang, Chi-Hung Lin, and Da-Wei Wang. Cost-sensitive
classification on pathogen species of bacterial meningitis
by Surface Enhanced Raman Scattering. In BIBM, 2011.
[Jia et al., 2014] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional
architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
[Krizhevsky and Hinton, 2009] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master’s thesis, Department of Computer Science,
University of Toronto, 2009.
[Krizhevsky and Hinton, 2011] Alex Krizhevsky and Geoffrey E. Hinton. Using very deep autoencoders for contentbased image retrieval. In ESANN, 2011.
[Krizhevsky et al., 2012] Alex Krizhevsky, Ilya Sutskever,
and Geoffrey E. Hinton. Imagenet classification with deep
convolutional neural networks. In NIPS, 2012.
[Kukar and Kononenko, 1998] Matjaz Kukar and Igor
Kononenko. Cost-sensitive learning with neural networks.
In ECAI, 1998.
[Langford and Beygelzimer, 2005] John Langford and Alina
Beygelzimer. Sensitive error correcting output codes. In
COLT, 2005.
[Larochelle et al., 2007] Hugo Larochelle, Dumitru Erhan,
Aaron Courville, James Bergstra, and Yoshua Bengio. An
empirical evaluation of deep architectures on problems
with many factors of variation. In ICML, 2007.
[Le Roux and Bengio, 2008] Nicolas Le Roux and Yoshua
Bengio. Representational power of restricted boltzmann
machines and deep belief networks. Neural Computation,
20:1631–1649, 2008.
[LeCun et al., 1998] Yann LeCun, Léon Bottou, Yoshua
Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE,
86:2278–2324, 1998.
[Lee and Mangasarian, 2001] Yuh-Jye Lee and O. L. Mangasarian. SSVM: A smooth support vector machine. Computational Optimization and Applications, 20:5–22, 2001.
[Lee et al., 2009] Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y. Ng. Convolutional deep belief networks for scalable unsupervised learning of hierarchical
representations. In ICML, 2009.
[Margineantu, 2001] Dragos D. Margineantu. Methods for
cost-sensitive learning. In IJCAI, 2001.
[Netzer et al., 2011] Yuval Netzer, Tao Wang, Adam Coates,
Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading
digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised
feature learning, 2011.
[Simonyan and Zisserman, 2014] Karen Simonyan and Andrew Zisserman.
Very deep convolutional networks
for large-scale image recognition.
arXiv preprint
arXiv:1409.1556, 2014.
[Tan, 1993] Ming Tan. Cost-sensitive learning of classification knowledge and its applications in robotics. Machine
Learning, 13:7–33, 1993.
[Tu and Lin, 2010] Han-Hsing Tu and Hsuan-Tien Lin. Onesided support vector regression for multiclass costsensitive classification. In ICML, 2010.
[Vincent et al., 2010] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful
representations in a deep network with a local denoising
criterion. JMLR, 11:3371–3408, 2010.
[Zadrozny and Elkan, 2001] Bianca Zadrozny and Charles
Elkan. Learning and making decisions when costs and
probabilities are both unknown. In KDD, 2001.
[Zadrozny et al., 2003] Bianca Zadrozny, John Langford,
and Naoki Abe.
Cost-sensitive learning by costproportionate example weighting. In ICDM, 2003.
[Zhang and Zhou, 2010] Yin Zhang and Zhi-Hua Zhou.
Cost-sensitive face recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 32:1758–1769,
2010.
[Zhou and Liu, 2006] Zhi-Hua Zhou and Xu-Ying Liu.
Training cost-sensitive neural networks with methods addressing the class imbalance problem. Knowledge and
Data Engineering, IEEE Transactions on, 18:63–77, 2006.
| 9cs.NE
|
Logical Methods in Computer Science
Vol. 4 (1:2) 2008, pp. 1–24
www.lmcs-online.org
λ-RBAC:
Submitted
Published
Nov. 21, 2006
Jan. 9, 2008
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
RADHA JAGADEESAN a , ALAN JEFFREY b , CORIN PITCHER c , AND JAMES RIELY d
a,c,d
b
CTI, DePaul University
e-mail address : {rjagadeesan, pit her,jriely} ti.depaul.edu
Bell Labs
e-mail address : ajereybell-labs. om
Abstra t. We study me hanisms that permit program omponents to express role on-
straints on lients, fo using on programmati se urity me hanisms, whi h permit a ess
ontrols to be expressed,
, as part of the ode realizing basi fun tionality. In this
setting, two questions immediately arise. (1) The user of a omponent fa es the issue of
safety: is a parti ular role su ient to use the omponent? (2) The omponent designer
fa es the dual issue of prote tion: is a parti ular role demanded in all exe ution paths of
the omponent? We provide a formal al ulus and stati analysis to answer both questions.
in situ
1.
Introdu tion
This paper addresses programmati
se urity me hanisms as realized in systems su h as
jaas
Java Authenti ation and Authorization Servi e (
) and
1
forms of a ess ontrol me hanisms . First, they permit
.net.
de larative
These systems enable two
a
ess
ontrol to des ribe
se urity spe i ations that are orthogonal and separate from des riptions of fun tionality,
e.g., in an interfa e
I,
a de larative a
possess a minimum set of rights.
dire tly permit the enfor ement of a
of the
ess
While
ess
ontrol me hanism
ould require the
ontrol that is sensitive to the
ode implementing the fun tionality
ontrol and data ow
onsider for example history sensitive se urity
poli ies that require runtime monitoring of relevant events. Consequently,
also in lude
programmati
with fun tionality
me hanisms that permit a
ode, e.g., in the
one hand, su h programmati
a,c
d
ode of a
ess
ontrol
jaas
.net
I.
On the
ess
ontrol
omponent implementing interfa e
approa h leads to the
and
ode to be intertwined
me hanisms permit the dire t expression of a
poli ies. However, the programmati
separate
aller to
on eptually elegant, su h spe i ations do not
ommingling of the
on eptually
on erns of se urity and fun tionality.
1998 ACM Subje t Classi ation: D.3, K.6.5.
Key words and phrases: role-based a ess ontrol, lambda- al ulus, stati
analysis.
Radha Jagadeesan and Corin Pit her were supported in part by NSF CyberTrust 0430175.
James Riely was supported in part by NSF CAREER 0347542.
1In this paper, we dis uss only authorization me hanisms, ignoring the authenti ation me hanisms that
are also part of these infrastru tures.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
©
DOI:10.2168/LMCS-4 (1:2) 2008
CC
R. Jagadeesan, A. Jeffrey, C. Pitcher, and J. Riely
Creative Commons
2
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
There is extensive literature on poli y languages to spe ify and implement poli ies
(e.g., [16, 28, 15, 5, 29, 13℄ to name but a few).
This resear h studies se urity poli ies
as separate and orthogonal additions to
ode, and is thus fo used on de larative
se urity in the parlan e of
In
jaas/.net.
ontrast, we study programmati
the se urity guarantees provided by a
omponent
•
The
ontrol
extra t
ode whi h has been written inline with
omponent fa es the issue of safety: is a parti ular set of rights su ient to
omponent? (ie. with that set of rights, there is no possible exe ution path that
would fail a se urity
use the
se urity me hanisms. Our motivation is to
ess
ode. We address this issue from two viewpoints:
The user of a
use the
•
omponent
he k. Furthermore, any greater set of rights will also be allowed to
omponent)
omponent designer fa es the dual issue of prote tion: is a parti ular set of rights
demanded in all exe ution paths of the
that set of rights.
omponent?
(ie. every exe ution path requires
Furthermore, any lesser set of rights will not be allowed to use the
omponent)
The main
ontribution of this paper is separate stati
to these two questions.
analyses to
al ulate approximations
An approximate answer to the rst question is a set of rights,
perhaps bigger than ne essary, that is
su ient
to use the
omponent. On the other hand,
an approximate answer to the se ond question, is a set of rights, perhaps smaller than what
is a tually enfor ed, that is
1.1.
ne essary
to use the
omponent.
An overview of our te hni al ontributions.
Based A
ess-Control (
rba
a textbook survey. The main motivation for
and frameworks su h as
There is extensive literature on Role-
nist standards for rba
rba , in software ar hite
) models in luding
jaas/.net,
[26, 12℄; see [11℄ for
tures (e.g., [22, 21℄)
is that it enables the enfor ement of se urity poli ies
at a granularity demanded by the appli ation. In these examples,
rba
allows permissions
to be de- oupled from users: Roles are the unit of administration for users and permissions
are assigned to roles.
Roles are often arranged in a hierar hy for su
of the mapping of permissions. Component programmers design
in t representation
ode in terms of a stati
olle tion of roles. When the appli ation is deployed, administrators map the roles de ned
in the appli ation to users in the parti ular domain.
In this paper, we study a lambda
dubbed
λ-RBAC.
al ulus enri hed with primitives for a
The underlying lambda
programming framework in a real system.
idioms in
jaas
and
.net,
avor of the basi
2
examples ,
In the
This
.net Framework lr, every thread has a Prin ipal obje
Prin ipal
obje t
obje t belongs to a familiar role. The
Prin ipalPermission
spe i
ontrol
obje t. The
t that
an be viewed as representing the user exe uting
the thread. In programming, it often needs to be determined whether a spe i
a
ess
losely based on [18℄, we give the reader a
programming idioms.
Example 1 ( [18℄).
arries its role.
.net
ontrol,
We draw inspiration from the programming
to determine the expressiveness required for the a
me hanisms. In a sequen e of
ess
al ulus serves as an abstra tion of the ambient
ode performs
Prin ipal
he ks by making a se urity
Prin ipalPermission
prin ipal needs to mat h. At the time of a se urity
all for
lass denotes the role that a
he k, the
lr
he ks whether
2
In order to minimize the synta ti barrage on the unsuspe ting reader, our examples to illustrate the
features are drawn solely from the .net programming domain. At the level of our dis ussion, there are no
real distin tions between jaas and .net se urity servi es.
λ-RBAC:
the role of the
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
Prin ipal obje
t of the
aller mat hes the role of the
3
Prin ipalPermission
obje t being requested. If the role values of the two obje ts do not mat h, an ex eption is
raised. The following
ode snippet illustrates the issues:
Prin ipalPermission usrPerm =
new Prin ipalPermission (null,"Manager");
usrPerm.Demand()
If the
urrent thread is asso iated with a prin ipal that has the the role of manager, the
Prin ipalPermission
obje ts are
reated and se urity a
ess is given as required. If the
redentials are not valid, a se urity ex eption is raised.
In this vein, the intuitive operation of
takes pla e in the
permissions.
ontext of a role, say
λ-RBAC is
r , whi h
as follows.
reation of roles.
λ-RBAC
supports run-time operations to
fa ilitated by operations that
rba
r
is at least as strong as the
he k fails.
The next example illustrates that boolean
lassi al
reate obje ts (i.e. higher-order
The use of su h guarded obje ts is
he k that the role- ontext
guarding role: an ex eption is raised if the
In
program exe ution
on retely as a set of
The set of roles used in a program is stati : we do not allow the dynami
fun tions) that are wrapped with prote ting roles.
grams.
λ-RBAC
an be viewed
ombinations of roles are permitted in pro-
terms, this is abstra ted by a latti e or boolean stru ture on
roles.
Example 2 ( [18℄). The Union method of the Prin ipalPermission
lass
ombines multiple
Prin ipalPermission obje ts. The following ode represents a se urity he k that su
only if the Prin ipal obje t represents a user in the CourseAdmin or BudgetManager
Prin ipalPermission Perm1 =
new Prin ipalPermission (null,"CourseAdmin");
Prin ipalPermission Perm2 =
new Prin ipalPermission(null,"BudgetManager');
// Demand at least one of the roles using Union
perm1.Union (perm2).Demand ()
Similarly, there is an Interse t method to represent a join
In
λ-RBAC,
role ordering is given by supersets, ie.
permissions; join of roles
roles
operation in the role latti e.
we assume that roles form a latti e: abstra ting the
se tion operations of these examples. In the
on rete union/inter-
on rete view of a role as a set of permissions,
a role is stronger than another role if it has more
orresponds to the union of the sets of permissions and meet of
orresponds to the interse tion of the sets of permissions. Some of our results assume
that the latti e is boolean, i.e. the latti e has a negation operation. In the
of the motivating examples, the negation operation is interpreted by set
respe t to a maximum
Our study is parametri
viewpoint, it is
on rete view
omplement with
olle tion of permissions
on the underlying role latti e.
The key operation in su h programming is
rights modulation.
From a programming
onvenient, indeed sometimes required, for an appli ation to operate under
the guise of dierent users at dierent times.
Rights modulation of
avors: rights weakening is overall a safe operation, sin e the
fewer rights. On the other hand, rights ampli ation is
In the
eeds
roles:
.net framework,
aller
ourse
omes in two
hooses to exe ute with
learly a more dangerous operation.
rights modulation is a hieved via a te hnique
alled impersonation.
4
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
Example 3.
Impersonation of an a
in the following
ount is a hieved using the a
ount's token, as shown
ode snippet:
WindowsIdentity stIdentity = new WindowsIdentity (StToken);
// StToken is the token asso iated with the Windows a t being impersonated
WindowsImpersonationContext stImp = stIdentity.Impersonate();
// now operating under the new identity
stImp.Undo(); // revert ba k
λ-RBAC
has
ombinators to perform s oped rights weakening and ampli ation.
We demonstrate the expressiveness of
λ-RBAC
by building a range of useful
ombinators
and a variety of small illustrative examples. We dis uss type systems to perform the two
analyses alluded to earlier: (a) an analysis to dete t and remove unne essary role- he ks in
a pie e of
ode for a
aller at a su iently high role, and (b) an analysis to determine the
(maximal) role that is guaranteed to be required by a pie e of
ode. The latter analysis a -
quires parti ular value in the presen e of rights modulation. For both we prove preservation
and progress properties.
Related work.
1.2.
Our paper falls into the broad area of resear h enlarging the s ope of
foundational, language-based se urity methods (see [27, 19, 3℄ for surveys).
Our work is
lose in spirit, if not in te hni al development, to edit automata [16℄, whi h
use aspe ts to avoid the expli it intermingling of se urity and baseline
The papers that are most dire tly relevant to the
ode.
urrent paper are those of Braghin,
Gorla and Sassone [7℄ and Compagnoni, Garalda and Gunter [10℄. [7℄ presents the rst
urrent
al ulus with a notion of
rba
on-
, whereas [10℄'s language enables privileges depending
upon lo ation.
Both these papers start o with a mobile pro ess-based
omputational model. Both
al-
uli have primitives to a tivate and dea tivate roles: these roles are used to prevent undesired
mobility and/or
ommuni ation, and are similar to the primitives for role restri tion and
ampli ation in this paper. The ambient pro ess
vides a dire t representation of the sessions of
al ulus framework of these papers pro-
rba
in
ontrast, our sequential
al ulus
is best thought of as modeling a single session.
[7, 10℄ develop type systems to provide guarantees about the minimal role required for
exe ution to be su
stati
essful our rst type system o
upies the same
analysis. However, our se ond type system that
on eptual spa e as this
al ulates minimum a
ess
ontrols
does not seem to have an analogue in these papers.
More globally, our paper has been in uen ed by the desire to serve loosely as a metalanguage for programming
rba
me hanisms in examples su h as the
Thus, our treatment internalizes rights ampli ation by program
plify role
onstru tor in role latti es.
In
jaas/.net
frameworks.
ombinators and the am-
ontrast, the above papers use external i.e.
not part of the pro ess language me hanisms (namely, user poli ies in [10℄, and
s hemes in [7℄) to enfor e
adapted to the pro ess
ontrol on rights a tivation.
al uli framework.
We expe t that our ideas
rba
-
an be
In future work, we also hope to integrate the
powerful bisimulation prin iples of these papers.
Our paper deals with a
ess
ontrol, so the extensive work on information ow, e.g.,
see [24℄ for a survey, is not dire tly relevant.
plays the same role in
λ-RBAC
However, we note that rights ampli ation
that de lassi ation and delimited release [9, 25, 20℄ plays in
λ-RBAC:
the
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
ontext of information ow; namely that of permitting a
5
ess that would not have been
possible otherwise. In addition, by supporting the internalizing of the ability to amplify
rights into the role latti e, our system permits a
ess
ontrol
ode
ode to a tively parti ipate in
managing rights ampli ation.
1.3.
Rest of the paper.
We present the language in Se tion 2, the type system in Se -
tion 3 and illustrate its expressiveness with examples in Se tion 4. We dis uss methods for
ontrolling rights ampli ation in Se tion 5. Se tion 6 provides proofs of the theorems from
Se tion 3.
2.
The Language
After a dis ussion of roles, we present an overview of the language design. The remaining
subse tions present the formal syntax, evaluation semanti s, typing system, and some simple
examples.
2.1.
Roles.
The language of roles is built up from
role onstru tors.
onstru tors is appli ation dependent, but must in lude the latti e
below. Ea h role
onstru tor,
κ,
has an asso iated arity,
arity (κ).
The
hoi e of role
onstru tors dis ussed
Roles
AE
have the form
κ(A1 , . . . , An ).
We require that roles form a boolean latti e; that is, the set of
the nullary
onstru tor
element.
1,
post x). 0
onstru tors
⋆ (written
0
and
⊓ and ⊔ are idempotent,
and join operations respe tively.
binary
onstru tors
⊔
and
⊓
onstru tors must in lude
(written in x), and unary
is the least element of the role latti e.
1
is the greatest
ommutative, asso iative, and mutually distributive meet
⋆ is the
omplement operator.
A role may be thought of as a set of permissions. Under this interpretation,
empty set, while
1
0
is the
is the set of all permissions.
The syntax of terms uses
role modi ers, ρ, whi
h may be of the form
role modi ers as fun tions from roles to roles, with
ρL A M
↑AL B M = A ⊔ B
↑A
or
↓A.
We use
de ned as follows:
↓AL B M = A ⊓ B
In summary, the syntax of roles is as follows.
κ ::= 0 | 1 | ⊔ | ⊓ |
⋆
| ···
Role
onstru tors
AE ::= κ(A1 , . . . , An )
Roles
ρ ::= ↑A | ↓A
Role modi ers
Throughout the paper, we assume that all roles (and therefore all types) are well-formed,
in the sense that role
onstru tors have the
orre t number of arguments.
The semanti s of roles is de ned by the relation A
provably equivalent. In addition to any appli ation-spe i
axioms of boolean algebra. We say that
(equivalently
for any
.
B = A ⊓ B)
A, B .
The role modi er
Dually, the modi er
↓A
↑A
A
dominates B
is derivable. Thus we
an
.
= B
stating that
A
B
are
axioms, we assume the standard
.
A > B ) if A = A ⊔ B
1 > A ⊔ B > A > A ⊓ B > 0,
(notation
on lude
reates a weaker role ( loser to 0 ), thus we refer to it as a
reates a stronger role ( loser to
and
1 ),
restri tion.
and thus we refer to it as an
6
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
ampli ation.
While this ordering follows that of the
nist rba
standard [12℄, it is dual to
the normal logi al reading; it may be helpful to keep in mind that, viewed as a logi ,
false,
0
is true,
⊔
is and,
Language overview.
2.2.
⊓
is or and
>
Our goal is to
stood, yet ri h enough to
omponents of the system. We have
hosen to
metalanguage be ause it is simple and well under-
apture the key
parti ularly well suited to studying
is
apture the essen e of role-based systems, where
roles are used to regulate the intera tion of
base our language on Moggi's monadi
1
is implies.
on epts. By design, the monadi
omputational side ee ts (or simply
entral to our work. (We expe t that our ideas
metalanguage is
ee ts ), whi
h are
an be adapted to both pro ess and obje t
al uli.)
The omponents in the monadi
them.
metalanguage are terms and the
To prote t terms, we introdu e guards of the form
{A}[M ℄,
ontexts that use
whi h
an only be
A. The notion of ontext role is formalized in
A ⊲ M → N indi ates that ontext role A is su ient to
redu e M to N . The term check M dis harges the guard on M . The evaluation rule allows
A ⊲ check {B }[M ℄ → [M ℄ only if A > B .
The ontext role may vary during evaluation: given ontext role A, the term ρ (M )
evaluates M with ontext role ρL A M. Thus, when ↓B (M ) is evaluated with ontext role A,
M is evaluated with ontext role A ⊓ B . A ontext may prote t itself from a term by pla ing
dis harged by a
ontext whose role dominates
the de nition of evaluation, where
the use of the term in su h a restri ted
role modi ers.) By
and thus
ontext. (The syntax enfor es a sta k dis ipline on
ombining upwards and downwards modi ers,
ode may assume any role
ir umvent an intended poli y. We address this issue in Se tion 5.
These
onstru ts are su ient to allow prote tion for both terms and
an be prote ted from
ontexts using guards, and
ontexts
ontexts: terms
an be prote ted from terms
using (restri tive) role modi ers.
Syntax.
2.3.
Let
x, y, z, f , g
range over variable names, and let
bv
range over base values.
Our presentation is abstra t with respe t to base values; we use the types
Unit
(with value
unit)
String, Int
and
in examples. We use the standard en odings of booleans and pairs
(see Example 14). The syntax of values and terms are as follows.
V, U, W ::=
bv | x
λx. M
{A}[M ℄
[M ℄
M, N, L ::=
V
M N | fix M
check M
let x = M ; N
ρ (M )
Notation.
In examples, we write
exa tly role
A.
A (M )
Values; Terms
Base Value
Abstra tion
Guard
Computation
Role Modi er
to abbreviate
↓0 (↑A (M )),
whi h exe utes
M
at
x is bound in the value λx. M (with s ope M ) and in the term let x= M ;
N ). If x does not appear free in M , we abbreviate λx. M as λ. M .
Similarly, if x does not appear free in N , we abbreviate let x = M ; N as M ; N . We
identify syntax up to renaming of bound variables and write N {x := M } for the aptureavoiding substitution of M for x in N .
The variable
N
(with s ope
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
In the presentation of the syntax above, we have paired the
on the left with the destru tors on
metalanguage distinguishes
latter are
2
from
the fun tion value
λx. M
onstru tors on values
omputations on the right. For example, the monadi
[2℄
and
[1+1℄:
the former is an integer, whereas the
omputations that, when bound, produ e an integer. The
must be dis harged in a binding
7
omputation value
[M ℄
ontext see the redu tion rule for let, below. Similarly,
must be dis harged by appli ation; in the redu tion semanti s
that follows, evaluation pro eeds in an appli ation till the term in fun tion position redu es
{A}[M ℄
to a lambda abstra tion.
onstru ts a guarded value; the asso iated destru tor is
check .
The monadi
and treats
unit
metalanguage distinguishes
omputations as rst
onstru tor
[M ℄.)
omputations from the values they produ e
lass entities. (Any term may be treated as a value via the
Both appli ation and the let
onstru t result in
omputations; how-
ever, the way that they handle their arguments is dierent. The appli ation (λx. N )
results in
2.4.
N {x := [M ℄},
whereas the binding let
Evaluation and role error.
x = [M ℄; M
results in
The small-step evaluation relation
de ned indu tively by the following redu tion and
[M ℄
N {x := M }.
A ⊲ M → M′
is
ontext rules.
( -app)
(r-app)
A ⊲ (λx. M ) N → M {x := N }
A ⊲ M → M′
A ⊲ M N → M′ N
( -fix)
(r-fix)
A ⊲ fix (λx. M ) → M {x := fix (λx. M )}
A ⊲ M → M′
A ⊲ fix M → fix M ′
( - hk)
(r- hk)
A ⊲ check {B }[M ℄ → [M ℄
A>B
A ⊲ M → M′
A ⊲ check M → check M ′
( -bind)
(r-bind)
A ⊲ let x = [M ℄; N → N {x := M }
A ⊲ M → M′
A ⊲ let x = M ; N → let x = M ′ ; N
( -mod)
ρL A M ⊲ M → M ′
A ⊲ ρ (M ) → ρ (M ′ )
(r-mod)
A ⊲ ρ (V ) → V
r/ -app for appli ation, r/ -fix for xed points and r/ -bind for let are
r- hk ensures that the ontext role is su ient before dis harging the relevant
guard.
-mod modi es the ontext role until the relevant term is redu ed to a value, at
whi h point r-mod dis ards the modi er.
The rules
standard.
The evaluation semanti s is designed to ensure a role-monotoni ity property. In reasing
the available role- ontext
annot invalidate transitions, it
Lemma 4. If B ⊲ M → M ′ and A > B then A ⊲ M
Proof.
(Sket h) The
ontext role is used only in
evaluation judgement.
an only enable more evolution.
→ M ′.
r- hk.
Result follows by indu tion on the
8
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
Via a series of
Su
M
the program
։
onse utive small steps, the nal value for the program
A⊲M ։V
with result V .
essful termination is written
to
ompletion,
whi h indi ates that
A
an be determined.
is authorized to run
Viewed as a role-indexed relation on terms,
is re exive and transitive.
De nition 5.
M0 evaluates to Mn at A (notation A ⊲ M0 ։ Mn ) if there exist terms
Mi su h that A ⊲ Mi → Mi+1 , for all i (0 ≤ i ≤ n − 1). (b) M diverges at A (notation
A ⊲ M →ω ) if there exist terms Mi su h that A ⊲ Mi → Mi+1 , for all i ∈ N.
(a)
Evaluation
an fail be ause a term diverges, be ause a destru tor is given a value of the
wrong shape, or be ause an inadequate role is provided at some point in the
We refer to the latter as a
A ⊲ check {B }[M ℄
A ⊲ M err
A ⊲ M N err
Example 6.
test<B>
err
role error
(notation
A 6> B
ρL A M ⊲ M
A ⊲ ρ (M )
A ⊲ M err
A ⊲ fix M err
A⊲M
err),
err
err
A ⊲ M err
A ⊲ let x = M ; N
Re all from Se tion 2.3 that
omputation.
de ned indu tively as follows.
B (M )
A ⊲ M err
A ⊲ check M
err
abbreviates
↓0 (↑B (M )),
err
and de ne
3
as follows .
test<B> = check {B}[unit℄
△
test<B> is a
omputation that requires
produ es a role error in any
the role
ontext role
B to evaluate.
For example,
↓B⋆ (test<B>)
⋆
ontext, sin e ↓B restri ts any role- ontext to the negation of
B.
Example 7.
We now illustrate how terms
an provide roles for themselves. Consider the
following guarded fun tion:
from<A,B> = {A}[λy. B (y)℄
△
from<A,B>
A, resulting in a fun tion
△
B. Let test<B> = check {B}[unit℄. No matter what the
B, the following evaluation su eeds:
is a guarded value that may only be dis harged by
that runs any
omputation at
relationship is between
A
and
A ⊲ let z = check from<A,B>; z test<B> ։ B (test<B>) ։ [unit℄
from<A,B>
is far too powerful to be useful. After the
fun tion will run
any
ode at role
B.
One
A-guard
an provide spe i
is dis harged, the resulting
ode, of
ourse, as in
Su h fun tions are inherently dangerous and therefore it is desirable
whi h su h fun tions are
su h as
λg. λy. B (g y),
reated. The essential idea is to atta h suitable
λy. B (M).
onstrain the way in
he ks to a fun tion
whi h takes a non-privileged fun tion and runs it under
are a number of subtleties to
B.
onsider in providing a general purpose infrastru ture to
terms with rights ampli ation. When should the guard be
be allowed to run, and in what
There
reate
he ked? What fun tions should
ontext? In Example 21, we dis uss the treatment of these
issues using the Domain and Type Enfor ement a
ess
ontrol me hanism.
3
We do not address parametri ity here; the bra kets in the names test<B> and from<A,B> are merely
suggestive.
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
3.
We present two typing systems that
Typing
ontrol role errors in addition to shape errors.
The rst typing system determines a
ontext role
su ient
is, with this role, there is no possible exe ution path that
high role.
The se ond system determines a
ontext role
Stated dierently, the se ond system
ne essary
to avoid role errors; that
auses a role error. This system
enables the removal of unne essary role- he ks in a pie e of
role that does not dominate this role will
9
ode for a
aller at a su iently
to avoid role errors; that is, any
ause every exe ution path to result in a role error.
al ulates the role that is
he ked and tested on every
exe ution path and thus determines the amount of prote tion that is enfor ed by the
allee.
Te hni ally, the two systems dier primarily in their notions of subtyping. In the absen e
of subtyping, the typing system determines a
ontext role that is both ne essary and su ient
to exe ute a term without role errors.
Be ause it
learly indi ates the point at whi h
omputation is performed, the monadi
metalanguage is attra tive for reasoning about se urity properties, whi h we understand as
omputational ee ts. The type
omputation type
[T ] to in
[T ]
is the type of
omputations of type
T.
We extend the
lude an ee t that indi ates the guards that are dis harged during
check {A}[1+1℄ has type hAi[Int] this type indi ates
Int) requires A. Guarded values inhabit
this type indi ates the prote tion of A around an underlying
be dis harged with a check, resulting in a term inhabiting the
evaluation of a term. Thus the term
that the redu tion of the term to a value (at type
types of the form
value at type
T.
{A}[T ]
These may
omputation type
hAi[T ].
The syntax of types is given below, with the
onstru tors and destru tors at ea h type
re alled from Se tion 2.3.
T, S ::=
Base
T S
{A}[T ]
hAi[T ]
3.1.
V, U, W ::=
bv | x
λx. M
{A}[M ℄
[M ℄
Subtyping.
whi h ranges over
M, N, L ::=
V
M N | fix M
check M
let x = M ; N
ρ (M )
Types; Values; Terms
Base Value
Abstra tion
Guard
Computation
Role Modi er
The judgments of the subtyping and typing relations are indexed by
{1, 2}.
The subtyping relation for
hAi[T ]
α
re e ts the dieren e between
the two type systems.
If role
A
su es to enable a term to evaluate without role errors, then any higher role
ontext also avoids role errors (using Lemma 4). This explains the subtyping rule for the
rst type system in parti ular,
su ient to run any
⊢1 hAi[T ] <: h1 i[T ],
re e ting the fa t that the top role is
omputation.
On the other hand, if a role
A of the role-
ontext is he ked and tested on every exe ution
path of a term, then so is any smaller role. This explains the subtyping rule for the rst
type system in parti ular,
va uously
he ked in any
⊢2 hAi[T ] <: h0 i[T ],
omputation.
re e ting the fa t that the bottom role is
10
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
⊢α
⊢α
⊢α
⊢α
⊢α Base <: Base
⊢α T ′ <: T ⊢α S <: S ′
⊢α T S <: T ′ S ′
T <: T ′
{A}[T ] <: {A′}[T ′ ]
T <: T ′
hAi[T ] <: hA′ i[T ′ ]
if α = 1 then A′ > A
if α = 2 then A > A′
if α = 1 then A′ > A
if α = 2 then A > A′
Lemma 8. The relations ⊢α T <: S are re exive and transitive.
Type systems.
3.2.
Typing is de ned using environments. An
environment,
Γ ::= x1 :T1 , . . . , xn :Tn
is a nite partial map from variables to types.
As usual, there is one typing rule for ea h synta ti
form plus the rule
t-sub
for sub-
sumption, whi h allows the use of subtyping. Upwards and downwards role modi ers have
separate rules, dis ussed below. The typing rules for the two systems dier only in their notion of subtyping and in the side
ondition on
t-mod-dn; we dis
uss the latter in Example 15.
(t-sub)
(t-base)
Γ ⊢α
(t-var)
bv : Base
Γ, x:T , Γ′
⊢α x : T
(t-abs)
Γ, x:T ⊢α M : S
x∈
/
Γ ⊢α λx. M : T S
Γ ⊢α M : T
⊢α T <: T ′
Γ ⊢α M : T ′
(t-app)
dom(Γ)
(t-fix)
Γ ⊢α M : T S
Γ ⊢α M N : S
Γ ⊢α N : T
(t-grd)
(t- hk)
Γ ⊢α M : T
Γ ⊢α {A}[M ℄ : {A}[T ]
Γ ⊢α M : {A}[T ]
Γ ⊢α check M : hAi[T ]
(t-unit)
(t-bind)
Γ ⊢α M : T T
Γ ⊢α fix M : T
Γ ⊢α M : T
Γ ⊢α [M ℄ : h0 i[T ]
Γ ⊢α M : hAi[T ] Γ, x:T ⊢α N : hBi[S]
x∈
/
Γ ⊢α let x = M ; N : hA ⊔ Bi[S]
(t-mod-up)
(t-mod-dn)
Γ ⊢α M : hBi[T ]
Γ ⊢α ↑A (M ) : hB ⊓ A⋆ i[T ]
Γ ⊢α M : hBi[T ]
if α = 1 then A > B
Γ ⊢α ↓A (M ) : hBi[T ]
The rules
t-base, t-var, t-sub, t-abs, t-app
and
t-fix
are standard. For example,
λx. x : T T, for any T .
⊢α fix (λx. x) : T, for any T .
the identity fun tion has the expe ted typing, ⊢α
omputations
an also be typed; for example,
Any term may be inje ted into a
inhabits
hAi[T ]
for every role
indi ating that no
A;
Nonterminating
omputation type at the least role using
in the light of the earlier dis ussion on subtyping, if ⊢α
M :T
dom(Γ)
t-unit.
in the se ond system, the term inhabits only type
he ks are required to su
Thus,
[M ℄
h0 i[T ],
then, in the rst system,
essfully evaluate the value
[M ℄.
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
Computations may be
hBi[S],
ombined using
then M ; N inhabits
hA ⊔ Bi[S].
t-bind
4
. If
M
inhabits
More generally, we
hAi[T ]
11
and
N
inhabits
an dedu e:
⊢α λx. let x′ = x; x′ : hAi[hBi[T ]] hA ⊔ Bi[T ]
In the rst type system, this rule is motivated by noting that the role
to su
essfully avoid role errors in the
(resp.
N ).
thus
C
For the se ond type system,
B.
M,
every
omputation path of
Thus, in role
ontext
C,
omputation if
onsider a role
is not bigger than at least one of
typing of
for
ombined
A, B . If it
M in role
every
C
A
A⊔B
ontext
(resp.
B)
M
A⊔B
that is not bigger than
is not greater than
C
ontext
A,
su es
su es for
by assumption on
leads to a role-error. Similarly
omputation path in the
ombined
omputation
leads to a role error. Furthermore, using the earlier subtyping dis ussion, the sequen e also
inhabits
h1i[S]
The rule
of
{A}[℄
in the rst system and
t-grd types basi
h0 i[S]
in the se ond.
values with their prote tion level. The higher-order version
has the natural typing:
⊢α λx. {A}[x℄ : T {A}[T ]
Re all that in the transition relation,
t- hk
typing rule
check {A}[N ℄
mirrors this behavior by
onstraints on role
he ks the role
ontext against
A.
The
onverting the prote tion level of values into
ontexts. For example, we have the typing:
⊢α λx. check x : {A}[T ] hAi[T ]
In the spe ial
Γ ⊢α check {A}[N ℄ : hAi[T ],
ase of typing
we
systems as follows. In terms of the rst type system, the role
is at least
will
A.
In terms of the se ond type system, any role
an further justify in the two
ontext passes this
he k if it
A
ontext that does not in lude
ause a role-error.
Role modi ers are treated by separate rules for upwards and downwards modi ers.
t-mod-up is justi ed for the rst type system as follows.
The rule for
that
B
su es to evaluate
⊓ A⋆ . This term
ontext B
(A ⊔ A⋆ ) = B
↑A (M )
without role-errors,
ontributes
for the evaluation of
not greater than
of
M
B,
in a role
pro eeds in role
assumption.
The rule for
M.
A
to role
C
C ⊔A
ontext
↑A (M ) in role
A ⊔ (B ⊓ A⋆ ) = (A ⊔ B) ⊓
ontext yielding
For the se ond type system, assume that if a role is
then the evaluation of
ontext
Under assumption
onsider evaluation of
N
leads to a role error. Consider the evaluation
B ⊓ A⋆ . Then,
eed B and hen e
M
that does not ex eed
the evaluation of
whi h does not ex
auses a role error by
t-mod-dn is justi ed for the rst type system as follows.
Under assumption
B su es to evaluate M without role-errors, and A is greater than B onsider evaluation
of ↓A (M ) in role ontext B . This term alters role- ontext B to B ⊓A = B for the evaluation
of M , whi h su es. For the se ond type system, assume that if a role is not greater than
B , then the evaluation of N leads to a role error. Consider the evaluation of ↓A (M ) in a
role ontext C that does not ex eed B . Then, C ⊓ A ertainly does not ex eed B and so the
evaluation of M auses a role error by assumption.
that
Example 16 and Example 15 dis uss alternate presentations for the rules of typing for
the role modi ers.
In stating the results, we distinguish
omputations from other types. Lemma 10 holds
trivially from the de nitions.
4
The distin tion between our system and dependen y-based systems an be see in t-bind, whi h in d
[1, 2, 30℄ states that ⊢ let x = M ; N : hBi[S] if B > A, where ⊢ M : hAi[T ] and x:T ⊢ N : hBi[S].
12
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
De nition 9.
or
T
is a
Role
A
dominates
omputation type
T (notation A ≥ T )
A > B.
type
hBi[S]
and
if
T
is not a
omputation type,
Lemma 10. (a) If A > B and B ≥ T then A ≥ T . (b) If ⊢1 T <: S and A ≥ S then A ≥ T .
( ) If ⊢2 T <: S and A ≥ T then A ≥ S .
The following theorems formalize the guarantees provided by the two systems.
The
proofs may be found in Se tion 6.
Theorem 11. If ⊢1 M : T and A ≥ T , then either A ⊲ M →ω or A ⊲ M ։ V for some V .
Theorem 12. If ⊢2 M : T and A 6≥ T , then either A ⊲ M →ω or there exists N su h that
and A ⊲ N
A⊲M ։N
err.
For the rst system, we have a standard type-safety theorem. For the se ond system,
su h a safety theorem does not hold; for example
check {11}[unit℄ → [unit℄
run with an insu ient
6⊢2 [unit℄ : h1 i[Unit].
but
⊢2 check {11}[unit℄ : h1 i[Unit]
and
1 ⊲
Instead Theorem 12 states that a term
ontext role is guaranteed either to diverge or to produ e a role
error.
Simple examples.
Example 13. We illustrate
3.3.
ombinators of the language with some simple fun tions. The
identity fun tion may be given its usual type:
⊢α λx. x : T T
The unit of
omputation
an be used to
reate a
omputation from any value:
⊢α λx. [x℄ : T h0 i[T]
The let
onstru t evaluates a
omputation
x′
fun tion:
The guard
omputation.
must itself be a
In this following example, the result of the
omputation be ause it is returned as the result of the
⊢α λx. let x′ = x; x′ : hAi[hBi[T]] hA ⊔ Bi[T]
onstru t
reates a guarded term:
⊢α λx. {A}[x℄ : T {A}[T]
The
he k
onstru t dis harges a guard, resulting in a
omputation:
⊢α λx. check x : {A}[T] hAi[T]
The upwards role modi er redu es the role required by a
omputation.
⊢α λx. ↑B (x) : hAi[T] hA ⊓ B i[T]
⋆
The rst typing system requires that any
ward role modi er
↓B ()
omputation performed in the
must not require more than role
⊢α λx. ↓B (x) : hAi[T] hAi[T]
(where
ontext of a down-
B:
B>A
if
α = 1)
In the rst type system, the last two judgments may be generalized as follows:
⊢1 λx. ρ (x) : hρL A Mi[T] hAi[T]
Thus a role modi er may be seen as transforming a
omputation that requires the modi er
into one that does not. For further dis ussion see Example 16.
λ-RBAC:
Example 14
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
.
(Booleans)
The Chur h Booleans,
tru = λt. λf. t
△
and
fls = λt. λf. f,
△
13
illustrate
the use of subtyping. In the two systems, these may be given the following types.
△
⊢1 tru, fls : Bool1
△
⊢2 tru, fls : Bool2
Bool1 = hAi[T ] hBi[T ] hA ⊔ Bi[T ]
Bool2 = hAi[T ] hBi[T ] hA ⊓ Bi[T ]
These types re e t the intuitions underlying the two type systems. The rst type system
re e ts a maximum over all paths typing, whereas the se ond re e ts a minimum over all
paths typing. The
onditional may be interpreted using the following derived rules.
Γ ⊢1 L : Bool1 Γ ⊢1 M : hAi[T ] Γ ⊢1 N : hBi[T ]
Γ ⊢1 if L then M else N : hA ⊔ Bi[T ]
Γ ⊢2 L : Bool2 Γ ⊢2 M : hAi[T ] Γ ⊢2 N : hBi[T ]
Γ ⊢2 if L then M else N : hA ⊓ Bi[T ]
Example 15 (t-mod-dn).
The side
t-mod-dn
ondition on
se ond typing system, but may unne essarily de rease the a
be seen from the following
Let
M, N
does not ee t typability in
ura y of the analysis, as
an
on rete example.
be terms of type
hBi[T ].
Γ ⊢α M : hBi[T ]
(t-sub)
Γ ⊢α M : hA ⊓ Bi[T ]
(t-mod-dn)
Γ ⊢α ↓A (M ) : hA ⊓ Bi[T ]
ondition, the term let x = ↓A (M ); N would have to be given a type of the
hA ⊓ Bi[T ], even though both M and N have type hBi[T ]. Without the side ondition,
better type hBi[T ] may be given to the entire let expression.
With the side
form
the
Example 16
and
.
(Alternative rule for role modi ers)
t-mod-dn
t-mod-up
t-mod-*.
In the rst typing system,
may be repla ed with the following rule, whi h we
all
Γ ⊢1 M : hρL B Mi[T ]
Γ ⊢1 ρ (M ) : hBi[T ]
ρ = ↑A.
Consider
Be ause
C > (A ⊔ C) ⊓ A⋆ ,
Γ ⊢1 M : hA ⊔ Ci[T ]
(t-mod-up)
Γ ⊢1 ↑A (M ) : h(A ⊔ C) ⊓ A⋆ i[T ]
(t-sub)
Γ ⊢1 ↑A (M ) : hCi[T ]
Γ ⊢1 M : hA ⊔ Ci[T ]
(t-mod-*)
Γ ⊢1 ↑A (M ) : hCi[T ]
Be ause
(D ⊓ A⋆ ) ⊔ A > D,
the following are equivalent.
the following are equivalent.
Γ ⊢1 M : hDi[T ]
(t-sub)
Γ ⊢1 M : h(D ⊓ A⋆ ) ⊔ Ai[T ]
(t-mod-*)
Γ ⊢1 ↑A (M ) : hD ⊓ A⋆ i[T ]
Consider
ρ = ↓A.
Be ause
A>A⊓C
and
Γ ⊢1 M : hDi[T ]
(t-mod-up)
Γ ⊢1 ↑A (M ) : hD ⊓ A⋆ i[T ]
C > A ⊓ C,
Γ ⊢1 M : hA ⊓ Ci[T ]
(t-mod-*)
Γ ⊢1 ↓A (M ) : hCi[T ]
the following are equivalent.
Γ ⊢1 M : hA ⊓ Ci[T ]
(t-mod-dn)
Γ ⊢1 ↓A (M ) : hA ⊓ Ci[T ]
(t-sub)
Γ ⊢1 ↓A (M ) : hCi[T ]
14
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
Suppose
A > D.
D ⊓ A > D,
Then
and the following are equivalent.
Γ ⊢1 M : hDi[T ]
(t-sub)
Γ ⊢1 M : hD ⊓ Ai[T ]
(t-mod-*)
Γ ⊢1 ↓A (M ) : hDi[T ]
Example 17 (A sublanguage).
omputational lambda
value types
T
Γ ⊢1 M : hDi[T ]
(t-mod-dn)
Γ ⊢1 ↓A (M ) : hDi[T ]
The following proper sublanguage is su ient to en ode the
al ulus.
and terms assigned
Here values and terms are disjoint, with values assigned
hAi[T ].
omputation types
T, S ::= Base | T hAi[S] | {A}[T ]
bv | x | λx. M | {A}[V ℄
V, U, W ::=
M, N, L ::= [V ℄ | V U | fix V | check V | let x = M ; N | ρ (M )
En oding the Chur h Booleans in this sublanguage is slightly more
tru and fls
type hAi[S].
Example 14;
blo ks of
must a
ept thunks of type
Operations on base values that have no
Unit hAi[S]
ompli ated than in
rather than the simpler
omputational ee t are pla ed in the language
of values rather than the language of terms. The resulting terms may be simpli ed at any
[1+2 == 3℄
time without ae ting the
omputation (e.g.,
Example 18 (Relation to
onferen e version)
than that of the
.
may be simpli ed to
[tru℄).
The language presented here is mu h simpler
onferen e version of this paper [14℄. In parti ular, the
onferen e version
{A}[λx. M ℄ with types of the form
{A} T hBi[S] : the immediate guard of the
applying the abstra tion is B .
ollapsed guards and abstra tions into a single form
T {A ⊲ B}[S],
abstra tion is A,
whi h translates here as
whereas the ee t of
In addition, the
appli ation
onferen e version
↓C V U
ontext modi ed by
↓C (x U ).
rst
↓C .
ollapsed role modi ation and appli ation:
he ked the guard of
In the
x = check V ;
Examples
In this se tion we assume nullary role
and
then performed the appli ation in a
urrent presentation, this translates as let
4.
Charlie, Admin,
V,
the
onstru tors for user roles, su h as
Alice, Bob,
Daemon.
Example 19
(ACLs)
.
(ACLs). One
an model su h a system as:
Consider a read-only lesystem prote ted by A
ess Control Lists
filesystem = λname. if name=="file1" then check {Admin}["data1"℄
else if name=="file2" then check {Alice ⊓ Bob}["data2"℄
else ["error: file not found"℄
△
If
Admin ≥ Alice ⊓ Bob
then
ode running in the
Admin
role
an a
ess both les:
Admin ⊲ filesystem "file1" ։ check {Admin}["data1"℄ ։ ["data1"℄
Admin ⊲ filesystem "file2" ։ check {Alice ⊓ Bob}["data2"℄ ։ ["data2"℄
If
Alice 6≥ Admin
then
ode running as
Alice
annot a
ess the rst le but
an a
se ond:
Alice ⊲ filesystem "file1" ։ check {Admin}["data1"℄ err
Alice ⊲ filesystem "file2" ։ check {Alice ⊓ Bob}["data2"℄ ։ ["data2"℄
ess the
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
Charlie 6≥ Alice ⊓ Bob
Finally, if
then
ode running as
Charlie
annot a
15
ess either le:
Charlie ⊲ filesystem "file1" ։ check {Admin}["data1"℄ err
Charlie ⊲ filesystem "file2" ։ check {Alice ⊓ Bob}["data2"℄ err
The lesystem
ode
an be assigned the following type, meaning that a
possess a role from ea h of the ACLs in order to guarantee that a
If, in addition,
Admin ≥ Alice ⊓ Bob
then the nal role is equal to
ess
aller must
he ks will not fail.
Admin.
⊢1 filesystem : String hAdmin ⊔ (Alice ⊓ Bob) ⊔ 0i[String]
In the above type, the nal role
an a
ess
0 arises from the unknown le
he k. The la k of an a
ess
bran h that does not require
he k explains the weaker
⊢2
type:
⊢2 filesystem : String hAdmin ⊓ (Alice ⊓ Bob) ⊓ 0i[String]
filesystem has the potential to expose some information to unprivi.
Admin ⊓ (Alice ⊓ Bob) ⊓ 0 = 0 , perhaps ausing the ode to be agged
This type indi ates that
leged
allers with role
for se urity review.
Example 20 (Web server).
tem des ribed above.
Consider a web server that provides remote a
The web server
lesystem (unless the web server's
an use the role assigned to a
ess to the lesysaller to a
ess the
aller withholds its role). To prevent an atta ker deter-
mining the non-existen e of les via the web server, the web server fails when an attempt is
made to a
ess an unknown le unless the
Debug
role is a tivated.
webserver = λname. if name=="file1" then filesystem name
else if name=="file2" then filesystem name
else check {Debug}["error: file not found"℄
△
For example,
ode running as Ali e
an a
ess
"file2"
via the web server:
Alice ⊲ webserver "file2" ։ filesystem "file2" ։ ["data2"℄
The a
ess he k in the web server does prevent the le not found error message leaking
Debug role is a tive, but, unfortunately, it is not possible to assign a role stri tly
than 0 to the web server using the se ond type system. The filesystem type does not
unless the
greater
re ord the dierent roles that must be
he ked depending upon the lename argument.
⊢2 webserver : String hAdmin ⊓ (Alice ⊓ Bob) ⊓ 0 i[String]
(derivable)
6⊢2 webserver : String hAdmin ⊓ (Alice ⊓ Bob) ⊓ Debugi[String]
(not derivable)
Example 21 illustrates how the
Domain-Type Enfor ement (dte) a
ess
ontrol me ha-
selinux) [17℄, an be modelled in λ-RBAC.
rba and dte an be found in [11, 13℄.
nism [6, 31℄, found in Se urity-Enhan ed Linux (
Further dis ussion of the relationship between
Example 21
denies a
.
(Domain-Type Enfor ement)
ess requests a
ording to the
The
urrent
dte
a
ess
ontrol me hanism grants or
domain of running
ode. The
urrent domain
hanges as new programs are exe uted, and transitions between domains are restri ted in
order to allow, and also for e,
ode to run with an appropriate domain. The restri tions upon
domain transitions are based upon a
example, the
dte poli
dte type asso
iated with ea h program to exe ute. For
y in [31℄ only permits transitions from a domain for daemon pro esses
to a domain for login pro esses when exe uting the login program, be ause
ode running in
the login domain is highly privileged. This ee t is a hieved by allowing transitions from
16
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
the daemon domain to the login domain only upon exe ution of programs asso iated with
a parti ular
dte
dte
type, and that
The essen e of
dte
types, and the
building upon the
fun tion
dte
an be
type is assigned only to the login program.
aptured in
λ-RBAC,
ontext role to model the
ode fragment
he king role
B
using roles to model both domains and
urrent domain of a system.
λg. λy. B (g y),
to be exe uted in the
We start by
dis ussed in Example 7, that allows a
ontext of
ode running at a dierent role.
We have the typing (for emphasis we use extra parentheses that are not stri tly ne essary
given the usual right asso iativity for the fun tion type
onstru tor):
⊢α λg. λy. B (g y) : (T hBi[S]) (T h0 i[S])
To aid readability, and xing types
T
and
S
for the remainder of this example, de ne:
R = R (T h0 i[S])
△
So that the previous typing be omes:
⊢α λg. λy. B (g y) : T hBi[S]
To restri t the use of the privileged fun tion
a ting as a
dte
in the sequel by
λg. λy. B (g y),
type, where the asso iation of the
ode that
an a tivate role
E.
dte
type
it
E
an be guarded by a role
E
with a fun tion is modelled
The guarded fun tion
an be typed as:
⊢α {E}[λg. λy. B (g y)℄ : {E}[T hBi[S]]
domtrans<A, E, B>
We now de ne a fun tion
A to domain
E. The fun
(role)
B
upon exe ution of a fun tion asso iated with
tion rst veri es that the
of the privileged fun tion
domtrans<A, E, B>
for a domain transition from domain (role)
λg. λy. B (g y)
ontext role dominates
by
ode that
A,
dte
type (also a role)
and then permits use
an a tivate role
E.
The fun tion
is de ned by:
domtrans<A, E, B> = λf. λx. check {A}[unit℄; f {E}[λg. λy. B (g y)℄ x
△
We have the typing:
⊢α domtrans<A, E, B> : {E}[T hBi[S]] (T hAi[S])
domtrans<A, E, B> an be used to turn a fun tion he king role
B into a fun tion he king role A, but only when the role E is availablein ontrast to the
type (T hBi[S]) (T hAi[S]) that does not require E.
In order to make use of domtrans<A, E, B>, we must also onsider ode that an a tivate
E. We de ne a fun tion assign<E> that takes a fun tion f and a tivates E in order to a ess
the privileged ode λg. λy. B (g y) from domtrans<A, E, B>. The fun tion assign<E> is
The above type shows that
de ned by:
assign<E> = λf. λx. λy. let g = E (check x); g f y
△
And we have the typing:
⊢α assign<E> : (T hBi[S]) {E}[T hBi[S]]
Therefore the fun tional
omposition of
assign<E>
and
domtrans<A, E, B>
has type:
(T hBi[S]) (T hAi[S])
To show that in the presen e of both
ontext
A
an exe ute
ode
assign<E>
he king for role
and
ontext
domtrans<A, E, B>, ode running with
B, we onsider the following redu tions
λ-RBAC:
in role
ontext
A,
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
where we take
F = λz. check {B}[unit℄
△
17
and underline terms to indi ate
the redex:
domtrans<A, E, B> (assign<E> F) unit
= (λf. λx. check {A}[unit℄; f ({E}[λg. λy. B (g y)℄) x) (assign<E> F) unit
→ (λx. check {A}[unit℄; (assign<E> F) ({E}[λg. λy. B (g y)℄) x) unit
→ check {A}[unit℄; (assign<E> F) ({E}[λg. λy. B (g y)℄) unit
→ (assign<E> F) ({E}[λg. λy. B (g y)℄) unit
= (λf. λx. λy. let g = E (check x); g f y) F ({E}[λg. λy. B (g y)℄) unit
→ (λx. λy. let g = E (check x); g F y) ({E}[λg. λy. B (g y)℄) unit
→ (λy. let g = E (check {E}[λg. λy. B (g y)℄); g F y) unit
→ let g = E (check {E}[λg. λy. B (g y)℄); g F unit
→ let g = E ([λg. λy. B (g y)℄); g F unit
։ let g = [λg. λy. B (g y)℄; g F unit
→ (λg. λy. B (g y)) F unit
→ (λy. B (F y)) unit
→ B (F unit)
= B ((λz. check {B}[unit℄) unit)
→ B (check {B}[unit℄)
→ B ([unit℄)
։ [unit℄
The strength of
dte lies in the ability
to fa tor a
ess
ontrol poli ies into two
dte
nents: the set of permitted domain transitions and the assignment of
ompo-
types to
ode.
We illustrate this by adapting the aforementioned login example from [31℄ to
λ − RBAC.
this example, the
ode (running
me hanism is used to for e every invo ation of user
In
ode (running at role Daemon) to o ur via trusted login ode
Login). This is a hieved by providing domain transitions from Login to
User, and Daemon to Login, but no others. Moreover, ode permitted to run at Login must
be assigned dte type LoginEXE, and similarly for User and UserEXE. Thus a full program
running daemon ode M has the following form, where neither M nor N ontain dire t rights
at role
User)
dte
from daemon
(running at role
ampli ation.
let dtLoginToUser = domtrans<Login, UserEXE, User>;
let dtDaemonToLogin = domtrans<Daemon, LoginEXE, Login>;
let shell = assign<UserEXE> (λ. M);
let login = assign<LoginEXE> (λpwd. if pwd=="se ret" then dtLoginToUser shell unit else . . .);
Daemon (N)
Be ause
the
login
provides the sole gateway to the role
User, the
User (in
orre t password in order to exe ute the shell at
User but not
dtDaemonToLogin makes
Daemon).
daemon
are available at role
at role
transition
it impossible for the daemon
at
User.
ode
order to a
N
must provide
ess resour es that
In addition, removal of the domain
ode to exe ute any
ode
18
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
Controlling rights amplifi ation
5.
Example 22.
of the form
Suppose that
↑A ( · ).
M
ontains no dire t rights ampli ation, that is, no subterms
Then, in
let priv = [λx. ↑A (V x)℄; ↓User (M)
V
we may view
as a
Trusted Computing Base (t b) a privileged fun tion whi h may
M as restri ted user ode. The fun tion priv is an entry point to
es alate rights and view
the
t b whi
User,
h is a
essible to user
ode; that is, user
and rights ampli ation may only o
Non-trivial programs have larger
t bs
ode is exe uted at the restri ted role
ur through invo ation of
priv.
with more entry points. As the size of the
t b
grows, it be omes di ult to understand the se urity guarantees oered by a system when
rights ampli ation is un onstrained, even if only in the
one may enfor e a
oding
he ks. As an example,
t b.
To manage this
omplexity,
onvention that requires rights in reases be justi ed by earlier
onsider the following, where
amplify
is a unary role
onstru tor.
let at<A>= [λf. check {amplify (A)}[λx. ↑A (f x)℄℄;
let priv = at<A> V;
↓User (M)
In a
ontext with role
amplify (A), this redu
es (using
r-bind, r-app
and
r- hk)
to
let priv = [λx. ↑A (V x)℄; ↓User (M)
In a
ontext without role
r- hk.
A)
amplify (A), evaluation be
The privileged fun tion returned by
is justi ed by the
he k for
at<A>
amplify (A) on any
omes stu k when attempting to exe ute
(whi h performs rights ampli ation for
aller of
at<A>.
One may also wish to expli itly prohibit a term from dire t ampli ation of some right
B; with su
h a
onvention in pla e, this
an be a hieved using the role modi er
↓amplify (B).
One may formalize the pre eding example by introdu ing the unary role
onstru tor
amplify , where amplify (A) stands for the right to provide the role A by storing ↑A in ode.
We require that amplify distribute over ⊔ and ⊓ and obey the following absorption laws:
.
.
A ⊔ amplify (A) = amplify (A)
A ⊓ amplify (A) = A
Thus amplify (A) > A for any role A.
To distinguish justi ed use of role modi ers from unjusti ed use, we augment the syntax
with
he ked role modi ers.
Whenever a
M, N ::= · · · | ρA (M )
M we mark role
he k is performed on role
indi ate that these modi ers have been justi ed by a
modi ers in the
he k.
De ne the fun tion
homomorphi ally over all terms but for role modi ers:
mark A (ρ (M )) = ρA (mark A (M ))
mark A (ρB (M )) = ρA⊔B (mark A (M ))
Modify the redu tion rule for
check
as follows.
A ⊲ check {B }[M ℄ → [mark B (M )℄
onsequent to
A>B
mark A
λ-RBAC:
Thus, the
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
19
he k in the example above will exe ute as follows.
amplify (A) ⊲ check {amplify (A)}[λx. ↑A (f x)℄ → λx. ↑Aamplify (A) (f x)
In the residual, the abstra tion
ontains a
ampli ation has been provided by
role modi ation errors
amplify (A).
We now de ne
not dominate
↑B (M )
↑B C (M )
moderr
M moderr
let x = M ; N
M moderr
M N moderr
he ked role modi er, indi ating that the role
ode that had the right to do so.
moderr
moderr
so that
C 6>
↑AB (M )
produ es an error if
B
does
amplify (B)
M moderr
check M moderr
M moderr
ρ (M ) moderr
Using this augmented language, unjusti ed rights ampli ation is noted as an error. To
prevent su h errors, we modify the typing system to have judgments of the form
T,
where
C
indi ates the a
the term may be exe uted; sin e
↑A ( · )
when
also modify
justi ed by
C >
t-grd
C.
amplify (A).
and
The rule
M
is guarded by
C,
it may in lude subterms of the form
In addition to adding rules for
he ked role modi ers, we
t-mod-up. The rule t-mod-up ensures that any ampli ation is
t-grd allows guards to be used in he king guarded terms; the rule
is sound sin e guarded terms must be
he ked before they are exe uted.
(t-grd′ )
(t-mod-up′ )
Γ; C ⊔ A ⊢α M : T
Γ; C ⊢α {A}[M ℄ : {A}[T ]
Γ; C ⊢α M : hBi[T ]
C>
Γ; C ⊢α ↑A (M ) : hB ⊓ A⋆ i[T ]
(t-mod-dn- he ked)
(t-mod-up- he ked)
Γ; C ⊢α M : hBi[T ]
if α = 1
Γ; C ⊢α ↓AD (M ) : hBi[T ] then A > B
Γ; C ⊢α M : hBi[T ]
C ⊔D >
Γ; C ⊢α ↑AD (M ) : hB ⊓ A⋆ i[T ]
amplify (A)
One may not assume that top level terms have been guarded; therefore, let
shorthand for
Γ; 0 ⊢α M : T .
Example 23.
Γ; C ⊢α M :
umulated guards on a term whi h must be dis harged before
The fun tions
domtrans
and
assign
amplify (A)
Γ ⊢α M : T
be
from Example 21 are not typable using
this more restri tive system. Re all the de nitions:
domtrans<A, E, B> = λf. λx. check {A}[unit℄; f {E}[λg. λy. B (g y)℄ x
△
assign<E> = λf. λx. λy. let g = E (check x); g f y
△
The ampli ation of
B in domtrans is not justi ed;
neither is the ampli ation of
E in assign.
The required form is:
domtrans<A, E, B> = {amplify (B)}[λf. λx. check {A}[unit℄; f {E}[λg. λy. B (g y)℄ x℄
△
assign<E> = {amplify (E)}[λf. λx. λy. let g = E (check x); g f y℄
△
20
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
The login example must now be modi ed in order to dis harge the guards.
Again the
modi ations are straightforward:
let dtLoginToUser = check domtrans<Login, UserEXE, User>;
let dtDaemonToLogin = check domtrans<Daemon, LoginEXE, Login>;
let assignXUser = check assign<UserEXE>;
let assignXLogin = check assign<LoginEXE>;
let shell = assignXUser (λ. M);
let login = assignXLogin (λpwd. if pwd=="se ret" then dtLoginToUser shell unit else . . .);
Daemon (N)
Thus modi ed, the program types
orre tly, but will only exe ute in a
amplify (User), amplify (UserEXE), amplify (Login),
the four roles
This ensures that domain transitions and assignments are
ontext that dominates
and
amplify (LoginEXE).
reated by authorized
ode.
Proposition 25 establishes that the typing system is su ient to prevent role modi ation errors. The proof of Proposition 25 relies on the following lemma, whi h establishes the
mark .
Lemma 24. If Γ; C ⊔ A ⊢α M : T then Γ; C ⊢α mark A (M ) : T .
Proof. By indu tion on the derivation of the typing judgment, appealing to the de nition
of mark .
Proposition 25. If ⊢α M : T and A ⊲ M ։ N then ¬(N moderr)
Proof Sket h. That ¬(M moderr) follows immediately from the de nition of role modi relation between typing and
ation error,
ombined with
t-mod-up′
and
t-mod-up- he ked.
It remains only to show
that typing is preserved by redu tion. We prove this for the type systems of Se tion 3 in
the next se tion. The proof extends easily to the type system
wrinkle is the evaluation rule for
6.
check,
onsidered here. The only
whi h is handled using the previous lemma.
Proof of Type Safety Theorems
The proofs for the rst and se ond systems are similar, both relying on well-studied
te hniques [23℄. We present proofs for the se ond system, whi h is the more
hallenging of
the two.
De nition 26
or
T = hAi[R]
(Compatibility)
and
S = hBi[R],
.
Types
T
and
for some type
S are
R.
ompatible
(notation
T ≈ S)
if
T =S
The following lemmas have straightforward indu tive proofs.
Lemma 27 (Compatibility). If ⊢α T <: T ′ then T ≈ S i T ′ ≈ S .
Lemma 28 (Substitution). If Γ ⊢α M : T and Γ, x : T ⊢α N : S , then Γ ⊢α N {x := M } : S .
Lemma 29 (Bound Weakening). If Γ, x:S ⊢α M : T and ⊢α S ′ <: S , then Γ, x : S ′ ⊢α M : T .
λ-RBAC:
Lemma 30
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
21
.
(Canoni al Forms)
If ⊢2 V : T S then V has form (λx. M ) where x:T ⊢2 M : S .
If ⊢2 V : hAi[T ] then V has form [M ℄ where ⊢2 M : T and A = 0 .
(3) If ⊢2 V : {A}[T ] then V has form {B }[M ℄ where ⊢2 M : T and B > A.
Proof.
(1)
(2)
(1) By indu tion on derivation of
⊢2 V : T S .
The only appli able
t-abs.
′
′
(t-sub) We know ⊢2 V : T S , where ⊢2 V : T S
and
⊢2 S <: S ′ .
t-abs)
(
V has form (λx. M )
x:T ′ ⊢2 M : S ′ .
By the IH,
and subsumption,
t-unit)
The only appli able
ases are
t-sub
By
and
and
⊢2 V : hA′ i[T ′ ], where ⊢2 V : hAi[T ] and ⊢2 hAi[T ] <: hA′ i[T ′ ],
A > A′ . By the IH, V has form [M ℄ where ⊢2 M : T and A = 0 ,
′
subsumption, ⊢2 M : T .
T ′ and
Immediate.
(3) By indu tion on derivation of
t-grd.
(t-sub)
⊢2 V : hAi[T ].
We know
⊢2 T <:
A′ = 0 .
(
where
t-sub
T S <: T ′ S ′ , so ⊢2 T ′ <: T
x:T ⊢2 M : S . By Lemma 29
Immediate.
(2) By indu tion on derivation of
t-unit.
(t-sub)
and ⊢2
ases are
We know
⊢2 V : {A}[T ].
⊢2 V : {A′}[T ′ ],
where
The only appli able
ases are
t-sub
so
so
and
⊢2 V : {A}[T ] and ⊢2 {A}[T ] <: {A′}[T ′ ], so
{B }[M ℄ where ⊢2 M : T and B > A,
⊢2 T <: T ′ and A > A′ . By the IH, V has form
′
′
so B > A . By subsumption, ⊢ M : T .
t-grd)
(
2
Immediate.
Proposition 31 (Preservation). If ⊢2 M
S≈T
: T and A ⊲ M → N then there exists S su h that
and ⊢2 N : S and if A ≥ S then A ≥ T .
Proof.
By indu tion on the derivation of
quanti ation over
A, N .
⊢2 M : T .
The indu tion hypothesis in ludes the
For values, the result is trivial; thus we
onsider only the rules for
non-values.
t-sub) We know ⊢2 M : T ′ , where ⊢2 M : T
T <: T ′ , and A ⊲ M → N . Applying the
IH to ⊢2 M : T and A ⊲ M → N yields S su h that ⊢2 N : S and S ≈ T and if A ≥ S
′
then A ≥ T . By Lemma 10 , this extends to if A ≥ S then A ≥ T . In addition, by
′
Lemma 27, we have S ≈ T .
(t-app) We know ⊢2 M N : T2 , where ⊢2 M : T1 T2 and ⊢2 N : T1 , and A ⊲ M N → L.
There are two sub ases depending on the redu tion rule used in A ⊲ M N → L.
′
′
(M is a value) By Lemma 30, M = λx. M and x:T1 ⊢2 M : T2 . The redu tion yields
′
L = M {x := N }. By Lemma 28, ⊢2 L : T2 . The remaining requirements on T2 are
(
and ⊢2
immediate.
A ⊲ M → M ′ and L = M ′ N . Applying the IH to
⊢2 M : T1 T2 and A ⊲ M → M ′ yields S su h that ⊢2 M ′ : S and S ≈ T1 T2 , whi h
implies that S = T1 T2 . Hen e ⊢2 L : T2 . The remaining requirements on T2 are
(M has a redu tion) Therefore
immediate.
t-fix)
(
We know
⊢2 fix M : T ,
where
⊢2 M : T T
sub ases depending on the redu tion rule used in
and A ⊲ fix M → L.
A ⊲ fix M → L.
There are two
22
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
M = λx. M ′ and x:T ⊢2 M ′ : T . The redu tion yields
Lemma 28, ⊢2 L : T . The remaining requirements on T are
(M is a value) By Lemma 30,
L=
M ′ {x
:= M }.
By
immediate.
A ⊲ M → M ′ and L = fix M ′ . Applying the IH to
⊢2 M : T T and A ⊲ M → M ′ yields S su h that ⊢2 M ′ : S and S ≈ T T , whi h
implies that S = T T . Hen e ⊢2 L : T . The remaining requirements on T are
(M has a redu tion) Therefore
immediate.
t- hk)
(
A ⊲ check M → L.
A ⊲ check M → L.
′
′
(M is a value) By Lemma 30, M = {A2 }[M ℄ and ⊢2 M : T and A2 > A1 . The
′
redu tion yields L = [M ℄ and from the redu tion we dedu e A > A2 , so A > A1
always holds. We assign type ⊢2 L : h0 i[T ], where h0 i[T ] ≈ hA1 i[T ], and we have
already shown that A > 0 implies A > A1 .
′
′
(M has a redu tion) Therefore A ⊲ M → M and L = check M . Applying the IH to
′
′
⊢2 M : {A1}[T ] and A ⊲ M → M yields S su h that ⊢2 M : S and S ≈ {A1}[T ],
so S = {A1}[T ]. Hen e ⊢2 L : hA1 i[T ]. The remaining requirements on hA1 i[T ] are
We know
⊢2 check M : hA1 i[T ],
where
⊢2 M : {A1}[T ],
and
There are two sub ases depending on the redu tion rule used in
immediate.
t-bind)
(
⊢2 let x = M ; N : hA1 ⊔ A2 i[T2 ], where ⊢2 M : hA1 i[T1 ] and x:T1 ⊢2 N :
A ⊲ let x= M ; N → L. There are two sub ases depending on the redu tion
rule used in A ⊲ let x = M ; N → L.
′
′
(M is a value) By Lemma 30, M = [M ℄ and ⊢2 M : T1 and A1 = 0 , so A1 ⊔ A2 = A2 .
′
The redu tion yields L = N {x := M }. By Lemma 28, ⊢2 L : hA2 i[T2 ]. The remaining
requirements on T2 are immediate.
′
′
(M has a redu tion) Therefore A ⊲ M → M and L = let x = M ; N . Applying the IH
′
′
to ⊢2 M : hA1 i[T1 ] and A ⊲ M → M yields S su h that ⊢2 M : S and S ≈ hA1 i[T1 ]
and if A ≥ S then A ≥ hA1 i[T1 ]. Hen e S = hA3 i[T1 ], for some A3 , and A > A3
implies A > A1 . We dedu e ⊢2 L : hA3 ⊔ A2 i[T2 ], where hA3 ⊔ A2 i[T2 ] ≈ hA1 ⊔ A2 i[T2 ].
Finally, suppose A ≥ hA3 ⊔ A2 i[T2 ], i.e., A > A3 ⊔ A2 , so A > A3 and A > A2 . By
the above, this entails A > A1 , so A > A1 ⊔ A2 . Therefore A ≥ hA1 ⊔ A2 i[T2 ], as
We know
hA2 i[T2 ],
and
required.
t-mod-up) We know ⊢2 ↑A (M ) : hA ⊓A⋆ i[T ], where ⊢2 M : hA i[T ], and A ⊲ ↑A (M ) →
(
1
2
1
2
1
L. There are two sub ases depending on the redu tion rule used in A ⊲ ↑A1 (M ) → L.
⋆
(M is a value) Therefore L = M and ⊢2 L : hA2 i[T ], where hA2 i[T ] ≈ hA2 ⊓ A1 i[T ]. By
′
′
Lemma 30, M = [M ℄ and ⊢2 M : T and A2 = 0 , and the remaining requirement on
hA2 i[T ], that A ≥ hA2 i[T ] implies A ≥ hA2 ⊓ A⋆1 i[T ], is immediate.
′
′
(M has a redu tion) Therefore ↑A1 L A M ⊲ M → M and L = ↑A1 (M ). Applying the
′
′
IH to ⊢2 M : hA2 i[T ] and ↑A1 L A M ⊲ M → M yields S su h that ⊢2 M : S and S ≈
hA2 i[T ], so S = hA3 i[T ] for some A3 , and if ↑A1 L A M ≥ S then ↑A1 L A M ≥ hA2 i[T ],
′
⋆
i.e., A ⊔ A1 > A3 implies A ⊔ A1 > A2 . We have ⊢2 ↑A1 (M ) : hA3 ⊓ A1 i[T ] and
⋆
⋆
⋆
hA3 ⊓ A1 i[T ] ≈ hA2 ⊓ A1 i[T ]. Finally, if A ≥ hA3 ⊓ A1 i[T ], then A > A3 ⊓ A⋆1 , so
A ⊔ A1 > (A3 ⊓ A1⋆ ) ⊔ A1 = (A3 ⊔ A1 ) ⊓ 1 = A3 ⊔ A1 . Hen e A ⊔ A1 > A3 , so
A ⊔ A1 > A2 , and A ⊓ A1⋆ = (A ⊓ A1⋆ ) ⊔ 0 = (A ⊔ A1 ) ⊓ A1⋆ > A2 ⊓ A⋆1 . Therefore
A > A2 ⊓ A1⋆ and A ≥ hA2 ⊓ A⋆1 i[T ], as required.
(t-mod-dn) We know ⊢2 ↓A1 (M ) : hA2 i[T ], where ⊢2 M : hA2 i[T ], and A ⊲ ↓A1 (M ) → L.
There are two sub ases depending on the redu tion rule used in A ⊲ ↓A1 (M ) → L.
(M is a value) Therefore L = M and ⊢2 L : hA2 i[T ], and we are done.
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
23
↓A1 L A M ⊲ M → M ′ and L = ↓A1 (M ′ ). By A > ↓A1 L A M
′
and Lemma 4, we have A ⊲ M → M . Applying the IH to ⊢2 M : hA2 i[T ] and
′
′
A ⊲ M → M yields S su h that ⊢2 M : S and S ≈ hA2 i[T ], so S = hA3 i[T ] for some
A3 , and if A ≥ S then A ≥ hA2 i[T ]. Hen e ⊢2 ↓A1 (M ′ ) : hA3 i[T ], whi h ompletes
(M has a redu tion) Therefore
the sub ase.
Corollary 32. If ⊢2 M
Proof.
ase,
:T
and A ⊲ M ։ V , then A ≥ T .
By indu tion on the length of the redu tion sequen e
M = V
and Lemma 30 implies that
A ≥ T,
A ⊲ M ։ V.
For the base
be ause every non- omputation type
T = hBi[S] Lemma 30 tells us that
N su h that A ⊲ M → N and A ⊲ N ։ V . By
Proposition 31, there exists S su h that ⊢2 N : S and if A ≥ S then A ≥ T . Applying the
IH to ⊢2 N : S and A ⊲ N ։ V yields A ≥ S , hen e A ≥ T as required.
is dominated by any role, and in a
B = 0.
omputation type
For the indu tive step, there exists
Proposition 33
(Progress)
. For all A, if ⊢2
M :T
or there exists N su h that A ⊲ M → N .
Proof. By indu tion on the derivation of ⊢2 M : T .
M
then either M is a value, A ⊲ M
We need only
onsider the
err,
ases when
is not a value.
t-sub) We know ⊢2 M : T ′ , where ⊢2 M : T and ⊢2 T <: T ′ . Immediate
(t-app) We know ⊢2 M N : T , where ⊢2 M : T T and ⊢2 N : T
(
2
1
by the IH.
1.
2
Apply the IH to
⊢2 M : T1 T2 and role A. If M is a value, then, by Lemma 30, M has form (λx. L), so
A ⊲ M N → L{x := N }. If A ⊲ M err, then A ⊲ M N err. Finally, if A ⊲ M → L,
then A ⊲ M N → L N .
(t-fix) We know ⊢2 fix M : T , where ⊢2 M : T T . Apply the IH to ⊢2 M : T T and role A.
If M is a value, then, by Lemma 30, M has form (λx. L), so A ⊲ fix M → L{x := (λx. L)}.
err, then A ⊲ fix M err. Finally, if A ⊲ M → L, then A ⊲ fix M → fix L.
If A ⊲ M
(t- hk) We know ⊢2 check M : hA1 i[T ], where ⊢2 M : {A1}[T ]. Apply the IH to ⊢2 M :
{A1}[T ] and role A. If M is a value, then, by Lemma 30, there exists B , L su h that
M = {B }[L℄, so either A ⊲ check M → [L℄ or A ⊲ check M err depending on whether
A > B holds or not. If A ⊲ M err, then A ⊲ check M err. Finally, if A ⊲ M → L,
then A ⊲ check M → check L.
(t-bind) We know ⊢2 let x = M ; N : hA1 ⊔ A2 i[T2 ], where ⊢2 M : hA1 i[T1 ] and x:T1 ⊢2
N : hA2 i[T2 ]. Apply the IH to ⊢2 M : hA1 i[T1 ] and role A. If M is a value, then, by
Lemma 30, M has form [L℄, so A ⊲ let x = M ; N → N {x := L}. If A ⊲ M
err, then
A ⊲ let x = M ; N err. Finally, if A ⊲ M → L, then A ⊲ let x = M ; N → let x = L; N .
⋆
(t-mod-up) We know ⊢2 ↑A1 (M ) : hA2 ⊓ A1 i[T ], where ⊢2 M : hA2 i[T ]. Apply the IH
to ⊢2 M : hA2 i[T ] and role ↑A1 L A M. If M is a value, then A ⊲ ↑A1 (M ) → M . If
↑A1 L A M ⊲ M
err, then A ⊲ ↑A1 (M )
err. Finally, if ↑A1 L A M ⊲ M → L, then
A ⊲ ↑A1 (M ) → ↑A1 (L).
(t-mod-dn) We know ⊢2 ↓A1 (M ) : hA2 i[T ], where ⊢2 M : hA2 i[T ]. Apply the IH to
⊢2 M : hA2 i[T ] and role ↓A1 L A M. If M is a value, then A ⊲ ↓A1 (M ) → M . If
↓A1 L A M ⊲ M
err, then A ⊲ ↓A1 (M )
err. Finally, if ↓A1 L A M ⊲ M → L, then
A ⊲ ↓A1 (M ) → ↓A1 (L).
Theorem (12). If ⊢2 M
A⊲M ։N
Proof.
and
We use a
: T and A 6≥ T ,
A ⊲ N err.
then either A ⊲ M →ω or there exists N su h that
oindu tive argument to
onstru t a redu tion sequen e that is either in nite
or terminates with a role
he k failure. When
⊢2 M : T
and
A 6≥ T ,
we know that
M
is
24
R. JAGADEESAN, A. JEFFREY, C. PITCHER, AND J. RIELY
A⊲M
not a value by Lemma 30. By Proposition 33, either
A ⊲ M → N . In the former
there exists S su h that ⊢2 N : S
so A 6≥ S , as required.
that
and if
7.
A≥S
then
A ≥ T.
or there exists
N
su h
ase, using Proposition 31,
However, we know that
A 6≥ T ,
Con lusions
The fo us of this paper is programmati
approa hes, su h as
From a software engineering approa h to the design of
ration of
err
ase, we are done. In the latter
on erns: the design of the system is
jaas/.net, that use rba .
rba fa ilitates a sepa-
omponents,
arried out in terms of a role hierar hy with
an asso iated assignment of permissions to roles, whereas the a tual assignment of users to
roles takes pla e at the time of deployment.
We have presented two methods to aid the design and use of
su h a
ess
ontrol
ode.
ode to dedu e the role at whi h
optimize
ode must be run.
The main use of this analysis is to
ode by enabling the removal of some dynami
more novel analysis
omponents that in lude
The rst admittedly standard te hnique enables users of
he ks. The se ond somewhat
al ulates the role that is veri ed on all exe ution paths.
This
analysis is potentially useful in validating ar hite tural se urity requirements by enabling
ode designers to dedu e the prote tion guarantees of their
ode.
We have demonstrated the use of these methods by modeling Domain Type Enfor ement, as used in SELinux. As future work, we will explore extensions to role polymorphism
and re ursive roles following the te hniques of [8, 4℄.
A knowledgments
The presentation of the paper has greatly improved thanks to the
omments of the
referees.
Referen es
SIGPLAN Not.
[1℄ M. Abadi. A ess ontrol in a ore al ulus of dependen y.
, 41(9):263273, 2006.
[2℄ M. Abadi, A. Banerjee, N. Heintze, and J. G. Rie ke. A ore al ulus of dependen y. In
,
pages 147160. ACM Press, 1999.
[3℄ M. Abadi, G. Morrisett, and A. Sabelfeld. Language-based se urity.
, 15(2):129, 2005.
[4℄ R. M. Amadio and L. Cardelli. Subtyping re ursive types.
, 15(4):575631, 1993.
[5℄ S. Barker and P. J. Stu key. Flexible a ess ontrol poli y spe i ation with onstraint logi programming.
, 6(4):501546, 2003.
[6℄ W. E. Boebert and R. Y. Kain. A pra ti al alternative to hierar hi al integrity poli ies. In
,
1985.
[7℄ C. Braghin, D. Gorla, and V. Sassone. A distributed al ulus for role-based a ess ontrol. In
,
pages 4860, 2004.
[8℄ M. Brandt and F. Henglein. Coindu tive axiomatization of re ursive type equality and subtyping.
, 33(4):309338, 1998.
[9℄ S. Chong and A. C. Myers. Se urity poli ies for downgrading. In
, pages 198209, 2004.
[10℄ A. Compagnoni, P. Garralda, and E. Gunter. Role-based a ess ontrol in a mobile environment. In
, 2005.
[11℄ D. F. Ferraiolo, D. R. Kuhn, and R. Chandramouli.
. Computer Se urity
Series. Arte h House, 2003.
[12℄ D. F. Ferraiolo, R. Sandhu, S. Gavrila, D. R. Kuhn, and R. Chandramouli. Proposed NIST standard
for role-based a ess ontrol.
, 4(3):224274, 2001.
ACM Trans. Inf. Syst. Se ur.
J. Fun t. Program.
ACM TOPLAS
dam. Inf.
Symposium on Trustworthy Global Computing
CCS '04
Role-Based A ess Control
ACM Trans. Inf. Syst. Se ur.
POPL '99
CSS '85
CSFW
Fun-
λ-RBAC:
PROGRAMMING WITH ROLE-BASED ACCESS CONTROL
25
Computer Se urity Appli ations (AC-
[13℄ J. Homan. Implementing RBAC on a type enfor ed system. In
, pages 158163, 1997.
[14℄ R. Jagadeesan, A. Jerey, C. Pit her, and J. Riely. λ-RBAC: Programming with role-based a ess
ontrol. In
, volume 4052 of
, pages 456467. Springer,
2006.
[15℄ S. Jajodia, P. Samarati, M. L. Sapino, and V. S. Subrahmanian. Flexible support for multiple a ess
ontrol poli ies.
, 26(2):214260, 2001.
[16℄ J. Ligatti, L. Bauer, and D. Walker. Edit automata: enfor ement me hanisms for run-time se urity
poli ies.
, 4(1-2):216, 2005.
[17℄ P. A. Los o o and S. D. Smalley. Meeting riti al se urity obje tives with Se urity-Enhan ed Linux.
In
, 2001.
[18℄ S. Malhotra.
. Premier Press, 2002.
[19℄ J. C. Mit hell. Programming language methods in omputer se urity. In
, pages 126, 2001.
[20℄ A. C. Myers, A. Sabelfeld, and S. Zdan ewi . Enfor ing robust de lassi ation. In
, pages 172
186, 2004.
[21℄ S. Osborn, R. Sandhu, and Q. Munawer. Con guring role-based a ess ontrol to enfor e mandatory
and dis retionary a ess ontrol poli ies.
, 3(2):85106, 2000.
[22℄ J. S. Park, R. S. Sandhu, and G.-J. Ahn. Role-based a ess ontrol on the web.
, 4(1):3771, 2001.
[23℄ B. Pier e.
. MIT Press, 2002.
[24℄ A. Sabelfeld and A. C. Myers. Language-based information- ow se urity.
, 21(1):519, Jan. 2003.
[25℄ A. Sabelfeld and A. C. Myers. A model for delimited information release. In
, pages 174191, 2003.
[26℄ R. Sandhu, E. Coyne, H. Feinstein, and C. Youman. Role-based a ess ontrol models.
,
29(2), 1996.
[27℄ F. B. S hneider, G. Morrisett, and R. Harper. A language-based approa h to se urity. In
, volume 2000 of
, pages 86101, 2000.
[28℄ F. Siewe, A. Cau, and H. Zedan. A ompositional framework for a ess ontrol poli ies enfor ement. In
, pages 3242, 2003.
[29℄ E. G. Sirer and K. Wang. An a ess ontrol language for web servi es. In
, pages 2330,
2002.
[30℄ S. Tse and S. Zdan ewi . Translating dependen y into parametri ity. In
, pages 115125, 2004.
[31℄ K. M. Walker, D. F. Sterne, M. L. Badger, M. J. Petka , D. L. Shermann, and K. A. Oostendorp.
Con ning root programs with Domain and Type Enfor ement (DTE). In
,
1996.
SAC '97)
ICALP '06
Le ture Notes in Computer S ien e
ACM Trans. Database Syst.
Int. J. Inf. Se .
Ottawa Linux Symposium
Mi rosoft .NET Framework Se urity
Se ur.
ACM Trans. Inf. Syst. Se ur.
Types and Programming Languages
Communi ations
10 Years Ba k, 10 Years Ahead
FMSE '03
LNCS
POPL '01
CSFW
ACM Trans. Inf. Syst.
IEEE J. Sele ted Areas in
ISSS
IEEE Computer
Informati s
SACMAT '02
ICFP
USENIX Se urity Symposium
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http:// reative ommons.org/li enses/by-nd/2.0/ or send a
letter to Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
| 6cs.PL
|
Exploring Geometric Property Thresholds For
Filtering Non-Text Regions In A Connected
Component Based Text Detection Application
arXiv:1709.03548v1 [cs.CV] 11 Sep 2017
Teresa Nicole Brooks
Seidenberg School of CSIS
Pace University New York, NY
[email protected]
Abstract—Automated text detection is a difficult computer
vision task. In order to accurately detect and identity text in
an image or video, two major problems must be addressed.
The primary problem is implementing a robust and reliable
method for distinguishing text vs non-text regions in images and
videos. Part of the difficulty stems from the almost unlimited
combinations of fonts, lighting conditions, distortions, and other
variations that can be found in images and videos. This paper
explores key properties of two popular and proven methods for
implementing text detection; maximum stable external regions
(MSER) and stroke width variation.
I.
I NTRODUCTION
Automated text detection is employed in modern software
systems to perform tasks such as detecting landmarks in
images and video, surveillance, visual sign translation mobile
apps and in more recent years self-driving cars. Many of these
applications require a robust automated text detection system
that can accurately detect text on signs in natural and unnatural
images.
One popular and proven approach to text detection is using
connected components (CC) [1] [2] to detect text. In connected
component based systems the goal is to separate non-text
and text region candidates. This is done by using learned or
manually established features (thresholds) for discriminating
geometric properties that are used to filter out non-text regions.
One of the most popular feature detectors used in conjunction
in connected component based systems is Maximally Stable
Extermal Regions (MSER) [3] [2].
The geometric property thresholds selected for filtering
non-text regions can vary from image to image and are often
affected by characteristics such as font style, image distortions
(blur, skew etc) and textures to name a few. The objectives of
this research is to explore and learn the effects of different
font styles in both natural scene and unnatural scene images
such as product boxes (i.e. cereal boxes) have on selecting
these geometric property thresholds and their implications for
the robustness of text detection in images. This research will
also explore which geometric properties work best across both
natural and unnatural scene images.
A. Maximally Stable Extermal Regions (MSER)
The exploratory text detection application implemented for
this project uses Maximally Stable Extermal Regions (MSER)
to perform the initial feature (region) detection in our test
images. MSER is a popular feature detection algorithm that
detects stable, connected component regions of gray level
images called MSERs. This algorithm was first introduced by
Matas et al. This paper introduced the use of MSER detection
to solve the “wide-baseline stereo problem [4].” This problem
seeks to find the correspondence between elements in two
images that were taken at different view points. One major
contribution of this paper is the introduction of “extermal
regions” [4]. Extermal regions are regions that are maximally
stable, where stability is defined by lack of or very low
variation in intensity values with respect to some established
thresholds. Though MSER is sensitive to noise and blur [5] it
offers important advantages over other region based detection
algorithms as it offers affine invariance, stability, and highly
repeatable results [6].
1) Region Geometric Properties: Once connected components regions (MSERs) are detected, we can use certain discriminating, geometric properties to filter out non-text region
candidates. The regions are filtered out via the use of thresholds. The property values are features whereby thresholds can
be manually established, calculated or learned using machine
learning techniques. These features are a fast and easy way
to discern non-text regions from text features in images [3].
Below is a brief discussion of the discriminating, geometric
properties used in the system implemented for this project.
•
Aspect ratio: The ratio of the width to the height of
bounding boxes.
•
Eccentricity: Is used to measure the circular nature of
a given region. It is defined as the distance between
the foci and/or the ellipse and its major axis.
•
Solidity: Is ratio of the pixels in the convex hull area
that are also in a given region. It is calculated by [3]
area
convex area
Extent: The size and location of the rectangle that
encloses text.
•
II.
BACKGROUND
This section discusses the techniques and algorithms that
are used in the exploratory text detection application implemented for this research project.
•
Euler Number: Is a feature of a binary image. It is the
number of connected components minus the number
of holes (in those components) [7].
IV.
B. Stroke Width Variation
Stroke width is another property that is used to distinguish
(and filter) non-text regions from text-regions. The approach
of using stroke width to aid in detecting text regions in images
was introduced by Epshtein et al [8]. The main idea behind
using stroke width to filter non-text region is based around
on the fact that usually text vs other elements in an image
have constant stroke width. This means that binary images
can be transformed into stroke width images (skeleton images)
and these skeletal images are used to calculate the stroke
width variation. This ratio can be used with a max variance
threshold to filter non-text regions. Because stroke widths are
calculated on a per pixel basis text detection systems that use
this technique can detect text in a way that is insensitive to
font, color, size, language and orientation [?]. Below is a brief
discussion of the stroke width related values that are calculated
in the text detection system implemented for this project.
•
•
Stroke Width Variation Metric:
Standard Deviation of Stroke Width Values
Mean of Stroke Width Values
Stroke Width Max Threshold
III.
R ELATED W ORK
The literature reflects the impact and popularity of the use
of Maximally Stable Extermal Regions (MSER) and Stroke
Width Variation for detecting text in images. This section
briefly discusses some key related works that influenced the
approach taken in the application implemented for this paper.
A. Text Location in Complex Images (2012)
Gonzalez et al, presented a three stage text detection system. This system combined “fast-to-compute” [3] connected
component features that were used to filter non-text regions,
segmentation to detect text character candidates, and a support
vector machine based text line classifier. Their segmentation
approach combines the use of MSER and ”locally adaptive
threshold” [3] methods and their ”fast-to-computer” features
includes discriminating geometric properties such as aspect
ration of bounding boxes, stroke width ration (metric) and
solidity amount others. This approach proved to be especially
effective in detected text in images with complex backgrounds.
Their system it outperformed various systems when tested
against the Robust Reading competitions datasets for IDCAR
2003, 2011 as well as the CoverDB test dataset.
B. Robust Text Detection In Natural Images With Edgeenhanced Maximally Stable Extremal Regions (2011)
Chen et al, proposed a robust connected component based
text detection system that implements an “edge-enhanced”
[2] version of Maximally Stable Extermal Regions to find
letter candidates. The enhanced version of MSER leveraged
Canny edges, a robust edge detector operator in order to
support detecting small text in low resolution images. This
approach also enabled their system to mitigate MSERś known
issues with detecting features in blurred images. This system
also employed stroke width information to filter out nontext regions. They evaluated their system against the ICDAR
competition dataset, and it performed well as or better than
state of the art systems at the time the paper was published.
A LGORITHM
The text detection applications created for this research
project implements the following workflow. This workflow
performs pre-processing, text detection in a given image and
optical character recognition when possible.
1)
2)
3)
4)
5)
6)
Pre-process images. This includes, converting images
to high contrast images, converting color images
to gray scale images and rotating images (or of
boudning boxes themselves) as needed.
Process each image with MSER feature detector
algorithm.
Perform first pass of non-text region filtering on the
features detected by MSER algorithm using discriminating geometric properties.
Perform second pass of non-text region filtering on
the remaining features using stroke width variation
filtering.
On the remaining regions (text regions), calculate
the overlapping bounding boxes of these regions and
combine them to create one bounded region of text.
Detect the text in the image and pass the detected text
bounding boxes to an ORC algorithm to determine
the text that was found in the image.
V.
E XPERIMENTS
The primary goal of all experiments was to compare
and contrast the discriminating geometric properties and their
thresholds that worked well for detecting text in both natural
and non-natural images. The secondary goal of these experiments was to understand the challenges if any of detecting text
when an image contains highly stylized fonts.
The dataset consists of both natural images and non-natural
images. The natural scene images in Fig 1 are from 2014
training dataset for the COCO Captioning Challenge competition. COCO is dataset for image recognition, segmentation and
captioning [9]. The non-natural images in Fig 2 were taken for
the purpose of this project.
(a) Rotated Stop Sign
(b) Simple Stop Sign
Fig. 1: Natural Images: Images from COCO Captioning
Challenge 2014 training set.
The main objective of the trials conducted was to find the
primary or dominate text regions in a given image. Primary
or dominate text regions defined as the largest text region
in the image. Each dataset image was processed individually
by being run through the workflow presented in Section IV.
During each iteration the geometric property thresholds, stroke
width variation max threshold as well as other properties were
(a) Box Image 1
(b) Box Image 2
Fig. 2: Non-Natural Images: Cereal Boxes
(a)
manually tuned until the primary or dominate text region were
detected. Note, all threshold and tuneable values started as the
defaults defined by libraries used to implement the system.
A. Object Character Recognition (OCR)
The primary objective of all experiments was to visually
verify text detection by plotting individual character and line
or word bounding boxes on processed images. An additional
verification step was employed using Matlab’s OCR library.
Matlab’s OCR library is implemented by a popular, robust
open source OCR engine called Tesseract [10]. Without any
additional training the OCR library’s default English language
model was able to recognize the text from the natural image
texts. One limitation encounter was that in order to recognize
custom, highly stylized fonts one must training customer OCR
models, because this was beyond the scope of the research
project this task will be deferred for future work.
VI.
(b)
Fig. 3: Detected Features & Text In Natural Image
R ESULTS
The text detection application used in these experiments
is not overly robust, but shows the power of the approach
employed for text detection. Even with a modest implementation the application was able detect primary and in some
cases secondary text in a given image with a few iterations of
manually tuning thresholds, as Fig 3 and Fig 4 show.
As expected though there are some discriminating geometric properties and other values that were more important
for tuning to detect text in the non-natural images than in
natural images all tuneable properties of this application were
applicable in detecting text in the given test images.
(a)
A. Issues With Detecting Text Using Custom, Highly Stylized
Fonts
In non-natural images there can be a lot variation among
the subjects in an image that can affect text detection. Characteristics such as high contrast, line spacing and the spacing
between the letter characters presented the most challenge for
the approaches employed in the implemented text detection
application. Despite these challenges the application with some
tuning was able to successfully detect with reasonable proficient, primary and secondary text regions in the non-natural
images.
The following geometric properties thresholds and related
values that appear to have some sensitivity to high contrast
images where text is highly stylized, custom fonts. Note, these
(b)
Fig. 4: Detected Features & Text In Non-Natural Image
are not the only values that needed to be tuned, but these are
the values that had to be tuned to process both non-natural
images.
1) Expansion Amount: This value is used to tune amount
by which the application expands neighboring bounding boxes.
Neighboring bounding box expansion is an important step that
is done in preparation of the final merge of individual character
bounding boxes. Note, this final merge is were the bounding
box is draw that suggests text candidates (words or lines of
text). For, the non-natural images processed this value had to
be lower which in terms increases the bounding box expansion.
This was necessary in order to detect a wide and varied areas
of text. Note, the size of the primary text regions in proportion
to the rest of the subjects in the image could also contribute
to this observation.
2) Stroke Width Variation (Max) Threshold: The value of
the stroke width max threshold had to be increased by a factor
of 2 in order to properly detect all individual text characters
in the non-natural images vs the natural images.
3) High Contrast Images: One the pre-processing steps
for most images is to increase the contrast of the image.
For images that are by default high contrast images it was
observed that increasing the image’s contrast didn’t contribute
to better text detection performance, in fact in some instances
the performance of text detection degraded.
Note,the application implemented for this project uses
a standard deviation based approach for performing image
stretching (increasing image contrast).
VII.
F UTURE W ORK
The next step is to automate the process of finding optimal
threshold values for each image, for each region geometric
property. By automating this process we will have a more
robust and scalable means of conducting further analysis on
different types of images and the properties that affected text
detection the most.
VIII.
C ONCLUSION
The approach of using Maximally Stable Extermal Regions
(MSER) based feature detection, stroke width variation and geometric property thresholding even in a modest text detection
application yields reasonably proficient results. Though highly
stylized, custom fonts have some affect on the geometric
properties thresholds used to filter non-text regions the affect
was not as dramatic and did not have anticipated impact.
This conclusion, reflects the power and robustness of all the
techniques employed especially the use of MSER.
R EFERENCES
[1]
L. Neumann and J. Matas, “Efficient Scene text localization and recognition with local character refinement,” Proceedings of the International
Conference on Document Analysis and Recognition, ICDAR, vol. 2015Novem, pp. 746–750, 2015.
[2] H. Chen, S. S. Tsai, G. Schroth, D. M. Chen, R. Grzeszczuk, and
B. Girod, “Robust text detection in natural images with edge-enhanced
maximally stable extremal regions,” Proceedings - International Conference on Image Processing, ICIP, pp. 2609–2612, 2011.
[3] L. M. Bergasa and J. J. Yebes, “Text Location in Complex Images,”
International Conference on Pattern Recognition, no. Icpr, pp. 617–
620, 2012.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
M. S. Extremal, J. Matas, O. Chum, M. Urban, and T. Pajdla, “Robust
Wide Baseline Stereo from,” In British Machine Vision Conference, pp.
384–393, 2002.
C. Republic, “FASText : Efficient Unconstrained Scene Text Detector.”
K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J. Matas,
F. Schaffalitzky, T. Kadir, and L. Van Gool, “A comparison of affine
region detectors,” International Journal of Computer Vision, vol. 65,
no. 1-2, pp. 43–72, 2005.
C. Republic, “Real-Time Scene Text Localization and Recognition,” pp.
3538–3545, 2012.
B. Epshtein, E. Ofek, and Y. Wexler, “Detecting text in natural scenes
with stroke width transform,” Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition, no. d,
pp. 2963–2970, 2010.
“COCO Common Objects.” [Online]. Available: http://mscoco.org/
home/
“Tesseract Open Source OCR Engine.” [Online]. Available: https:
//github.com/tesseract-ocr/tesseract
| 1cs.CV
|
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
ARE GENERATED IN BOUNDED DEGREE
arXiv:1803.04259v2 [math.AC] 23 Mar 2018
ROBERT P. LAUDONE
Abstract. Over a field of characteristic 0, we prove that for each r ≥ 0 there exists a constant C(r) so
that the prime ideal of the rth secant variety of any Plücker-embedded Grassmannian Gr(d, n) is generated
by polynomials of degree at most C(r), where C(r) is independent of d and n. This bounded generation
ultimately reduces to proving a poset is noetherian, we develop a new method to do this. We then translate
the structure we develop to the language of functor categories to prove the ith syzygy module of the
coordinate ring of the rth secant variety of any Plücker-embedded Grassmannian Gr(d, n) is concentrated
in degrees bounded by a constant C(i, r), which is again independent of d and n.
1.
Introduction
Given a vector space V of dimension n over a field k of characteristic 0, recall that Gr(d, V) is the space
that parametrizes all dimension d subspaces of V called the Grassmannian. We will omit the choice of V
and just write Gr(d, n). A classical result in algebraic geometry realizes Gr(d, n) as a projective variety
Vd
via the Plücker embedding. Specifically, we can define a map Gr(d, n) ֒→ P( kn ) as follows. Given a
d-dimensional subspace spanned by v1 , . . . , vd in Gr(d, n) we send
span(v1 , . . . , vd ) 7→ v1 ∧ · · · ∧ vd .
This choice of basis is not unique, but when we apply a change of basis we scale the wedge product by
the determinant and so this map is well defined on projective space. The rth secant variety of the Plücker
V
embedding
of Gr(d, n) denoted Secr (ρ(Gr(d, n)), is the Zariski closure in P( d kn ) of the set of expressions
Pr
i=0 xi where xi is in the embedded Grassmannian. Our convention is that the zeroth secant variety is the
original variety.
Secant varieties have long been a topic of interest in algebraic geometry. Despite this, very little is known
about their algebraic structure. Many results about secant varieties focus on the dimension of the space or
finding bounds on the degrees of set theoretic generators [DE, DK]. Ideal-theoretic generators are hard to
find [MM, LM, LO] and accordingly are not well understood.
Specifically for the Plücker embedding, a good amount is known about the dimensions of these secant
varieties [CGG, BDdG]. Some set-theoretic results are also known. For example, in [KPRS] the authors
prove that all Plücker embeddings are generated set theoretically by pullbacks of the Klein quadric.
Recently, in [DE] the authors greatly expand the scope of [KPRS] to show that for any fixed r, the rth
secant variety of the Plücker-embedded Gr(d, n) is defined set theoretically by polynomials of bounded degree
independent of d and n. They pose a question at the end of their paper about whether the ideal-theoretic
version of their theorem holds. Furthermore, they mention that the ideas present in their paper will not
suffice to address the ideal-theoretic version.
The purpose of this paper is to answer this question in the affirmative in characteristic 0. We ultimately
prove the following:
Theorem 1.1. Assume char(k) = 0. For each r ≥ 0, there is a constant C(r) such that the prime ideal of
the rth secant variety of the Plücker-embedded Gr(d, n), is generated by polynomials of degree ≤ C(r), where
C(r) does not depend on the choice of d or n.
This theorem has an immediate corollary resulting from the proof techniques. Exact descriptions of ∗g
and ·, can be found in §3.
Date: March 28, 2018.
2010 Mathematics Subject Classification. 13E05, 13D02, 14M15, 15A69, 16T15.
RL was supported by NSF grant DMS-1502553.
1
2
ROBERT P. LAUDONE
Corollary 1.2. Assume char(k) = 0. For r ≥ 0, the equations for the rth secant variety of the Plücker
embedding of any Gr(d, n) can be built out of finitely many equations f1 , . . . , fN of degree bounded by C(r)
under the operations ∗g and ·.
The main idea in proving Theorem 1.1 is to combine all of the ideals of the Plücker-embedded Grassmannians into a Hopf ring PΣ which we define in §3. We then prove noetherianity results with respect to the
additional structure on this ring.
Once we define the Hopf ring PΣ and show it is noetherian, it is natural to ask if all finitely generated
modules over PΣ are noetherian. To address this question, we must abstract the structure we develop in
proving Theorem 1.1 to the language of functor categories as seen in [Sa2, CEF, SS1]. After we transition to
this language, we use the new tools available to develop a syzygy theory for Plücker embedded Grassmannians
analogous to the ∆-modules seen in [Sn] and the Veronese theory in [Sa2], in particular we prove the following:
Theorem 1.3. There is a function C(i, r), depending on i, r, but independent of d, n, such that the ith syzygy
module of the coordinate ring of the rth secant variety of the Plücker-embedded Gr(d, n) is concentrated in
degrees bounded by C(i, r).
This theorem ultimately encapsulates Theorem 1.1 when we take i = 1, but the structure we develop to
prove Theorem 1.1 is crucial in proving Theorem 1.3.
1.1
Outline of Argument
The proof of Theorem 1.1 breaks into the following steps:
(1) For fixed r ≥ 0, we reduce to considering Gr(d, (r + 2)d) as d varies. This will allow us to bound
the degrees of the ideal generators for the rth secant varieties of any Gr(d, n). For this we use [MM,
Proposition 5.7] as explained in §5.
Vd
L
(2) We now consider all values of d via the space PΣ = n,d Symn ( k(r+2)d ), where k is a field of
characteristic 0. If V is a (r + 2)d dimensional vector space over k, we know V ∼
= k(r+2)d , so it
suffices to consider PΣ . As in [Sa1], we observe that there are two products on this space: the
usual “external” product that multiplies outside symmetric powers and a new “internal” product
that multiplies inside exterior powers up to an increasing change of index. We show that subspaces
of this space which are ideals for both products are finitely generated. The key insight is that in
an infinite antichain of monomials in this space, both n and d cannot be unbounded, this is seen
in §2. The internal product involves symmetrizations and so we must assume that the field k has
characteristic 0. This step is done in §3 with the key preparations in §2.
(3) Finally, we notice that the two products are compatible with the standard comultiplication on the
symmetric algebra PΣ . We can define secant varieties in terms of comultiplication. Using this
structure, we prove the essential fact the ideal of the rth secant variety of the direct sum of all the
Plücker ideals corresponding to Gr(d, (r + 2)d) as d varies, is an ideal in PΣ with respect to both
products. So using the above, because the (d, n)-bigraded component of this ideal corresponds to all
degree n polynomials in the rth secant variety of the Plücker embedding of Gr(d, (r + 2)d), we can
deduce finite generation. This result is stated in §4 with most of the preparation and work done in
§3.
In the last two sections, the proof of Theorem 1.3 breaks into the following steps:
(1) We translate the structure of PM from §2 to the language of functor categories, by developing a
category GM whose principal projective generated in degree (0, 0) corresponds exactly to PM and
whose other morphisms (d, m) → (e, n) encapsulate multiplication from the (d, m) bigraded piece of
PM to the (e, n) bigraded piece. We then use the results from §2 to show GM is a Gröbner category
as defined in [SS1]. The bulk of this is done in §6.
(2) We then define a symmetrized version of GM called GM whose principal projective generated in
degree (0, 0) corresponds to (PΣ )M as seen in §3. At the end of §6, we use the fact that GM is
Gröbner to prove that every finitely generated GM -module is noetherian.
(3) With this structure we can study free resolutions of secant ideals of Plücker embedded Grassmannians. In §7, we find a particular free resolution using the principal projectives in GM which allows us
to ultimately deduce Theorem 1.3.
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
1.2
3
Relation to previous work
• Sec0 (ρ(Gr(d, n)) is just the Plücker-embedded Grassmannian. It is well known that its ideal is
generated by quadratic polynomials (the Plücker equations), so C(0) = 2. The case for d = 2 is well
known, in particular the ideal of Secr (ρ(Gr(2, n))) is generated in degree r + 2 by sub-Phaffians of
size 2r + 4 [LO, §10]. This implies a lower bound, C(r) ≥ r + 2, but outside of this we know very
little about C(r).
• As mentioned the Veronese case was addressed in [Sa1]. We address the Plücker case in this paper.
Snowden developed ∆-modules in [Sn] to prove a boundedness result about the syzygies of the Segre
embeddings. The question still remains, are the ideals of the secant varieties of the Segre embeddings
defined in bounded degree? Can these techniques be used to address the Segre case and ultimately
prove results about the syzygies of secant varieties as well?
• Rather than look at all Plücker embeddings of Grassmannians, one can consider all Segre embeddings
of products of projective spaces or Veronese embeddings of projective space. If a Segre analogue of
these methods can be developed, could it also apply to Segre-Veronese embeddings?
• In general, computing the ideals of secant varieties is difficult. We refer the reader to [MM, LO, LW]
for references concerning these explicit computations for some cases of the Segre, Veronese and
Plücker embeddings.
• The idea for showing ideals in PΣ are finitely generated was motivated mainly by work in [Sa1].
The underlying idea in most of this work is noetherianity up to symmetry. For a nice introduction
we recommend [D]. Ultimately, one works with a space or object on which a group acts and proves
finite generation up to the action of this group. This idea is essential in [SS1, SS2, SS3, NSS], where
the authors explore various manifestations of this idea to prove finite generation results for various
representations of categories and twisted commutative algebras. These ideas are also present in
[CEF, Sn, DE, DK, Hi, HS, To] where they were used to prove more surprising stability theorems.
1.3
Conventions
For the most part, k will denote a field of characteristic 0. In §2, this assumption is not necessary and so
we let k be any commutative noetherian ring, but the assumption is needed in the following sections. We
always tensor over k.
We always denote by Σn the symmetric group on n letters, and we denote the set {1, . . . , n} by [n].
Vd
Given a vector space V,
V denotes its dth exterior power. Similarly, Symd V denotes the dth symL
metric power and Sym(V) = d≥0 Symd V.
Acknowledgements. I thank Steven Sam for directing me towards this problem, and for his constant
guidance and helpful conversations.
2.
Shuffle-Star Algebra
Fix a commutative noetherian ring k (for our purposes taking k to be a field suffices, but the general case
has the same proof and could be useful in the future).
For a fixed M ∈ Z≥0 , consider the following algebra:
PM =
d
M^
( kMd )⊗n .
d,n
In general, we will suppress the subscript M , only using it when the value of M will affect the definition
Vd Md
k
and
or result. A monomial of P is an element of the form w1 ⊗ · · · ⊗ wn where each wi ∈
wi = α(vj1 ∧ · · · ∧ vjd ) where vji come from the standard basis for kMd and α ∈ k.
We define two multiplication structures on P. The first is the same as the shuffle product in [Sa1], we
recall it for sake of completeness. Pick a subset {i1 < · · · < in } of [n + m] and let {j1 < · · · < jm } be its
complement; denote this pair of subsets by σ and call it a split of [n + m]. A split defines a shuffle product
d
d
d
^
^
^
( kMd )⊗n ·σ ( kMd )⊗m → ( kMd )⊗(n+m) ,
4
ROBERT P. LAUDONE
where (u1 ⊗ · · · ⊗ un ) ·σ (v1 ⊗ · · · ⊗ vm ) = w1 ⊗ · · · ⊗ wn+m with wik = uk and wjk = vk . Whenever we write
f ·σ g, we are implicitly assuming that σ is a split of the correct format, otherwise define it to be 0.
Notice there is an action of Σ∞ on P by permuting all possible indices. We recall that for f ∈ P, the
width of f , denoted w(f ), is the smallest integer n such that for every σ ∈ Σ∞ that fixes {1, . . . , n}, σ also
fixes f . Every element of P must be a finite linear combination of monomials, so only finite many indexes can
appear. This means every f ∈ P has finite width, or equivalently that P satisfies the finite width condition.
We recall the definition of the monoid of increasing functions:
Inc(N) = {ρ : N → N | ∀a < b, ρ(a) < ρ(b)}.
Since P carries an action of Σ∞ , there is a natural action of Inc(N) on P as follows. Fix f ∈ P, for any
σ ∈ Σ∞ , σf only depends on σ|[w(f )] considering σ as a function N → N. For any ρ ∈ Inc(N), there exists
some σ ∈ Σ∞ such that ρ|[w(f )] = σ|[w(f )] , define ρf = σf . The same argument presented in [DK, Pages 6-7]
shows that this gives a well defined action of Inc(N) on P.
We define a new product ∗g where g ∈ Inc(N). For monomials we define,
d+e
e
d
^
^
^
Me ⊗n
Md ⊗n
kM(e+d) )⊗n .
( k ) ∗g ( k ) → (
as follows. We first require g([M d]) ⊆ [M (e + d)], otherwise we define the product to be zero. Whenever
we use this product, we will always implicitly assume that g satisfies this property. Suppose g([M d]) =
{α1 , . . . , αMd } ⊂ [M (e + d)] with g(i) = αi , also let {β1 , . . . , βMe } = [M (e + d)] \ g([M d]). For ease of
notation, we define g c ∈ Inc(N) to be g c (i) = βi , we call this the complement of g. Then for monomials,
(vi1 ∧ · · · ∧ vid ) ⊗ · · · ⊗ (vind−n+1 ∧ · · · ∧ vind ) ∗g (vj1 ∧ · · · ∧ vjd ) ⊗ · · · ⊗ (vjnd−n+1 ∧ · · · ∧ vjnd )
= g(vi1 ∧ · · · ∧ vid ) ∧ g c (vj1 ∧ · · · ∧ vjd ) ⊗ · · · ⊗ g(vind−n+1 ∧ · · · ∧ vind ) ∧ g c (vjnd−n+1 ∧ · · · ∧ vjnd )
= (vαi1 ∧ · · · ∧ vαid ∧ vβj1 ∧ · · · ∧ vβjd ) ⊗ · · · ⊗ (vαind−n+1 ∧ · · · ∧ vαind ∧ vβjnd−n+1 ∧ · · · ∧ vβjnd )
Where we view the vαij and vβij as the standard basis vectors in kM(d+e) . For general f ∈ Pd,n , h ∈ Pe,n
and g ∈ E extend bilinearly. Extend to the rest of P by declaring all other products to be 0. We first notice
a few properties of these products. There is a modified associativity.
Lemma 2.1. Given f ∈ Pd,n , b ∈ Pd,m and a ∈ Pe,n+m . and a split σ of [n + m] and g ∈ E, there exist
p1 , . . . , pr ∈ Pe,n and h1 , . . . , hr ∈ Pd+e,m so that,
(f ·σ b) ∗g a =
r
X
hi ·σ (f ∗g pi ).
i=1
Proof. Both ·σ and ∗g are bilinear, so assume without loss of generality that both a and b are monomials.
Write a = α1 ⊗ · · · ⊗ αn+m and b = β1 ⊗ · · · ⊗ βm . Suppose σ is the split {i1 , . . . , im }, {j1 , . . . , jn } of [n + m].
Furthermore, suppose [M (d + e)] \ g([M d]) = {γ1 , . . . , γMe }, let α′i be the image of αi under identifying kMe
with the subspace of kM(e+d) spanned by the standard basis vectors {vγi }. So if αi = wj1 ∧ · · · ∧ wje , with
{wi } the standard basis for kMe , then α′i = vγj1 ∧ · · · ∧ vγjd .
Taking hi = g(β1 ) ∧ α′i1 ⊗ · · · ⊗ g(βm ) ∧ α′im and pi = αj1 ⊗ · · · ⊗ αjn makes the identity valid.
We can use the two products to define an ideal in P as follows,
Definition 2.2. A homogeneous subspace I ⊂ P is an ideal if f ∈ I implies that f ∗g h ∈ I and h ·σ f ∈ I
for all h ∈ P. A subset of elements of P generates an ideal I if I is the smallest ideal that contains the
subset.
With this new language, we get an immediate corollary of Lemma 2.1:
Corollary 2.3. If f1 , f2 , . . . generate an ideal I, then every element of I can be written as a sum of elements
of the form h ·σ (fi ∗g a) where a, h ∈ P.
We wish to use Gröbner methods to prove noetherianity of P, to do this we will work with monomial
ideals. In our context, an ideal is a monomial ideal if it has a generating set of monomials. Notice, the
product of two monomials under our operations is still a monomial. We will show that all monomial ideals
are finitely generated. We will then use this to show that all ideals in P are finitely generated. To do this
we must first make some definitions and reformulations. Each monomial m ∈ Pd,n can be encoded as a
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
5
reading list (RL), Sd,n = (S 1 , . . . , S n ). Where each S i is an increasing word of length d created from the
finite alphabet [M d]. In particular, |S i | = d with i = 1, . . . , n. S i records the indices in tensor position i.
The subscript indicates which bigraded piece of P the monomial is in. Let RL denote the set of reading
lists.
Example 2.4. To get an idea of what this looks like we give some examples encoding in both directions. A
basic example would be the monomial v1 ∧ v2 ⊗ v1 ∧ v3 corresponds to,
((1, 2), (1, 3)).
As a more complicated example consider,
((1, 5, 6), (1, 2, 4), (1, 2, 6)).
This corresponds to the monomial (v1 ∧ v5 ∧ v6 ) ⊗ (v1 ∧ v2 ∧ v4 ) ⊗ (v1 ∧ v2 ∧ v6 ).
Suppose we have a monomial ideal J of P that is not finitely generated. Then there is an infinite list of
monomials m1 , m2 , . . . such that the ideal generated by m1 , . . . , mi does not contain mi+1 . This list translates
to an infinite list of RLs that are incomparable under the following order. Given a RL Sd,n = (S 1 , . . . , S n ),
this corresponds uniquely to a monomial
vS 1 ⊗ · · · ⊗ vS n .
We say that Sd,n ≤ Te,m if and only if vT 1 ⊗ · · · ⊗ vT m is in the ideal generated by vS 1 ⊗ · · · ⊗ vS n . Call this
the monomial order on RLs. Notice the following:
Lemma 2.5. For RLs, Sd,n ≤ Te,m in the monomial order is equivalent to the existence of a map f : Sd,n →
Te,m with the following properties. If Sd,n = (S 1 , . . . , S n ) and Te,m = (T 1 , . . . , T m ), f = (f1 , . . . , fn ) where
fi : Si → Tki with the following properties:
(1) k1 < k2 < · · · < kn .
(2) fi (Si ∩ Sj ) = fj (Si ∩ Sj ), so they agree on overlap.
(3) Each fi ∈ Inc(N).
Proof. This follows easily from definition. The first property means we must map the tensor positions in
order. The last property is necessary because we have g ∈ Inc(N).
We will use this equivalence often. We define a new relation on the set RL ≤RL , where Sd,n ≤RL Te,m
if a map as described in Lemma 2.5 exists from Sd,n → Te,m . This is easily seen to be a partial order. The
above discussion can be summarized as follows:
Corollary 2.6. An infinite list of monomials m1 , m2 , . . . in P such that mi+1 is not contained in the ideal
generated by m1 , . . . , mi induces an infinite chain of incomparable RLs with respect to ≤RL .
Given this infinite antichain of RLs from Corollary 2.6 we will show that either n or d must be bounded.
Lemma 2.7. Given an infinite antichain of reading lists under ≤RL ,
Sd1 ,n1 , Sd2 ,n2 , . . .
either the ni or the di must be bounded.
Proof. Suppose this is not the case. Fix Sd1 ,n1 in this antichain, i.e. the first element. If both dj and nj
are unbounded, we can assume dj > 2M d1 . For any RL Tdj ,nj = (T 1 , T 2 , . . . , T nj ) with dj > 2M d1 in our
dj
unique sub-lists of size M d1 in each list T i of size dj in Tdj ,nj . There are a total
antichain, there are Md
1
i
j
of Md
Md1 possible lists of this size that could occur in any T . Hence if
nj > n1
M dj
(M
d1 )
,
dj
(M d1 )
by the pigeon hole principal we will have at least n1 tensor positions (lists) whose intersection has size greater
than M d1 .
( M dj )
dj
different sub-lists of size M d1 . If
To see this, notice if nj > n1 Mddj 1 , for each list of size dj , we have Md
1
( M d1 )
Mdj
dj
we have a hole for each list of size d1 , there are Md
holes.
For
each
list
of
side
d
,
we
place
pigeons
j
Md
1
1
into distinct holes. The holes are distinct because there are no repeated numbers. Each of these pigeons
6
ROBERT P. LAUDONE
( M dj )
represents a tensor position that the list appears in. If there are more than n1 Mddj 1 lists, this implies we
(M d1 )
Mdj
j
holes.
The
pigeon
hole
principal
implies that there is one
pigeons
in
the
place more than n1 Md
Md1
Md1
hole with at least n1 distinct elements in it.
Hence there is some list of size M d1 that occurs in at least n1 different tensor positions. Such a list cannot
occur twice in the same tensor position because tensor positions cannot contain repeated numbers. As a
result, there are at least n1 tensor positions whose intersection contains at least M d1 numbers.
We will show this lower bound is independent of dj . In particular,
n1 M Md1 2Md1 −1 > n1
M dj
(M
d1 )
,
(Mddj 1 )
This is because
M dj
(M
d1 )
=
dj
(M d1 )
(Mdj )(Mdj −1)···(Mdj −Md1 +1)
(dj )(dj −1)···(dj −Md1 +1)
If we pair each of these terms as
a
dj − M
dj −a
=
Md
1
1
M M d1 (dj )(dj − M )···(dj − M 1 + M )
.
(dj )(dj −1)···(dj −Md1 +1)
for a = 0, . . . , (M d1 − 1), we see that the largest of the terms is
Md1 −1
dj − M
dj −(Md1 −1) .
This can easily be checked because
a
dj − M
dj −a
<
a+1
dj − M
dj −(a+1) ,
reduces to the inequality
dj
M
< dj .
Which is clearly true so long as dj is positive, which it is. We now claim,
1≤
Md1 −1
dj − M
dj −(Md1 −1)
< 2.
To see this we clear denominators since dj > 2M d1 we know the denominator is nonzero. So this inequality
is equivalent to
1
1
dj − M d1 + 1 ≤ dj − Md
(2.8)
M + M < 2dj − 2M d1 + 2.
The first inequality reduces to
M−1
M ≤ (M − 1)d1 ,
which is clearly true. The second inequality reduces to
(2M − 1)d1 −
2M−1
M
< dj .
However we took dj > 2M d1 , so this is also true. This implies that every term in the product is in the
interval [1, 2). Clearly every term of the form
a
dj − M
dj −a
,
is greater than 1 for a = 1, . . . , (d1 − 1) and every term is less than
2. Hence
1≤
Md
1
1
(dj )(dj − M )···(dj − M 1 + M )
(dj )(dj −1)···(dj −Md1 +1)
d −1
dj − 1M
dj −(d1 −1)
which we showed is less than
< 2Md1 −1 .
Notice this bound applies regardless of dj so long as dj > 2M d1 . This implies
d
1
1
M M d1 (dj )(dj − 2 )···(dj − 21 + 2 )
(dj )(dj −1)···(dj −d1 +1)
≤ M Md1 2Md1 −1 .
So if nj > n1 M Md1 2Md1 −1 and dj > 2M d1 , we can find n1 tensor positions whose common intersection has
size at least M d1 . If nj and dj are unbounded for any sequence if we fix some d1 and n1 , this will always
occur.
Say the n1 tensor positions we find are i1 , . . . , in1 and T i1 ∩ · · · ∩ T in1 contains the M d1 elements j1 <
· · · < jMd1 . Then we have a clear map Sd1 ,n1 → Tdj ,nj where we map S n → T in and send i 7→ ji . This is an
order preserving injection that satisfies all the properties of Lemma 2.5 and implies Sd1 ,n1 ≤ Tdj ,nj which is
a contradiction.
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
7
Remark 2.9. We believe the idea in Lemma 2.7 could have further applications in showing various posets
are noetherian. In particular, the idea is to fix a small element in any given antichain and assume that the
size of the elements in this antichain grow arbitrarily. With this assumption we prove the resulting elements
are forced to eventually contain a structure that resembles the fixed element. We then deduce that the size
of the elements must be bounded in some sense. This often drastically simplifies the problem as we will see
below.
This leaves us with two cases. Either di or ni is bounded. We will show that both lead to a contradiction.
Lemma 2.10. Given an infinite antichain in (RL, ≤RL ),
Sd1 ,n1 , Sd2 ,n2 , . . .
ni cannot be bounded.
Proof. Assume ni is bounded. Then since our antichain is infinite, this implies that there must be some
infinite subchain of our antichain with ni constant. Restrict our attention to this sub-antichain with ni = n.
We will now embed each Sdi ,n of this antichain as a labeled tree and derive a contradiction via Kruskal’s
tree theorem.
Send each Sdi ,n to the tree TSdi ,n with a root vertex labeled (0, 0, (n + 1, n + 1, . . . , n + 1)). Define the
function ψ(j, Sdi ,n ), which takes as an input some RL Sdi ,n and some element j ∈ [M di ] and returns the
finite list of size n with a k in position k if j appears in S k and a zero if j does not appear in S k . This
encodes which of the tensor positions j appears in.
There will be n branches off of the root. Branch j will have M di vertices. Vertex k of branch j will be
labeled by (k, j, ψ(k, Sdi ,n )). Order the first label with the standard order on Z≥0 . Order the last two labels
with the componentwise subsequence order, in this case the quasi-well-order will be equality. This product
is a quasi-well-order by Dickson’s lemma because the alphabets for the last two labels are finite. Hence, each
component is a quasi-well order and Dickson’s lemma tells us that the finite product of quasi-well orders
compared componentwise is also a quasi-well order.
Notice that this is an injective mapping from RLs to trees because we can easily recover Sd,n from TS by
reading off vertices.
Furthermore, using the order described in Kruskal’s tree theorem if TS ≤ TW , this implies that S ≤ W .
We must send S i to W i because of the second label. Also, for any k, ψ(k, Sdi ,n ) is fixed and we have that
TS ≤ TW if every vertex v maps to some vertex F (v) with v ≤ F (v). In combination with the first label, this
implies that every number k maps to some number m ≥ k with ψ(k, Sdi ,n ) = ψ(m, Sdi ,n ). As a result, in
each of the branches where k occurs, there is some number m ≥ k that it can map to because m will occur
in all of the remaining branches in which k occurs.
Additionally, we must map branches to branches. Hence if a vertex v with first label k maps to a vertex
F (v) with first label mk , we have m1 < m2 < · · · < mMdi , so the map on indices is in the monoid of
increasing functions.
Define a mapping inductively form S → W . Begin by sending 1 to the minimal first label m that occurs
for all the vertices corresponding to a vertex with first label 1. That is, let {v1 , . . . , vℓ } be all the vertices in
the tree TS with first entry 1. Each vi has an image F (vi ) in TW . Out of all the vertices {F (v1 ), . . . , F (vℓ )},
send 1 to the minimum first entry that occurs, call it α1 . By construction if 1 occurs in any branch, so must
α1 . Hence if 1 occurs in S j it has a well defined image in the corresponding W j . Put f (1) = α1 .
Now repeat the same procedure for 2. The element we send 2 to cannot be α1 because even if 1 and 2
occur in all the same branches, any vertex corresponding to a 1 occurs earlier in the branch that the vertex
corresponding to a 2 and we preserve this order. Hence if α1 is the minimal element for 2, this would imply
α1 is not minimal for 1. Continue in this way until we have a mapping of all the numbers occurring in S.
By construction, two numbers could map to the same βj if and only if they occur in exactly the same
branches. So F being well defined implies that f is well defined, i.e. that every element has an image.
Furthermore, f is a map of RLs because if a vertex with first entry j is mapped to another vertex with
first entry m, m must occur in all of the branches that j does. Hence when we map j to m, j has an image
in each restriction. The map also satisfies property (4) in Lemma 2.5 because we must stay within a branch
and we can only map a number to another number that is greater than or equal to it. Finally, S j must map
into W j because the third label on each vertex must be equal.
8
ROBERT P. LAUDONE
The contrapositive implies that our infinite antichain of RLs yields an infinite antichain of trees. This
contradicts Kruskal’s tree theorem.
Remark 2.11. We use Kruskal’s tree theorem because it provides a more intuitive picture, but Higman’s
lemma [D, Theorem 1.3] would suffice. We could encode each branch as an element of a quasi-well-ordered
set and then view the trees as words of length n over this quasi-well-ordered set. The mapping of trees in
this context is equivalent to one of these words being a subsequence of the other.
Example 2.12. To see this proof in action, consider the following example. Suppose n = 3 is fixed and
M = 2. Given the two trees
(1, 1, (1, 0, 3))
✐
✐
✐
✐
✐✐✐✐
(0, 0, (4, 4, 4))
(1, 2, (1, 0, 3))
❯❯❯❯
❯❯❯❯
(1, 3, (1, 0, 3))
(2, 1, (1, 2, 0))
(3, 1, (0, 2, 0))
(4, 1, (0, 0, 3))
(2, 2, (1, 2, 0))
(3, 2, (0, 2, 0))
(4, 2, (0, 0, 3))
(2, 3, (1, 2, 0))
(3, 3, (0, 2, 0))
(4, 3, (0, 0, 3))
(2.13)
and
(1, 1, (1, 2, 0))
(2, 1, (1, 0, 3))
(3, 1, (1, 2, 0))
(4, 1, (0, 2, 0))
(5, 1, (0, 0, 3))
(6, 1, (0, 0, 3)) .
(2, 2, (1, 0, 3))
(3, 2, (1, 2, 0))
(4, 2, (0, 2, 0))
(5, 2, (0, 0, 3))
(6, 2, (0, 0, 3))
(2, 3, (1, 0, 3))
(3, 3, (1, 2, 0))
(4, 3, (0, 2, 0))
(5, 3, (0, 0, 3))
(6, 3, (0, 0, 3))
(2.14)
❧
❧❧❧
❧❧❧
(0, 0, (4, 4, 4))
(1, 2, (1, 2, 0))
❘❘❘
❘❘❘
❘
(1, 3, (1, 2, 0))
Call these T1 and T2 . It is easy to read off their corresponding RLs. T1 has RL S1 = ((1, 2), (2, 3), (1, 4))
and T2 has RL S2 = ((1, 2, 3), (1, 3, 4), (2, 5, 6)). Notice T1 ≤ T2 in the order described in Kruskal’s tree
theorem where we map the vertices as below
(1, 1, (1, 0, 3)) 7→ (2, 1, (1, 0, 3))
(2, 1, (1, 2, 0)) 7→ (3, 1, (1, 2, 0))
(3, 1, (0, 2, 0)) 7→ (4, 1, (0, 2, 0))
(4, 1, (0, 0, 3)) 7→ (5, 1, (0, 0, 3))
(1, 2, (1, 0, 3)) 7→ (2, 1, (1, 0, 3))
(2, 2, (1, 2, 0)) 7→ (3, 2, (1, 2, 0))
(3, 2, (0, 2, 0)) 7→ (4, 2, (0, 2, 0))
(4, 2, (0, 0, 3)) 7→ (6, 2, (0, 0, 3))
(1, 3, (1, 0, 3)) 7→ (2, 3, (1, 0, 3))
(2, 3, (1, 2, 0)) 7→ (3, 3, (1, 2, 0))
(3, 3, (0, 2, 0)) 7→ (4, 3, (0, 2, 0))
(4, 3, (0, 0, 3)) 7→ (5, 3, (0, 0, 3))
There can be multiple embeddings, but we choose one of them. This gives us a map from S1 → S2 inductively
as described in the proof. We see that 1 only maps to a vertex with first label 2. So we let f (1) = 2. Now
we see that 2 only maps to vertices with first label 3. So f (2) = 3. Continuing we have f (3) = 4. Now 4
maps to vertices with different first labels, the set of first labels is {5, 6}. We then map 4 to the minimal
such label that has not yet been used, so f (4) = 5.
This map f induces a map from each component of S1 to each component of S2 , so we can easily define
fi : S1i → S2i by restriction. Clearly this satisfies property (1) of Lemma 2.5. Furthermore, f satisfies
properties (2) and (3) of Lemma 2.5 because we defined it via restriction and f satisfies property (4) by
construction.
As described above this gives us
[(v1 ∧ v2 ) ⊗ (v2 ∧ v3 ) ⊗ (v1 ∧ v4 )] ∗f (v1 ⊗ v1 ⊗ v2 ) = (v2 ∧ v3 ∧ v1 ) ⊗ (v3 ∧ v4 ∧ v1 ) ⊗ (v2 ∧ v5 ∧ v6 )
Where f is the map found above.
Lemma 2.10 implies that in our infinite antichain we must have di bounded. Furthermore, because this
is an infinite chain and only finitely many di can occur, we can find an infinite subchain with a fixed d.
This implies that for this d, we have an infinite antichain in
!⊗n
d
M ^
Md
.
k
n≥0
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
9
Vd Md
However there are βd = Md
basis vectors for
k , order them in some way as {e1 , . . . , eβd }. Then
d
every monomial in this algebra is a word in the ei . As there are finitely many basis vectors, Higman’s lemma
implies that for any infinite sequence of such words, two are comparable. Hence we cannot possibly have an
infinite antichain.
This discussion proves the following,
Lemma 2.15. Given an infinite antichain in (RL, ≤RL ),
Sd1 ,n1 , Sd2 ,n2 , . . .
di cannot be bounded.
The results above cumulatively show,
Theorem 2.16. The poset (RL, ≤RL ) is noetherian.
Proof. Suppose there were an infinite anti-chain of RLs. From Lemma 2.7 we know that either the n or d
that appear in this list must be bounded otherwise it could not be an antichain. However this cannot be the
case via Lemmas 2.15 and 2.10. So no such antichain exists.
An immediate consequence of this theorem is the following,
Theorem 2.17. All monomial ideals of P are finitely generated.
Proof. From Corollary 2.6 if there is an infinite anti-chain of monomials, this leads to an infinite anti-chain
of RLs under the order described in Lemma 2.5. This contradicts Theorem 2.16. Hence all monomial ideals
of P are finitely generated.
Now place a total ordering on monomials (ignoring coefficients) of the same bidegree (d, n) as follows.
Encode each monomial by (Zd )n via the dth index in the nth tensor position. First, define on Zd using
lexicographic ordering, i.e. (a1 , . . . , ad ) (b1 , . . . , bd ) if the first nonzero element of (b1 − a1 , . . . , bd − ad ) is
positive. Then compare tensors using lexicographic ordering. We only work with bihomogeneous elements,
so it is not necessary to find a way to compare elements of different bidegrees.
Lemma 2.18. Let m, m′ , n be monomials. For any g ∈ Inc(N) and σ a splitting,
(1) If m m′ , then m ∗g n m′ ∗g n.
(2) For any n, if m m′ then n ·σ m n ·σ m.
Proof. The proof of the first claim follows because g ∈ Inc(N). The second follows because we never change
any indices.
Given f ∈ Pd,n we define init(f ) as the largest monomial along with its coefficient with respect to that
has a nonzero coefficient in f . Given an ideal I, let init(I) be the k-span of {init(f ) | f ∈ I homogeneous}.
Lemma 2.19. If I is an ideal, then init(I) is a monomial ideal.
Proof. If m ∈ init(I), we have m = init(f ) for f ∈ I. For any monomial n ∈ P, we have n ·σ m = init(n ·σ f )
by Lemma 2.18 part (2) Furthermore, we claim that m ∗g n = init(f ∗g n). Using the bilinearity of ∗g , the
result follows from Lemma 2.18 part (1).
So m ∗g n, n ·σ m ∈ init(I). To generalize to any h ∈ P, because ∗g is bilinear it suffices to work with
monomials and the above implies the desired result. Hence, init(I) is a monomial ideal.
Lemma 2.20. If I ⊂ J are ideals and init(I) = init(J), then I = J. In particular, if f1 , f2 , · · · ∈ J and
init(f1 ), init(f2 ), . . . generate init(J) then f1 , f2 , . . . generate J.
Proof. Suppose I does not equal J. Pick f ∈ J \ I where init(f ) is minimal with respect to . Then because
init(I) = init(J), we have init(f ) = init(f ′ ) for some f ′ ∈ I. But init(f − f ′ ) is strictly smaller than init(f )
and f − f ′ ∈ J \ I. This is a contradiction.
For the other statement, let I be the ideal generated by f1 , f2 , . . . .
Corollary 2.21. Every ideal of P is finitely generated.
Proof. Combine Corollary 2.17, Lemma 2.19 and Lemma 2.20.
10
ROBERT P. LAUDONE
3.
Symmetrizing
L
Vd
What we are really interested in is n,d≥0 Symn ( kMd ) for fixed M . So we must symmetrize. Assume
k is a field of characteristic 0. Much of this section is translating the structure of [Sa1, §3] to suit P. Define
!Σn
!
d
d
M
^
M
^
Σ
Md ⊗n
Md ⊗n
PM =
( k )
(PΣ )M =
( k )
,
n,d≥0
n,d≥0
Σn
where Σn acts by permuting the tensor factors, and the superscript and subscript respectively denote taking
invariants and coinvariants. We generally suppress the additional M subscript and when it matters explicitly
mention which M we are working with, for ease of notation writing just P Σ or PΣ .
Our internal product ∗g respects the structure of P Σ and so P Σ is a subalgebra of P with respect to it.
Unfortunately the shuffle product does not respect the symmetric invariance and so P Σ is not closed under
it. We remedy this by defining
X
f ·g =
f ·σ g.
σ
Averaging over all splits produces symmetric elements and so P Σ is naturally closed under this new ·.
Additionally, · is both commutative and associative. Both of these algebras are naturally bigraded by (d, n).
Σ
We denote these bigraded pieces by Pd,n
and (PΣ )d,n . We will only consider bi-homogeneous subspaces of
Σ
P and PΣ .
For each d, n define a linear projection
Σ
π : Pd,n → Pd,n
1 X
wσ(1) ⊗ · · · ⊗ wσ(n) .
w1 ⊗ · · · ⊗ wn 7→
n!
σ∈Σn
Σ
Pd,n
,
If f ∈
we have π(f ) = f , so π is surjective. We let π ′ = n!π. We also denote the direct sum of these
maps by π : P → P Σ and π ′ : P → P Σ . Next, define
Σ
G : (PΣ )d,n → Pd,n
X
1
wσ(1) ⊗ · · · ⊗ wσ(n) .
w1 · · · wn 7→ n!
σ∈Σn
Then G is a linear isomorphism, since G is the inverse of the composition
Σ
Pd,n
→ Pd,n → (PΣ )d,n ,
where the first map is the natural injection and the second is the natural projection. Denote the direct sum
of these maps by G : PΣ → P Σ .
3.1
Properties of P Σ
Σ
Lemma 3.1.1.
(1) If f ∈ Pd,n
and h ∈ Pe,n are homogeneous, then π(f ∗g h) = f ∗g π(h) and π(h∗g f ) =
π(h) ∗g f .
(2) If f ∈ Pd,n and g ∈ Pd,m , then n+m
π(f ·σ g) = π(f ) · π(g) for any split of [n + m].
n
Proof.
(1) Both π(f ∗g h) and f ∗g π(h) are bilinear in h and f , so without loss we assume h = h1 ⊗· · ·⊗hn
Vd Md
Ve Me
P
k . Then
for hi ∈
k
and that f = σ∈Σn fσ(1) ⊗ · · · ⊗ fσ(n) for fi ∈
f ∗g π(h) =
1 X
gfσ(1) ∧ g c hτ (1) ⊗ · · · ⊗ gfσ(n) ∧ g c hτ (n)
n!
σ,τ ∈Σn
1 X
gfστ (1) ∧ g c hτ (1) ⊗ · · · ⊗ gfστ (n) ∧ g c hτ (n) .
π(f ∗g h) =
n!
σ,τ ∈Σn
These sums are identical. In the second, perform the change of variables σ 7→ στ −1 . This shows
π(f ∗g h) = f ∗g π(h).
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
11
We can apply similar reasoning to the second claim by considering
1 X
π(h) ∗g f =
ghτ (1) ∧ g c fσ(1) ⊗ · · · ⊗ ghτ (n) ∧ g c fσ(n)
n!
σ,τ ∈Σn
1 X
π(h ∗g f ) =
ghτ (1) ∧ g c fστ (1) ⊗ · · · ⊗ ghτ (n) ⊗ g c fστ (n) .
n!
σ,τ ∈Σn
(2) This is the same proof as in [Sa1, Lemma 3.1] because it is the same product.
Definition 3.1.2. A homogeneous subspace I ⊆ P Σ is an ideal if f ∈ I implies that f · h ∈ I and f ∗g h ∈ I
for all h ∈ P σ and g ∈ Inc(N).
We now pass the noetherianity of P to P Σ .
Proposition 3.1.3. Every ideal of P Σ is finitely generated.
Proof. Let J be an ideal of P Σ . As P Σ naturally lies inside P, we can consider the ideal in P generated by
J, call it I. By Corollary 2.21, I is finitely generated, say by f1 , . . . , fN . As J generates I, we may assume
that each of the fi belong to J. We now claim that the fi also generate J as an ideal in P Σ . By Corollary
2.3, every f ∈ J can be written as a sum of terms of the form h ·σ (fi ∗g a) where h, a ∈ P, g ∈ Inc(N). By
Lemma 3.1.1 we have
−1
n+m
π(h ·σ (fi ∗g a)) =
π(h) · (f ∗g π(a)),
n
here h ∈ Pd,n and f ∗g a ∈ Pd,m . By the surjectivity of π, we conclude that every element of J can be written
as a sum of terms of the form h′ · (fi ∗g a′ ) where h′ , a′ ∈ P Σ and g ∈ Inc(N).
L Σ
is a free divided power algebra under ·, and hence is freely generated in degree n = 1.
For fixed d, n Pd,n
Σ
So we can define a comultiplication ∆ : P Σ → P Σ ⊗ P Σ by w 7→ 1 ⊗ w + w ⊗ 1 when w ∈ Pd,1
and requiring
that it is an algebra homomorphism for ·.
We ultimately wish to show that the two products · and ∗g respect this comultiplication structure. To do
this, we first extend the products to P Σ ⊗ P Σ componentwise.
Σ
Σ
Σ
Lemma 3.1.4. Pick x ∈ Pd,n
and y ∈ Pe,m
, and v ∈ Pe,n
. Then
∆(y · v) = ∆(y) · ∆(v)
∆(x ∗g v) = ∆(x) ∗g ∆(v)
Proof. The first identity follows from the definition of ∆.
For the second identity, we once again notice ∆(x ∗g v) and ∆(x) ∗g ∆(v) are both bilinear in x and v.
As a result, we may assume
X
x=
xσ(1) ⊗ · · · ⊗ xσ(n) ,
σ∈Σn
for some x1 ⊗ · · · ⊗ xn ∈ Pd,n and that
v=
X
vσ(1) ⊗ · · · ⊗ vσ(n) ,
σ∈Σn
for some v1 ⊗ · · · ⊗ vn ∈ Pe,n . Then by definition v = v1 · · · vn and x = x1 · · · xn , here we are using the ·
product. We defined ∆ so that it is an algebra homomorphism for ·, so we have
X
π ′ (vS ) ⊗ π ′ (v[n]\S )
∆(v) = ∆(v1 ) · · · ∆(vn ) =
S⊆[n]
∆(x) = ∆(x1 ) · · · ∆(xn ) =
X
S⊆[n]
π ′ (xS ) ⊗ π ′ (x[n]\S )
12
ROBERT P. LAUDONE
where the sum is over all subsets S = {s1 < · · · < sj } of [n] and vS = vs1 ⊗ · · · ⊗ vsj and xs = xs1 ⊗ · · · ⊗ xsj .
This gives
X
X
π ′ (vS ) ⊗ π ′ (v[n]\S )
π ′ (xT ) ⊗ π ′ (x[n]\T ) ∗g
∆(x) ∗g ∆(v) =
T ⊆[n]
=
X
S⊆[n]
π ′ (xT ∗g π ′ (vS )) ⊗ π ′ (x[n]\T ∗g π ′ (v[n]\S )).
(3.1.5)
S,T ⊆[n]
Where in the second equality we use Lemma 3.1.1(1) to write π ′ (xT ) ∗g π ′ (vS ) = π ′ (xT ∗g π ′ (vS )). On the
other hand,
X
gxσ(1) ∧ g c vτ (1) ⊗ · · · ⊗ gxσ(n) ∧ g c vτ (n)
x ∗g v =
σ,τ ∈Σn
=
X
(gxσ(1) ∧ g c v1 ) · · · (gxσ (n) ∧ g c vn ),
σ∈Σn
where in the second sum we are using the · product. In particular,
X X
∆(x ∗g v) =
π ′ (gx ∧ g c v)σ,S ⊗ π ′ (gx ∧ g c v)σ,[n]\S
(3.1.6)
σ∈Σn S⊆[n]
Where (gx ∧ g c v)σ,S = gxσ(s1 ) ∧ g c vs1 ⊗ · · · gxσ(sj ) ∧ g c vsj if S = {s1 < · · · < sj }. We also write [n] \ S =
{sj+1 < · · · < sn }.
The same proof as in [Sa1, Lemma 3.4] shows that the expressions (3.1.5) and (3.1.6) for ∆(x) ∗g ∆(v)
and ∆(x ∗g v) are equal.
We can now present a symmetrized version of Lemma 2.1:
Σ
Σ
Σ
Lemma 3.1.7. Given f ∈ Pd,n
, b ∈ Pd,m
, and a ∈ Pe,n+m
, we have
(f · b) ∗g a = (f ⊗ b) ∗g ∆(a)
Proof. The proof is essentially identical to the proof of Lemma 2.1. Assume that
X
a=
αρ(1) ⊗ · · · ⊗ αρ(n+m) ,
ρ∈Σn+m
and
b=
X
βτ (1) ⊗ · · · ⊗ βτ (m) .
τ ∈Σm
Write aρ = αρ(1) ⊗ · · · ⊗ αρ(n+m) and β τ = βτ (1) ⊗ · · · βτ (m) . Pick a split σ = {i1 , . . . , in }, {j1 , . . . , jm } of
[n + m]. Then
(f ·σ bτ ) ∗g aρ = (f ∗g (αρ(i1 ) ⊗ · · · ⊗ αρ(in ) )) ·σ ((βτ (1) ⊗ · · · ⊗ βτ (m) ) ∗g (αρ(i1 ) ⊗ · · · ⊗ αρ(im ) )),
and (f · b) ∗g a is the sum of these over all choices of ρ, τ, σ. This is exactly (f ⊗ b) ∗g ∆(a).
Remark
Expanding on this componentwise definition of our products in P Σ ⊗ P Σ , if we write
P3.1.8.
(1)
∆(a) = a ⊗ a(2) . Then
X
(f ∗g a(1) ) · (b ∗g a(2) ).
(f · b) ∗g a = (f ⊗ b) ∗g ∆(a) =
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
3.2
13
Properties of PΣ
We are ultimately interested in PΣ . It is much easier to work directly with P Σ and most of this subsection
will be devoted to transferring results from P Σ to PΣ via G.
h under
First we notice that PΣ has an algebra structure. We can define f · h to be the image of fe ·σ e
B → BΣ for any split σ and lifts fe, e
h ∈ P of f, h ∈ PΣ . It is not hard to see that this is independent of the
choice of lists and the choice of split. To define a ∗g -product on PΣ we must rely heavily on P Σ and the fact
that G is a linear isomorphism. We define
f ∗g h = G−1 (G(f ) ∗g G(h)).
(3.2.1)
The reason for constructing this algebra is to prove a bounded generation result about the secant ideals
of Plücker-embedded Grassmannians. We will show how to get this bounded generation from the finite
generation in PΣ for the sum of the Plücker ideals as a consequence of what we have developed. This
example is complicated and skipping it will not detract from ones understanding of the paper. We include
it to both explicitly illustrate the techniques we are developing and demonstrate the need for these more
general techniques due to the difficulty of explicitly working through the easiest possible case, i.e, the 0th
secant ideal.
Example 3.2.2. First, we have the Pieri decomposition,
p
^
⊗
p
^
The symmetric square has a simple description,
!
p
^
2
=
Sym
=
p
M
S(2i ,12(p−i) ) .
i=0
M
S(2i ,12(p−i)) ) .
0≤i≤p,
i≡p (mod 2)
The Plücker equations of the Grassmannian Gr(p, n) span the sum of the representations where i < p.
Taking Hödge dual isomorphisms, we can generate all the Plücker equations from the basic ones f1 , f2 , . . . ,
V2n
where fi ∈ Sym2 (
) is defined as follows,
X
1
(−1)sgn(S) xS x[4n]\S .
fn =
2
S⊂[4n], |S|=2n
To show the Plücker equations are all generated from finitely many equations it suffices to prove that the
ideal generated by the fi in PΣ is finitely generated. To see this explicitly we will show all the fi are generated
by f1 under our products. In particular, we claim
X
(3.2.3)
(−1)sgn(T ) fn ∗g (xσ(1)σ(2) xσ(3)σ(4) ) = γ(n)
2 fn+1
g,σ
Where
γ(n) =
4n
2n
4n+3
2n+1
4n+4
2n+2
·6·
.
The numerator is the total number of terms in the sum. The denominator is the number of terms in fn+1 .
We then divide by 2 because of G. It is not hard to see γ(n) this is an even integer.
We sum over all g ∈ Inc(N), g|[4n] : [4n] → [4n + 4] such that g(1) = 1 and only one g acts trivially
on fn . We also sum over all σ ∈ Σ4 /(Σ2 × Σ2 ) permutations such that σ(1) < σ(2) and σ(3) < σ(4).
Let (j1 < j2 < j3 < j4 ) = [4n + 4] \ g([4n]) where jk is the image of k in [4n + 4]. In this sum,
T = (g(1), g(2), . . . , g(2n), jσ(1) , jσ(2) , g(2n + 1), g(n + 2), . . . , g(4n), jσ(3) , jσ(4) ) and sgn(T ) is the sign of
the permutation that orders T . Tracing through definitions, we find that on monomials
xi1 i2 xi3 i4 ∗g xj1 j2 xj3 j4 = 12 (xg(i1 )g(i2 )j1 j2 xg(i3 )g(i4 )j3 j4 + xg(i1 )g(i2 )j3 j4 xg(i3 )g(i4 )j1 j2 ).
To see this we explicitly calculate what (3.2.1) does to monomials f and h. First we know that
G(xi1 i2 xi3 i4 ) = 12 (xi1 i2 ⊗ xi3 i4 + xi3 i4 ⊗ xi1 i2 ).
14
ROBERT P. LAUDONE
So
G(xi1 i2 xi3 i4 ) ∗g G(xj1 j2 xj3 j4 ) = 14 ((xi1 i2 ⊗ xi3 i4 + xi3 i4 ⊗ xi1 i2 ) ∗g (xj1 j2 ⊗ xj3 j4 + xj3 j4 ⊗ xj1 j2 )).
Expanding by linearity this becomes
1
4
xg(i1 )g(i2 )j1 j2 ⊗ xg(i3 )g(i4 )j3 j4 + xg(i3 )g(i4 )j3 j4 ⊗ xg(i1 )g(i2 )j1 j2
+
1
4
xg(i1 )g(i2 )j3 j4 ⊗ xg(i3 )g(i4 )j1 j2 + xg(i3 )g(i4 )j1 j2 ⊗ xg(i1 )g(i2 )j3 j4 .
Now when we apply G−1 we notice this is exactly
1
2 (xg(i1 )g(i2 )j1 j2 xg(i3 )g(i4 )j3 j4
+ xg(i1 )g(i2 )j3 j4 xg(i3 )g(i4 )j1 j2 ).
Now to justify our claim, we will show explicitly that
X
f1 ∗g (xiσ (1)iσ (2) xiσ (3)iσ (4) ) = 9f2
(3.2.4)
g,σ
and the proof in the general case is exactly the same but with more indices. Referring to the definition,
f1 = x12 x34 − x13 x24 + x14 x23 .
We will also drop the
1
2
in the definition of fi without loss of generality because if we show
X
2·
(−1)sgn(T ) fn ∗g (xσ(1)σ(2) xσ(3)σ(4) ) = γ(n)
2 (2 · fn+1 )
g,σ
this is equivalent to proving (3.2.3). We will compute an element of the sum in (3.2.4) for one choice of g
to illustrate how we would proceed in general. We will then show that every term of f2 appears. If we fix
g = id and σ = id. Then [8] \ g([4]) = (5, 6, 7, 8). So this element of our summand is
(x12 x34 − x13 x24 + x14 x23 ) ∗g (x12 x34 ).
From what we computed above we know this is
1
2 (x1256 x3478
+ x1278 x3456 − x1356 x2478 − x1378 x2456 + x1456 x2378 + x1478 x2356 )
So we see that each term in our summand gives us 6 unique terms. In what follows we will compute without
carrying through the 12 and add it back in at the end to simplify the exposition. That is, we will prove
X
2
f1 ∗g (xiσ (1)iσ (2) xiσ (3)iσ (4) ) = 18f2
g,σ
and this implies (3.2.4). So every time we apply ∗g we will not include the resulting 12 . We will now show
that we can get any term β = xi1 i2 i3 i4 xi5 i6 i7 i8 by choosing the correct summand of f1 , and choice of g and
σ. We can assume that i1 < i2 and i5 < i6 . Under these conditions there are three possible orders that can
occur if we fix i1 = 1. We could have i1 < i2 < i5 < i6 , i1 < i5 < i2 < i6 or i1 < i5 < i6 < i2 . This choice
will determine which of the monomials in f1 we will use to get β.
In particular, suppose we have i1 < i5 < i2 < i6 , then we will use the term x13 x24 because we know g is
order preserving and so this term that will have this order if we insert the numbers as the first two indices
in each monomial. Choose g with
g(1) = i1 ,
g(2) = i5 ,
g(3) = i2 ,
g(4) = i6 .
Now we have four numbers remaining {j1 < j2 < j3 < j4 } = [8] \ g([4]). Suppose τ is the permutation
with τ (i3 ) < τ (i4 ) < τ (i7 ) < τ (i8 ). Realize τ as an element of Σ4 /(Σ2 × Σ2 ) and let σ = τ −1 . After we
apply ∗g we know i 7→ ji . We also have τ (i3 ) = j1 , τ (i4 ) = j2 , τ (i7 ) = j3 and τ (i8 ) = j4 . So σ(i) maps to
σji = στ (ik ) = ik , where the permutations act on the indices.
Then in this summand if we focus on the monomial coming from x13 x24 with our selected g and σ, we
have
x13 x24 ∗g xσ(1)σ(2) xσ(3)σ(4) = xi1 i2 i3 i4 xi5 i6 i7 i8 .
We only mention this once more, but here we recall that we have multiplied the entire sum by 2. Furthermore,
notice this term has the correct sign. The monomial x13 x24 is negative in f1 because the permutation
ordering this sequence is odd. The monomial above should have sign (−1)sgn(τ ) where τ orders the set
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
15
(i1 , i2 , i3 , i4 , i5 , i6 , i7 , i8 ). By construction this set is (g(1), g(3), i3 , i4 , g(2), g(4), i7, i8 ). In our equation this
monomial will have sign
(−1)sgn(T ) sgn(σ)+1 .
If we apply the odd permutation swapping g(2) with g(3) we have sgn(τ ) + 1 is the sign of the permutation
ordering (g(1), g(2), i3 , i4 , g(3), g(4), i7 , i8 ). But by construction i3 = σ(j1 ) etc where {j1 , j2 , j3 , j4 } = [8] \
g([4]). So sgn(τ ) + 1 is the sign of the permutation ordering
(g(1), g(2), σ(j1 ), σ(j2 ), g(3), g(4), σ(j3 ), σ(j4 )).
This is precisely sgn(T ) + 1. So the signs match. Notice this argument will work for every other term as well
because we apply the permutation to correct the g(i) terms, this sign will cancel with the sign the monomial
has in fi .
So every term that appears in this sum has the correct sign and every term of f2 appears in the sum. It
remains to show that each term appears exactly γ(n) times. In this case n = 1, γ(1) = 18. So we want to
show each term appears 18 times. However, it suffices to show this for one term by symmetry.
Suppose β = x1457 x2368 . We will show β appears 18 times in our sum. f1 has three terms in it, we
consider each separately. If we first consider the term x12 x34 . There are two choices of g and σ which yield
β. Recalling that g(1) = 1 for all g, the two choices are,
g(2) = 4
g(3) = 6
g(4) = 8
and
g(2) = 5
g(3) = 6
g(4) = 8.
For the first map we need σ to be the permutation (1, 3)(2, 4) with (j1 < j2 < j3 < j4 ) = (2, 3, 5, 7). For the
second σ is also the permutation (1, 3)(2, 4). In both cases
x12 x34 ∗g xσ(1)σ(2) xσ(3)σ(4) = x1457 x2368 .
Now if we focus on the second monomial in f , x13 x24 we will see there are 11 pairs (g, σ) which yield β. If
we fix g(3) = 4, then there are 2 choices for g(2) and 2 choices for g(4), each paired with the appropriate σ.
Similarly if we fix g(3) = 5, there are once again 4 total options. If g(3) = 7, there are now 3 options for
g(2) and only 1 for g(4) because g must be increasing. This gives us 11 total options.
Finally if we focus on the monomial x14 x23 , if we fix g(4) = 4, then g(2) and g(3) must also be fixed. If
g(4) = 5, once again we can only have g(2) = 2 and g(3) = 3. If g(4) = 7, then there are 32 options for
where to send 2 and 3. Hence in this case there are 5 total ways to get β.
In all, we found 2 + 11 + 5 = 18 ways β could appear in our sum. This shows that
X
2
f1 ∗g (xiσ (1)iσ (2) xiσ (3)iσ (4) ) = 18f2 .
g,σ
Which is equivalent to (3.2.4). This same argument generalizes to any fn . We are essentially appending
missing terms on in all possible ways to account for the symmetry present in fn .
In particular, this shows that if fn is in our ideal, fn+1 is as well. So f1 generates all the fi . By the above
the Plücker ideal is finitely generated under the operations we have defined.
Definition 3.2.5. An ideal J ⊆ PΣ is a di-ideal if G(J ) is closed under both ∗g and ·.
For ease of notation we make the following definition
Definition 3.2.6. For any fixed M ≥ 2, call the sum of the Plücker ideals corresponding to Gr(d, M d) as
d varies, SM . So that (SM )d,n is the space of degree n polynomials in the ideal of ρ(Gr(d, M d)) where ρ is
the Plücker embedding.
We ultimately wish to show that for any r ≥ 0, Sr is a di-ideal. We will see this can be deduced from the
following lemma:
Lemma 3.2.7. For any fixed M ≥ 2, SM , is closed under ∗g and · in (PΣ )M .
Proof. Using Weyman’s construction of the Plücker equations in [We, Proposition 3.1.2], any Plücker equation f in the graded coordinate ring of the Plücker embedding of Gr(d, M d) which is contained in (PΣ )M
will be of the form
X
f=
sgn(β)xi1 ...iu jβ(1) ...jβ(d−u) xjβ(d−u+1) ...jβ(2d−u−v) l1 ...lv
β
16
ROBERT P. LAUDONE
where we sum over all permutations β of {1, . . . , d − u − v} such that β(1) < β(2) < · · · < β(d − u) and
β(d − u + 1) < β(d − u + 2) < · · · < β(2d − u − v). And we choose i1 , . . . , iu , j1 , . . . , j2d−u−v , l1 , . . . , lv as
distinct elements from [M d].
Each term in f is a product xI xK with I 6= K, |I| = |K| = d as seen in [We, Proposition 3.1.6]. An
alternative way to see this is to refer back to the description of the image of the Grassmannian in Example
3.2.2. The Plücker equations span the sum of the representations
!
d
M
^
2
S(2i ,12(r−i)) ) ,
=
Sym
0≤i≤d,
i≡d (mod 2)
with i < d. If I = K, this would imply we only use d distinct numbers as indices in xI xK . However, this only
occurs in the representation S2d , which means the corresponding element could not be a Plücker equation.
Fix an element n ∈ (PΣ )M of bidegree (a, m). For any Plücker equation f , to get f ∗g n, from Example
3.2.2, we append additional distinct indices to the monomials appearing in f . Hence, each term in f ∗g n
consists of xI xK where I 6= K, with |I| = |K| = d + a.
As in Weyman, we may identify each xI with the element e∗I ∈ ∧d+a (kM(d+a) )∗ where e∗i denotes the dual
basis element to the standard basis element ei of kM(d+a) .
To show that f ∗g n ∈ J it suffices to show that it vanishes on all decomposable elements of kM(d+a) . To
do this consider all of the equations given by f ∗g n for all f a Plücker equation, g ∈ Inc(N) and n ∈ (PΣ )M .
The following computation is not that enlightening, but the result is very important so we will state it as a
claim and skipping the proof will not take away from the proof of this lemma.
Claim 3.2.8. The collection of equations f ∗g n with f a Plücker equation, g ∈ Inc(N) and n ∈ (PΣ )M of
fixed bidegree (a, m) is GL(kM(d+a) )-invariant.
Proof. Indeed, given h ∈ GL(kM(d+a) ), we have
h(f ∗g n) = [g −1 hgf ] ∗g (g ′ )−1 hg ′ n.
Where g ′ ∈ Inc(N) is the map induced by g which sends [2a] to [M (d + a)] \ g([M d]) in increasing order. It
is easy to check on indices that this is valid.
The set of Plücker equations is GL-invariant [We, Proposition 3.1.2] and we can view g −1 hg ∈ GL(kMd ),
so (g −1 hg)f is another Plücker equation. Also, (g ′ )−1 hg ′ n ∈ (PΣ )M and has the same bidegree (a, m).
Furthermore, GL(kM(d+a) ) acts transitively on Gr(d + a, M (d + a)) and so acts transitively on the
Vd+a M(d+a)
k
. As a result, it suffices to show that all the equations f ∗g n as f ,
decomposable elements of
g and n vary as described in Claim 3.2.8 vanish on one decomposable element, say ei1 ∧ · · · ∧ eid+a , where
i1 < · · · < id+a are chosen from [M (d + a)].
Indeed, any decomposable element is in the GL-orbit of this element, but the set of f ∗g n with f , g and
n varying is GL-invariant. Accordingly, if every equation in this set vanishes on ei1 ∧ · · · ∧ eid+a , they will
vanish on every decomposable element and so will be in SM .
However, it must be the case that the set of equations f ∗g n for any f , g and n vanish on ei1 ∧ · · · ∧ eid+a ,
because only e∗i1 ∧ · · · ∧ e∗id+a does not vanish on ei1 ∧ · · · ∧ eid+a . As mentioned above, f ∗g n is a sum of
products of terms of the form xI xK where I 6= K, so one of xI or xJ will vanish. This means every term in
the sum vanishes, i.e. f ∗g n vanishes for any choice of f and n.
The case for f · n is clearer because we just multiply f and n in PΣ . By definition f will vanish on any
decomposable element and so f · n will vanish as well.
We also notice that PΣ has a natural comultiplication ∆ defined in the following way. If v1 · · · vn ∈
Symn (∧d kMd ), then
X
vS ⊗ v[n]\S ,
∆(v1 · · · vn ) =
S⊆[n]
Vd
where vS = i∈S vi ∈ Sym ( kMd ). Less explicitly, but still importantly, this is defined in the usual
Vd
way by letting w 7→ w ⊗ 1 + 1 ⊗ w when w ∈ Sym1 ( kMd ) and extending uniquely while requiring that
∆ : PΣ → PΣ ⊗ PΣ be an algebra homomorphism.
We will now show that Lemma 3.2.7 implies J is a di-ideal. To do this, we need the following results.
Q
|S|
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
17
Proposition 3.2.9. The symmetrization map G : PΣ → P Σ is an isomorphism of bigraded bialgebras under
the · product. More precisely, the following two diagrams commute:
PΣ ⊗ PΣ
·
PΣ
G⊗G
G
PΣ ⊗ PΣ
·
PΣ
PΣ
∆
PΣ ⊗ PΣ
G⊗G
G
PΣ
∆
PΣ ⊗ PΣ
Proof. This is the same proof as in [Sa1, Proposition 3.8].
With this proposition, we can prove the following important fact:
Proposition 3.2.10. If J ⊆ PΣ is closed under ∗g and · in PΣ , then J is a di-ideal.
Proof. Proposition 3.2.9 shows that if J ⊂ BΣ is an ideal under ·, the same is true for G(J ) ⊆ P Σ .
It remains to show G(J ) is closed under the ∗g product in P Σ . We have f ∗g n ∈ J for all monomials n
and f ∈ J . In PΣ , we defined
f ∗g n = G −1 (G(f ) ∗g G(n)),
so
G(f ∗g n) = G(f ) ∗g G(n).
As G is a linear isomorphism, we can write any n′ ∈ P Σ as G(n′ ) for some n′ ∈ PΣ . To see G(J ) is closed
under ∗g , we just need to check G(f ) ∗g n′ ∈ G(J ) for any n′ ∈ P Σ . Find n ∈ PΣ such that G(n) = n′ , then
we know
f ∗g n ∈ J ,
hence
G(f ) ∗g G(n) ∈ G(J ),
but G(n) = n′ .
Theorem 3.2.11. For any fixed M ≥ 2, SM , is a di-ideal.
Proof. Combine Lemma 3.2.7 with Propositions 3.2.9 and 3.2.10.
4.
Joins and Secants
Let V be a vector space and Sym(V ) be its symmetric algebra. Given ideals I, J ⊂ Sym(V ), their join
I ⋆ J is the kernel of
∆
Sym(V ) −→ Sym(V ) ⊗ Sym(V ) → Sym(V )/I ⊗ Sym(V )/J,
where the first map is the standard comultiplication. Note that ⋆ is an associative and commutative operation
since ∆ is coassociative and cocommutative. Set I ⋆0 = I and I ⋆r = I ⋆ I ⋆(r−1) for r > 0.
Proposition 4.1. [Sa1, Proposition 4.1] Assume k is an algebraically closed field. If I and J are radical
ideals, then I ⋆ J is a radical ideal. If I and J are prime ideals, then I ⋆ J is a prime ideal.
These definitions make sense for ideals I, J ⊆ PΣ , so we can define the join I ⋆J . To be precise, (I ⋆J )d,n
is the kernel of the map
n
M
∆
(PΣ /I)d,i ⊗ (PΣ /J )d,n−i .
(PΣ )d,n −→
i=0
Since G is compatible with ∆ (Proposition 3.2.9), we deduce that
G(I ⋆ J ) = G(I) ⋆ G(J ).
Proposition 4.2. If I, J ⊆ PΣ are di-ideals, then I ⋆ J is a di-ideal.
Proof. Pick v ∈ G(I ⋆ J ). By definition, v is in the kernel of the map
∆ : P Σ → P Σ /G(I) ⊗ P Σ /G(J ).
Since both G(I) and G(J ) are ideals under ∗g via Proposition 3.2.10, it gives a well-defined multiplication
on P Σ /G(I) ⊗ P Σ /G(J ). By Lemma 3.1.4, given x ∈ P Σ , we have ∆(x ∗g v) = ∆(x) ∗g ∆(v). But ∆(v) = 0,
so x ∗g v ∈ G(I ⋆ J ).
18
ROBERT P. LAUDONE
Let V be a vector space. A subscheme X ⊆ V is conical if its defining ideal IX is homogeneous. The rth
⋆r
secant scheme of X is the subscheme of V defined by the ideal IX
. We wish to consider secant varieties
of Plücker embeddings, we make the following definition,
Definition 4.3. For any fixed r ≥ 0 and M ≥ 2, let SM (r) = (SM )⋆r . Where SM is defined in Definition
3.2.6.
An immediate corollary of what we have just shown
Corollary 4.4. SM (r) is a di-ideal for any r ≥ 0 and M ≥ 2.
Proof. This follows immediately from combining Proposition 4.2 and Theorem 3.2.11.
5.
Proof of Theorem 1.1
Before we can get to the main result, we need the following lemma
Lemma 5.1. For any fixed r ≥ 0, to bound the degrees of the ideal generators of SM (r) for any M , it
suffices to consider M = (r + 2).
Proof. This follows immediately from [MM, Proposition 5.7]. Rephrasing, in our case for a given GrassVd
mannian Gr(d, V ), the number of nonzero rows in λ with Sλ (V ) =
V is d. So from the quoted result,
to bound the degrees of the ideal generators of the rth secant variety of a given Grassmannian it suffices
to consider a vector space of dimension (r + 2)d. It is then clear that to bound the degrees of the ideal
generators of SM (r) it suffices to consider Sr+2 (r) in Pr+2 .
Remark 5.2. The result, [MM, Proposition 5.7], specifically says that if the degrees of the ideal generators
of the rth secant variety of a given Grassmannian for a vector space of dimension (r + 2)d are bounded by C,
then the same is true for the degrees of the ideal generators of the rth secant variety of a given Grassmannian
for a vector space of higher dimension. One might be concerned that we are not considering vector spaces of
dimension less than (r + 2)d. However if we work with any lower dimensional vector space say of dimension
c < (r + 2)d, the ideal of the rth secant variety of Gr(d, c) is contained in the ideal of the rth secant variety
of Gr(d, (r + 2)d) by the exact argument seen in the proof of [MM, Proposition 5.7]. So for any given d, we
can fix the dimension of the vector spaces as (r + 2)d to bound the ideal generators of the rth secant variety
of Gr(d, n) for any n.
Proof of Theorem 1.1 and Corollary 1.2. From Lemma 5.1 it suffices to consider Gr(d, (r + 2)d). The ideal
of Secr (ρ(Gr(d, (r + 2)d))) is contained in Sr+2 (r), where again ρ is the Plücker embedding. In particular,
(Sr+2 (r))d,n is precisely the space of degree n polynomials in the ideal of Secr (ρ(Gr(d, (r + 2)d))).
Corollary 4.4 implies Sr+2 (r) is a di-ideal. By Proposition 3.1.3, G(Sr+2 (r)) is generated by finitely many
elements f1 , . . . , fN under · and ∗g . Every element of G(Sr+2 (r)) can be written as a linear combination of
elements of the form h · (fi ∗g a) and so for fixed d a set of generators for Secr (ρ(Gr(d, (r + 2)d))) can be
Σ
taken to be the set of all G −1 (fi ∗g a) such that fi ∗g a ∈ Pd,n
for some n. The degree of fi ∗g a is the same
Σ
as that of fi (if f ∈ Pd,n then its degree is n). So we can take C(r) = max(deg(f1 ), . . . , deg(fN )).
The proof of Corollary 1.2 follows immediately from the above.
Remark 5.3. In the case where r = 0, this theorem tells us that the homogeneous ideal of the Plücker
image of all Grassmannians can be generated by finitely many polynomials of a finite degree bounded by
C(0) under the operations ∗g and ·. In this case we know C(0) = 2 and Example 3.2.2 shows that all Plücker
equations can be obtained from the Klein quadric.
6.
The Plücker Category
In this section, we will translate the above work into the language of functor categories in the spirit of
[Sa2]. Once we transition to this language, we can study free resolutions of secant ideals of Plücker embedded
Grassmannians. In particular, we will show that the ith syzygy module of the coordinate ring of the rth
secant variety of the Plücker embedded Gr(d, n) (whose space of generators is the ith Tor group with the
residue field) is generated in bounded degree with bound independent of d and n. The case i = 1 corresponds
to the above results.
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
19
Vd
Let k be a commutative ring and fix M ≥ 0. Recall that Pd,n = Symn ( kMd ). We will now encode the
morphisms from Pd,m to Pe,n as the space of morphisms from an objects (d, m) to another object (e, n) in
the abstract category kGM . The operations · and ∗g tell us how to do this when d = e or m = n respectively.
More explicitly, when d = e an operation Pd,m to Pd,n is given by a partition σ of [n] and an element of
Pd,n−m . A basis for these operations can be encoded by an order preserving injection [m] → [n] together
with a list of monomials.
When m = n, an operation Pd,n to Pe,n consists of a choice of an element of Pe−d,n as well as a map
g ∈ inc(N) with g([d]) ⊂ [e]. It has a basis given by the monomials, which are represented by an ordered
Ve−d M(e−d)
k
. Once again, we prefer to represent these lists of monomials by reading
list of n monomials in
lists, denote the poset of readings lists by RL, and the poset of readings lists with n entries of size e by
RLn,e . Explicitly, if S ∈ RLn,e , then S = (S 1 , . . . , S n ) with |S i | = e. Where the readings lists are defined
above in §2, below Corollary 2.3. In particular, each S i consists of distinct numbers selected from [M e].
When d 6= e and m 6= n, it is harder to describe a basis for the space of operations. Given a map
α : [d] → [e], suppose [e] \ α([d]) = {a1 , . . . , ae−d }, define αc : [e − d] → [e] as αc (i) = ai . We call this the
complement of α.
Definition 6.1. Define the Plücker category GM as follows. The objects of GM are pairs (d, m) ∈ Z2≥0
and a morphism α : (d, m) → (e, n) consists of the following data:
• An order-preserving injection α1 : [m] → [n].
• A function α2 : [n] \ α1 ([m]) → RL1,e
• A function α3 : [m] → RL1,e−d
• An order-preserving injection α4 : [d] → [e].
In particular, HomGM ((d, m), (e, n)) = ∅ if d > e. Given another morphism β : (e, n) → (f, p), the composition β ◦ α = γ : (d, m) → (f, p) is defined by
• γ1 = β1 ◦ α1 .
• γ2 : [p] \ γ1 ([m]) → RL1,f is defined by:
– if i ∈ [p] \ β1 ([n]), then γ2 (i) = β2 (i), and
– if i ∈ β1 ([n] \ α1 ([m])), then γ2 (i) = β4 (α2 (i′ )) + β4c (β3 (i′ )) where i′ is the unique preimage of i
under β1 .
• γ3 : [m] → RL1,f −d is defined by γ3 (i) = (γ4c )−1 (β4 αc4 α3 (i) + β4c β3 (α1 (i))).
• γ4 = β4 ◦ α4 .
When d = e, the functions α3 and α4 are superfluous and the pair (α1 , α2 ) encodes an operation as
discussed above. Similarly, when n = m, the functions α1 and α2 are superfluous, and α3 also encodes an
operation as discussed above.
Remark 6.2. Each of these maps α2 and α3 can be represented by reading lists in RLn−m,e and RLm,e−d
respectively. To explicitly see this, α3 can be represented as Sm,e−d with S i exactly the image of i. We do
not take this perspective for ease of composition in the above definition. However, taking the reading list
perspective will be important in Proposition 6.4.
Lemma 6.3. Composition as defined above is associative.
Proof. Suppose we are give three morphisms
α
β
γ
(d, m) −
→ (e, n) −
→ (f, p) −
→ (g, q).
We will verify that all three components of both ways of interpreting γβα are the same.
• The associativity of the first and fourth maps follows by the associativity of function composition.
• Consider [q] \ (γ1 β1 α1 ([m])) → RL1,g .
– If i ∈ [q] \ γ1 ([p]) then i 7→ γ2 (i) under both compositions.
– If i ∈ γ1 ([p] \ β1 ([n])), let i′ be the unique premiage of i under γ1 and let i′′ be the unique
preimage of i′ under β1 .
Under γ(βα), we have i 7→ γ4 ((βα)2 (i′ )) + γ4c ((γ)3 (i′ )) = γ4 β2 (i′ ) + γ4c γ3 (i′ )
Under (γβ)α, we have i 7→ (γβ)2 (i′ ) = γ4 β2 (i′ ) + γ4c γ3 (i′ ). This is because i ∈ [q] \ γ1 β1 ([n]) by
assumption.
20
ROBERT P. LAUDONE
– If i ∈ γ1 β1 ([n] \ α1 ([m])), let i′ be the unique preimage of i under γ1 and let i′′ be the unique
preimage of i′ under β1 .
Under γ(βα), we have i 7→ γ4 ((βα)2 (i′ )) + γ4c γ3 (i′ ) = γ4 β4 α2 (i′′ ) + α4 β4c β3 (i′′ ) + γ4c γ3 (i′ ).
Under (γβ)α, we have i 7→ (γβ)4 (α2 (i′′ )) + (γβ)c4 ((γβ)3 (i′′ )) = γ4 β4 α2 (i′′ ) + α4 β4c β3 (i′′ ) +
γ4c γ3 (β1 (i′′ )).
But β1 (i′′ ) = i′ , so these are equal.
• Now consider the map [m] → RL1,g−d .
– Under γ(βα), we have i 7→ ((γ4 β4 α4 )c )−1 (γ4 (β4 α4 )c (βα)3 (i) + γ4c γ3 (βα)1 (i)) = ((γ4 β4 α4 )c )−1
(γ4 β4 αc4 α3 (i) + γ4 β4c β3 α1 (i) + γ4c γ3 β1 α1 (i)).
– Under (γβ)α, we have i 7→ ((γ4 β4 α4 )c )−1 ((γβ)4 αc4 α3 (i) + (γ4 β4 )c (γβ)3 (α1 (i))) = ((γ4 β4 α4 )c )−1
(γ4 β4c β3 α1 (i) + γ4c γ3 β1 α1 (i) + γ4 β4 αc4 α3 (i)).
We have equality in each case. As these are all the ways the components interact, this implies associativity.
Let kGM be the k-linearization of GM , i.e., HomkGM (x, y) = k[HomGM (x, y)]. A kGM -module is a functor
form GM to the category of k-modules. Equivalently, a kGM -modules is a k-linear functor from kGM to the
category of k-modules. Morphisms of kGM -modules are natural transformations, and kGM -modules form an
abelian category where submodules, kernels, cokernels, etc. are computed component pointwise.
Given (d, m) ∈ Z2≥0 , define a kGM -module Pd,m by
Pd,m (e, n) = k[HomGM ((d, m), (e, n))].
This is the principal projective kGM -module generated in bidegree (d, m), and they give a set of projective
generators for the category of kGM -modules. That is, every kGM -modules is a quotient of a direct sum of
principal projectives. For further exposition on principal projectives we refer the reader to [SS1, §3.1]. Then
Pd,m (e, n) is the space of operations from Pd,m to Pe,n which we discussed at the beginning of the section,
so Pd,m is a P-module freely generated in bidegree (d, m).
GM
To emphasize the category we may sometimes write Pd,m
. With these definition we can now make sense
of what it means for modules to be finitely generated. A kGM -modules N is finitely generated if there is
a surjection
g
M
Pdi ,mi → N → 0,
i=1
with g finite. A kGM -modules is noetherian if all of its submodules are finitely generated. For a definition
of a Gröbner category, see [SS1, Definition 4.3.1]. We only need this definition for the next result and it is
lengthy, so we choose to omit it so as not to distract.
Proposition 6.4. GM is a Gröbner category. In particular, if k is noetherian, then every finitely generated
kGM module is noetherian.
Proof. We will use [SS1, Theorem 4.3.2]. Fix (d, m) ∈ Z2≥0 . Let
d
Σ = RL × RL ×Zm
≥0 × Z≥0
By Dickson’s Lemma and Theorem 2.16, the finite product of noetherian posets is also noetherian with the
componentwise order. Hence Σ is noetherian.
From Remark 6.2 we can associate to each α2 and α3 a RL, Sαi , for i = 2, 3. Given a morphism
α : (d, m) → (e, n) encode it as w(α) ∈ Σ
w(α) = (Sα2 , Sα3 , im(α1 ), im(α4 )).
We can recover α from w(α), so this is an injection. Define α ≤ γ if there exists some β such that γ = β ◦ α.
Then, it follows from the definition of composition that the set of morphisms α : (d, m) → (e, n) with (d, m)
fixed and (e, n) varying is naturally a subposet of Σ, i.e. α ≤ α′ if and only if w(α) ≤ w(α′ ). Since
noetherianity is inherited by subposets, we conclude that this partial order on morphisms with source (d, m)
is noetherian.
It remains to prove that the set of morphisms with source (d, m) is orderable, i.e., for each (e, n) there
exists a total ordering on the set of morphisms (d, m) → (e, n) so that for any β : (e, n) → (f, p), we have
α < α′ implies that βα < βα′ . To do this first put the lexicographic order on RL. That is given two RLs,
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
21
Sd,n = (S 1 , . . . , S n ) and Te,m = (T 1 , . . . , T m ) we say Sd,n ≤ Te,m if n < m or if n = m and d < e, if m = n
and d = e we compare the lists in (Zd )n using the natural lexicographic order described in §2. In particular,
we first compare S 1 and T 1 lexicographically, if they are equal we consider S 2 and T 2 , etc.
d
This defines a total order on RL. Now put a lexicographic order on Zm
≥0 and Z≥0 in the natural way.
Totally order Σ by declaring all of the elements of the first RL to be larger than the second RL which is
d
larger than Zm
≥0 which is larger than Z≥0 . This is just another lexicographic order. This orders Σ, which in
turn gives the desired ordering.
This proves GM is Gröbner, in particular this also implies Theorem 2.17.
6.1
Symmetrized versions.
V
In kGM , the space of morphisms (0, 0) → (d, m) is identified with the tensor power ( d kMd )⊗m . For
V
our applications, we need symmetric powers, Symm ( d kMd ), so we now define symmetrized versions of the
Gröbner category kGM .
Definition 6.1.1. Given α : (d, m) → (e, n) and σ ∈ Σn , there is a unique τ ∈ Σm so that σα1 τ −1 is
order-preserving; we refer to τ as the permutation induced by σ with respect to α1 . Define σ(α) by
•
•
•
•
σ(α)1
σ(α)2
σ(α)3
σ(α)4
= σα1 τ −1 ,
= α2 σ −1 ,
= α3 τ −1 ,
= α4 .
This defines an action of Σn on HomGM ((d, m), (e, n)), and we set
Σn
Σ ((d, m), (e, n)) = k[HomG
HomkGM
M ((d, m), (e, n))]
where the superscript denotes taking invariants.
Lemma 6.1.2. Given α : (d, m) → (e, n) and β : (e, n) → (f, p), and σ ∈ Σp , we have σ(β ◦ α) = σ(β) ◦ τ (α)
Σ
where τ ∈ Σn is the permutation induced by σ with respect to β1 . In particular, kGM
is a k-linear subcategory
of kGM .
Proof. Let ρ ∈ Σℓ be the permutation induced by τ with respect to α1 . Then (σβ1 τ −1 )(τ α1 ρ−1 ) is order
preserving, so ρ is also the permutation induced by σ with respect to β1 α1 . Hence σ(βα)1 = σ(β)1 τ (α)1 .
Next, we show that σ(βα)2 = (σ(β)τ (α))2 . If i ∈ [p] \ σ(β)1 ([n]), then
σ(βα)2 (i) = (βα2 )σ −1 (i) = β2 σ −1 (i) = σ(β)2 (i) = (σ(β)τ (α))2 (i).
Else, i ∈ σ(β)1 ([n] \ τ (α)1 ([m])), let i′ be the unique preimage of i under σβ1 τ −1 . Then τ −1 (i′ ) is the unique
preimage of σ −1 (i) under β1 , and we have
σ(βα)2 (i) = (βα)2 σ −1 (i) = β4 (α2 (τ −1 (i′ ))) + β4c (β3 (τ −1 (i′ )))
= β4 (τ (α)2 (i′ )) + β4c (σ(β)3 (i′ )) = (σ(β)τ (α))2 (i).
Now, we show that σ(βα)3 = (σ(β)τ (α))3 . For i ∈ [ℓ], we have
(σ(β)τ (α))3 (i) = ((σ(β)4 τ (α)4 )c )−1 (σ(β)4 τ (α)c4 τ (α)3 (i) + σ(β)c4 σ(β)3 τ (α)1 (i))
= ((σ(β)4 τ (α)4 )c )−1 (β4 αc4 α3 ρ−1 (i) + β4c β3 τ −1 τ α1 ρ−1 (i))
= ((σ(β)4 τ (α)4 )c )−1 (β4 αc4 α3 + β4c β3 α1 )(ρ−1 (i))
= (βα)3 ρ−1 (i)
= σ(βα)3 .
Finally, we show that σ(βα)4 = (σ(β)τ (α))4 . This is clear because σ acts trivially on this map, so
σ(βα)4 = β4 α4 = σ(β)4 τ (α)4 = (σ(β)τ (α))4 .
22
ROBERT P. LAUDONE
Σ
Σ
A kGM
-module is a k-linear functor from kGM
to the category of k-modules. For each (d, m), the principal
Σ
projective kGM -module is defined by
kG Σ
Pd,mM (e, n) = HomkGM
Σ ((d, m), (e, n)),
Σ
and we say that a kGM
-module N is finitely generated if there is a surjection
g
M
kG Σ
M
→N →0
Pdi ,m
i
i=1
with g finite.
Σ
Proposition 6.1.3. If k contains a field of characteristic 0, then every finitely generated kGM
-module is
noetherian.
kG Σ
GM
Proof. Set Pd,m = Pd,m
and Qd,m = Pd,mM ; we have a natural inclusion Qd,m (e, n) ⊆ Pd,m (e, n) for all
Σ
(e, n). Given a kGM -submodule M of Qd,m , let N be the GM -submodule of Pd,m that it generates. Given a
list of generators of N coming from M , Proposition 6.4 shows Pd,m is noetherian and so some finite subset
γ1 , . . . , γg of them already generates N . Let π be the symmetrization map
k[HomGM ((d, m), (e, n))] → k[HomGM ((d, m), (e, n))]Σn
X
1
α 7→ n!
σ(α).
σ∈Σn
Σn
If α ∈ k[HomGM ((d, m), (e, n))] , then π(α) = α; given β ∈ k[HomGM ((e, n), (f, p))], then π(βα) = π(β)α
by Lemma 6.1.2.
P
Given any element
P γ of M , we have an expression γ = i δi γi ,Σwhere δi ∈ kGM . So, applying π, we
get γ = π(γ) = i π(δi )γi , so γ1 , . . . , γg also generate M as a kGM -module. In particular, the principal
Σ
projectives of kGM
are noetherian, so the same is true for any finitely generated module.
Now assume that k contains a field of characteristic 0. We define the symmetrized Plücker category
GM = (kGM )Σ as follows. First, set
HomGM ((d, m), (e, n)) = k[HomGM ((d, m), (e, n))]Σn
where the subscript denotes coinvariants under Σn . As in §3, we have an isomorphism
∼
=
k[HomGM ((d, m), (e, n))]Σn −
→ k[HomGM ((d, m), (e, n))]Σn
X
1
σ(α),
α 7→ n!
σ∈Σn
Σ
and as above we use this to transfer the k-linear category structure from kGM
to GM . Now we notice
V
d (r+2)d
m
), which was our goal. This isomorphism in
that HomGM ((0, 0), (d, m)) is identified with Sym ( k
combination with Proposition 6.1.3 give us the following:
Proposition 6.1.4. Suppose k is a field of characteristic 0. Every finitely generated GM -module is noetherian.
Remark 6.1.5. These definitions parallel the constructions in §3. In particular, we can identify P Σ and
Σ
PΣ from this section with the principal projectives generated in degree (0, 0) in kGM
and GM respectively.
Furthermore, the notions of ideal and di-ideal translate to submodules in both cases. So Proposition 3.1.3
is a special case of Proposition 6.1.3.
7.
Syzygies of Secant Ideals
In this section, k is a field of characteristic 0. For this section fix some M ≥ 0. The principal projective
P0,0 in GM is the algebra PΣ from §3 and each principal projective Pd,m is a module over it. We use
PΣ (−d, −m) to denote this module; by Proposition 6.1.4 these are all noetherian modules.
In Definition 3.2.6, we defined SM to be the sum of the Plücker ideals corresponding to Gr(d, M d) as
d ≥ 0 varies and SM (r) = (SM )⋆r . By Corollary 4.4, SM (r) is a GM -submodule of P0,0 for all r.
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
For d fixed,
L
m
23
Vd
SM (r)d,m is an ideal in Sym( kMd ). So we can define an alegbra
Secd,r (M ) =
M
m≥0
d
^
Symm ( kMd )/SM (r)d,m
Vd
which is a quotient of Sym( kMd ). Notice this is exactly the rth secant ideal of the Plücker-embedded
L
Vd
Gr(d, M d). More generally, if M is a GM -module, then for d fixed, m Md,m is a Sym( kMd )-module.
Lemma 7.1. Fix d, e, n. Then
M
PΣ (−e, −n)d,m
m≥0
Vd−e M(d−e)
Vd
k
)).
is a free Sym( kMd )-module generated in degree n whose rank is dimk (Symn (
Proof. We have
M
PΣ (−e, −n)d,m =
m≥0
M
d−e
^
(Symn (
m≥0
d−e
^
= Symn (
d
^
kM(d−e) ) ⊗ Symm−n ( kMd ))
d
^
kM(d−e) ) ⊗ Sym( kMd )(−n).
V
As follows from the definitions, the action of Sym( d kMd ) on this space corresponds to the usual multipliVd Md
cation on Sym( k )(−n).
Theorem 7.2. There is a function CM (i, r), depending on i, r, M , but independent of d, such that
Sym(
Tori
Vd
kM d )
(Secd,r (M ), k)
is concentrated in degrees ≤ CM (i, r).
Proof. We know SM (r) is a finitely generated submodule of PΣ , and hence has a projective resolution
· · · → Fi → Fi−1 → · · · → F0 ,
such that each Fi is a finite direct sum of principal projective modules by Proposition 6.1.4. For d fixed, we
V
get an exact complex of Sym( d kMd )-modules
M
M
M
··· →
(Fi )d,m →
(Fi−1 )d,m → · · · →
(F0 )d,m → Secd,r (M ) → 0.
m
If Fi =
Lk
j=1
m
m
PΣ (−dj , −mj ), then set CM (i, r) = max(m1 , . . . , mk ). In particular, by Lemma 7.1, this
Sym(
gives a free resolution which can be used to compute Tori
concentrated in degrees ≤ CM (i, r).
Sym(
Vd
Vd
kM d )
(Secd,r (M ), k) which we conclude is
kM d )
(Secd,r (M ), k). As used above, this is Z-graded, so we
Remark 7.3. If we write Ti;d,r (M ) = Tori
denote the mth graded component by Ti;d,r (M )m . For fixed i.m, r, we get a functor on the full subcategory
kGM on objects of the form (d, m) by
(d, m) 7→ Ti;d,r (M )m .
From the results above, we conclude that this is a finitely generated functor. In particular, as we allow d
to vary, this means that Ti;d,r (M )m is “built out” of Ti;d′ ,r (M )m where the d′ range over some finite list of
integers. This can be thought of as the Plücker analogue of ∆-modules from [Sn].
As above, we would like to find a bound independent of M , i.e., independent of the chosen vector space
for the Plücker embedded Grassmannian.
Theorem 7.4. The function CM (i, r) is independent of M once M ≥ r + 1 + i. In particular, there is a
bound C(i, r) that works for all M simultaneously.
24
ROBERT P. LAUDONE
Proof. First we notice that Secd,r (M ) is a direct sum of Schur functors Sλ ((kMd )∗ ) with ℓ(λ) ≤ d(r + 1)
Vd Md ∗
by [MM, §5.1]. The ith Tor module of Secd,r (M ) is the ith homology of the Koszul complex of
(k )
Vd Md ∗
tensored (over Sym( (k ) )) with Secd,r (M ), so is a subquotient of
d
i ^
d
^
^
(kMd )∗ ⊂ ( (kMd )∗ )⊗i
tensored (over k) with Secd,r (M ). So all Schur functors Sµ ((kMd )∗ ) that appear in the ith Tor module
satisfy ℓ(µ) ≤ d(r + 1) + di by the subadditivity of ℓ. In particular, no information is lost by specializing to
the case M d = d(r + 1 + i) [SS2, Corollary 9.1.3]. So it suffices to take M = r + 1 + i.
Remark 7.5. This is a special case of Lemma 5.1. There, i = 1 and we can take M = r + 2, so that
Cr+2 (1, r) = C(r) from Theorem 1.1.
Proof of Theorem 1.3. Combine Theorem 7.2 and Theorem 7.4.
References
[BDdG] Karin Baur, Jan Draisma, Willem A. de Graaf, Secant dimensions of minimal orbits: computations and
conjectures, Exp. Math. 16 (2007), no. 2, 239-250.
[CEF]
Thomas Church, Jordan Ellenberg, Benson Farb, FI-modules and stability for representations of symmetric
groups, Duke Math. J. 164 (2015), no. 9, 1833-1910, arXiv:1006.5248v4
[CGG] Maria V. Vatalisano, Anthony V. Geramita, Alessandro Gimigliano, Secant varieties of Grassmann varieties,
Proc. Am. Math. Soc. 133 (2005), no. 3, 633-642.
[D]
Jan Draisma, Noetherianity up to symmetry, Combinatorial algebraic geometry, Lecture Notes in Math.
2108, Springer, 2014, arXiv:1310.1705v2
[DE]
Jan Draisma, Rob H. Eggermont, Plücker varieties and higher secants of Sato’s Grassmannian, J. Reine
Angew. Math., to appear, arXiv:1402.1667v2.
[DK]
Jan Draisma, Jochen Kuttler, Bounded-rank tensors are defined in bounded degree, Duke Math. J. 163
(2014), no. 1, 35-63, arXiv:1103.5336v2.
[Hi]
Graham Higman, Ordering by divisibility in abstract algebras, Proc. London Math. Soc. (3) 2 (1952), 326336.
[HS]
Christopher J. Hillar, Seth Sullivant, Finite Gröbner bases in infinite dimensional polynomial rings and
applications, Adv. Math. 221 (2012), 1-25, arXiv:0908.1777v2.
[KPRS] Alex Kasman, Kathryn Pedings, Amy Reiszl and Takahiro Shiota, Universality of Rank 6 Plücker Relations
and Grassmann Cone Preserving Maps, The Proceedings of the American Mathematical Society, 136 (2008),
77-87.
[LM]
J.M. Landsberg, L. Manivel, On the ideals of secant varieties of Segre varieties, Found. Comput. Math. 4
(2004), no. 4, 397-422, , arXiv:math/0311388v1.
[LO]
J.M. Landsberg, Giorgio Ottoviani, Equations for secant varieties of Veronese and other varieties, Ann. Mat.
Pura Appl. (4) 192 (2013), no. 4, 596-606, arXiv:1111.4567v1.
[LW]
J.M. Landsberg, Jerzy Weyman, On the ideals and singularities of secant varieties of Segre varieties, Bull.
Lond. Math. Soc. 39 (2007), no. 4, 685-697, arXiv:math/0601452v2.
[MM]
Laurent Manivel, Mateusz Michalek, Secants of minuscule and cominuscule minimal orbits, Linear Algebra
Appl. 481 (2015), 288-312.
[NSS]
Rohit Nagpal, Steven V Sam, Andrew Snowden, Noetherianity of some degree two twisted commutative
algebras, Selecta Math. (N.S.), 22 (2016), no.2, 913-937.
[Sa1]
Steven Sam, Ideals of bounded rank symmetric tensors are generated in bounded degree, Invent. Math. 207
(2017), no. 1, 1-21, arXiv:1510.04904.
[Sa2]
Steven Sam, Syzygies of bounded rank symmetric tensors are generated in bounded degree, Math. Ann. 368
(2017), no. 3, 1095-1108, arXiv:1608.01722
[SS1]
Steven Sam, Andrew Snowden, Gröbner methods for representations of combinatorial categories, J. Amer.
Math. Soc. 30 (2017), 159-203, arXiv:1409.1670
[SS2]
Steven V Sam, Andrew Snowden, Introduction to twisted commutative algebras, arXiv:1209.5122v1
SYZYGIES OF SECANT IDEALS OF PLÜCKER-EMBEDDED GRASSMANNIANS
25
[SS3]
Steven V Sam, Andrew Snowden, Stability patterns in representation theory, Forum. Math. Sigma 3 (2015),
e11, 108pp., arXiv:1302.5859v2
[Sn]
Andrew Snowden, Syzygies of Segre embeddings and ∆-modules, Duke Math. J. 162 (2013), no.2, 225-277,
arXiv:1006.5248v4.
[To]
Philip Tosteson, Stability in the homology of Deligne-Mumford compactifications, arXiv:1801.03894
[We]
Jerzy Weyman, Cohomology of vector bundles and syzygies, Cambridge Tracts in Mathematics 149, Cambridge University Press, Cambridge, 2003.
Department of Mathematics, University of Wisconsin, Madison
E-mail address: [email protected]
URL: https://www.math.wisc.edu/~ laudone/
| 0math.AC
|
Tumor Classification and Segmentation of MR Brain
Images
Tanvi Guptaa,∗, Pranay Manochab , Tapan K. Gandhia , R.K. Guptac , B.K.
Panigrahia
a Department
of Electrical Engineering, Indian Institute of Technology, Delhi, India, 110016
of Electrical Engineering, Indian Institute of Technology, Guwahati, India,
781039
c Fortis memorial research institute, Gurgaon, India,122002
b Department
Abstract
The diagnosis and segmentation of tumors using any medical diagnostic tool
can be challenging due to the varying nature of this pathology. Magnetic Resonance Imaging (MRI) is an established diagnostic tool for various diseases and
disorders, and plays a major role in clinical neuro-diagnosis. Supplementing
this technique with automated classification and segmentation tools is gaining
importance, to reduce errors and time needed to make a conclusive diagnosis.
In this paper a simple three step algorithm is proposed; (1) identification of
patients that present with tumors, (2) automatic selection of abnormal slices
of the patients, and (3) segmentation and detection of tumor.
Features were
extracted by using discrete wavelet transform on the normalized images, and
classified by support vector machine (for step (1)) and random forest (for step
(2)). The 400 subjects were divided in a 3:1 ratio between training and test with
no overlap. This study is novel in terms of use of data, as it employed the entire
T2 weighted slices as a single image for classification and a unique combination
of contralateral approach with patch thresholding for segmentation, which does
not require a training set or a template as is used by most segmentation studies. Using the proposed method, the tumors were segmented accurately with a
classification accuracy of 95% with 100% specificity and 90% sensitivity.
∗ Corresponding
author
Email address: [email protected] (Tanvi Gupta)
Preprint submitted to Medical Image Analysis
May 13, 2017
Keywords:
MRI, SVM, T2 weighted, FLAIR, tumor, segmentation,
contralateral
1. Introduction
The human brain is formidably complex entailing a host of factors such as
age, gender, ethnicity and personal medical history. Diagnosis of brain abnormalities such as degenerative, infectious, ischemic or malignant are done using the Magnetic Resonance Imaging (MRI) which is an effective standardized
neuro-imaging tool. A routine brain imaging protocol includes T1-weighted, T2weighted, Fluid attenuated inversion recovery (FLAIR), Gadolinium-enhanced
T1-weighted images. The mode of data acquisition is gradually shifting from
two-dimensional to three-dimensional imaging. This results in a large volume
of data per patient, for which the analysis is both time consuming and prone to
error. This makes computer-aided detection desirable as an aide to the radiologist.
Tumors are atypical cells multiplying out of control. These may vary in size,
location and type. They show a spectrum of atypia from benign to malignant.
It is usually variegated with high grade and low grade tumor cells, necrosis and
edema. Therefore, it is daunting to train a computational system to identify
and segment the region of interest, making it the pathology of interest.
Various methods are used for automated disease classification and tumor segmentation, each with their own restrictions. The pathologies that are more commonly studied for classification purposes are degenarative diseases like Alzheimer’s
[1, 2, 3, 4], Parkinson’s and Schizophrenia [5]. These affect the entire brain posing less of a challenge as far as classification is concerned as the effected brain
varies significantly from physiology. The studies are primarily limited by a small
dataset [6, 7, 8, 9, 10] mostly taken from medical libraries that are available on
the internet, like the Harvard School Medical Library [11, 12, 13, 14, 15]. Moreover, the data sets often lack header information. Mostly, the images used for
classification and segmentation are taken as one slice per patient from a two
dimensional scan set [16, 10, 15]. However, the lesion usually does not appear
only in one slice and hence limits the training and test data to certain slices of
the brain, making it difficult to use in a clinical scenario.
Methods like Multi-Geometric Analysis (MGA) [15] and entropy based features using discrete wavelet transform [14], have been used for feature extraction.
Detailed extraction methodology and normalization techique used for images is
not discussed in most of the studies. Principal Component Analysis (PCA) is a
commonly used unsupervised learning method for feature selection and reduction which increases the computation time and can eliminate certain data that is
essential for classification. Most studies use PCA without justification or mention of the computational time [15, 10, 17, 18]. Many classification techniques
such as support vector machine (SVM) [15, 1, 3, 13, 19, 2, 20, 21, 4], Artificial
neural network (ANN) [16, 11, 17], Probablistic neural network (PNN) [22],
Linear discriminant analysis (LDA) [10], k-Nearest neighbour (k-NN) [17] have
been employed based on their own merits. Template comparison to a normal
atlas of the human brain has also been used for classification [23], which would
be limited to certain data and may not be able to correctly classify variants
from normal physiology. The patient data under analysis is linear in nature and
therefore, linear classifiers like SVM and random forest are useful. The success
of the classification techniques is dependent on the feature set available and
pre-processing of images used.
Convolutional neural network (CNN) [24, 25, 26], k-NN, Fuzzy networks [8]
and other feature based techniques are used for segmentation which requires a
large data base usually acquired from Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) 2013. Methods like fuzzy c-means algorithm
produce multiple segmented images depending on the number of clusters chosen
and thus the final segmented image must be manually selected. Other methods
use a normal template which may vary in intensity, could be a normal variant
and may not be very accurate for tumor segmentation purposes.
The motivation of this study was to combine classification and segmentation
techniques to form a robust system for identification of tumors using all the
slices of each patient as a single component to classify patients into two classes
of pathology and physiology. A multivariate data set comprising T2-weighted
and FLAIR images from 400 patients, classified by the radiologist into normal
and abnormal (etiology being tumor) has been considered.
Tumors exist in
multiple locations and vary between Glioblastomamultiforme (GBMs) and Low
Grade Gliomas (LGGs). T2 and FLAIR sequences for these patients have been
processed to classify and segment tumor.
No feature selection method is em-
ployed and the algorithm is executed with a short computation time of 375 secs
that includes training, testing and segmentation of all patients in question. It
should be noted that the total number of images being processed are roughly
300 images for 3D data and 30 images for a 2D transverse data. Since 12 slices
are used for classification and segmentation, after pre-processing of the data,
the total images are 4800 (12 x 400). The patients that were identified to have
tumors were further filtered to select only the slices with tumor, upon which
tumor segmentation was performed. The segmentation is done without the use
of a training set or a normal template.
7
This paper is structured as follows: Section 2 gives the step by step method
followed for classification and segmentation. Section 3 analyses and discusses
the results obtained and Section 4 concludes the work done.
2. Methodology
This work has been divided into three sections to systemically segment tumors. The first section filters patients with tumors (group1) in contrast to patients without tumors (group2), the second further filters group1 into the slices
with tumors, and the third segments the tumor and stores the 3D data with
the marked region of interest. The process flow for the proposed methodology
is shown in figure 1.
2.1. Data used
Data was collected from a single setup with all imaging performed on a
3-Tesla MR scanner (Philips Healthcare, Netherlands) under the expert super-
Figure 1: Process for classification and tumor segmentation.
vision of a neuroradiologist (R.K.G.) with more than 30 years of experience.
Test subjects were classified into normal, referred to as group2 (n=200) and
with findings, referred to as group1 (n=200). Training and testing groups were
randomly selected from each group in the ratio of 3:1, respectively and without
duplication.
Age of group1 ranged from 2 to 82 years with a mean of 43.64 ± 18.41 years;
male:female ratio of 1.82. Among these, the training group had a male:female
ratio of 1.68 with a mean age of 32.66 ± 18.01 and test group had a male:female
ratio of 2.33 with a mean age of 43.92 ± 21.44. Age of group2 ranged from 4 to
71 years with a mean of 35.36 ± 14.44 years; male:female ratio was 1.27. Among
these, the training group had a male:female ratio of 1.17 with a mean age of
27.01 ± 14.71 and test group had a male:female ratio of 1.63 with a mean age of
33.42 ± 16. Tumors varied in size, shape, location, tumor grade and multiplicity.
Volumetric T2-FLAIR, and volumetric T2-weighted images (or 2D) were
taken. T2-weighted MR images use long TE (echo time) and TR (relaxation
time) to resolve water from fat; the former has longer time span for both. FLAIR
is a T2-based pulse sequence which nullifies fluid and displays the pathology
prominently [27].
All images were normalized to 64 x 64 matrix of 10 mm slice thickness
yielding a homogeneous element of 16 slices.
2.2. Image Post-processing
For all the three steps the image post-processing steps followed were the
same and are described in this section.
2.2.1. Normalization
The patient data, either volumetric or 2-dimensional, was normalized using
SPM8 toolbox in MATLAB using the respective T2.nii template. The size
taken for normalization was [2,2,10] which indicates that the image size for each
slice is 64x64 with 10 mm slice gap, thus generating a total of 16 slices per
patient. The volumetric normalization was done as per the equations 1 to 4.
Volumetric normalization maximizes the overlap of voxels of the images being
processed X and the template X0 as seen in equation 1, where T is a rigid body
transformation.
X0 = {X0 : X0 ∈ X ∩ T (X 0 )}
(1)
If F (X0 ) is the set of intensities for the overlapped voxels in X0 with mean f¯
and G(X0 ) is the set of intensities for the overlapped voxels of X with mean ḡ
the normalized correlation coefficient (NCC) is given in equation 2.
1 Σx0 ∈X0 (f(x0 ) −f¯)(g(x0 ) −ḡ)
σf σg
N0 2
q
1
2
Σx0 ∈X0 (f (x0 ) − f¯)
σf =
N0
N C C (F, G) =
σg =
1 p
Σx0 ∈X0 (g(x0 ) − ḡ)2
N0
(2)
(3)
(4)
The first and last 2 slices are discarded as they contain very little useful information. Intensity normalization was performed for each slice of each patient,
where the intensity range is changed to 0-1 by dividing the value of each pixel
by the highest pixel value of the slice as seen in equation 5 where x is the pixel
value and i and j represent row and column of the image respectively.
Then
the data was compiled to yield a homogeneous element of 12 slices for further
processing and classification.
x i,j =
xi,j
max(X )
(5)
2.2.2. Feature Extraction
Each slice was divided by a 2x2 grid and the mean of the voxel values in the
grid was considered as the feature of the grid as shown in equation 6. Therefore,
32x32 (64x64/2x2) features were obtained per slice and 32x32x12 features were
obtained for each patient.
xi,j =
xi,j + xi+1,j + xi,j+1 + xi+1,j+1
4
(6)
2.3. Patient classification by Support Vector Machine
Support Vector Machines (SVMs) are based on the concept of decision planes
that define boundaries between different classes of objects. A decision plane is
one that separates a set of objects having different class memberships. There are
different types of decision planes like Linear, Quadratic and Polynomial which
fit data into different classes for classification [28, 2, 29, 30, 31, 32, 33, 34]. For
the linear kernel, the equations for solving the hyper plane equations are as
given in equations 7 and 8.
wT + b = +1 for d = +1
(7)
wT + b = −1 for d = −1
(8)
Where, w is a weight vector, x is the input vector and b is a bias. d is the
margin of separation between the hyper plane and the closest data point for a
given weight w and bias b. Optimal decision plane is the one which maximizes
the margin of separation d.
2.4. Slice Identification
Group 1 patients were processed at this stage to identify the slices of interest
with each slice considered as a separate feature set. The number of rows are
equal to the product of number of patients and slices (12) and columns are equal
to the square of 32, for the resulting feature matrix.
2.4.1. Random Forest Classifier
Random forests is a learning method for classification, regression and other
tasks.
It operates by constructing a multitude of decision trees for training
and gives the class as the output. It is a combination of multiple classification
or regression trees combined to improve the accuracy of training and therefore
classification [35, 36, 37]. The slices that are used differ from each other significantly thus linear SVM was insufficient for this classification. Random forest
with 25 trees was used to obtain a high sensitivity.
2.5. Tumor Segmentation
With the abnormal slices known for each patient the data is processed for
tumor segmentation. The data used is the normalized T2 data which has undergone intensity normalization and has been re-sized to 64 x 64 for ease of
processing.
2.5.1. Discrete Wavelet Transform (DWT)
DWT helps visualize images in the time-frequency domain using low and
high pass filters to decompose it. The function in its discrete form is given by
the following equations:
hX
cAj,k (n) =
f (n)l∗j (n − 2j k)
i
(9)
n
hX
cDj,k (n) =
f (n)h∗j (n − 2j k)
i
(10)
n
l(n) and h(n) are the low and high pass filters, respectively, and cAj,k and
cDj,k represent the approximation components containing the low frequency
Figure 2: Flow chart for tumor segmentation.
information and detailed components containing the high frequency details of
the image which basically are the edge of the image k-space [38, 39, 40, 41, 14].
Tumor segmentation requires an approximation component image as a base
contrast of the tumor. Eliminating the high frequency components removes the
edges including skull patterns which might cause errors in segmentation. The
approximation image is re-sized to 64x64 and used for tumor segmentation.
2.5.2. Thresholding by Contralateral Comparison
The Vertical Symmetry or the Contralateral approach is dependent on the
fact that the bilateral cerebral hemispheres are comparable. The presence of a
tumor distorts the symmetry of the brain, and hence this method is appropriate.
Only the slices with tumors as classified by the Random Forest classifier are
taken and analyzed further for tumor segmentation. The steps for segmentation
are as follows and are shown in figure 2:
1. If any slice demonstrates tumor, the neighboring 3 slices are analyzed to
locate tumor margins and hence verify the prediction.
2. Next all stray tumor slices are removed i.e those slices with no adjoining
slices, and therefore, removing the confounding features.
3. Then the remaining slices are made continuous i.e. for example if slices
6-7 and 11-12 were found to qualify as above, then all slices from 6 to 12
are considered for segmentation.
4. To segment the tumor (voxels identified as tv), the right and left half of
the brain are compared to find the points of intensity difference above a
threshold as shown in equation 11. This is advantageous as the contralateral side serves as the control and training set is not required.
tvi,j = xi,j+n/2 > xi,n/2−j + threshold
(11)
where n is the total number of columns and j=0,1,2...n/2.
5. A 4x4 section is created around the selected points and if, the number of
points in the patch is less than a threshold, the patch is removed. This
is done assuming that pathology shall be larger than the patch size considered, and our methodology removes smaller asymmetries which qualify
for normal variants.
6. The remaining sections that are considered to be tumors are delineated.
3. Results
3.1. Patient Classification
In the first step, all patients were classified into group1 and group2 using
SVM. The accuracy, sensitivity and specificity were calculated as per equations
12, 13 and 14 [17, 11, 42, 43]. Testing with T2WI and FLAIR yielded accuracy, sensitivity and specificity of 92.00%, 90.00%, 94.00% and 88.78%, 84.91%,
92.60%, respectively.
Group2 patients were subjected to the next step. The algorithm was run on
MATLAB R2014a on a 4GB RAM, 2.3GHz i5 processor and took 97 seconds to
process the algorithm.
Accuracy =
Correctly classified data
× 100%
Total data
(12)
Correctly identified abnormal data
× 100%
Total abnormal data
Correctly identified normal data
Specificity =
× 100%
Total normal data
Sensitivity =
(13)
(14)
3.2. Slice Selection
Testing with T2WI yielded sensitivity of 77.52%. The algorithm outed the
slice numbers that are abnormal. This result acts as the input to the next stage.
3.3. Tumor Segmentation
These slices are thereafter, presented to the next step where they undergo
segmentation and the output is a nifti file of the T2 weighted images which include a white line which demarcates the tumor region, as exemplified in figure 3.
The false positive from the first stage do not have any tumor regions segmented,
thus increasing the specificity of the overall algorithm to 100% and accuracy to
95%. Figure 3 shows an example of the corresponding slice for a normal patient,
patient with tumor and the post-segmentation image of the respective patient.
3.4. Discussion
Automated classification of abnormal images and tumor segmentation is no
longer a pre-clinical research tool. Our algorithm demonstrates that it can be applied to clinical data. To be applicable to clinical data all slices must be included
for classification with a standardized normalization procedure. As compared to
work done in previous literature, in the present approach, a combination of classification and tumor segmentation is used to achieve a higher accuracy in terms
of identifying patients with tumors and identifying the slice and area in which
the tumor occurs [18, 44, 22, 45, 46]. A multi-variate data set of 400 patients
(T2-weighted images) are used where all the images taken in the scan are considered unlike most studies which are limited by data [18, 22, 44, 45, 46, 47, 12]
and use only one image per patient [18, 22]. Each patient set is treated as
an image thus increasing accuracy and reliability. If only one slice is given as
training, it implies that a diagnostician would be required to supply the tumor
Figure 3: Tumor segmentation: (a) Slice of normal patient, (b) Slice of patient with tumor,
(c) Slice with segmented tumor.
image for classification, thus defeating the classification purpose. Here the entire data taken from a scan is fed into the program to give a result of pathology
or physiology as well as slice and area in which the tumor appears. The data is
taken from a clinical setup and the parameters of the scanner and protocols are
known. The images are normalized and then intensity normalization is done to
ensure that all images are scaled. The details of normalization, sequence proto-
col used, details of feature extraction and size of feature matrix are limited in
past studies [48, 47, 12].
A linear SVM classifier was used for patient classification in the first step,
as the features when plotted appeared linear. This is a basic classifier which
reduces computational time as it is used without a feature reduction method
namely PCA. With a large feature matrix, computation time for covariance
matrix and eigen vectors is high. Therefore the time increases on use of PCA.
Most studies that use PCA do not justify the use of PCA in terms of accuracy
or feature reduction and do not report computation time [18]. We tested the
algorithm with and without PCA and found that there was no significant change
in accuracy whereas the computation time increased significantly with its use.
For slice selection, a high sensitivity is required. This was achieved with a simple
random forest classifier. As the image data is large, SVM is unable to converge.
Also the features differ significantly between slices and the data is not linear.
Thresholding techniques are used to remove outliers in slice selection. DWT
removes high frequency components for tissue-type segmentation. The contralateral sides are compared to find differences based on intensity thresholding.
The intensity and features are symmetric across a patient image, which might
not be true if a normal template been used. This methodology does not require
a training set or a normal template thus making the segmentation process faster
in terms of computation time and also more reliable [24, 25, 26]. Using a normal template can result in errors in identification as some normal variants may
appear different in terms of intensity and be identified as tumors [7]. Also a
training set may not incorporate all the different types of tumors that are being
tested for. Many studies use spatial fuzzy c means clustering for segmentation
[44, 48]. This method is not fully automated as the use of fuzzy gives a number
of segmented images as per the number of clusters and the tumor segmented
image must be chosen manually.
Using this three step process, false positive cases were eliminated, increasing
the accuracy to 95%. This paper reports a higher accuracy than most studies
with a larger dataset and the use of classification and segmentation jointly
[18, 22, 45, 44].
4. Conclusion
In past, many groups have attempted to automate the classification and
segmentation of brain tumor using different MR image sequences.
However,
the reported results have restrictions in terms of images used and use of training sets or templates for segmentation. In this work, a combined model has
been proposed for both classification and segmentation subsequently to obtain
a higher accuracy. The data set is large, varied and the training and testing
set do not overlap. The algorithm is comprehensive and effective within a short
computation time. It is a three step process which involves identifying patients
with tumor, then extracting the abnormal slices followed by segmentation of the
tumor. The entire patient data set is used for classification treating each scan
set as a single image comprising of 12 slices. Passing all slices considered as
abnormal through segmentation allows for normal patients that have been misclassified to be correctly classified in the third step. Approximation components
of the original images are mapped contralaterally for tumor segmentation which
unlike previous studies does not require a normal template or a training set.
The overall accuracy of the method proposed is 95% with 100% specificity and
90% sensitivity. Future work could include improving accuracy and segmenting
the various cell aggregates that differ in composition, within the tumor.
References
[1] B. Magnin, L. Mesrob, S. Kinkingnhun, M. Plgrini-Issac, O. Colliot,
M. Sarazin, B. Dubois, S. Lehricy, H. Benali, Support vector machine-based
classification of alzheimers disease from whole-brain anatomical MRI, Neuroradiology 51 (2009) 73–83.
[2] A. Ortiz, J. M. Grriz, J. Ramrez, F. Martnez-Murcia, LVQ-SVM based
CAD tool applied to structural MRI for the diagnosis of the alzheimers
disease, Pattern Recognition Letters 34 (2013) 17251733.
[3] S. Kloppel, C. M. Stonnington, C. Chu, B. Draganski, R. I. Scahill, J. D.
Rohrer, N. C. Fox, C. R. J. Jr, J. Ashburner, R. S. J. Frackowiak, Automatic classification of MR scans in alzheimers disease, Brain 131 (2008)
681–689.
[4] M. Liu, D. Zhang, E. Adeli-Mosabbeb, D. Shen, Inherent structure
based multi-view learning with multi-template feature representation for
alzheimers disease diagnosis.
[5] M. Nieuwenhuis, N. E. van Haren, H. E. H. Pol, W. Cah, R. S. Kahn, H. G.
Schnack, Classification of schizophrenia patients and healthy controls from
structural MRI scans in two large independent samples, NeuroImage 61
(2012) 606–612.
[6] Y. Sun, B. Bhanu, S. Bhanu, Automatic symmetry-integrated brain injury
detection in MRI sequences, in: Computer Vision and Pattern Recognition Workshops, 2009. CVPR Workshops 2009. IEEE Computer Society
Conference on, IEEE, 2009, pp. 79–86.
[7] M. Prastawa, E. Bullitt, S. Ho, G. Gerig, A brain tumor segmentation
framework based on outlier detection, Medical image analysis 8 (3) (2004)
275–283.
[8] E. Abdel-Maksoud, M. Elmogy, R. Al-Awadi, Brain tumor segmentation
based on a hybrid clustering technique, Egyptian Informatics Journal 16 (1)
(2015) 71–81.
[9] M. B. Cuadra, C. Pollo, A. Bardera, O. Cuisenaire, J.-G. Villemure, J.-P.
Thiran, Atlas-based segmentation of pathological MR brain images using
a model of lesion growth, Medical Imaging, IEEE Transactions on 23 (10)
(2004) 1301–1314.
[10] V. Rathi, S. Palani, Brain tumor MRI image classification with feature
selection and extraction using linear discriminant analysis, arXiv preprint
arXiv:1208.2128.
[11] E.-S. A. El-Dahshan, T. Hosny, A.-B. M. Salem, Hybrid intelligent techniques for MRI brain images classification, Digital Signal Processing 20
(2010) 433–441.
[12] N. Zhang, S. Ruan, S. Lebonvallet, Q. Liao, Y. Zhu, Kernel feature selectionto
fuse multi-spectral MRI images for brain tumor segmentation, Computer
Vision and Image Understanding 115 (2) (2011) 256–269.
[13] S. C. ana L.M. Patnaik, N. Jagannathan, Classification of magnetic resonance brain images using wavelets as input to support vector machine and
neural network, Biomedical Signal Processing and Control 1 (2006) 86–92.
[14] M. Saritha, K. P. Joseph, A. T. Mathew, Classification of MRI brain images
using combined wavelet entropy based spider web plots and probabilistic
neural network, Pattern Recognition Letters 34 (2013) 2151–2156.
[15] S. Das, M. Chowdhury, M. K. Kundu, Brain MR image classification using multi-scale geometric analysis of ripplet, Progress In Electromagnetics
Research 137 (2013) 1–17.
[16] W. H. Ibrahim, A. A. A. Osman, Y. I. Mohamed, MRI brain image classification using neural networks, in: Computing, Electrical and Electronics
Engineering (ICCEEE), 2013 International Conference on, IEEE, 2013, pp.
253–258.
[17] N. H. Rajini, R. Bhavani, Classification of MRI brain images using knearest neighbor and artificial neural network, in: IEEE-ICRTIT, 2011.
[18] M. F. Othman, M. A. M. Basri, Probabilistic neural network for brain
tumor classification, in: Intelligent Systems, Modelling and Simulation
(ISMS), 2011 Second International Conference on, IEEE, 2011, pp. 136–
138.
[19] N. K. Focke, M. Yogarajah, M. R. Symms, O. Gruber, W. Paulus, J. S.
Duncan, Automated MR image classification in temporal lobe epilepsy,
NeuroImage 59 (2012) 356–362.
[20] K. Machhale, H. B. Nandpuru, V. Kapur, L. Kosta, MRI brain cance classification using hybrid classifier (SVM-KNN), in: International conference
on Industrial instrumentation and control, 2015.
[21] S. Vidhusha, K. Anandhan, Analysis and evaluation of autistic brain MR
images using learning vector quantization and support vector machines, in:
International conference on Industrial instrumentation and control, 2015.
[22] D. Sridhar, M. Krishna, Brain tumor classification using discrete cosine
transform and probablistic neural network, in: ICSIPR, 2013.
[23] M. Kaus, S. K. Warfield, A. Nabavi, E. Chatzidakis, P. M. Black, F. A.
Jolesz, R. Kikinis, Segmentation of meningiomas and low grade gliomas in
MRI, in: Medical Image Computing and Computer-Assisted Intervention–
MICCAI99, Springer, 1999, pp. 1–10.
[24] S. Pereira, A. Pinto, V. Alves, C. A. Silva, Brain tumor segmentation using
convolutional neural networks in mri images, IEEE transactions on medical
imaging 35 (5) (2016) 1240–1251.
[25] M. Havaei, A. Davy, D. Warde-Farley, A. Biard, A. Courville, Y. Bengio,
C. Pal, P.-M. Jodoin, H. Larochelle, Brain tumor segmentation with deep
neural networks, Medical Image Analysis.
[26] L. Zhao, K. Jia, Multiscale cnns for brain tumor segmentation and diagnosis, Computational and mathematical methods in medicine 2016.
[27] V. S. Lee, Cardiovascular MR Imaging: Physical Principles to Practical
Protocols, 1st Edition, Lippincott Williams and Wilkins, Philadelphia, PA,
2006.
[28] J. Zhou, K. Chan, V. Chong, S. M. Krishnan, Extraction of brain tumor
from MR images using one-class support vector machine, in: Engineering
in Medicine and Biology Society, 2005. IEEE-EMBS 2005. 27th Annual
International Conference of the, IEEE, 2006, pp. 6411–6414.
[29] R. Cuingnet, J. A. Glaunès, M. Chupin, H. Benali, O. Colliot, T. A. D. N.
Initiative, Spatial and anatomical regularization of SVM: A general framework for neuroimaging data, IEEE Trans. on Pattern Analysis and Machine
5
Intelligence 35 (2013) 682–696.
[30] D. R. J. Ramteke, K. M. Y, Automatic medical image classification and
abnormality detection using k-nearest neighbour, International Journal of
Advanced Computer Research 2.
[31] V. P. Rathi, Dr.S.Palani, Brain tumor MRI image classification with feature selection and extraction using linear discriminant analysis, CoRR
abs/1208.2128.
[32] A. J. Smola, B. Scholkopf, A tutorial on support vector regression, Statistics
and computing 14 (3) (2004) 199–222.
[33] C. J. Burges, A tutorial on support vector machines for pattern recognition,
Data mining and knowledge discovery 2 (2) (1998) 121–167.
[34] C.-C. Chang, C.-b. Lin, Training v-support vector classifiers: theory and
algorithms, Neural computation 13 (9) (2001) 2119–2147.
[35] L. Breiman, Random forests, Machine learning 45 (1) (2001) 5–32.
[36] S. Bernard, L. Heutte, S. Adam, On the selection of decision trees in random forests, in: Neural Networks, 2009. IJCNN 2009. International Joint
Conference on, IEEE, 2009, pp. 302–307.
[37] B. S. Wade, S. H. Joshi, T. Pirnia, A. M. Leaver, R. P. Woods, P. M.
Thompson, R. Espinoza, K. L. Narr, Random forest classification of depression status based on subcortical brain morphometry following electroconvulsive therapy, in: Biomedical Imaging (ISBI), 2015 IEEE 12th International Symposium on, IEEE, 2015, pp. 92–96.
[38] M. Sifuzzaman, M. Islam, M. Ali, Application of wavelet transform and its
advantages compared to fourier transform.
[39] M. Kociolek, A. Materka, M. Strzelecki, P. Szczypiński, Discrete wavelet
transform-derived features for digital image texture analysis, in: International Conference on Signals and Electronic Systems, 2001, pp. 99–104.
[40] S. K. Mohideen, S. A. Perumal, M. M. Sathik, Image de-noising using
discrete wavelet transform, International Journal of Computer Science and
Network Security 8 (1) (2008) 213–216.
[41] C.-L. Yang, W.-R. Gao, L.-M. Po, Discrete wavelet transform-based structural similarity for image quality assessment, in: Image Processing, 2008.
ICIP 2008. 15th IEEE International Conference on, IEEE, 2008, pp. 377–
380.
[42] I. Krashenyi, A. Popov, J. Ramirez, J. M. Gorriz, Application of fuzzy
logic for alzheimer’s disease diagnosis, in: Signal Processing Symposium
(SPSympo), 2015, IEEE, 2015, pp. 1–4.
[43] A. Osareh, M. Mirmehdi, B. Thomas, R. Markham, Automated identification of diabetic retinal exudates in digital colour images, Br J
Opthalmal
87 (2003) 1220–1223.
[44] A. P. Kumar, J. K. Chaithanya, Automatic classification and segmentation
of tumors from skull stripped images using pnn & sfcm, Global Journal of
Computer Science and Technology 15 (1).
[45] S. D. S. Al-Shaikhli, M. Y. Yang, B. Rosenhahn, Brain tumor classification
using sparse coding and dictionary learning, in: Image Processing (ICIP),
2014 IEEE International Conference on, IEEE, 2014, pp. 2774–2778.
[46] M. M. Letteboer, O. F. Olsen, E. B. Dam, P. W. Willems, M. A. Viergever,
W. J. Niessen, Segmentation of tumors in magnetic resonance brain images
using an interactive multiscale watershed algorithm 1, Academic Radiology
11 (10) (2004) 1125–1138.
[47] B. H. Menze, K. Van Leemput, D. Lashkari, M.-A. Weber, N. Ayache,
P. Golland, A generative model for brain tumor segmentation in multi-
modal images, in: Medical Image Computing and Computer-Assisted
Intervention–MICCAI 2010, Springer, 2010, pp. 151–159.
[48] J. Selvakumar, A. Lakshmi, T. Arivoli, Brain tumor segmentation and its
area calculation in brain MR images using K-mean clustering and fuzzy
C-mean algorithm, in: Advances in Engineering, Science and Management
(ICAESM), 2012 International Conference on, IEEE, 2012, pp. 186–190.
| 1cs.CV
|
Constrained Bayesian Networks:
Theory, Optimization, and Applications
arXiv:1705.05326v1 [cs.AI] 15 May 2017
Paul Beaumont and Michael Huth
Department of Computing, Imperial College London
London, SW7 2AZ, United Kingdom
tpaul.beaumont, [email protected]
Abstract
We develop the theory and practice of an approach to modeling and probabilistic inference in causal networks that is suitable when application-specific or analysis-specific
constraints should inform such inference or when little or no data for the learning
of causal network structure or probability values at nodes are available. Constrained
Bayesian Networks generalize a Bayesian Network such that probabilities can be symbolic, arithmetic expressions and where the meaning of the network is constrained by
finitely many formulas from the theory of the reals. A formal semantics for constrained
Bayesian Networks over first-order logic of the reals is given, which enables non-linear
and non-convex optimization algorithms that rely on decision procedures for this logic,
and supports the composition of several constrained Bayesian Networks. A non-trivial
case study in arms control, where few or no data are available to assess the effectiveness of an arms inspection process, evaluates our approach. An open-access prototype
implementation of these foundations and their algorithms uses the SMT solver Z3 as
decision procedure, leverages an open-source package for Bayesian inference to symbolic
computation, and is evaluated experimentally.
Keywords: Bayesian Belief Network. Imprecise Probabilities. Lack of Prior Data. NonLinear Optimization. Confidence Building in Nuclear Arms Control.
1
Introduction
Bayesian Networks (BN) [36, 37, 35] are a prominent, well established, and widely used formalism for expressing discrete probability distributions in terms of directed, acyclic graphs
(DAG) that encode conditional independence assumptions of distributions. Bayesian Networks have a wide range of applications – for example, trouble shooting [12], design of
experiments [30, 28], and digital forensics to support legal reasoning [38]. Their graph-based
formalism and automated support for probabilistic inference seem to lower adaption hurdles
for a diverse set of users with different technical backgrounds. Bayesian Networks are also
appealing since we may combine, within the same Bayesian Network, different aspects such
as subjective beliefs expressed in probabilities, implicit trust assumptions reflected in a bias
of information processing or the combinatorial logic of a process. Probabilistic inference for
such combinations is supported, including belief updates based on observed evidence.
Bayesian Networks also come with methodological support for learning an appropriate
graph structure as well as appropriate prior probability values at nodes in such graphs from
1
pre-existing data (see for example [29, 19]). The appropriateness of chosen prior probability
values may depend on a variety of factors: the quality and quantity of data used for learning
these values or the trust we place in experts who determine such values subjectively – to give
two examples. We would therefore like reassurance that the prior distributions represented
by such values are robust enough in that small changes to such values only result in small
changes of posterior distributions of interest. This naturally leads to the consideration of
robust Bayesian statistics [9, 10].
A popular idea here is to approximate prior probabilities with intervals and to then calculate – somehow – the intervals that correspond to posterior probabilities. A good conceptual
explanation of this is Good’s black box model [26, 27], in which interval information of priors
is submitted into a black box that contains all the usual methods associated with precise
computations in Bayesian Networks, and where the box then outputs intervals of posteriors
without limiting any interpretations or judgments on those output intervals.
Our engagement with a problem owner in arms control made us realize the benefits of
Good’s black box model and made us identify opportunities for extending it to increase the
confidence that users from such problem domains can place in models and their robustness.
Specifically, we want to be able to
R1 re-interpret compactly a BN as a possibly infinite set of BNs over the same graph, with
robustness being analyzable over that re-interpretation
R2 add logical constraints to capture domain knowledge or dependencies, and reflect constraints in robustness analyses in a coherent manner
R3 compare models, within a composition context, to determine any differences in the
robustness that they may offer for supporting decision making
R4 parametrize the use of such a box so that it can produce outputs for any quantitative
measure of interest definable as an arithmetic term
R5 retain the “blackness” of the box so that the user neither has to see nor has to understand its inner workings
R6 interpret outputs of the black box within the usual methodology of Bayesian Networks
in as far as this may be possible.
We believe that these requirements are desired or apt in a wide range of problem domains,
in addition to the fact that they should enhance usability of such a methodology in practice. We develop constrained Bayesian Networks in this paper and show that they meet the
above requirements. This development is informed by advances made in areas of Symbolic
Computation [34] and Automated Theorem Proving [23] (with its applications to Non-Linear,
Non-Convex Optimization [8]), Three-Valued Model Checking [13], and Abstract Interpretation [15]. Concretely, we allow prior probabilities to be arithmetic expressions that may
contain variables, and we enrich this model with logical constraints expressed in the theory
of the reals.
We draw comparisons to related work, including Credal Networks [16, 22, 33] and Constraint Networks [24]. We then highlight similarities, differences, and complementary value
of our approach to this previous work.
2
Contributions and methodology of the paper We develop a formal syntax and semantics of constrained Bayesian Networks which denote an empty, finite or infinite set of
Bayesian Networks over the same directed acyclic graph. We support this concept with a
composition operator in which two or more constrained Bayesian Networks with different or
“overlapping” graphs may be combined for cross-model analysis, subject to constraints that
are an optional parameter of that composition. We formulate a three-valued semantics of the
theory of the reals over constrained BNs that captures the familiar duality of satisfiability and
validity but over the set of Bayesian Networks that a constrained Bayesian Network denotes.
This semantics is used to reduce the computation of its judgments to satisfiability checks in
the first-order logic over the reals. We then apply that reduction to design optimization algorithms that can compute, for any term definable in that logic, infima and suprema up to a
specified accuracy – for example for terms that specify the meaning of marginal probabilities
symbolically. These optimization algorithms and their term parameter allow us to explore
or verify the robustness of a constrained Bayesian Network including, but not limited to, the
robustness of posterior distributions. We demonstrate the use of such extended robustness
analyses on a non-trivial case study in the domain of arms control. We also report a tool
prototype that we have implemented and used to conduct these analyses; it uses an SMT
solver as a feasibility checker to implement these optimization algorithms; and it adapts an
open-source package for Bayesian inference to symbolic computations.
Our principal theoretical contribution is the introduction of the concept of a Constrained
Bayesian Network itself, as well as its intuitive yet formal semantics. Our theoretical results,
such as those for computational complexity and algorithm design, follow rather straightforwardly from these definitions. This is because the latter allow us to appeal directly to existing
results from the existential theory of the reals and optimization based thereupon.
Our main practical contribution is the successful integration of a number of disparate
techniques and approaches into a coherent semantic framework and tool prototype that supports a range of modelling and analysis capabilities, and does so in a highly automated
manner.
2
Background on Bayesian Networks
A Bayesian Network (BN) is a graph-based statistical model for expressing and reasoning
about probabilistic uncertainty. The underlying graph is directed and acyclic, a DAG. Nodes
in this DAG represent random variables defined by discrete probability distributions that are
also a function of the random variables represented by the parent nodes in the DAG. In other
words, a random variable is conditioned on the random variables of its parent nodes.
We can use a BN to compute probabilities for events of interest over these random variables. Bayesian inference also allows us to revise such probabilities when additional observations of “evidence” have been made.
Figure 1 shows a simple BN, which is part of the folklore of example Bayesian Networks.
It depicts the possible causes of wet grass on two neighbours’ lawns. For example, the
probabilities of Holmes’ Grass Wet is conditioned on its parents’ output – whether It Rains
and Holmes’ Sprinkler is turned On or Off. The probability of Holmes’ Grass Wet = T, given
that Holmes’ Sprinkler = Off, Rain = F, for instance, is computed to be 0.05, and is formally
3
Sprinkler
On Off
0.1
F
T
F
T
It Rains
0.9
Sprinkler/It Rains
Off
Off
On
On
It Rains
T
F
Holmes’
Sprinkler
Holmes’s
T
F
0.05
0.7
0.95
0.99
Holmes’
Grass Wet
0.95
0.3
0.05
0.01
0.2
0.8
Watson’s
Grass Wet
It Rains
Watson’s
T
F
F
T
0.05 0.95
0.7 0.3
Figure 1: A BN with a 4-node DAG in the center and probability tables next to the respective
nodes. This BN models beliefs about possible causes of wet grass on two neighbours’ lawns.
This BN allows us to reason about, for example, whether Holmes’ lawn being wet is due to
rain or Holmes’ sprinkler – using observed evidence about a neighbour’s lawn.
stated as:
ppHolmes’ Grass Wet = T | Sprinkler = Off, Rain = Fq “ 0.05
This approach naturally gives cause to computations of the “overall” probability of an
event happening, referred to as the marginal probability. In the Bayesian approach, the
Junction Tree Algorithm (JTA) (see e.g. Chapter 6 in [6] for further details) may be used
to revise a marginal of a BN because of “hard”, respectively “soft”, evidence – the definite,
respectively probabilistic, observation of an additional or new event.
We now formalize BNs and use this below to enrich BNs with modeling and reasoning
capabilities that realize the aforementioned requirements.
Definition 1
1. A Bayesian network (BN) is a pair pG, πq where G is a finite, directed,
acyclic graph G “ pN, Eq of nodes N and edge relation E Ă N ˆ N , and where π is a
tuple pπn qnPN of formal probability tables.
2. The formal probability table πn is defined as follows. Let pntpnq “ tn1 P N | pn1 , nq P Eu
be the (possibly empty) set of parents of n in DAG G and On the set of outcomes of the
random variable
`ś at node ˘n. Then πn is a discrete probability
ř distribution, a function
πn of type
n1 Ppntpnq On1 ˆ On Ñ r0, 1s such that its mass
e πn peq equals 1.
ś
Above, it is understood that n1 PH On1 equals t˚u; in that case, πn has type isomorphic
to On Ñ r0, 1s.
4
Sprinkler
On
0.5˚x
Off
T
1 ´ 0.5˚x
x
Rain
F
1´x
C0
Figure 2: Probability tables that make a constrained BN BX
out of the BN B in Figure 1 by
0
entering non-constant terms into tables Sprinkler and Rain. Marginal probabilities of interest
are those given in (1) and (2) and their meaning is defined as constraints in (3) and (4)
3
Constrained Bayesian Networks
Informally, a constrained BN is obtained from a BN by replacing one or more probabilities
in its probability tables with symbolic expressions, and by adding constraints for variables
used in these expressions or in quantitative terms of interest, and for variables that refer to
C
for constrained BN with set of constraints C
marginal probabilities of interest. We write BX
and variable set X.
To illustrate, in Figure 2 the probability tables for two nodes Sprinkler and Rain of the
BN in Figure 1 are made symbolic with a variable x to obtain a constrained BN. This allows
us to model strict uncertainty (also known as Knightian uncertainty) in the actual value of
such probabilities. Our approach allows variables to be shared across such tables, as x is
shared across the tables for Sprinkler and Rain. This is certainly useful, e.g., to express that
a certain subjective probability is twice as likely as another one.
We use variables mpH and mpW to refer to marginal probabilities
ppHolmes’ Grass Wet = Trueq
ppHolmes’ Grass Wet = True | Watson’s Grass Wet = Trueq
(1)
(2)
respectively. The constraints we then consider are 0 ď x ď 1, to ensure that symbolic
expressions still specify probability distributions, as well as the symbolic meaning of the
marginal probabilities mpH and mpW which are captured in two non-linear equations in x
as
mpH “ 0.495˚x˚x ` 0.5˚x˚p´0.95˚x ` 0.95q `
0.7˚x˚p´0.5˚x ` 1q ` 1.0˚p´0.5˚x ` 1q˚p´0.05˚x ` 0.05q
(3)
mpW ˚p0.35˚x˚x ` 0.025˚x˚p´0.95˚x ` 0.95q ` 0.7˚x˚p´x˚0.5 ` 1q `
0.025˚x˚p´0.05˚x ` 0.05q ` 0.05˚p´0.95˚x ` 0.95q˚p´x˚0.5 ` 1q `
0.05˚p´x˚0.5 ` 1q˚p´0.05˚x ` 0.05qq “
p0.3465˚x˚x ` 0.025˚x˚p´0.95˚x ` 0.95q ` 0.49˚x˚p´x˚0.5 ` 1q `
0.05˚p´x˚0.5 ` 1q˚p´0.05˚x ` 0.05qq
(4)
The above equations are constructed through symbolic interpretations of computations
of marginals, for example of the Junction Tree Algorithm, and subsequent elimination of
5
t ::“ c | mp | x | t ` t | t ˚ t
ϕ ::“ true | t ď t | t ă t | ϕ | ϕ ^ ϕ
φ ::“ ϕ | Dx : φ | φ | φ ^ φ
Figure 3: BNF grammars for real-valued terms t, constraints ϕ, and queries φ where x are
variables from set Xx , c are constant reals, mp are variables from set Xmp denoting marginal
probabilities, and true denotes logical truth
division operators. The latter computes a normal form of rational terms from which these
equations are easily derived.
3.1
Theoretical Foundations
We begin the formal development by defining grammars for symbolic expressions that occur
in probability tables and for properties that contain such expressions as arguments. Figure 3
shows definitions for real-valued terms t, where c ranges over real constants, and x and
mp are real variables ranging over variable sets Xx and Xmp , respectively. The distinction
between mp and x is one of modelling intent. Variables mp refer to marginal probabilities
C
. The meaning of these symbolic marginals is defined via constraints
of a constrained BN BX
in C. Variables in Xx may occur in symbolic expressions in probability tables of nodes or
denote any quantitative measures of interest. We write X “ Xx Y Xmp for the disjoint union
of such variable sets.
Constraints ϕ are quantifier-free formulas built from inequalities over terms t, logical
truth constant true, and propositional operators. Queries φ are built out of constraints and
first-order quantifiers.
Definition 2 We write T rXs for the set of all terms t, CrXs for the set of all constraints
ϕ generated in this manner, and we write QrXs for the set of all queries φ generated in this
manner from variable set X.
We write T , C, and Q whenever X is clear from context and write _, ´, “, ą and so forth
for derived logical, arithmetic, and relational operators.
C
We may think of a constrained BN BX
as a BN B in which entries in probability tables
of nodes may not only be concrete probabilities but terms t of the grammar in Figure 3 over
variable set Xx , and where the BN is enriched with a finite set of constraints C “ tϕi | 1 ď
C
i ď nu. The intuition is that BX
denotes a set of BNs that all have the same graph and the
same structure of probability tables but where probability values may be uncertain, modelled
as arithmetic terms, and subject to application-specific or analysis-specific constraints. The
only difference in two BNs from that set may be in the real number entries in those probability
tables, and those real numbers are instantiations of the specified arithmetic terms such that
all constraints are met. We formalize this:
Definition 3 A constrained BN of type pXx , Xmp q – denoted as X “ Xx Y Xmp by abuse
of notation – is a triple pG, C, πq where G “ pN, Eq is a finite DAG, C a finite set of
constraints from CrXs, and π a tuple pπn qnPN of symbolic probability tables with On as the
6
set of outcomes of random variable at node n:
` ź
˘
πn :
On1 ˆ On Ñ T rXx s
n1 Ppntpnq
Note that a symbolic probability table has the same input type as a formal probability
table, but its output type is a set of terms not the unit interval. Let us first define syntactic
restrictions for constrained BNs.
Definition 4
1. A constrained BN pG, C, πq of type X is well-formed if
(a) X “ Xx Y Xmp equals the set of variables that occur in C
(b) all mp in Xmp have exactly one defining equation mp “ t or mp ˚ t “ t1 in C where
neither t nor t1 contain variables from Xmp .
2. When G and π are determined by context, we refer to a well-formed, constrained BN
C
.
pG, C, πq of type X as BX
Item 1(a) says that all variables in X occur in some constraint from C. Item 1(b) ensures
all variables mp that model marginal probabilities have a defined meaning in C. Note that
item 1(b) is consistent with having other constraints on such variables in C, for example a
constraint saying that 0.1 ď mp ď x ˚ y. These items create a two-level term language, with
variables in Xx informing meaning of variables in Xmp .
A sound, constrained BN has a semantic requirement about its concretizations, which we
now formalize using assignments for quantifier-free formulas.
Definition 5
1. An assignment α is a function α : X Ñ R. For c in R and x in X,
assignment αrx ÞÑ cs equals α except at x, where it outputs c.
2. The meaning αptq of term t in T under α, as well as the judgment α |ù φ for all φ in
Q, are defined in Figure 4.
Note that αptq extends α : X Ñ R to type T rXs Ñ R. The judgment α |ù φ is satisfaction
of first-order logic over the reals. We use these judgments to define the set of concretizations
of a well-formed, constrained BN:
C
Definition 6 Let BX
“ pG, C, πq be a well-formed, constrained BN where G “ pN, Eq. Let
α : X Ñ R be an assignment.
C
C
1. We write BX
rαs for the BN pG, πrαsq that forgets C from BX
and has formal probability
table πrαsn for each node n with πrαsn “ λe : αpπn peqq.
“ C‰
C
2. The set | BX
| of BNs that BX
denotes, its set of concretizations, is
ľ
“ C‰
C
| BX | “ tBX
rαs | α : X Ñ R and α |ù
ϕ1 u
(5)
ϕ1 PC
Note that the formal probability table πrαsn computes πrαsn peq as αptq where t is the
term πn peq in T rXx s. We can now define sound constrained BNs.
7
αpcq
αpxq
αpt1 ` t2 q
αpt1 ˚ t2 q
α |ù true
α |ù t1 ď t2
α |ù t1 ă t2
α |ù ϕ
α |ù ϕ1 ^ ϕ2
α |ù Dx : φ
α |ù φ
α |ù φ1 ^ φ2
“
“
“
“
c
constants denote themselves
αpxq
for all x in X, overloading of notation
αpt1 q ` αpt2 q
αpt1 q ˚ αpt2 q
iff
iff
iff
iff
iff
iff
iff
holds
αpt1 q ď αpt2 q
αpt1 q ă αpt2 q
pnot α |ù ϕq
pα |ù ϕ1 and α |ù ϕ2 q
αrx ÞÑ cs |ù φ for some real number c
pnot α |ù φq
pα |ù φ1 and α |ù φ2 q
Figure 4: Top: meaning of terms for an assignment α, where we identify constants with their
meaning, and use the same symbol ` for syntax and semantics, and similarly for ˚. Bottom:
Semantics of judgment α |ù φ for an assignment α : X Ñ R and φ from Q. This uses the
meaning αptq from Figure 4 and uses ď both as syntax and semantics, similarly for ă
C
C
is sound if
“ pG, C, πq be a well-formed, constrained BN. Then BX
Definition 7 Let BX
C
C
we
have,
for
all
nodes
n
and
inputs
e of πrαsn ,
rαs that are concretizations
of
B
for all BX
X
ř
that πrαsn peq is in r0, 1s and e πrαsn peq “ 1.
Soundness is saying that all concretizations of a well-formed, constrained BN are actually
C
rαs and node n in it, πrαsn is a discrete probability distribution.
BNs: for each such BX
Assumption 1 All constrained BNs used in this paper are sound.
“ C‰
| is non-empty.
It is important to know whether | BX
“ C‰
C
Definition 8 A constrained BX
is consistent iff | BX
| “ H.
The techniques developed in the next Section 3.2 will also allow us to decide whether a
constrained BN is consistent.
3.2
Semantic Judgments
“ C‰
How should we best reason about a set of BNs | BX
| ? We propose two semantic judgments
C
that allow us to explore worst-case and best-case properties of BX
. A judicious combination of
these judgments also enables us to express optimizations over the imprecision and probabilisC
tic uncertainty inherent in BX
, whilst reflecting any application-specific or analysis-specific
constraints. Both semantic judgments rest on a satisfaction relation between concretization
BNs and queries. We define this formally.
8
C
Definition 9 Let BX
be a constrained BN. For all φ in Q, the two semantic judgments |ùmust
and |ùmay are defined as
C must
C
C
BX
|ù φ iff for all BX
rαs in BX
we have α |ù φ
C may
C
C
BX |ù φ iff for some BX
rαs in BX
we have α |ù φ
(6)
(7)
C must
The definition in (6) allows us to discover
truth of BX
|ù φ implies that φ
“ invariants:
‰
C
holds no matter what concrete instance in | BX | the modeller may face, a form of worst-case
C may
reasoning.
the truth of BX
|ù φ in (7) implies it is possible that the modeller faces
“ Dually,
‰
C
a BN in | BX | that satisfies φ, a form of best-case reasoning. We may formalize this duality.
C
Theorem 1 For all constrained BNs BX
and φ in Q we have
C must
C may
1. BX
|ù φ iff not BX
|ù
φ
C may
C must
2. BX
|ù φ iff not BX
|ù
φ
C must
C must
C must
3. BX
|ù φ1 ^ φ2 iff (BX
|ù φ1 and BX
|ù φ2 )
C may
C may
C may
|ù φ2 )
|ù φ1 or BX
|ù φ1 _ φ2 iff (BX
4. BX
We illustrate this formalization of constrained BNs with an example.
Example 1 The constrained BN from Figures 1 and 2 is sound and consistent. Consider the query ϕH being mpH ă 0.3 where we use variable mpH to denote the marginal
C must
probability ppHolmes’ Grass Wet = Trueq. We mean to compute whether BX
|ù ϕH and
C must
C may
BX |ù ϕH hold for this constrained BN. We conclude that BX |ù ϕH does not hold:
ppHolmes’ Grass Wet = Trueq
equals 0.35255 for αpxq “ 0.3 and so αpmpH q “ 0.35255 as
Ź
C may
|ù ϕH holds
0.35255 ę 0.3. On the other hand, BX
well, and we have α |ù ϕ1 PC ϕ1 and
Ź
1
1
1
1
since for α pxq “ 0.1 we have α |ù ϕ1 PC ϕ and α pmpH q “ 0.15695 is less than or equal to
0.3.
Observing additional hard evidence that Watson’s grass is wet, we similarly evaluate judgC must
C may
|ù ϕ when ϕ contains mpW which refers to marginal
ments BX
|ù ϕ and BX
ppHolmes’ Grass Wet = True | Watson’s Grass Wet = Trueq
3.3
Consistent constrained BNs
It is important to understand how the semantic judgments |ùmay and |ùmust relate to consistent
or inconsistent constrained BNs. We can characterize consistency through properties of these
semantic judgments:
C
Theorem 2 Let BX
be a constrained BN. Then the following are all equivalent:
C may
1. BX
|ù true holds.
C
2. BX
is consistent.
C must
C may
3. For all φ in Q, we have that BX
|ù φ implies BX
|ù φ.
9
C may
4. For all φ in Q, we have that BX
|ù φ _ φ holds.
C must
5. For all φ in Q, we have that BX
|ù φ ^ φ does not hold.
C
We stress that it is vital to check the consistency
of B
X prior to relying on any findings
“
‰
C
C
C must
of its further analysis. If BX
is inconsistent, then | BX
| is empty and so BX
|ù φ holds
trivially for all φ in Q since the universal quantification of its defining semantics in (6)
ranges over the empty set. Not detecting such inconsistency may thus lead to unintended
and flawed reasoning. In our tool, this is a non-issue as it uses these judgments within
optimization algorithms that either report a concretization as witness or report a discovered
inconsistency.
Consistency checking is NP-hard: checking the satisfiability of constraints in logic C is
NP-hard. And so this hardness is inherited for any notion of size of a constrained BN that
includes the sum of the sizes of all its constraints.
3.4
Reducing |ùmay and |ùmust to satisfiability checking
Our case studies involving constrained BNs suggest that it suffices to consider elements of C,
i.e. to consider formulas of Q that are quantifier-free. The benefit of having the more expresC may
sive logic Q, however, is that its quantifiers allow us to reduce the decisions for BX
|ù ϕ
C must
and BX |ù ϕ for quantifier-free formulas ϕ to satisfiability checking, respectively validity
C
, the
checking in the logic Q – which we now demonstrate. For sound, constrained BNs Ź
BX
C may
judgment BX |ù ϕ asks whether there is an assignment α : X Ñ R such that α |ù ϕ1 PC ϕ1
and α |ù ϕ both hold. Since C is contained in Q, we may capture this meaning within the
logic Ź
Q itself as a satisfiability check. Let set X equal tx1 , . . . , xn u. Then, asking whether
α |ù ϕ1 PC ϕ1 and α |ù ϕ hold, asks whether the formula in (8) of logic Q is satisfiable:
ľ
Dx1 : . . . : Dxn : ϕ ^
ϕ1
(8)
ϕ1 PC
C
C
Definition 10 For a constrained BN BX
and ϕ in C, we write ExpBX
, ϕq to denote the
formula defined in (8).
C
C
Note that ExpBX
, ϕq depends on BX
: namely on its set of variables X and constraint set
C, the latter reflecting symbolic meanings of marginal probabilities. Let us illustrate this by
revisiting Example 1.
Example 2 For ϕH as in Example 1 with type X “ txu Y tmpH u, the formula we derive for
C may
BX
|ù ϕH is
DmpH : Dx : pmpH ă 0.3q ^ p0.1 ď xq ^ px ď 0.3q ^
p0.5 ˚ x ` p1 ´ 0.5 ˚ xq “ 1q ^ px ` p1 ´ xq “ 1q ^ pmpH “ tq
(9)
where t is the term on the righthand side of the equation in (3).
We can summarize this discussion, where we also appeal to the first item of Theorem 1
to get a similar characterization for |ùmust .
10
C
Theorem 3 Let BX
be a constrained BN and ϕ in C. Then we have:
C
1. Formula ExpBX
, ϕq in (8) is in Q and in the existential fragment of Q.
C may
C
2. Truth of BX
|ù ϕ is equivalent to the satisfiability of ExpBX
, ϕq in Q.
C may
C
3. BX
|ù ϕ can be decided in PSPACE in the size of formula ExpBX
, ϕq.
C must
C
4. BX
|ù ϕ can be decided in PSPACE in the size of formula ExpBX
, ϕq.
This result of deciding semantic judgments in polynomial space pertains to the size of
C
C
formulas ExpBX
, ϕq and ExpBX
, ϕq, and these formulas contain equations that define the
meaning of marginals symbolically. There is therefore an incentive to simplify such symbolic expressions prior to their incorporation into C and these formulas, and we do such
simplifications in our implementation.
3.5
Constrained Union Operator
We also want the ability to compare two or more constrained BNs or to discover relationships
between them. This is facilitated by a notion of composition of constrained BNs, which we
C1
C2
now develop. Consider two constrained BNs BX
and BX
. Our intuition for composition is
1
2
to use a disjoint union of the graphs of each of these constrained BNs such that each node
in this unioned DAG still has its symbolic probability table as before. This union operator
renames nodes that appear in both graphs so that the union is indeed disjoint. As a set of
constraints for the resulting constrained BN, we then consider C1 Y C2 .
It is useful to make this composition depend on another set of constraints C. The idea
is that C can specify known or assumed relationships between these BNs. The resulting
C
composition operator Y defines the composition
C
C1
C2
BX
Y BX
1
2
(10)
as the constrained BN with graph and probability tables obtained by disjoint union of the
C1
C2
graphs and symbolic probability tables of BX
and BX
, where the set of constraints for this
1
2
resulting constrained BN is now C1 Y C2 Y C.
This composition operator has an implicit assumption for being well defined, namely
that C does not contain any equations that (re)define the (symbolic) meaning of marginal
probabilities given in C1 Y C2 .
We give an example of such a union of constrained BNs that already illustrates some
reasoning capabilities to be developed in this paper:
C1
C0
Example 3 Figure 5 specifies a constrained BN BX01 that is similar to constrained BN BX
0
0
defined in Figure 2 but that models rain with more specificity. Variables y and z are used
in symbolic probabilities, and variables mp1H and mp1W refer to the marginals in (1) and (2)
respectively. The constraint 0.1 ď 5˚y ď 0.3 in C01 corresponds to the constraint 0.1 ď x ď 0.3
C0
in C0 and so term 5 ˚ y in some way reflects x, that it rains according to BX
.
0
The constraint set C that binds the two models together is t2 ˚ z “ xu, which ensures
that the probability for the sprinkler to be on is the same in both models. In the constrained
11
Sprinkler
On
Off
z
Heavy
y
Sprinkler/Rain
1´z
Rain
Light
4y
Off
Off
Off
On
On
On
None
1 ´ 5y
N
L
H
N
L
H
Holmes’
T
F
0.05
0.65
0.9
0.95
0.95
0.99
0.95
0.35
0.1
0.05
0.05
0.01
Rain
Watson’s
T
F
N
L
H
0.05 0.95
0.65 0.35
0.9 0.1
C1
Figure 5: Symbolic probability tables for a constrained BN BX01 that has the same DAG
0
C0
as BX
of Figure 2 but is more complex: rain is modelled to a greater degree of specificity.
0
Variable set X01 is ty, zu Y tmp1H , mp1W u. The constraint set C01 includes 0.1 ď 5 ˚ y ď 0.3,
equations that define the meaning of marginals mp1W and mp1W in terms of y and z (not
shown), and equations that ensure that all tables specify probability tables.
C
C1
C0
BN BX
Y BX01 , we want to understand the difference in the marginal probabilities mpW and
0
0
mp1W , expressed by term diff “ mpW ´ mp1W .
Subtraction ´ and equality “ are derived operations in Q. The methods we will develop
in this paper allow us to conclude that the maximal value of diff is in the closed interval
r0.134079500198, 0.134079508781s, with diff being 0.134079500198 when
x “ 0.299999999930
z “ 0.149999999965
y “ 0.020000000003
mpW “ 0.663714285678 mp1W “ 0.529634782614
Similarly, we may infer that the minimum of diff is in the closed interval
r´0.164272228181, ´0.164272221575s
with diff being ´0.164272221575 when
x “ 0.100000000093
z “ 0.050000000046
y “ 0.059999999855
mpW “ 0.472086956699 mp1W “ 0.636359183424
In particular, the absolute value of the difference of the marginal probability (2) in those
constrained BNs is less than 0.1643, attained for the values just shown.
C
C
C1
C2
C2
C1
These union operators are symmetric in that BX
Y BX
and BX
Y BX
satisfy the same
1
2
2
1
must
may
judgments |ù
and |ù
for all φ in Q. Idempotency won’t hold in general as unions may
introduce a new set of constraints C. Associativity holds, assuming all compositions in (11)
give rise to sound constrained BNs:
C
C1
C
C1
C1
C2
C3
C1
C2
C3
pBX
Y BX
q Y BX
is equivalent to BX
Y pBX
Y BX
q
1
2
3
1
2
3
C
(11)
C1
C2
Assumption 2 All composed, constrained BNs BX
Y BX
used in this paper are sound.
1
2
12
3.6
Non-Linear Optimization
We next relate the judgments |ùmay and |ùmust to optimization problems that seek to minimize
C
C
or maximize values of terms t of interest in a constrained BN BX
, and where BX
itself may
well be the result of a composition of constrained BNs as just described. We define the set
C
of “concretizations” of term t for BX
:
C
Definition 11 Let t be a term whose variables
are
all in X for a constrained BN BX
. Then
“
‰
C
C
t| t |u Ď R is defined as set tαptq | BX rαs P | BX | u.
Note that t| t |u does depend on C and X as well, but this dependency will be clear from
context. We can compute approximations of sup t| t |u and inf t| t |u, assuming that these
values are finite. To learn that sup t| t |u is bounded above by a real high, we can check
C must
whether BX
|ù t ď high holds. To learn whether sup t| t |u is bounded below by a real
C may
low, we can check whether BX
|ù low ď t holds. Gaining such knowledge involves both
judgments |ùmust and |ùmay . So we cannot compute approximations of sup t| t |u directly
in the existential fragment of Q but search for approximations by repeatedly deciding such
judgments.
We want to do this without making any assumptions about the implementation of a decision procedure for logic Q or its existential fragment. This can be accommodated through the
use of extended binary search, as seen in Figure 6, to derive an algorithm Sup for computing
a closed interval rlow, highs of length at most δ ą 0 such that sup t| t |u is guaranteed to be
C
with X as set of variables
in rlow, highs. This algorithm has as input a constrained BN BX
for constraint set C, a term t in T rXs, and a desired accuracy δ ą 0. This algorithm assumes
that BCX is consistent and that 0 ă sup t| t |u ă 8. We explain below how we can weaken
those assumptions to sup t| t |u ă 8.
Algorithm Sup first uses
witness α to compute a real value αptq that t can
“ aCsatisfiability
‰
C
attain for some BX
rαs in | BX
| such that αptq ą 0. It then stores this real value in a cache
and increases the value of cache each time it can find a satisfiability witness that makes the
value of t at least twice that of the current cache value. Since sup t| t |u ă 8, this while loop
terminates. The subsequent assignments to low and high establish an invariant that there
is a value in t| t |u that is greater or equal to low, but that there is no value in t| t |u that is
greater or equal to high.
The second while statement maintains this invariant but makes progress using bisection
of the interval rlow, highs. This is achieved by deciding whether there is a value in t| t |u
that is greater or equal to the arithmetic mean of low and high. If so, that mean becomes
the new value of low, otherwise that mean becomes the new value of high. By virtue of these
invariants, the returned closed interval rlow, highs contains sup t| t |u as desired. We capture
this formally:
C
Theorem 4 Let BX
be a consistent constrained BN and δ ą 0. Let 0 ă sup t| t |u ă 8.
Then we have:
C
1. Algorithm Suppt, δ, BX
q terminates, sup t| t |u is in the returned closed interval rl, hs of
C may
length ď δ, and BX |ù t ě l is true.
2. Let c be the initial value of cache. Then the algorithm makes at most t2¨log2 psup t| t |uq´
C
C
log2 pcq ´ log2 pδq ` 1u satisfiability checks for formulas ExpBX
, t ě rq or ExpBX
, t ą rq,
13
C
Suppt, δ, BX
qt
C
let α : X Ñ R make ExpBX
, t ą 0q true;
cache “ αptq;
C
while pExpBX
, t ě 2 ˚ cacheq satisfiableq t
1
C
let α : X Ñ R make ExpBX
, t ě 2 ˚ cacheq true;
1
cache “ α ptq;
u
low “ cache; high “ 2 ˚ cache;
C
C
assert ppExpBX
, t ě lowq satisfiableq && pExpBX
, t ě highq unsatisfiableqq;
while p| high ´ low | ą δq t
C
if pExpBX
, t ě low ` | high ´ low | { 2qq satisfiableq t
low “ low ` | high ´ low | { 2;
C
C
assert ppExpBX
, t ě lowq satisfiableq && pExpBX
, t ě highq unsatisfiableqq;
u else t
high “ low ` | high ´ low | { 2;
C
C
, t ě highq unsatisfiableqq;
, t ě lowq satisfiableq && ExpBX
assert ppExpBX
u
u
return rlow, highs;
u
Figure 6: Algorithm for approximating sup t| t |u up to δ ą 0 for a consistent, constrained
C
and term t with variables in X when 0 ă sup t| t |u ă 8. The returned closed interval
BN BX
rlow, highs has length ď δ and contains sup t| t |u. Key invariants are given as asserts
and these formulas only differ in the choice of comparison operator and in the value of
real constant r.
We now give an example of using algorithm Sup. Our specifications of optimization
algorithms such as that of algorithm Sup in Figure 6 do not return witness information, we
omitted such details for sake of simplicity.
C0
C0
Example 4 For constrained BN BX
of Figure 2, SuppmpW , δ, BX
q terminates for δ “
0
0
0.000000001 with output r0.663714282364, 0.663714291751s. The value 0.663714282364 is
attained when x equals 0.299999999188.
C
C
An algorithm Infpt, δ, BX
q is defined in Figure 7. It assumes that BX
is consistent and
`
that inf t| t |u is a subset of R0 and contains a positive real – conditions we will weaken
below. In that case, it terminates and returns a closed interval rl, hs such that inf t| t |u is in
rl, hs. We prove this formally:
C
Theorem 5 Let BX
be a consistent constrained BN and δ ą 0. Let t| t |u Ď R`
0 contain a
positive real. Then we have:
C
1. Algorithm Infpt, δ, BX
q terminates and inf t| t |u is in the returned interval rl, hs such
C may
that h ´ l ď δ and BX |ù t ď h are true.
14
C
Infpt, δ, BX
qt
C
let α : X Ñ R make ExpBX
, t ą 0q true;
cache “ αptq;
C
while pExpBX
, t ď 0.5 ˚ cacheq satisfiable and 0.5 ˚ cache ą δq t
1
C
let α : X Ñ R make ExpBX
, t ď 0.5 ˚ cacheq true;
1
cache “ α ptq;
u
C
if pExpBX
, t ď 0.5 ˚ cacheq satisfiableq t return r0, 0.5 ˚ caches; u
low “ 0.5 ˚ cache; high “ cache;
C
C
assert pExpBX
, t ď lowq unsatisfiableq && pExpBX
, t ď highq satisfiableqq;
while p| high ´ low | ą δq t
C
if pExpBX
, t ď low ` | high ´ low | { 2qq satisfiableq t
high “ low ` | high ´ low | { 2;
C
C
assert ppExpBX
, t ď lowq unsatisfiableq && pExpBX
, t ď highq satisfiableqq;
u else t
low “ low ` | high ´ low | { 2;
C
C
, t ď highq satisfiableqq;
, t ď lowq unsatisfiableq && pExpBX
assert ppExpBX
u
u
return rlow, highs;
u
Figure 7: Algorithm for approximating inf t| t |u up to δ ą 0 for a consistent, constrained
C
and term t in T rXs when t| t |u Ď R`
BN BX
0 contains a positive real.
2. Let c be the initial value of cache. Then the algorithm makes one satisfiability check
C
, t ą 0q and at most t2 ¨ log2 pcq ´ log2 pminpδ, inf t| t |uqqu satisfiability checks for
ExpBX
C
formulas ExpBX
, t ď rq, and these formulas only differ in the size of real constant r.
C
being consistent and of 0 ă
We now show how we can relax the conditions of BX
sup t| t |u ă 8 to sup t| t |u ă 8. In Figure 8, we see this modified algorithm Sup˚ which
relies on both Sup and Inf. It returns a closed interval with the same properties as that
returned by Sup but where sup t| t |u only need be finite. We state the correctness of this
algorithm formally:
C
C
Theorem 6 Let BX
be a constrained BN, δ ą 0, and sup t| t |u ă 8. Then Sup‹ pt, δ, BX
q
terminates and its calls to Sup and Inf meet their preconditions. Moreover, it either correctly
C
identifies that BX
is inconsistent, that 0 is the maximum of t| t |u or it returns a closed
interval rl, hs such that sup t| t |u is in that interval, h ´ l ď is less than or equal to δ, and
C may
BX
|ù t ě l holds.
We conclude this section by leveraging Sup‹ to an algorithm Inf ‹ , seen in Figure 9. Algorithm Inf ‹ relaxes that t| t |u contains a positive real and is a subset of R`
0 to a more general
pre-condition ´8 ă inf t| t |u, and it has correct output for inconsistent, constrained BNs.
We formalize this:
15
C
Sup‹ pt, δ, BX
qt
C
if pExpBX , t ą 0q satisfiableq t
C
return Suppt, δ, BX
q; u
C
elseif pExpBX , t “ 0q satisfiableq t
return 0 as maximum for t; u
C
elseif pExpBX
, t ă 0q satisfiableq t
C
let rl, hs “ Infp´t, δ, BX
q;
return r´h, ´ls;
u
C
return BX
is inconsistent;
u
Figure 8: Algorithm Sup‹ uses algorithms Sup and Inf and terminates whenever sup t| t |u ă
8. It either recognizing that 0 is the maximum of t| t |u, returns a closed interval rl, hs with
C
is inconsistent
h ´ l ď δ such that sup t| t |u is in rl, hs, or it detects that BX
C
Inf ‹ pt, δ, BX
qt
C
let x “ Sup‹ p´t, δ, BX
q;
C
C
if px reports that BX is inconsistentq t return BX
is inconsistent; u
elseif px reports 0 as maximum for ´tq t return 0 as minimum for t; u
elseif px reports interval rl, hsq t return r´h, ´ls; u
u
Figure 9: Algorithm Inf ‹ uses algorithm Sup‹ and terminates whenever ´8 ă inf t| t |u. It
either recognizes that 0 is the minimum of t| t |u, returns a closed interval rl, hs with h ´ l ď δ
C
such that inf t| t |u is in rl, hs, or it detects that BX
is inconsistent
16
C
Theorem 7 Let BX
be a constrained BN, δ ą 0, and t a term with ´8 ă inf t| t |u. Then
‹
C
C
Inf pt, δ, BX q terminates and either correctly identifies that BX
is inconsistent, that 0 is the
minimum of t| t |u or it returns a closed interval rl, hs of length ď δ such that inf t| t |u is in
C may
rl, hs and BX
|ù t ď h holds.
Let us revisit Example 3 to illustrate use of Sup‹ .
C
C1
C̃0
Example 5 Let C̃0 be C0 Y t0.1 ď x ď 0.2u. For constrained BN BX
Y BX01 , we maximise
0
0
diff using Sup
‹
C1
C̃0 C
pdiff, 0.000000001, BX
Y BX01 q,
0
0
which returns the interval
r´0.055219501217, ´0.0552194960809s
C
C
arising from the third case of Sup‹ as both ExpBX
, t ą 0q and ExpBX
, t “ 0q are unsatisfiable,
C
but formula ExpBX , t ă 0q is satisfiable. It shows that marginal mpW is always smaller than
marginal mp1W in this constrained BN, in contrast to the situation of Example 3.
4
Detailed Case Study
We now apply and evaluate the foundations for constrained BNs on a case study in the context
of arms control. Article VI of the Treaty on the Non-Proliferation of Nuclear Weapons (NPT)
[1] states that each treaty party
“undertakes to pursue negotiations in good faith on effective measures relating to
cessation of the nuclear arms race at an early date and to nuclear disarmament,
and on a treaty on general and complete disarmament under strict and effective
international control.”
One important aspect of meeting such treaty obligations may be the creation and execution
of trustworthy inspection processes, for example to verify that a treaty-accountable item
has been made inoperable. Designing such processes is challenging as it needs to guarantee
sufficient mutual trust between the inspected and inspecting party in the presence of potentially conflicting interests. Without such trust, the parties might not agree to conduct such
inspections.
The potential benefit of mathematical models for the design and evaluation of such inspection processes is apparent. Bayesian Networks can capture a form of trust – through
an inherent bias of processing imperfect information – and different degrees of beliefs –
expressed, e.g., in subjective probabilities. Bayesian Networks can also represent objective
data accurately, and their graphical formalism may be understood by domain experts such as
diplomats. These are good reasons for exploring Bayesian Networks for modeling and evaluating inspection processes. But Bayesian Networks do not seem to have means of building
confidence in their adequacy and utility, especially in this domain in which prior data for
learning both graph structure and probabilities at nodes in such a graph are hard to find.
We now show how constrained BNs can be used to build such confidence in mathematical
models of an inspection process.
17
Figure 10: Prototype of an information barrier (IB), photo taken from [3] , that the inspecting
nation might build. The IB would output either a green or red light to inspectors based on
physical measurements made within the IB that verify the presence of nuclear material
4.1
An Arms Control Inspection Process
Consider the situation of two fictitious nation states. The inspecting nation is tasked with
identifying whether an item belonging to the host nation, available to inspect in a controlled
inspection facility and declared by the host nation to be a nuclear weapon, is indeed a nuclear
weapon. This situation is similar to a scenario that had been explored in the UK/Norway
initiative in 2007 [3, 4].
Given the nations’ non-proliferation obligations and national security concerns, the design
details of the inspected item must be protected: the inspecting nation will have no visual
access to the item. Instead the nations agree that the to-be-inspected item contain Plutonium
with the isotopic ratio 240Pu:239Pu below a certain threshold value, which they set at 0.1.
In order to draw conclusions about whether an item presented for inspection is a weapon,
the inspecting nation uses an information barrier (IB) system comprising a HPGe detector
and bespoke electronics with well-understood performance characteristics (see Figure 10,
[3]) to conduct measurements on the item while the item is concealed in a box. The IB
system displays a green light if it detects a gamma-ray spectrum indicative of the presence
of Plutonium with the appropriate isotopic ratio; if it does not detect this spectrum for
whatever reason, it shows a red light. No other information is provided, and weapon-design
information is thus protected [4].
The inspecting nation believes that it may be possible for the host nation to spoof a
radioactive signal – or in some way provide a surrogate – to fool the detector, or that the
host nation may have just placed Plutonium with the appropriate isotopic ratio in the box
rather than a weapon. These subjective assessments should be reflected in the mathematical
model alongside the error rates of the IB system that reflect the reliability of that device.
In order to deter cheating, the inspecting nation is allowed to choose the IBs used in the
verification from a pool of such devices provided by the host nation, and may choose one or
two IBs to that end. From that same pool of devices, the inspecting nation may take some
unused IBs away for authentication – activities designed to assess whether the host nation
tampered with the IBs. But the inspecting nation must not inspect any used IBs, to prevent
it from exploiting any residual information still present in such used IBs to its advantage.
This selection process of IBs is therefore designed to ensure that a nefarious host nation
18
is held at risk of detection should it decide to tamper with the IBs used in verification: it
would run the risk of one or more tampered IBs being selected for authentication by the
inspecting nation. Although such authentication cannot be assumed to be perfect – and this
fact, too, should be modelled – the prospect of detection may deter such a host.
We model this inspection process through constrained BNs that are abstracted from a
sole BN with DAG shown in Figure 11 and based on a design developed by the Arms Control
Verification Research group at AWE. This DAG depicts different aspects of the verification
procedure in four key areas:
• the selection of the IBs for inspection or authentication purposes,
• the workings of the IB in the inspection itself,
• authentication of (other) IBs, and
• the combination of these aspects to assess any possibility of cheating overall, be it
through IB tampering, surrogate nuclear sources, and so forth.
The selection of the IBs starts with the IB pool size; a selection of IBs built by the host
nation, from which there will be a Number of IBs picked for authentication and Number of
IBs picked for use by the inspecting nation. Should a Number of tampered IBs exist, then
the selection process (blind to such a tamper) follows a Hypergeometric distribution and will
probabilistically determine whether such tampered IBs make it into use in the verification
process, authentication process or neither. The choice of distribution reflects that IBs – once
chosen for either verification or authentication – cannot be used for any other purpose.
The IBs picked for either authentication or verification help the inspecting nation to judge
whether the item under inspection Is a weapon. A weapon or a Surrogate Pu source determine
physical nuclear properties about the Isotopic ratio of Plutonium elements. Our mathematical
model captures a possible inspector judgment that a surrogate source would only be used if
the host felt that it was extremely likely to pass the IBs verification tests. Therefore, any
surrogate source would have isotopic properties at least as good as those of a real weapon.
We stress that the probabilities chosen for each isotopic ratio, conditioned on whether
the item under test is or is not a weapon, are not derived from real-world weapons data, but
instead reflect in broad terms that Plutonium with a higher isotopic ratio than the chosen
threshold is less likely to be found in a nuclear weapon. A bespoke algorithm is used by the
IB system on the collected gamma-ray spectrum to test whether both the Peaks are in the
expected locations and the Peak aspect ratio are as expected. If all 5 peaks are present and
the Ratio of 240/239 isotopes is acceptable, then one or both of the First IB result or Second
IB result are reported, conditional on any tampering and depending on whether or not two
IBs are used to test the same item.
A mathematical model cannot hope to reflect each potential tamper. Therefore, we model
authentication as an assessment of the Inspector’s authentication capabilities: the better these
are, the more likely the Tamper will be found, and this requires that at least one tampered
IB exists and was selected for authentication. This is controlled by the parent nodes: the
aforementioned Hypergeometric distribution, and a node Chance of picking a tampered IB for
authentication.
19
Figure 11: A BN [7] which details aspects of an arms inspection process. Aspects of the
1-2 IB devices used for verification are modeled in blue and green nodes, respectively. The
assessment of cheating and the operation of the inspection in other ways are shown in orange
nodes, and authentication procedures are modeled in purple nodes. Mathematical or logical
computations are represented in grey nodes.
20
The mathematical model is drawn together by the overarching question of “Is the Host
cheating?”. If so, we then determine a Cheating method, which reflects the understanding of
the inspecting nation about the possible ways that the host nation could try to cheat, as
outlined above, and the prior beliefs of the inspecting nation about the relative likelihood of
the use of each method if the host nation were to be cheating.
Finally, we check whether a Portal monitor is used to stop transportation of radioactive
material – which could be used as a surrogate source – in and out of the facility, although
we do not model this aspect in greater detail.
The probabilities used in this BN stem from a variety of sources. Some are somewhat
arbitrarily selected, as described above, and therefore need means of building confidence in
their choice. Probabilities relating to the performance of the IB system are derived from
experimental analysis of the UKNI IB [4, 3].
The size of the probability tables for nodes of the BN in Figure 11 range from small (a few
or tens of entries), to medium (hundreds of entries) and larger ones (thousands of entries).
Given that complexity, we refrain from specifying more details on these tables within the
paper itself.
Our evaluation of the methods developed in Section 3 will abstract the BN described
above (see Figure 11) into constrained BNs, and demonstrate that these abstractions can
inform decision support given the sparsity or lack of prior data that informed its choices of
probabilities.
Assumption 3 For convenience, this case study will not explicitly list or show the constraints that define the meaning of marginals symbolically. These meanings are included in
the open-access research code cited on page 34.
4.2
Impact of Cheating Method on Tamper Detection
We want to understand how the choice of cheating method can impact the probability of
detecting a tamper. The uncertainty about what cheating method the host nation will adopt
C1
is modelled in a constrained BN BX
that takes the BN from Figure 11 and replaces the
1
probability table for its node Cheating Method as specified in Figure 12. We use variables x,
y, and u to denote, respectively, the probability of IB tamper only, Surrogate source tamper
only, and both IB tamper and surrogate source tamper. The variable mptf refers to the
marginal probability ppWill tamper be found? = Yesq.
We compute the interval rl, hs “ r0.197137598548, 0.197137608314s as output of the funcC1
tion call Sup‹ pmptf , 0.00000001, BX
q. The witness information for the existentially quantified
1
variables x, y, u, and mptf pertains to value l “ 0.197137598548:
x “ 0.000000010001153 y “ 0.000000010001153 u “ 0.999999979997693
We compute the interval r˜l, h̃s “ r5.875158e´09, 1.1750316e´08s as output of function call
C1
Inf ‹ pmptf , 0.00000001, BX
q. The witness information is now for the value h̃ “ 1.1750316e´08
1
of mptf and we get
x “ 0.000000030265893 y “ 0.999999939468212 u “ 0.000000030265893
21
Cheating Method
None
IB tamper only
Surrogate source only
IB tamper & surrogate source
Is cheating
0
x
y
u
Is not cheating
1
0
0
0
C1
Figure 12: Probability table for node Cheating Method in constrained BN BX
where C1
1
contains t0 ă x, y, u ă 1.0, x ` y ` u “ 1.0u, Xx “ tx, y, uu, Xmp “ tmptf u, X1 “ Xx Y Xmp ,
C1
and the BN graph and all other probability tables for BX
are as for the BN in Figure 11
1
We may combine this information, for example to bound the range of values that mptf can
possibly attain, as the interval
r˜l, hs “ r0.000000000587, 0.197137608314s
We therefore conclude that this marginal probability can only vary by less than 0.19714 in
the given strict uncertainty of the model.
Let us now ask for what values of x can mptf be within 0.01 of the lower bound l “
C1
0.197137598548 returned for Sup‹ above. To that end, we consider the constrained BN BX11
where C11 “ C1 Y t| mptf ´ 0.197137598548 | ď 0.01u and compute lower and upper bounds
for x in this constrained BN:
C1
rlx , hx s “ Sup‹ px, 0.00000001, BX11 q “ r0.999999994824, 1.00000000196s
1
C
r˜lx , h̃x s “ Inf ‹ px, 0.00000001, BX11 q “ r7.4505805e´09, 1.4901161e´08s
From this we can learn that
ľ ‰
“
@x : p1.4901161e´08 ď x ď 0.999999994824q ^
C1 Ñ
| mptf ´ 0.197137598548 | ď 0.01
(12)
is logically valid: whenever x is in that value range and all constraints in C1 are satisfied
C1
(which is true for all concretizations of BX
), then the marginal mptf is within 0.01 of the
1
lower bound for its maximal value.
Repeating these optimizations above for variables y and u, we determine similar formulas
that are logically valid:
ľ ‰
“
@y : p1.209402e´08 ď y ď 0.0507259986533q ^
C1 Ñ
| mptf ´ 0.197137598548 | ď 0.01
ľ ‰
@u : p1.4901161e´08 ď u ď 0.999999998164q ^
C1 Ñ
“
| mptf ´ 0.197137598548 | ď 0.01
These results say that the marginal mptf is insensitive to changes to x, which is able
to vary across the whole range p0.0, 1.0q without having much impact on the results; the
22
Authentication Capabilities
Good
x
Medium
0.3333
Poor
0.6667 ´ x
C2
Figure 13: Probability table for node Authentication Capabilities in constrained BN BX
that
2
is like that BN in Figure 11 except that the symbolic probability table for node Authentication
Capabilities is as above, Xx “ txu, Xmp “ tmptf2 u, and C2 contains t0 ď x ď 0.6667, x `
0.3333 ` p0.6667 ´ xq “ 1u. Variable mptf2 denotes the marginal in (14). Constrained BN
C1
C2
BX21 is a “copy” of BX
that replaces all occurrences of x with y and has Xmp “ tmptf21 u
2
2
where mptf21 denotes the marginal in (15)
situation is very similar for variable u. For variable y, the range at which mptf is not too
sensitive on changes of y is much smaller – just over 0.05. Overall, we conclude that the
model remains in the area of highest probability for detecting tampering as long as x or u
are large.
Our analysis shows that the “tamper” cheating method is the one for which there is the
highest chance of detecting cheating. However, our results also highlight that unless both
tamper and surrogate source, or tamper on its own are used, there are limited ways in which
to detect cheating through these nodes. From this we learn that use of a portal monitor is
advisable, as any increase in y moves the marginal out of the region of highest probability of
detecting cheating, and decreases the chance of cheating being detected otherwise. Related to
this is that the range of y gives potential insight into future work required on tamper detection
for the inspecting nation. Despite contributing neither to an IB tamper nor detection, y can
vary by over 0.05 – over five times that of the movement away from the marginal mptf ’s
maximum point by only 0.01. This suggests there are other limiting factors to tamper
detection, such as capability, that could be better reflected in a mathematical model.
4.3
Comparing two BN models
We now illustrate the benefits of composing two constrained BNs (see Section 3.5). Two
C1
C2
constrained BNs, BX
and BX21 , are defined in Figure 13. Both have symbolic and equivalent
2
2
probability tables for node Authentication Capabilities but consider different hard evidence
C2
for the probability of a tamper to be found. In BX
, there is 1 IB machine picked for
2
C1
authentication whereas in BX21 there are 5 IB machines picked to that end, resulting in the
2
respective marginals
ppWill tamper be found? = Yes | Host cheating = Yes,
Number of IBs picked for authentication = 1q
ppWill tamper be found? = Yes | Host cheating = Yes,
Number of IBs picked for authentication = 5q
(13)
(14)
In both models, the probability for state “Good” is bounded by 0.6667 so that there is a
“gradient” pivoting around Medium capabilities fixed at 0.3333.
23
Figure 14: The surface in blue shows the values of diff in (15) for the sample points defined
in (16). The plane diff “ 0 is shown in red. Its intersection with the blue surface marks
the boundary of where decision support would favor running 1 IB authentication (above the
red plane) and 5 IB authentications (below the red plane). The linear equation defining this
boundary line is 0 “ ´0.212884507399337˚x ` 0.354426098468987˚y ´ 0.16515544651.
We seek decision support on how much to prioritise research into IB authentication caC21
C2
pabilities, each of BX
and
B
X21 representing a different capability scenario. Of interest here
2
is the change in the likelihood that a tamper will be found. We can simply model this by
defining a new term
diff “ mptf2 ´ mptf21
(15)
C
C1
C2
Variable diff is in Xx for the constrained BN BX
Y BX21 where the constraint set C for this
2
2
combination is tdiff “ mptf2 ´ mptf21 u. We compute the value of diff for each combination of
values px, yq from set
S “ tp0.0 ` 0.01 ¨ a, 0.0 ` 0.01 ¨ bq | 0 ď a, b ď 67u
(16)
and linearly interpolate the result as a surface seen in Figure 14. The linear relationship
between the symbolic probabilities of node Authentication Capability to that of its child node
Will the tamper be found? make this surface flat.
We can now use the method familiar from our earlier analyses to assess the value range of
C1
C2 C
term diff in this composed, constrained BN. The function call Sup‹ pdiff, 0.00000001, BX
Y BX21 q
2
2
returns the interval
rl, hs “ r0.0711404333363, 0.0711404338663s
C
C1
C2
Next, Inf ‹ pdiff, 0.00000001, BX
Y BX21 q is computed as the interval
2
2
r˜l, h̃s “ r´0.307085548061, ´0.307085547533s
C
C1
C2
In particular, the values of diff for all concretizations of BX
Y BX21 lie in the interval
2
2
r´0.307085548061, 0.0711404338663s
24
The blue surface of diff in Figure 14 is mostly negative (below the red plane). This shows that
the case of testing 5 IBs for tampers is nearly always better, irrespective of the confidence
one may have in one’s ability to find a tamper. This is true, other than for the most extreme
cases when there is the least confidence in authentication capabilities when testing five IBs
(for y “ 0) and most confidence when testing one (for x “ 0.667).
Let us next explore a situation in which the inspector believes to have high authentication
capabilities, regardless of whether 1 or 5 IBs are picked for authentication. We can easily
model this by setting C 1 “ C Y t0.467 ď x, y ď 0.667u and refining the composed model
C 1
C1
C2
using C 1 . We compute the output of Sup‹ pdiff, 0.00000001, BX
Y BX21 q to be the interval
2
2
rl, hs “ r´0.0282766319763, ´0.0282766314489s
C 1
C1
C2
and Inf ‹ pdiff, 0.00000001, BX
Y BX21 q to be the interval
2
2
r˜l, h̃s “ r´0.141568560141, ´0.141568559299s
Now diff is in r´0.141568560141, ´0.0282766314489s and the largest absolute difference between picking 1 and 5 IBs for authentication is greater than 0.14, witnessed when the inspector has a particularly high capability in authenticating 5 IBs, (when y “ 0.667 and
mptf21 “ 0.24932) compared with only inspecting one IB with more moderate capability
(when x “ 0.467, mptf 2 “ 0.10696).
A decision maker could vary the use of the above approach in order to weigh the cost of
IB production against the cost of developing and employing more advanced authentication
capabilities. He or she could also query in detail how the results of such cost-benefit analyses
might change as new information is learned or new techniques deployed. This capability
might help decision makers to balance their priorities and to gain the best assurance possible
within a cost budget that the verification regime they implement is effective.
4.4
Determining equivalent decision support
We assess the consistency of two different constrained BNs of equal intent of decision support.
C31
C3
C1
Constrained BNs BX
and
B
X31 are identical to BX1 and its symbolic probability table for
3
node Cheating Method as in Figure 12, except that th is an additional variable used to model
C3
decision support. Variable set Xmp also changes. For BX
we have Xmp “ tmptf3 u and for
3
C1
BX31 we set Xmp “ tmptf31 u instead. Variable mptf3 denotes marginal probabilities for hard
3
evidence that Initial Pool Size = 10 IBs in (17), whereas variable mptf31 denotes a marginal
for hard evidence Initial Pool Size = 20 IBs in (18):
ppWill tamper be found? = Yes | Initial Pool Size = 10q
ppWill tamper be found? = Yes | Initial Pool Size = 20q
(17)
(18)
These are marginal probabilities that the nation which is authenticating IBs will find a
tamper. A decision – for example that an IB has been tampered with – may then be supported
if such a marginal is above a certain threshold th. We now want to understand whether the
25
two constrained BNs would support decisions in the same manner, and for what values or
value ranges of th.
For any value th, consider the constraint ϕth in Q given by
“`
˘ `
˘‰
pth ă mptf3 q ^ pmptf31 ď thq _ pth ă mptf31 q ^ pmptf3 ď thq
We can now analyze whether both constrained BNs will always support decisions through
threshold th by evaluating
C31
C3 C
must
BX
Y
B
ϕth
(19)
X 1 |ù
3
3
where C equals t0 ă th ă 1u. By Theorem 1, judgment (19) is equivalent to
`
˘ `
˘
C31
C3 C
may
1 ď thq _ pth ă mptf 1 q ^ pmptf ď thq
not BX
Y
B
|ù
pth
ă
mp
q
^
pmp
1
tf
tf
3
3
X
3
3
3
3
(20)
Setting ϕ1 ” pth ă mptf3 q ^ pmptf31 ď thq and ϕ2 ” pth ă mptf31 q ^ pmptf3 ď thq, the same
theorem tells us that (20) is equivalent to
“
‰
‰
“
C1
C1
C3 C
C3 C
Y BX31 |ùmay ϕ2
not BX
Y BX31 |ùmay ϕ1 and not BX
3
3
3
3
(21)
C1
C
C3
Using our tool, we determine that ExpBX
Y BX31 , ϕ1 q is unsatisfiable and so – by appeal to
3
3
Theorem 3 – the first proof obligation of (21) holds. Similarly, we evaluate the satisfiability
C
C1
C3
of ExpBX
Y BX31 , ϕ2 q. Our tool reports this to be satisfiable and so the two constrained BNs
3
3
do not always support the same decision. We now want to utilize our non-linear optimization
method to compute ranges of the th itself for which both models render the same decision.
Understanding such a range will be useful to a modeller as both models are then discovered
to be in agreement for all values of th in such a range.
C1
C3 C
Since ExpBX
Y BX31 , ϕ1 q is unsatisfiable, we use C 1 “ t0 ă th ă 1, ϕ2 u which forces truth
3
3
C1
C1
C3
of ϕ2 , and compute Sup‹ pth, 0.00000001, BX
Y BX31 q to maximise expression th. This obtains
3
3
the interval
rl, hs “ r0.259147588164, 0.259147588909s
C1
C1
C3
Computing Inf ‹ pth, 0.00000001, BX
Y BX31 q outputs the interval
3
3
r˜l, h̃s “ r´9.31322e´10, ´4.65661e´10s
For the given accuracy δ, the interval r˜l, h̃s may be interpreted as 0. Thus, we can say that for
C1
C3
all th in r0, 0.259147588909s the use of either BX
or BX33 could support different decisions.
3
More importantly, we now know that both constrained BNs always support the same decision
as described above when the value of the threshold th for decision making is greater or equal
to 0.2592, say.
The range of th for which both models can support different decisions may seem rather
large and it may be surprising that it goes down to zero. But this is a function of the chance
and capability of finding a tamper in an IB. Intuitively, the models tend to disagree most in
situations where the chance of cheating by tampering is highest, when x “ 1, and thus where
26
Cheating Method
None
IB tamper only
Surrogate source only
IB tamper & surrogate source
Is cheating
0
x
0.6666 ´ x
0.3334
Is not cheating
1
0
0
0
C4
Figure 15: Probability table for node Cheating Method in constrained BN BX
where C4
4
contains t0 ă x ă 0.6666, 0 ` x ` p0.6666 ´ xq ` 0.3334 “ 1.0u, Xx “ txu, Xmp “ tmptf u,
X4 “ Xx Y Xmp , and where the BN graph is that of Figure 11. All other symbolic probability
C4
tables for BX
are as for the BN in Figure 11
4
authenticating the IB has benefit. Our approach gave a decision maker safe knowledge that
any threshold for decision making outside the range r´9.31322e´10, 0.259147588909s would
statistically agree and lead to the same decision regarding finding tampers, irrespective of
the initial number of IBs – either 10 or 20 – in the pool. Dependent on the nations involved,
and the tolerances for decision making they are willing to set, it could be decided – for
instance – that building only 10 IBs per inspection would be enshrined in the treaty to avoid
unnecessary expense and so forth. This would undoubtedly be an important data-driven
decision for diplomats and negotiators to make.
4.5
Symbolic sensitivity analysis
It is well known that BNs may be sensitive to small changes in probability values in tables
of some nodes. Sensitivity analyses have therefore been devised as a means for assessing the
degree of such sensitivities and the impact this may have on decision support. See, e.g., the
sensitivity value defined in [32, 31].
We now leverage such analyses to our approach by computing such sensitivity measures
symbolically as terms of the logic Q. Then we may analyze such terms using the methods
Sup‹ and Inf ‹ as before to understand how such sensitivity measures may vary across conC4
cretizations of a constrained BN. We illustrate this capability for constrained BN BX
, which
4
C1
is similar to BX1 but has probability table for node Cheating Method as shown in Figure 15.
The sensitivity value describes the change in the posterior output of the hypothesis for
small variations in the likelihood of the evidence under study. The larger the sensitivity
value, the less robust the posterior output of the hypothesis. In other words, a likelihood
value with a large sensitivity value is prone to generate an inaccurate posterior output. If
the sensitivity value is less than 1, then a small change in the likelihood value has a minimal
effect on the result of the posterior output of the hypothesis.
A modeller may be uncertain about the sensitivity of event Will tamper be found? =
Yes to the authentication of IBs if probabilities in node Authentication Capability of the
IB were to change. Our tool can compute such a sensitivity value s symbolically for the
sensitivity of event Will tamper be found? = Yes to small perturbations in probabilities of
node Authentication Capability.
27
s = 250.0*(0.000333567254313619*x + 0.171472799414097)*
(0.000400681386562896*x + 0.205973332629545)**2*
(0.000400681386562899*x + 0.205973332629545)*
(0.00133627242418726*x + 0.686921064319534)/
((0.19713762029366*x + 0.0638269490499437)*
(4.01363933844916e-6*x**2 + 0.00412648402564936*x + 1.06062534386303)**2)
C4
Figure 16: Symbolic sensitivity value s in constrained BN BX
for node Will tamper be found?
4
in IB with respect to event Authentication Capability, as a function of sole variable x in Xx
The sensitivity value [32, 31] is defined in this instance as
s“
P O ˚ p1 ´ P Oxq ˚ P x
pP O ˚ P xO ` p1 ´ P Oxq ˚ P xq2
(22)
where terms P O, P x, P Ox and P xO are defined as
PO
Px
P Ox
P xO
”
”
”
”
ppAuthentication Capability = Lowq
ppFinding a tamper in IB = Yesq
ppAuthentication Capability = Low | Finding a tamper in IB = Yesq
ppFinding a tamper in IB = Yes | Authentication Capability = Lowq
(23)
All three marginals of Authentication Capability are considered using just two functions P Ox
and 1 ´ P Ox. In (23), P O is a modelling choice that combines the states of Medium and
High into one state (1 ´ P Ox). Term 1 ´ P Ox accounts for situations in which an inspector
is relatively good at authentication, with P Ox representing situations in which they are less
capable. Other modelling choices would lead to a marginally small difference in s.
Our tool can compute an explicit function of s in variable x, as defined in (22). This
symbolic expression for s is depicted in Figure 16 and shown as a function of x in Figure 17.
This confirms that as the value of x increases, and thus the probability of “IB tamper only”
seen in Figure 15 decreases, the marginal of interest for Will tamper be found? = Yes becomes
less sensitive to changes in the probabilities of the node Authentication Capability of IB.
We can now determine the worst-case sensitivity value by computing the interval returned
C4
by function call Sup‹ pts , 0.00000001, BX
q as r3.5838265468, 3.5838265475s where ts is the
4
term in the righthand side of the equation in Figure 16 that describes s as a function of x.
Thus we learn that this sensitivity cannot be larger than 3.5838265475 for all concretizations
C4
of constrained BN BX
. As is evident from the graph, there are no valid values of x where
4
the sensitivity value drops below 1.0 – the aforementioned bound at which a sensitivity score,
and therefore its corresponding marginal probability, is deemed to be robust.
C4
The output r1.17313380051, 1.17313380116s of Inf ‹ ps, 0.00000001, BX
q confirms this, and
4
C4
shows that s is always above 1.17313380051 for all concretizations of constrained BN BX
.
4
Knowing this may indicate to a decision-maker that potential deviations in the real domain
from the model of node Authentication Capability will require close attention, irrespective of
the value of x and thus of the perceived marginal probabilities of the states of node Cheating
Method.
28
Figure 17: Sensitivity value s of Figure 16 and (22) as a function of x
5
5.1
Implementation and Evaluation
Software Engineering
The numerical results reported in previous sections were computed by a prototype implementation of the approach developed in this paper. This implementation uses Python to capture
a data model for Bayesian Networks and constraints, to formulate marginals of interest, and
to interface with the SMT solver Z3 [23]. The latter we use as a decision procedure for logic Q
that also returns witness information for all variables. The computation of symbolic meaning
of marginals relies on the Junction Tree Algorithm and is achieved through software from an
open-source Python package provided in [5].
Python also supports a lightweight and open-source library for symbolic computation,
sympy [34], which we can employ to run the Junction Tree Algorithm in [5] fully symbolically.
The generated symbolic expressions are then simplified using a method of sympy before they
are put into constraints such as in (3) and added to the SMT solver for analysis.
5.2
Validation and Evaluation
Some symbolic marginals that we generated for analyses but not reported in our case study
were too large to be handled by the SMT solver we used: the string representation of the
symbolic meaning was about 25 Megabytes. We performed linear regression on those symbolic
expressions and then validated that this approximation has higher precision than the accuracy
δ, before defining the meaning of marginal variables as these regressed expressions. Our openaccess research data, discussed on page 34, contains details on these analyses.
We evaluated the performance of the symbolic interpretation of the JTA as implemented in
29
[5] on randomly generated constrained BNs. This does not evaluate our approach per se, but
the manner in which we interpreted an existing inference implementation symbolically. We
refer to our open-access data repository for more details on model generation: key parameters
are the number of nodes | N |, the number of variables | Xx |, and a random choice of the
number of states for each node (between 1 and 10 uniformly at random). Terms in probability
tables have form c, x or 1 ´ x for constants c or x in Xx . In generated models, a random node
was picked to determine hard evidence – its first state having probability 1. The JTA was
run for that hard evidence, and the time to complete it was recorded. These automated test
suites ran on an institutional server with 64 Intel Xeon E5/Core i7 processors, on Ubuntu
14.04.
Many of these tests terminated very quickly. Though, as the number of nodes per graph
and the number of states per node increased, the running times increased on some but not all
tests. The size of Xx seemed to have a limited effect, possibly indicating that the additional
overhead of our approach to running the JTA implementation of [5] symbolically in Python
is not huge.
Figure 18 shows plots for the computation times (in seconds) of 1000 such test cases
against the number of nodes, the size of Xx , the number of node states in total (a summation
over all nodes) and the average length over all nodes of the outputted marginal text string
in characters.
For this randomized test suite, there was a small trend for the running times to increase
with the size of the DAG. But computations were still quicker for many of the BNs of larger
size compared to smaller ones. The size of Xx appears to have little impact on computation
time, nor any strong correlation to the length of the computed symbolic marginal. This
suggests that use of symbolic probabilities may not in and of itself increase such empirical
complexity.
6
Discussion
Our approach advocates the use of constrained Bayesian Networks as a means of gaining
confidence into Bayesian Network modelling and inference in the face of little or no data. A
modeler may thus start with a BN, turn it into several constrained BNs and subject them to
analysis, and perhaps modify the BN based on such findings. Witness information computed
in analyses could, in principle, be fed back into a BN modeling tool so that users can see a
concrete BN that would, for example, explain how a marginal of interest can attain a certain
value in a constrained BN.
The ability to represent witness information as a concrete BN is also a means of testing
whether the computation of symbolic meaning of marginals is free of errors. We have indeed
conducted such tests to gain confidence into the correctness of our tool and the packages that
it depends upon. Note also that errors in the symbolic meaning of marginals are likely to create numerical inconsistencies, so our analyses would detect such an inconsistent, constrained
BN.
The algorithms that we devised for non-linear optimization made no assumptions about
the internal workings of the decision procedure used and its witness information apart from
that such results would be semantically correct. Knowledge of such internal details could,
30
Figure 18: Plots for the time the symbolic Junction Tree Algorithm takes to run (in seconds,
on the y-axis) against properties of various randomly generated BNs, on the x-axis. We
assess (from left-to-right, top-to-bottom) the effect on the computation time by the number
of nodes, the number of variables in Xx , the number of states (a summation of the number of
states in each of the nodes) and the average length of the text string in the resulting marginal
computation. In this last graph, we embed the closeup of the datapoint in the larger graph.
31
however, be exploited to speed up computation. For example, such a method is used in
the SMT solver Z3 to optimize linear objective functions. One could therefore run different
methods in parallel or even let them share information in between search iterations.
Our tool prototype interprets the JTA implementation provided in [5] symbolically, and
symbolic meanings of marginal variables may contain divisions. Of course, we could translate
away all division operators without changing meaning – to match this with the formal setting
of Section 3. We did not do this since our foundations apply equally to Q extended with
division, such translations would increase the size of these terms, and the SMT solver we
used, Z3, was able to process and reason with such or suitably simplified terms.
7
Related Work
In [21], it is shown how probabilistic inference in Bayesian Networks can be represented
through the evaluation and formal differentiation of a “network polynomial”. The size of the
polynomial can be reduced by its representation in an arithmetic circuit, in which evaluation
and differentiation are more efficient. It would be of interest to determine whether this work
can be extended to make the computation of symbolic marginals generated in our approach
more efficient.
For Bayesian Networks there are methods for learning the structure of a DAG and for
learning the probabilities within nodes of such a graph (see e.g. [29, 19]) – based on existing
empirical data. We assumed in this paper that little or no data are available, ruling out the
effective use of such learning methods. But our approach is consistent with settings in which
plenty of data are available.
Bayesian Networks have tool support such as the software JavaBayes [2], which is able
to perform robustness analysis. But this software can neither cope with the Knightian uncertainty of our approach, nor fuse networks of different structures together with non-trivial
constraints.
Our work in [8] reported early attempts of developing the approach presented in this
paper: in [8], a much simpler Bayesian Network of a nuclear inspection process is presented
and some analyses with preliminary versions of our tool are discussed; but that work offered
neither formal foundations nor greater technical details for the methods it used. The more
detailed Bayesian Network we studied in Section 4 was discussed in [7], along with a nontechnical summary of our general approach and some of its analysis findings.
Credal networks – see e.g. [16] – refer to the theory and practice of associating a convex
set of probability measures with directed, acyclic graphs. Credal networks are also referred
to as the Theory of Imprecise Probabilities [39] or as the Quasi-Bayesian Theory [25].
The generalization of probability measures to sets of such measures can accommodate a
formal notion of probabilistic independence, rooted in axioms of preferences as developed in
[16]. The approach is based on constraints for such convex sets of probability measures. Inference algorithms and their approximations are bespoke for an interpretation of constraints;
an interpretation is called an “extension” in [16].
To compare this to our approach, we follow Good’s black box model in that our semantics
and optimizations reflect Bayesian inference – even though this is done symbolically. Another
difference is that a constrained Bayesian Network may have nodes with non-convex sets of
32
probability measures as meaning, for example when logical constraints on variables rule out
certain points in intervals. Our approach is also more practical in outlook, since we rely on
reductions to known and tried techniques, such as satisfiability checking for the existential
theory of the reals. In contrast, theoretical results for Credal Networks range from different
evidence propagation and inference methods (see e.g. [17, 20]) to deep relationships to logic
programming and its semantics [18].
In [11], a methodology is developed for assessing sensitivity of lower and upper probabilities in Credal networks. It is shown that for some classes of parameters in Bayesian networks
one may replace the Credal sets of probability measures associated with such parameters with
a sole such measure. It would be of interest to determine whether these or similar results are
attainable for suitable classes of constrained Bayesian Networks.
Constraint Networks [24] are graphical representations that are meant to guide solution
strategies for constraint satisfaction problems. In our tool prototype, we decoupled the choice
of graph structure for a constrained Bayesian Network from the use of strategies for solving
satisfiability problems over the existential theory of the reals. It may be beneficial to couple
graph structure and satisfiability checking in tool support of our approach that relies on
constraint satisfaction solvers.
8
Conclusions
This work was motivated by the fact that some problem domains have little or no data that
one could use to learn the structure of a causal network or the probabilities for nodes within
that structure – whatever the reasons for such sparsity of data may be in such a domain.
This led us to consider suitable generalizations of Bayesian Networks. Ideally, we wanted a
formalism that those who already use Bayesian Networks for modeling and analysis would
find easy to adopt. In particular, we sought to preserve – as much as possible – the manner
in which probabilistic inference is done in Bayesian Networks. Crucially, we wanted a set of
methods whose use could help us to build sufficient confidence into the quality, suitability or
robustness of models expressed in such a formalisms in the face of little or no empirical data.
We propose constrained Bayesian Networks as such a formalism. The derivation of that
concept is a contribution in and of itself, and it used first-order logic and its semantics as
well as syntactic criteria for wellformedness. But it also required methods from three-valued
logic to define a precise yet intuitive semantics for a constrained BN.
We also developed meta-properties of this semantics, including checks for the consistency
of a constrained Bayesian Network. These properties were needed to prove the correctness of
our optimization algorithms, which can compute suprema or infima of bounded arithmetic
terms up to a specific accuracy. These optimization algorithms are non-standard in that they
rely on a decision procedure for the theory of reals and in that the optimization problems
are generally non-linear and non-convex.
The marginals in a constrained Bayesian Network are computed symbolically, but computed in the same manner as the marginals for a Bayesian Network – a concretization of that
constrained Bayesian Network. This is appealing as it allows reuse of known and trusted
methods such as the Junction Tree Algorithm. But it also creates a potential computational
bottleneck with scope for future work that may extend an approach in [21] to our setting.
33
We implemented our approach in a tool prototype, which benefitted from the significant
advances in symbolic computation and in the implementation of theorem provers such as
SMT solvers. We evaluated this prototype through stress tests and a non-trivial case study
in the domain of nuclear arms control. The latter is a domain in which the availability of
data is very limited and where any means of building confidence into the trustworthiness of
mathematical models are expected to have positive impact on arms reduction efforts.
We used this case study to illustrate some pertinent types of analyses of a constrained
Bayesian Network that our approach can accommodate: a range analysis that computes
infima and suprema for a term of interest to determine their robustness, the comparison
of two or more constrained Bayesian Networks to assess modeling impact, the ability to
determine ranges of threshold values that would render equivalent decision support, and the
symbolic computation of a sensitivity measure for a given node – with the ability to optimize
this to understand worst-case sensitivities. We trust that the approach presented in this
paper will be useful for other applications in the arms-control domain, as well as in other
domains – particularly those with a lack of data.
Acknowledgements: This work was supported by AWE plc, and in part by the UK Engineering and Physical Sciences Research Council grants EP/N020030/1 and EP/N023242/1.
Open Access: The Python and SMT code for the queries and models of this paper and
raw SMT analysis results are found in the public data repository
bitbucket.org/pjbeaumont/beaumonthuthcbns/
References
[1] International Atomic Energy Agency. Treaty on the Non-Proliferation of Nuclear
Weapons, 22 April 1970.
[2] JavaBayes:
Bayesian Networks in Java,
http://www.cs.cmu.edu/„javabayes/.
1998.
Software
available
at
[3] UK/Norway Initiative, 2007. Website http://ukni.info/.
[4] Ministry of Defence of the United Kingdom: the UK/Norway initiative; report on the
UKNI nuclear weapons states workshop, March 2010.
[5] ebay: Python Bayesian belief network framework; open-source package available at
https://github.com/ebay/bayesian-belief-networks, Last commit seen on 9 August 2014 2014.
[6] David Barber. Bayesian Reasoning and Machine Learning. Cambridge University Press,
2012.
34
[7] Paul Beaumont, Edward Day, Neil Evans, Sam Haworth, Michael Huth, Tom Plant,
and Catherine Roberts. An in-depth case study: modelling an information barrier in
Bayesian belief networks. In Journal of the Institute of Nuclear Materials Management,
INMM57, 24-28 July, Atlanta, USA, 2016. Open access version at institutional repository: https://spiral.imperial.ac.uk:8443/handle/10044/1/41828.
[8] Paul Beaumont, Neil Evans, Michael Huth, and Tom Plant. Confidence analysis for
nuclear arms control: SMT abstractions of Bayesian Belief Networks. In Computer Security - ESORICS 2015 - 20th European Symposium on Research in Computer Security,
Vienna, Austria, September 21-25, 2015, Proceedings, Part I, pages 521–540, 2015.
[9] James O. Berger. An Overview of Robust Bayesian Analysis. Test, 3(1):5–124, 1994.
[10] James O. Berger, David Rı́os Insua, and Fabrizio Ruggeri. Bayesian Robustness, pages
1–32. Springer New York, New York, NY, 2000.
[11] Janneke H. Bolt, Jasper De Bock, and Silja Renooij. Exploiting Bayesian Network
Sensitivity Functions for Inference in Credal Networks. In ECAI 2016 - 22nd European
Conference on Artificial Intelligence, 29 August-2 September 2016, The Hague, The
Netherlands - Including Prestigious Applications of Artificial Intelligence (PAIS 2016),
pages 646–654, 2016.
[12] John S. Breese and David Heckerman. Decision-theoretic troubleshooting: A framework
for repair and experiment. In UAI ’96: Proceedings of the Twelfth Annual Conference
on Uncertainty in Artificial Intelligence, Reed College, Portland, Oregon, USA, August
1-4, 1996, pages 124–132, 1996.
[13] Glenn Bruns and Patrice Godefroid. Model checking partial state spaces with 3-valued
temporal logics. In Computer Aided Verification, 11th International Conference, CAV
’99, Trento, Italy, July 6-10, 1999, Proceedings, pages 274–287, 1999.
[14] J. F. Canny. The complexity of Robot Motion Planning. PhD thesis, MIT, Cambridge,
1988.
[15] Patrick Cousot and Radhia Cousot. Abstract interpretation: past, present and future.
In Joint Meeting of the Twenty-Third EACSL Annual Conference on Computer Science Logic (CSL) and the Twenty-Ninth Annual ACM/IEEE Symposium on Logic in
Computer Science (LICS), CSL-LICS ’14, Vienna, Austria, July 14 - 18, 2014, pages
2:1–2:10, 2014.
[16] Fábio Gagliardi Cozman. Credal networks. Artif. Intell., 120(2):199–233, 2000.
[17] Fabio Gagliardi Cozman and Jose C. F. da Rocha. Inference with Separately Specified
Sets of Probabilities in Credal Networks. Proceedings of the Eighteenth conference on
Uncertainty in artificial intelligence, 2002.
[18] Fábio Gagliardi Cozman and Denis Deratani Mauá. The Structure and Complexity of
Credal Semantics. In Proc. of the Third International Workshop on Probabilistic Logic
Programming, pages 3–14, 2016.
35
[19] James Cussens, David Haws, and Milan Studeny. Polyhedral aspects of score equivalence
in Bayesian network structure learning. arXiv.org, (1503.00829), 2015.
[20] José Carlos Ferreira da Rocha and Fábio Gagliardi Cozman. Evidence Propagation in
Credal Networks: An Exact Algorithm Based on Separately Specified Sets of Probability. In Advances in Artificial Intelligence, 16th Brazilian Symposium on Artificial
Intelligence, SBIA 2002, Porto de Galinhas/Recife, Brazil, November 11-14, 2002, Proceedings, pages 376–385, 2002.
[21] Adnan Darwiche. A differential approach to inference in Bayesian networks. J. ACM,
50(3):280–305, 2003.
[22] Cassio Polpo de Campos and Fábio Gagliardi Cozman. The inferential complexity of
Bayesian and credal networks. In IJCAI-05, Proceedings of the Nineteenth International
Joint Conference on Artificial Intelligence, Edinburgh, Scotland, UK, July 30 - August
5, 2005, pages 1313–1318, 2005.
[23] Leonardo De Moura and Nikolaj Bjørner. Z3: An Efficient SMT Solver. In Proceedings
of the Theory and Practice of Software, 14th International Conference on Tools and
Algorithms for the Construction and Analysis of Systems, TACAS’08/ETAPS’08, pages
337–340, Berlin, Heidelberg, 2008. Springer-Verlag.
[24] Rina Detcher. Constraint networks. Encyclopedia of Artif. Intell., page 276285, 1992.
[25] F. J. Giron and S. Rios. Quasi-Bayesian behaviour: A more realistic approach to decision
making? Trabajos de Estadistica Y de Investigacion Operativa, 31(1):17–38, 1980.
[26] Irving John Good. Could a machine make probability judgments? Computation and
Automation, 8:14–16 and 24–26, 1959.
[27] Irving John Good. Degrees of Belief. In Encyclopedia of Statistical Sciences, volume 2.
Wiley, New York, 1982.
[28] Alexander J. Hartemink, David K. Gifford, Tommi S. Jaakkola, and Richard A. Young.
Bayesian methods for elucidating genetic regulatory networks. IEEE Intelligent Systems,
17(2):37–43, 2002.
[29] David Heckerman. A Tutorial on Learning With Bayesian Networks. Technical Report MSR-TR-95-06, Microsoft Research, Advanced Technology Division, March 1995.
Revised November 1996.
[30] C. David Page Jr. and Irene M. Ong. Experimental design of time series data for
learning from dynamic Bayesian networks. In Biocomputing 2006, Proceedings of the
Pacific Symposium, Maui, Hawaii, USA, 3-7 January 2006, pages 267–278, 2006.
[31] Michael Y. K. Kwan, Richard E. Overill, Kam-Pui Chow, Hayson Tse, Frank Y. W.
Law, and Pierre K. Y. Lai. Sensitivity analysis of Bayesian networks used in forensic
investigations. In Advances in Digital Forensics VII - 7th IFIP WG 11.9 International
Conference on Digital Forensics, Orlando, FL, USA, January 31 - February 2, 2011,
Revised Selected Papers, pages 231–243, 2011.
36
[32] Kathryn Blackmond Laskey. Sensitivity analysis for probability assessments in Bayesian
networks. IEEE Trans. Systems, Man, and Cybernetics, 25(6):901–909, 1995.
[33] Denis Deratani Mauá, Cassio Polpo De Campos, Alessio Benavoli, and Alessandro Antonucci. Probabilistic inference in credal networks: New complexity results. J. Artif.
Int. Res., 50(1):603–637, May 2014.
[34] Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ondrej Certı́k, Matthew Rocklin, Amit Kumar, Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rathnayake,
Sean Vig, Brian E. Granger, Richard P. Muller, Francesco Bonazzi, Harsh Gupta, Shivam
Vats, Fredrik Johansson, Fabian Pedregosa, Matthew J. Curry, Ashutosh Saboo, Isuru
Fernando, Sumith Kulal, Robert Cimrman, and Anthony M. Scopatz. Sympy: Symbolic
computing in python. PeerJ PrePrints, 4:e2083, 2016.
[35] Richard E. Neapolitan. Probabilistic Reasoning in Expert Systems: Theory and Algorithms. John Wiley & Sons, Inc., New York, NY, USA, 1990.
[36] J. Pearl. Bayesian networks: A model of self-activated memory for evidential reasoning.
Technical Report CSD-850021, R-43, UCLA Computer Science Department, June 1985.
[37] Judea Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1988.
[38] Hayson Tse, Kam-Pui Chow, and Michael Y. K. Kwan. Reasoning about evidence
using Bayesian networks. In Advances in Digital Forensics VIII - 8th IFIP WG 11.9
International Conference on Digital Forensics, Pretoria, South Africa, January 3-5,
2012, Revised Selected Papers, pages 99–113, 2012.
[39] Peter Walley. Statistical Reasoning with Imprecise Probabilities. Chapman and Hall,
London, 1991.
A
Mathematical Proofs
Proof of Theorem 1:
C must
C
C
1. We have that BX
|ù φ holds iff for all concretizations BX
rαs of BX
we have that
C
C
α |ù φ holds iff for all concretizations BX rαs of BX we have that α |ù φ does not hold
C may
iff BX
|ù
φ does not hold.
C may
C
C
2. We have that BX
|ù φ holds iff there is some concretization BX
rαs of BX
such that
C
C
α |ù φ holds iff there is some concretization BX
rαs of BX
such that α |ù φ does not
C must
hold iff BX
|ù
φ does not hold.
C must
C
C
3. (a) Let BX
|ù φ1 ^ φ2 hold. Let BX
rαs be a concretization of BX
. Then we know
that α |ù φ1 ^ φ2 holds. This implies that α |ù φi holds for i “ 1, 2. But then both
C must
C must
C
BX
|ù φ1 and BX
|ù φ2 hold since BX
rαs was an arbitrary concretization of
C
BX .
37
C must
C must
C
C
(b) Let both BX
|ù φ1 and BX
|ù φ2 hold. Let BX
rαs be a concretization of BX
.
C must
Then BX |ù φi implies that α |ù φi holds for i “ 1, 2. Therefore, we get that
C
C
α |ù φ1 ^ φ2 holds as well. Since BX
rαs was an arbitrary concretization of BX
,
C must
this gives us that BX |ù φ1 ^ φ2 holds.
C may
C
C
4. (a) Let BX
|ù φ1 _ φ2 hold. Then there is some concretization BX
rαs of BX
such
that α |ù φ1 _ φ2 holds. This implies that α |ù φi holds for some i “ 1, 2. But
C may
then BX
|ù φi holds as claimed.
C may
C may
C may
(b) Let one of BX
|ù φ1 and BX
|ù φ2 hold, say BX
|ù φi . Then there is some
C
C
concretization BX rαs of BX such that α |ù φi holds. This implies that α |ù φ1 _φ2
C
C
C may
holds as well. Since BX
rαs is a concretization of BX
, we get that BX
|ù φ1 _ φ2
holds.
Proof of Theorem 2:
C may
|ù true hold. By definition of |ùmay , there then is
• Item 1 implies item 2: Let BX
C
C
some concretization BX rαs of BX
such that α |ù true holds. Therefore the set of
C
C
concretizations of BX is non-empty and so BX
is consistent.
C
• Item 2 implies item 3: Let BX
be consistent. Suppose that φ is in Q such that
C
C
C
C must
.
rαs of BX
BX |ù φ holds. Since BX is consistent, there is some concretization BX
C may
C must
|ù φ be
|ù φ holds, we get that α |ù φ holds. But then we have BX
Since BX
definition of |ùmay .
C must
C may
• Item 3 implies item 4: Let BX
|ù φ imply BX
|ù φ for all φ in Q. Let ψ be in
C may
Q. We claim that BX |ù ψ _ ψ holds. By Theorem 1.4, it suffices to show that
C may
C may
BX
|ù ψ or BX
|ù
ψ holds. If the former holds, we are done. Otherwise, we have
C must
C may
|ù
ψ holds.
that BX |ù ψ does not hold. By Theorem 1.2, this implies that BX
C must
C
|ù true holds by
has to be consistent: note that BX
– Next, we show that BX
must
C may
the definitions of |ù
and |ù. Therefore, we get that BX |ù true holds by
C
item 3 – and we already showed that this implies that BX
is consistent.
C
C
C
Let BX
rαs be a concretization of BX
, which exists as BX
is consistent. Since we showed
C must
C may
BX |ù
φ, the latter implies that α |ù φ. But then BX
|ù
φ follows given the
may
definition of |ù .
C may
• Item 4 implies item 5: Let BX
|ù φ _ φ hold for all φ in Q. Let ψ be in Q. Then we
C may
C must
have that BX |ù ψ _ ψ holds, and we need to show that BX
|ù ψ ^ ψ does not
C must
C may
hold. Proof by contradiction: assume that BX |ù ψ^ ψ holds. Since BX
|ù ψ_ ψ
C
C
holds, we know that there is some concretization BX rαs of BX such that α |ù ψ _ ψ
C must
C must
C must
holds. Since BX
|ù ψ ^ ψ holds, we know that BX
|ù ψ and BX
|ù
ψ hold by
Theorem 1.3. We do a case analysis on the truth of judgment α |ù ψ _ ψ:
C must
– Let α |ù ψ hold . Since BX
|ù
ψ holds, this implies that α |ù
contradicts that α |ù ψ holds.
38
ψ holds. This
C must
– Let α |ù ψ hold. Since BX
|ù ψ holds, this implies that α |ù ψ holds. This
contradicts that α |ù ψ holds.
C must
• Item 5 implies item 1: Let BX
|ù φ ^ φ not hold, for all φ in Q. Since true is in
C must
Q, we know that BX |ù true ^ true does not hold. By definition of |ùmust and |ù,
C must
C must
we have that BX
|ù true holds. By Theorem 1.3 and since BX
|ù true ^ true does
C must
not hold, we infer that BX |ù
true does not hold. By Theorem 1.2, this implies that
C may
BX
|ù true holds as claimed.
Proof of Theorem 3:
1. Constraints ϕ1 in C and ϕ are quantifier-free formulas of Q with variables contained in
X, which equals tx1 , x2 , . . . , xn u. Therefore, the formula in (8) is in Q, and contains
only existential quantifiers and all in front of the formula.
2. We prove this claim by structural induction over ϕ:
Ź
C
1
• Let ϕ be true. Then ExpBX
, trueq equals Dx1 : . . . : Dxn : true
Ź ^ ϕ1 1 PC ϕ and this
is satisfiable iff there is an assignment α such that α |ù ϕ1 PC ϕ and α |ù true
C
rαs of
both hold (the latter holding by definition) iff there is a concretization BX
C may
C
C
BX iff BX is consistent iff (by Theorem 2) BX |ù true holds.
Ź
C
, t1 ď t2 q equals Dx1 : . . . : Dxn : pt1Źď t2 q ^ ϕ1 PC ϕ1
• Let ϕ be t1 ď t2 . Then ExpBX
and this is satisfiable iff there is an assignment α such that α |ù ϕ1 PC ϕ1 and α |ù
C
C
t1 ď t2 both hold iff there is a concretization BX
rαs of BX
such that α |ù t1 ď t2
C may
holds iff BX |ù t1 ď t2 holds.
Ź
C
• Let ϕ be t1 ă t2 . Then ExpBX
, t1 ă t2 q equals Dx1 : . . . : Dxn : pt1Źă t2 q ^ ϕ1 PC ϕ1
and this is satisfiable iff there is an assignment α such that α |ù ϕ1 PC ϕ1 and α |ù
C
C
such that α |ù t1 ă t2
rαs of BX
t1 ă t2 both hold iff there is a concretization BX
C may
holds iff BX |ù t1 ă t2 holds.
Ź
C
• Let ϕ be ψ. Then ExpBX
, ψq equals Dx1 : . . . : Dxn : Ź
ψ ^ ϕ1 PC ϕ1 and this
is satisfiable iff there is an assignment α such that α |ù ϕ1 PC ϕ1 and α |ù ψ
C
C
both hold iff there is a concretization BX
rαs of BX
such that α |ù ψ holds iff
C may
BX
|ù
ψ holds.
Ź
C
1
• Let ϕ be ϕ1 ^ ϕ2 . Then ExpBX
, ϕ1 ^ ϕ2 q equals Dx1 : . . . : Dxn : ϕŹ
1 ^ ϕ2 ^
ϕ1 PC ϕ
and this is satisfiable iff there is an assignment α such that α |ù ϕ1 PC ϕ1 and α |ù
C
C
ϕ1 ^ ϕ2 both hold iff there is a concretization BX
rαs of BX
such that α |ù ϕ1 ^ ϕ2
C may
holds iff BX
|ù ϕ1 ^ ϕ2 holds.
C may
C
3. By the previous item, we may decide BX
|ù ϕ by deciding whether formula ExpBX
, ϕq
is satisfiable. By item 1 above, that formula is in the existential fragment of Q. By
[14], deciding the satisfiability (truth) of such formulas is in PSPACE in the size of
such formulas.
39
C must
C may
4. By Theorem 1.1, we have that BX
|ù ϕ holds iff BX
|ù
ϕ does not hold. By item 2
C
above, the latter is equivalent to ExpBX , ϕq not being satisfiable. By [14], this can be
C
decided in PSPACE in the size of formula ExpBX
, ϕq.
Proof of Theorem 4: The arguments below make use of Theorems 1 and 3 without explicit
C
reference to them. Note that consistency of BX
and 0 ă sup t| t |u guarantee that the first
let statement in Sup can find such a α. In particular, we see that 0 ă cache becomes an
invariant and so cache ă 2 ˚ cache is another invariant.
1. First, we show that the asserts hold prior to the execution of the second while loop.
C
Note that cache is always assigned reals of form ηptq for some concretization BX
rηs of
C
C may
BX . So when low is initialized with the last updated value of cache, then BX |ù t ď
low clearly holds after the first assignment to low (witnessed by the assignment that
gave rise to the last value of cache) and prior to its reassignment. By definition of the
C may
initial value of high, we have that BX
|ù t ě high does not hold after that initial
assignment and prior to the reassignment of high. Therefore, both asserts in front of
the second while loop hold, and we get that low ď high is an invariant.
Second, we show that each iteration of the second while loop preserves the asserts.
This is clear as the Boolean guard of the if statement tests for preservation of these
asserts, and makes the correct, invariant-preserving assignment accordingly.
Third, let rl, hs be the returned closed interval. It is clear that h ´ l ď δ holds as
required. We argue that sup t| t |u is in rl, hs. Since the asserts hold for l and h, we
C may
C may
know that BX
|ù t ě l holds, but BX
|ù t ě h does not hold. Let c be in t| t |u.
C may
|ù t ě h does not hold, we get that
Then there is some α with c “ αptq. Since BX
αptq ă αphq “ h. Therefore, h is an upper bound of t| t |u which implies sup t| t |u ď h.
C
C may
rα1 s with α1 |ù t ě l. This
|ù t ě l holds, we have some concretization BX
Since BX
1
1
means α ptq ě l. But sup t| t |u ě α ptq as the latter is an element of t| t |u. Thus,
l ď sup t| t |u follows.
2. Let s be sup t| t |u. For the first while loop, we have at least k iterations if s ě 2k ¨ c,
i.e. if s ¨ c´1 ě k, i.e. if k ď log2 psq ´ log2 pcq. So the real number log2 psq ´ log2 pcq is an
upper bound on the number of iterations of the first while loop.
To get an upper bound for the number of iterations of the second while loop, we know
that high is of form 2l`1 ¨ c and so low equals 2l ¨ c. But then high ´ low equals 2l ¨ c.
Since this is monotone in l, we may use the upper bound for the number of iterations
of the first while loop as an upper bound of l, to get 2log2 psq´log2 pcq ¨ c “ s ¨ c´1 ¨ c “ s as
an upper bound on the value of | high ´ low | before the Boolean guard of the second
while is first evaluated. This allows us to derive an upper bound on the number of
iterations of the second while loop, since the larger that value is, the more iterations
take place. Based on the bisection in each iteration, there are at least k iterations if
s ¨ 2´k ą δ, i.e. if k ă log2 psq ´ log2 pδq. Therefore, the total number of iterations of
both while loops combined is plog2 psq ´ log2 pcqq ` plog2 psq ´ log2 pδqq. The claim now
follows given that each iteration makes exactly one satisfiability check and since there
is an initial satisfiability check as well.
40
Proof of Theorem 5:
1. The argument is similar to the one for Theorem 4 but there are important differences.
Note that cache ą 0 is also here an invariant, guaranteed by the fact that t| t |u contains
a positive real. We know that p0.5n ˚ cacheqnPN converges to 0 for any positive constant
cache. Since 0 ă δ and since α1 ptq ď 0.5 ˚ cache for the α1 ptq assigned to cache, there
is some n0 such that 0.5n0 ˚ cache ď δ. This proves that the first while statement
terminates.
(a) Suppose that the return statement in the line after the first while loop is exeC
cuted. Then ExpBX
, t ď 0.5 ˚ cacheq is satisfiable and so there is some concretizaC
tion BX rαs such that αptq ď 0.5 ˚ cache. But then inf t| t |u ď 0.5 ˚ cache as well.
From t| t |u Ď R`
0 , we get 0 ď inf t| t |u. Therefore, inf t| t |u is in the returned
C may
interval r0, 0.5 ˚ caches and BX
|ù t ď 0.5 ˚ cache is true. Moreover, the length of
the interval is 0.5 ˚ cache, which must be less than or equal to δ as the first while
loop just terminated and the first conjunct of its Boolean guard is true – forcing
0.5 ˚ cache ą δ to be false.
C
C
(b) Otherwise, ExpBX
, t ď 0.5 ˚ cacheq is not satisfiable but the formula ExpBX
, t ď cacheq
is satisfiable. From that, it should then be clear that the asserts in front of the
second while statement hold when they are reached. That each iteration of the
second while statement maintains these two asserts is reasoned similarly as for
Sup.
C must
C may
|ù t ą l are invariants. This means
|ù t ď h and BX
So we have that BX
that l is a lower bound of t| t |u and αptq ď h for some αptq in t| t |u. But then
l ď inf t| t |u ď αptq ď h shows that inf t| t |u is in rl, hs.
2. Let i be inf t| t |u. We derive an upper bound on the number of iterations for the first
while loop. Because we are interested in upper bounds, we may assume that the α1 ptq
assigned to cache equals 0.5¨cache for the current value of cache. We then have at least
k iterations if δ ă c ¨ 2´k and i ď c ¨ 2´k . Since we are interested in upper bounds on
that number of iterations, we get at least k iterations if both δ ď c ¨ 2´k and i ď c ¨ 2´k
hold, i.e. if minpδ, iq ď c ¨ 2´k . But this is equivalent to k ď log2 pcq ´ log2 pminpi, δqq.
We now derive an upper bound on the number of iterations of the second while loop.
The initial value of high´low equals cache´0.5¨cache “ 0.5¨cache for the current value
of cache when entering that loop. The value of cache is monotonically decreasing during
program execution and so c{2 is an upper bound of high ´ low. We may therefore use
c{2 as initial value of high ´ low since this can only increase the number of iterations,
for which we seek an upper bound. There are now at least k iterations if pc{2q ¨ 2´k ą δ
which is equivalent to k ă log2 pcq ´ 1 ´ log2 pδq.
The total number of iterations for both while loops is therefore plog2 pcq´log2 pminpi, δqqq`
plog2 pcq ´ 1 ´ log2 pδqq “ 2 ¨ log2 pcq ´ log2 pminpi, δqq ´ 1. From this the claim follows
since each iteration has exactly one satisfiability check of the stated form, and there is
one more satisfiability check between the first and second while loop.
41
Proof of Theorem 6:
1. We do a case analysis:
C
(a) If algorithm Sup is called, then consistency of BX
and 0 ă sup t| t |u follow from
the Boolean guard that triggered the call. Since sup t| t |u ă 8 is assumed, we
get 0 ă sup t| t |u ă 8 and so Sup terminates by Theorem 4.
(b) If 0 is returned as a maximum, the algorithm clearly terminates and no preconditions are needed.
(c) If Inf is called, we have to show that t| ´t |u is a subset of R`
0 that contains a
positive real. Since the first two return statements were not reached, we know
C
that BX
is consistent and t| t |u is a subset of R´ . But then t| ´t |u is a subset of
`
R .
2. If the algorithm reports that 0 is the maximum for t, then we know that t| t |u cannot
contain a positive real (first if-statement), and that it contains 0 (second if-statement).
Clearly, this means that 0 is the supremum of t| t |u and so also its maximum as 0 is
in t| t |u.
C
q return an interval r´h, ´ls. Then rl, hs is the interval returned by a
3. Let Sup‹ pt, δ, BX
C may
C
|ù
´t ď h
q. By the first item and Theorem 5, we get that BX
call to Infp´t, δ, BX
C may
|ù t ě ´h
holds, inf t| ´t |u is in rl, hs, and h´l ď δ . Therefore, we conclude that BX
holds as claimed. Moreover, since inf t| ´t |u equals ´ sup t| t |u, this implies that
sup t| t |u is in the closed interval r´h, ´ls, whose length is that of rl, hs and so ď δ.
C
C
, t ą 0q,
is inconsistent, then all three formulas ExpBX
4. If the algorithm returns saying that BX
C
C
ExpBX , t “ 0q, and ExpBX , t ă 0q are unsatisfiable. But then we know that the three
C may
C may
C may
|ù t ă 0 do not hold, by Theo|ù t “ 0, and BX
|ù t ą 0, BX
judgments BX
C
C
rem 3. This means that BX is inconsistent: for all concretization BX
rαs we have that
α |ù pt ą 0q _ pt “ 0q _ pt ă 0q holds as that query is a tautology over the theory of
C
reals; and then Theorem 3.4 yields a contradiction to BX
being consistent.
Proof of Theorem 7: The correctness of the first two claims in that theorem (inconsistency
and minimum) for ´8 ă inf t| t |u follows from the corresponding items of Theorem 6. The
general identity inftxi | i P Iu “ ´ supt´xi | i P Iu shows that ´8 ă inftxi | i P Iu iff
´ supt´xi | i P Iu ă 8 and so preconditions are also met. Finally, to see the correctness of
Inf ‹ when interval rl, hs is returned, note that this means that interval r´h, ´ls is returned
C
C may
for the call Sup‹ p´t, δ, BX
q and so BX
|ù ´ t ě ´h holds by Theorem 6. But this implies
C may
that BX |ù t ď h holds as claimed.
B
Quantitative Information about the BN of Figure 11
Table 1 shows quantitative information about the size and complexity of the BN in Figure 11.
42
Table 1: The probability tables for the BN in Figure 11 are too large to be specified explicitly
in the paper. Here we want to convey the structural and resulting computational complexity
of this BN in tabular form. For each node shown in the leftmost column we list its number
of parentsś
(# Parents) in the BN graph, its number of input combinations (# Rows) which
is maxp1, ki“1 ni q where ni is the number of outputs any of the 0 ď k parents i can have,
whereas in # Columns we list the number of output values that node itself can have. Rightmost column Table Size depicts the size of the support of the probability distribution of that
node.
43
| 2cs.AI
|
Distributed Evaluation and Convergence of
Self-Appraisals in Social Networks
arXiv:1503.08175v3 [cs.SY] 27 Oct 2015
Xudong Chen, Ji Liu, M.-A. Belabbas, Zhi Xu, Tamer Başar
Abstract—We consider in this paper a networked system of
opinion dynamics in continuous time, where the agents are able
to evaluate their self-appraisals in a distributed way. In the model
we formulate, the underlying network topology is described by a
rooted digraph. For each ordered pair of agents (i, j), we assign
a function of self-appraisal to agent i, which measures the level
of importance of agent i to agent j. Thus, by communicating only
with her neighbors, each agent is able to calculate the difference
between her level of importance to others and others’ level of
importance to her. The dynamical system of self-appraisals is
then designed to drive these differences to zero. We show that
for almost all initial conditions, the trajectory generated by this
dynamical system asymptotically converges to an equilibrium
point which is exponentially stable.
I. I NTRODUCTION
Social network is a social structure made up of actors, such
as agents and organizations, and the relationships between
these actors, particularly those that are neighbors to each
other. The concept of social networks is familiar to most
people because of the emergence of online social networking services such as Facebook, Twitter, and Google+. Many
social behaviors spread through social networks of interacting
agents. Examples are opinion dynamics [1], adoption of new
technology or products [2], voting [3], and demonstrations [4].
In this paper, we will focus on an important issue in opinion
dynamics which is about how each agent in the social network
evaluates her importance in a distributed way.
Over the past decades, there has been considerable attention
paid to understanding how an agent’s opinion evolves over
time. In social science, various models have been proposed
to illustrate opinion dynamics. Notable among them are the
four classical models: the DeGroot model [1], the FriedkinJohnsen model [5], the Hegselmann-Krause model [6], and
the Deffuant-Weisbuch model [7]. In the DeGroot model,
each agent has a fixed set of neighbors and takes a convex
combination of her own opinion and the opinions of her
neighbors. The Friedkin-Johnsen model is a variation of the
DeGroot model in which each agent adheres to her initial
opinion to a certain degree, which can be interpreted as her
level of “stubbornness”. The Hegselmann-Krause model and
the Deffuant-Weisbuch model define the neighbor sets in a
different way: each agent takes a set of agents as her neighbors
only if the opinions of these agents differ from her by no
more than a certain confidence level. With the defined set
Xudong Chen, Ji Liu, M.-A. Belabbas and Tamer Başar are with the
Coordinated Science Laboratory, University of Illinois at Urbana-Champaign,
emails: {xdchen, jiliu, belabbas, basar1}@illinois.edu.
Zhi Xu is with the Department of Electrical Engineering and Computer
Science, Massachusetts Institute of Technology, email: [email protected].
This research was supported in part by the U.S. Air Force Office of
Scientific Research (AFOSR) MURI grant FA9550-10-1-0573.
of neighbors, each agent takes a convex combination of her
own opinion with either (i) all of her current neighbors in
the Hegselmann-Krause model, or (ii) only one of her current
neighbors in the Deffuant-Weisbuch model. Other important
opinion models are, for example, the Sznajd model [8], which
uses the Ising spin model to describe a mechanism of making
a decision in a closed community, and the voter model, which
is a continuous-time Markov process defined over a lattice of
integers, proposed by [9], [10].
Recently, with the rapid expansion of large-scale, online
networks, there has been an increased interest in the analysis
of opinion formation, with the objective of extending the
classical models by taking into account more factors of social
interactions [11]–[19]. In the work of [11] and [12], the effects
of the existence of stubborn agents–the agents who never
update their opinions–are investigated in a randomized pairwise updating process. In [13], the opinion formation process
is reformulated into a local interaction game, and the concept
of stubbornness of an agent regarding her initial opinion is
introduced. The Krause model and its variations are studied
in [14]–[17]. For example, a game-theoretic analysis of the
Krause model is studied in [16]. The work of [17] takes into
account exogenous factors, such as the influence of media,
and assumes that each agent updates her opinion via the
opinions of the population inside her “confidence range” and
the information from an exogenous input in that range. In the
literature, both discrete-time [1], [5] and continuous-time [20],
[21] approaches have been adopted to model the update rule
of opinions of agents.
Recently, Jia et al proposed the so-called DeGroot-Friedkin
model [18], [19]. This model uses the concept of reflected appraisal from sociology [22], [23], and studies the evolution of
self-confidence, i.e., how confident an agent is in her opinions
on a variety of issues. Briefly speaking, reflected appraisal
describes the phenomenon that agents’ self-appraisals are influenced by the appraisals of other agents on them. Following
the work of [18], [19], a modified DeGroot-Friedkin model
is proposed in [24] in which each agent updates her level of
self-confidence in a more frequent manner. Specifically, all
the agents in the network update their own levels of selfconfidence immediately after each time of discussion, instead
of waiting for the opinion process to reach a consensus on
any particular issue, which generally takes infinite time. The
analysis of the modified DeGroot-Friedkin model is carried out
only in the special case when the so-called relative interaction
matrix is doubly stochastic. Yet, a complete understanding of
the system behavior for the most general case has remained
open.
We introduce in this paper a continuous-time self-appraisal
model whereby the agents in a social network are able to
evaluate their self-appraisals in a distributed way. For each
ordered pair of agents i and j, we assign a function of selfappraisal to agent i, which measures the level of importance
of agent i to agent j. With local interaction with her neighbors
only, each agent is able to calculate the difference between her
level of importance to others and others’ level of importance
to herself. The proposed dynamical system of self-appraisals
aims to drive these differences to zero.
Although many opinion dynamics are built on discrete-time
scales, a continuous-time model holds its merits in many ways:
(i) a continuous-time model would be a natural choice if
the opinions of individuals evolve gradually over time, and
moreover, can be used to describe the limiting behavior of
a discrete-time system if the exchange of the opinions are
frequent enough; (ii) often, the analysis of a continuous-time
system would be easier to carried out than a discrete-time system; and (iii) a complete analysis of a continuous-time system
provides valuable insight into the dynamics of the discretetime counterpart. We should note that the asymptotic behavior
of the trajectories of the continuous-time self-appraisal model
studied in this paper matches numerical simulations of the
discrete-time modified DeGroot-Friedkin model, where the
latter does not permit a thorough analysis.
We note here that the analysis of the self-appraisal model
has been carried out earlier in our conference paper [25], but
for the case when the network topology is a strongly connected graph, for which the convergence of the corresponding
dynamical system has been shown, with several of proofs left
out. In this paper, we generalize the result to the case when
the network topology is a rooted graph, and show that under
some mild assumptions, for almost all initial conditions, the
trajectory generated by the dynamical system asymptotically
converges to an equilibrium point which is exponentially
stable. We provide a complete analysis, as well as proofs for
establishing this result.
The remainder of the paper is organized as follows. In section II, we describe in detail the self-appraisal model as well
as the motivation behind it. We also state the main theorem
of the paper on the convergence of the self-appraisal model.
In particular, the main theorem says that there is only one
stable equilibrium point in the unit simplex (as the underlying
space of the dynamical system), and moreover, for almost all
initial conditions in the unit simplex, the trajectories of the
dynamical system converge to that stable equilibrium point.
Sections III-V are devoted to establishing several properties
that are needed for the proof of the main theorem. A detailed
organization of these three sections will be given after the
statement of Theorem 1. We provide conclusions in the last
section. The paper ends with an Appendix.
II. T HE S ELF -A PPRAISAL M ODEL AND M AIN R ESULT
In this section, we introduce the continuous-time selfappraisal model, and state the main result of this paper.
A. Background and notations
By convention, the neighbor relation among the n agents in
the network is characterized by a directed graph (or digraph)
G = (V, E), with V = {1, . . . , n} the vertex set and E
the edge set. We consider in this paper only simple directed
graphs, that is directed graphs with no self loops, and with at
most one edge between each ordered pair of vertices. Denote
by i → j a directed edge of G in which i is the start-vertex and
j is the end-vertex, and we say that j is an outgoing neighbor
of i and i is an incoming neighbor of j. Denote by Vi+ and Vi−
the sets of incoming and outgoing neighbors of i, respectively.
A directed path (or simply path) of G is a sequence of edges
i1 → i2 → . . . → ik
connecting vertices of G, and all the vertices in the path are
distinct from each other. A vertex i is said to be a root if for
any other vertex j, there is a directed path from j to i in G.
The digraph G is said to be rooted if it contains at least a
root, and strongly connected if each vertex of G is a root.
For a subset V 0 ⊂ V , we call G0 a subgraph of G induced by
V 0 if G0 = (V 0 , E 0 ) and E 0 contains any edge of E whose
start-vertex and end-vertex are in V 0 . Denote by Vr ⊆ V the
collection of roots of G, and denote by Gr = (Vr , Er ) the
subgraph of G induced by Vr . It is well known that the digraph
Gr is strongly connected.
Denote by Sp[V ] the (n − 1)-simplex contained in Rn with
vertices the standard basis vectors e1 , . . . , en ∈ Rn . For V 0 ⊂
V , we define similarly Sp[V 0 ] as the convex hull of {ei | i ∈
V 0 }, i.e.,
(
)
X
X
0
Sp[V ] :=
αi ei | αi ≥ 0,
αi = 1 .
i∈V 0
i∈V 0
A point x ∈ Sp[V ] is said to be a boundary point of Sp[V ]
if xi = 0 for some i ∈ V . If V 0 is a proper subset of V , then
each x ∈ Sp[V 0 ] is a boundary point of Sp[V ].
For an arbitrary dynamical system ẋ = f (x) in Rn , a subset
Q ⊆ Rn is said to be positive invariant if for any initial
condition x(0) ∈ Q, the solution x(t) of the dynamical system
is contained in Q for all time t ≥ 0.
B. The self-appraisal model
To introduce the self-appraisal model, we first consider the
following opinion consensus process, as a variation of the
DeGroot model [1], in continuous-time: Let G = (V, E) be
a rooted graph; each vertex i of G has at least one outgoing
neighbor. The opinion consensus process is described by
X
cij zj (t) .
(1)
żi (t) = (1 − xi (t)) −zi (t) +
j∈Vi−
Each zi (t) is a real number (or vector), representing the
opinion of agent i on certain ongoing issue(s) at time t. The
number xi (t) ∈ [0, 1] represents the current self-appraisal of
agent i in the social network. The coefficients cij ’s, termed
as relative inter-personal weights in [18], are positive real
numbers, satisfying the following condition:
X
cij = 1, ∀i ∈ V.
(2)
j∈Vi−
Each cij can be set by agent i herself, encoding the willingness
of agent i to accept the opinion of agent j. Alternatively, cij
can be also viewed as a measure of the influence of the opinion
of agent j on the opinion of agent i.
We then recognize that system (1), with condition (2), is
a continuous-time consensus process [26], with the dynamics
of zi (t) scaled by the non-negative factor (1 − xi (t)). Note
that in (1), a larger value of xi (t) implies a smaller value of
kżi (t)k. So then, (1 − xi (t)) can be viewed as a measure of
the total amount of opinions agent i accepts from others at
time t, and cij (1 − xi (t)) is the corresponding portion agent i
accepts from agent j.
With the opinion consensus-process (1) in mind, we propose
the following dynamics for the evolutions of self-appraisals:
X
ẋi = −(1 − xi )xi +
cji (1 − xj )xj , ∀i ∈ V.
(3)
j∈Vi+
Similar to the works [18], [24], the self-appraisals are scaled
so that they sum to one, i.e.,
x := (x1 , . . . , xn ) ∈ Sp[V ].
Note that such a scaling will be effective along the evolution;
indeed, we show in the next section that Sp[V ] is a positive
invariant set for system (3).
We now justify the self-appraisal model introduced above.
From
(3), the evolution of xi is determined by two terms:
P
+ cji (1 − xj )xj and (1 − xi )xi . Explanations of these
j∈Vi
two terms are given below.
Each summand cji (1 − xj )xj in the first term is a product
of two factors: one is cji (1 − xj ) which measures the amount
of opinion agent j accepts from agent i during the consensus
process (1), and the other factor xj is the self-appraisal of
agent j which reflects the importance of agent j in the
network. Thus, their product cji (1 − xj )xj can be viewed as
the measure of importance of agent i to agent j. Of course,
there are numerous ways of modeling these two factors, the
rationale behind the choice of using the product is given below.
We first consider the case where xj = 0; then agent j is not
important at all in the social network, and thus, agent i will
not increase her self-appraisal regardless of how much opinion
agent j accepts from her. On the other hand, consider the case
where xj = 1; then from the consensus process, we see that
agent j will not accept any opinion from agent i. Thus, in this
situation, agent i will not increase her self-appraisal either
regardless of how important agent j is in the network. By
taking these two cases into account, we realize that cji (1 −
xj )xj may be one of the simplest expressions that realistically
models how the neighbor
j affects the self-appraisal of agent i.
P
The summation j∈V + cji (1 − xj )xj can then be viewed
i
as the measure of importance of agent i to others. Conversely,
by (2), the other term (1 − xi )xi can be expressed as follows:
X
(1 − xi )xi =
cij (1 − xi )xi ,
j∈Vi−
We thus interpret (1 − xi )xi as the measure of importance
of others to agent i. Note that by communicating with her
neighbors, each agent i is able to compute these two terms by
herself.
The self-appraisal model is then designed so that agent i
measures the difference of the two terms, and drives it to
zero. In particular, note that if x∗ is an equilibrium point
of system (3), then for each agent i, we have the balance
equation:
X
(1 − x∗i )x∗i =
cji (1 − x∗j )x∗j .
j∈Vi+
In other words, at an equilibrium point, the importance of
agent i to others is equal to the importance of others to agent i.
C. The main result
In this subsection, we state the main result of the paper. We
first recall that the coefficient cij can be viewed as a measure
of the influence of the opinion of agent j on the opinion of
agent i. The larger the value of cij is, the more influential is
agent j on agent i. We call agent j a dominant neighbor of
agent i if cij > 1/2. Note that in this case, because of (2), we
have
X
cij >
cij 0 ;
j 0 ∈Vi− −{j}
in other words, the influence of agent j on agent i exceeds the
influences of all the other neighbors of agent i together. Note
that if an agent i has agent j as its unique outgoing neighbor;
then, by (2), we have cij = 1, and hence agent j is a dominant
neighbor of agent i; indeed, in this case, agent i can only take
opinions from agent j. On the other hand, if agent i has at
least two outgoing neighbors, then, the coefficients cij , for
j ∈ Vi− , can be set by agent i in a way so that there is no
dominant neighbor of agent i.
We assume in this paper that agents in the network act
cautiously when taking opinions from others so that there is
no dominant neighbor in the network. Specifically, we assume
that each vertex i of G has at least two outgoing neighbors.
Moreover, the coefficients cij , for i → j ∈ E, are such that
cij ≤ 1/2,
∀i → j ∈ E.
We formalize below the following facts about the self-appraisal
model: (i) There is a unique stable equilibrium point x∗ :=
(x∗1 , . . . , x∗n ) of system (3) in Sp[V ]; (ii) Almost all trajectories
x(t) converge to x∗ , and hence x∗ can be interpreted as the
steady state of the self-appraisals of the agents in the network,
independent of their initial conditions; (iii) The self-appraisal
x∗i is zero if vertex i is not a root of the graph G. Furthermore,
we have x∗i < 1/2, for all i ∈ V . This,P
in particular, implies
that in the steady state, we have x∗i < j6=i x∗j , i.e., there is
no single agent whose self-appraisal is greater than or equal
to the sum of the self-appraisals of the remaining agents.
We now state the main result in precise terms. We first
summarize the key conditions for system (3), which will be
assumed in the remainder of the paper:
Assumption 1. The digraph G = (V, E) is simple and rooted,
with Vr the set of roots. Each vertex i of G has at least two
outgoing neighbors. The coefficients cij of (3), for i → j ∈ E,
satisfy the following conditions:
X
cij = 1, ∀i ∈ V,
j∈Vi−
and
cij ≤ 1/2,
∀i → j ∈ E.
We now have the main result, captured by the theorem
below:
Theorem 1. Under Assumption 1, the self-appraisal model (3)
satisfies the following properties:
1) The unit simplex Sp[V ] is a positive invariant set.
2) There are (n + 1) equilibrium points of system (3). Each
of the n vertices ei of Sp[V ] is an unstable equilibrium
point. The remaining equilibrium point x∗ lies in Sp[Vr ],
and satisfies the following condition:
0 < x∗i < 1/2,
∀i ∈ Vr .
Moreover, x∗ is exponentially stable.
3) For any initial condition x(0) other than a vertex of
Sp[V ], the solution x(t) of system (3) converges to x∗ .
Remark 1. Note that the cascaded system (1) and (3) admits
a triangular structure (i.e., (3) feeds into (1) but not the other
way around), and hence the convergence of system (3) implies
the convergence of the consensus process (1).
We note here that the self-appraisal model (3) is similar, in
its format, to the replicator dynamics (see, for example, [27],
[28]), which is also defined over the unit simplex. Yet,
these are two different types of dynamical systems. Indeed,
a replicator dynamics equation is given by
n
X
xj gj (xj ) , ∀ i ∈ V.
(4)
ẋi = xi gi (x) −
j∈V
First, note that the dynamics of each xi in (4) depends on a
global information of x. Yet, in the self-appraisal model (3),
the dynamics of xi depends only on xj for j ∈ Vi+ . Second,
note that for the replicator dynamics, each subset Sp[V 0 ], for
V 0 ⊂ V , is a positive invariant set; indeed, from (4), if xi =
0, then ẋi = 0. Yet, such an invariance property does not
hold in the self-appraisal model; indeed, from Theorem 1, if
Sp[V 0 ] is positive invariant, then, either V 0 = {ei } for some
i ∈ V , or V 0 contains the root set Vr (we prove this fact
formally in the next section). We further note that if G is
a strongly connected graph, then Vr = V , and hence from
Theorem 1, there is a unique stable equilibrium point of (3)
in the interior of the simplex. For the replicator dynamics, a
stable equilibrium point may or may not lie in the interior of
the simplex (see, for example, [28]). All these facts point to
the intrinsic differences between the two types of models.
On the other hand, the self-appraisal model (3) can be
viewed as a prototype of a general class of nonlinear dynamical systems defined over the unit simplex:
X
ẋi = −gi (xi ) +
cji gj (xj ).
j∈Vi+
with gi (x) ≥ 0 and gi (0) = gi (1) = 0 for all i ∈ V . Hence,
the analysis of system (3) provided in the paper, for locating
the set of equilibrium points and for establishing the global
convergence, might be of independent interest.
The remainder of the paper is now devoted to establishing
properties of system (3) that are needed to prove Theorem 1. In
section III, we focus on some basic properties of system (3). In
particular, we construct a family of subsets of the unit-simplex
each of which is a positive invariant set for system (3). In
section IV, we focus on the system behavior around a vertex
of the simplex. We show that there is a closed neighborhood
around each vertex in the simplex such that if the initial
condition is not the vertex, then the solution of system (3)
will be away from that neighborhood after a finite amount of
time. In section V, we establish the global convergence of the
self-appraisals towards the unique stable equilibrium point.
III. BASIC P ROPERTIES OF THE S ELF -A PPRAISAL M ODEL
In this section, we compute time-derivatives along trajectories of system (3) at boundary points of the unit simplex.
We then exhibit a family of subsets of Sp[V ] (including
the simplex itself), each of which is positive invariant for
system (3).
A. Time-derivatives at boundary points
In this subsection, we assume that xi (t) = 0, for some
i ∈ V , and evaluate the time-derivatives dk xi (t)/dtk , for k ≥
1. We first establish in this subsection the following fact: let
dk xi (t)/dtk be the first non-vanishing time derivative, i.e.,
dk xi (t)/dtk 6= 0 and dl xi (t)/dtl = 0 for all l < k; then,
we have dk xi (t)/dtk > 0. Toward that end, we first introduce
some definitions about the graph G:
Definition 1 (Supporting set). For a vertex i ∈ V , a set Si ⊆
V is a supporting set, if a vertex j is in Si whenever there
is a path from j to i. The vertex i is in Si by default.
Note that if i is a root of G, then Si = V . Indeed, if j ∈ Si ,
then so is any vertex in Vj+ .
Let γ be a path in G, and l(γ) denote its length. For each
non-negative integer k, we define Si (k) ⊆ Si as follows: if
j ∈ Si (k), then there is a path γ from j to i with l(γ) ≤ k.
We assume that i ∈ Si (k) for all k ≥ 0, and Si (0) = {i}. By
definition, we have Si (k − 1) ⊆ Si (k) for all k ≥ 1. Denote
by Di (k) the difference between Si (k) and Si (k − 1), i.e.,
Di (k) := Si (k) − Si (k − 1);
(5)
for k = 0, we set Di (0) := {i}. Note that in the case k = 1,
we have Si (1) = {i} ∪ Vi+ , and hence Di (1) = Vi+ . We
further note the following fact:
Lemma 1. Let j ∈ Di (k) for some k ≥ 1. Then, Vj− intersects
Si (k − 1) and their intersection lies in Di (k − 1).
Proof. First note that if j ∈ Di (k) and γ is a path from j
to i of shortest length, then l(γ) = k. Since j ∈ Di (k) for
k ≥ 1, there exists at least one vertex j 0 ∈ Vj− such that
j 0 ∈ Si (k − 1), because otherwise the length of any path from
j to i is greater than k. Now let j 0 ∈ Vj− ∩ Si (k − 1), and γ 0
be a path from j 0 to i; it suffices to show that l(γ 0 ) ≥ (k − 1).
We prove by contradiction. Suppose that l(γ 0 ) < (k − 1); then
by concatenating j → j 0 with γ 0 , we get a path γ from j to
i with l(γ) < k, which contradicts the fact that j ∈ Di (k).
This completes the proof.
Let j ∈ Di (k), and denote by Γji (k) the collection of paths
from j to i of length k. Choose any path γ ∈ Γji (k), and write
γ = jk → jk−1 → . . . → j1 → j0
with jk = j and j0 = i. Then, each jl lies in the intersection
of Si (l) and Vj−
. Thus, by Lemma 1, we have
l+1
jl ∈ Di (l),
∀l = 0, . . . , k.
For each γ ∈ Γji (k), define a positive number as follows:
With the preliminaries above, we compute below the time
derivatives of xi (t). For convenience, let
(k)
xi (t) := dk xi (t)/dtk .
We then have the following fact:
Proposition 1. Let the initial condition x(0) of system (3) be
in Sp[V ], other than a vertex. Suppose that xi (0) = 0 for
some i ∈ V . Then, the following properties hold:
1. If there does not exist a supporting path for i at x(0),
then, i ∈
/ Vr and xi (t) = 0 for all t ≥ 0.
2. If there is a critical supporting path for i at x(0), and
the length of the path is k, then,
αγ := Πkl=1 cjl jl−1 .
(l)
xi (0) = 0
We further define
αji :=
X
for all l < k, and
αγ .
(6)
(k)
Note that αji can be defined recursively as follows: For the
base case, we assume that j ∈ Di (1) = Vi+ ; then, Γji (1) is
a singleton, comprised only of the edge j → i, and hence
αji = cji . For the inductive step, we assume that k > 1 and
αj 0 i , for j 0 ∈ Di (k − 1), are well defined. Let j ∈ Di (k), then
X
αji =
cjj 0 αj 0 i
(7)
j0
where the summation is over Vj− ∩ Di (k − 1), which is
nonempty by Lemma 1.
To state the main result of this subsection, we further need
the following definition:
Definition 2 (Supporting path). Let x ∈ Sp[V ] with xi = 0
for some i ∈ V . A path
If, in addition, k is the least integer among the lengths of all
supporting paths for i at x, then γ is a critical supporting
path.
It should be clear that if there exists a supporting path for i
at x, then there will be a critical supporting path. Note that
all critical supporting paths for i at x have the same length.
We have, however, also the following fact:
Lemma 2. Let x ∈ Sp[V ] with xi = 0. If there is no
supporting path for i at x, then xj = 0 for all j ∈ Si .
Proof. We prove the result by contradiction. Suppose that
there exists some j ∈ Si such that xj 6= 0. Then, by the
construction of the set Si , we can choose a path γ from j to
i:
γ = j → jk−1 → . . . → j1 → i.
By truncating the path if necessary, we can assume that xjl =
0 for all l = 1, . . . , k − 1. Then, γ is a supporting path for i
at x, which is a contradiction.
αji (1 − xj (0))xj (0) > 0
j∈Di (k)
with Di (k) defined in (5) and αji defined in (6).
We refer to the Appendix for a proof of Proposition 1.
B. Positive invariant sets
Now, for each subset V 0 ⊆ V , we define
SV 0 := ∪i∈V 0 Si .
(8)
If V 0 is empty, then so is SV 0 . Note that if V 0 contains a root
of G, then SV 0 = V . Denote by SVc 0 the complement of SV 0
in V , i.e.,
SVc 0 := V − SV 0 .
Then, the set Sp[SVc 0 ] can be described by
γ := jk → . . . → j1 → i
is a supporting path for i at x if
xjk > 0
xjl = 0 ∀l < k.
X
xi (0) =
γ∈Γji (k)
Sp[SVc 0 ] = {x ∈ Sp[V ] | xi = 0, ∀i ∈ SV 0 }.
We next establish the following fact:
Proposition 2. Let V 0 be any subset of V . Then, Sp[SVc 0 ] is
a positive invariant set for system (3).
Proof. First, note that if x(0) ∈ Sp[SVc 0 ], then there is no
supporting path for i at x(0) for all i ∈ SV 0 . Then, by
Proposition 1, we have
xi (t) = 0, ∀t ≥ 0 and ∀i ∈ SV 0 .
Pn
Next, we show that i=1 xi is invariant along the evolution.
To see this, note that
Pn P
Pn
i=1 (1 − xi )xi = Pi=1 Pj∈Vi− cij (1 − xi )xi
n
=
i=1
j∈V + cji (1 − xj )xj .
i
Hence, we have
n
X
ẋi =
i=1
Thus,
n
X
i=1
X
i=1
Pn
cji (1 − xj )xj − (1 − xi )xi = 0.
j∈Vi+
xi (t) = 1 for t ≥ 0. It now suffices to show that
xi (t) ≥ 0,
∀t ≥ 0 and ∀i ∈ V.
We prove this by contradiction. Suppose that, to the contrary,
there exists a vertex i ∈ V , an instant t > 0 and an > 0
such that xi (t) = 0 and
xi (t0 ) < 0,
∀ t0 ∈ (t, t + ).
(9)
There are two cases:
Case I. Suppose that there does not exist a supporting path
for i at x(t); then, by Proposition 1, xi (t0 ) = 0 for all t0 ≥ t,
which contradicts (9).
Case II. Suppose that there is a supporting path for i at
x(t); then, appealing again to Proposition 1, we have that there
(k)
(l)
exists a k such that xi (t) > 0 and xi (t) = 0 for all l < k.
This, in particular, implies that there is an 0 > 0 such that
xi (t0 ) > 0 for all t0 ∈ (t, t + 0 ), which contradicts (9). This
completes the proof.
We state below some implications of Proposition 2. First,
let V 0 be an empty set; then SVc 0 = V . Thus, we have the
following fact as an immediate consequence of Proposition 2:
Corollary 1. The unit simplex Sp[V ] is a positive invariant
set for system (3).
Recall that Vr is the root set of G; we establish the following
fact as another corollary to Proposition 2:
Corollary 2. The subset Sp[Vr ] of Sp[V ] is a positive invariant set for system (3).
Proof. Let V 0 := V − Vr ; we prove the result by showing
that SVc 0 = Vr , and then, appealing to Proposition 2. It should
be clear that SVc 0 ⊆ Vr , and we prove that SVc 0 ⊇ Vr . The
proof is carried out by contradiction. Suppose that there is a
root i ∈ SV 0 . From (8), we know that there is a vertex j ∈ V 0 ,
together with a path γ from i to j. But then, since i is a root,
each vertex of γ is a root. In particular, we have that j ∈ Vr ,
which contradicts the fact that j ∈ V 0 . This completes the
proof.
IV. V ERTICES A RE R EPELLERS
In this section, we focus on the system behavior around a
vertex of the simplex. First, observe that each vertex ei ∈ Rn
of the simplex is an equilibrium point of system (3). We show
below that each ei is unstable, and in fact, it is a repeller. To
this end, we choose an ∈ (0, 1), and define Pi () ⊂ Sp[V ]
as follows:
Pi () := x ∈ Sp[V ] | xi ≥ .
It should be clear that each Pi () is a closed neighborhood of
ei in Sp[V ], and
Pi () ) Pi (0 ),
if < 0 .
Denote by f (x) the vector field of system (3), and by fi (x)
the i-th component of f (x), i.e.,
X
fi (x) := −(1 − xi )xi +
cji (1 − xj )xj .
(10)
j∈Vi+
We next establish the following fact:
Proposition 3. Suppose that for a vertex i ∈ V , there exists
∈ (0, 1) such that cji ≤ for all j ∈ Vi+ . Then, there exists
an αi ∈ (0, ) such that fi (x) < 0 for all x ∈ Pi (αi ) − {ei }.
Remark 2. We recall that cji measures the influence of agent i
on agent j. Proposition 3 then says that if the influence of
agent i on each of her incoming neighbors is less than ,
then the self-appraisal xi (t) of agent i decreases along the
evolution and stays below .
Proof of Proposition 3. First, note that if vertex i does not
have any incoming neighbor, then
fi (x) = −xi (1 − xi ),
and hence, fi (x) < 0 as long as xi ∈ (0, 1). Hence, in this
case, Proposition 3 holds for any αi ∈ (0, ). We thus assume
that Vi+ is nonempty for the rest of the proof.
To prove the result, we first apply the condition that cji ≤ ,
and obtain
P
fi (x) ≤ −(1 − xi )xi + j∈V + (1 − xj )xj .
(11)
i
Now, fix the value of xi , and consider the following optimization problem:
P
max j∈V + (1 − xj )xj
i
P
+ xj ≤ 1 − xi
s.t.
and xj ≥ 0.
j∈Vi
P
Since the function j∈V + (1 − xj )xj is strictly concave, the
i
maximum is achieved uniquely at
xj = (1 − xi )/di ,
∀j ∈ Vi+
where di is the cardinality of Vi+ . Thus,
X
max
(1 − xj )xj = (1 − xi )(1 − (1 − xi )/di ).
(12)
j∈Vi+
We then combine (11) and (12), and get
di − 1
.
fi (x) ≤ −(1 − /di )(1 − xi ) xi −
di −
For the first term on the right hand side of the expression, we
have (1 − /di ) > 0. For the second, we have that if x 6= ei ,
then (1 − xi ) > 0. For the last term, we let
αi := −
1 (1 − )
.
2 di −
(13)
Then, we have
xi −
di − 1
1 (1 − )
≥
> 0,
di −
2 di −
∀xi ≥ αi .
Combining these facts, we arrive at the result that fi (x) < 0,
for all x ∈ Pi (αi ) − {ei }.
Recall that cij ≤ 1/2 for all i → j ∈ E. Thus, if we let
= 1/2, then following (13), we have
αi =
1
1
1
1
−
≤ −
.
2 4(2di − 1)
2 4(2n − 3)
The equality holds if and only if Vi+ = V − {i}, and hence
di = n − 1. For convenience, let α be defined as follows:
1
1
α := −
.
(14)
2 4(2n − 3)
It is the sharp upper-bound for all αi ’s. This number α will
be fixed for the remainder of the paper.
Following Proposition 3, we now describe some relevant
properties of system (3). First, we define a subset Q ⊂ Sp[V ]
as follows:
Q := {x ∈ Sp[V ] | xi ≤ α, ∀i ∈ V }.
Note that Q can be expressed as follows
\
Q=
cl Sp[V ] − Pi (α)
(15)
By the fact that cl(Sp[V ] − Pi (α)) is positive invariant, we
can choose Ti = 0. We now assume that xi (0) > α. Then,
Pi (α) ) Pi (xi (0)),
and hence the set Pi (α) − Pi (xi (0)) is nonempty. Let
vi := inf |fi (x)| | x ∈ Pi (α) − Pi (xi (0))
Since cl (Pi (α) − Pi (xi (0))) is compact over which fi (x) is
strictly negative, we have vi > 0. Let
(16)
i∈V
where cl(·) denotes the closure of a set in Rn . An illustration
of Q in the case n = 3 is given in Figure 1.
Ti :=
xi (0) − α
+1
vi
Then, by construction, the trajectory x(t) will enter cl(Sp[V ]−
Pi (α)) in no more than Ti units of time. This completes the
proof.
V. T HE G LOBAL C ONVERGENCE OF
T HE S ELF -A PPRAISAL M ODEL
In this section, we establish the global convergence of
system (3). The section is divided into four parts: In subsection V-A, we show that the self-appraisals of non-root agents
converge to zero. In subsection V-B, we establish the second
part of Theorem 1, i.e., there is only one stable equilibrium
point x∗ of system (3), which lies in Sp[Vr ], and moreover,
0 < x∗i < 1/2, for all i ∈ Vr . In subsection V-C, we show
that if the initial condition x(0) of system (3) is such that
x(0) ∈ Q, and xi (0) = 0 for i ∈
/ Vr , then, the trajectory
x(t) converges to the unique stable equilibrium point x∗ .
In subsection V-D, we combine all the results and prove
Theorem 1.
Fig. 1. Illustration of how Pi (α)’s and Q are defined and how these sets are
distributed over the simplex in the case n = 3.
We obtain some results as corollaries to Proposition 3:
Corollary 3. If x is a non-vertex equilibrium point of system (3), then x ∈ Q.
Proof. If x ∈ Sp[V ] − Q, then x ∈ Pi (α) for some i ∈ V .
Since x is not a vertex, by Proposition 3, we have fi (x) < 0.
Thus, x can not be an equilibrium point if x ∈
/ Q.
We now show that the set Q is also positive invariant:
Corollary 4. The set Q is positive invariant for system (3).
For any initial condition x(0) ∈ Sp[V ] other than a vertex,
the trajectory x(t) enters Q in a finite amount of time.
Proof. We first show that Q is positive invariant. From (16),
it suffices to show that each cl(Sp[V ] − Pi (α)) is positive invariant. By construction, xi ≤ α if and only if x ∈ cl(Sp[V ]−
Pi (α)). On the other hand, by Proposition 3, fi (x) < 0 for all
x ∈ Pi (α). It then follows that cl(Sp[V ] − Pi (α)) is positive
invariant.
We now show that if x(0) 6= ei for any i ∈ V , then there
exists an instant T ≥ 0 such that x(t) ∈ Q for all t ≥ T .
Similarly, it suffices to show that for each i ∈ V , there exists
an instant Ti ≥ 0 such that x(t) ∈ cl(Sp[V ]−Pi (α)) for all t ≥
Ti . First, note that if xi (0) ≤ α, then x ∈ cl(Sp[V ] − Pi (α)).
A. Convergence of self-appraisals of non-root agents
An agent i is said to be a non-root agent if i is not a
root of G. We establish in this subsection the result that the
self-appraisals of non-root agents decay to zero.
Proposition 4. Let x(0) ∈ Sp[V ] be other than a vertex. Then,
xi (t) converges to 0 for all i ∈
/ Vr .
To prove Proposition 4, we first etablish the following fact:
Lemma 3. If xi (t) converges to 0, then so does xj (t) for all
j ∈ Vi+ .
Proof. We prove Lemma 3 by contradiction. Suppose that
there is a vertex j ∈ Vi+ such that xj (t) does not converge to
0. Then, there is an ∈ (0, 1/2), and an infinite time sequence
{tk }k∈N with limk→∞ tk = ∞ such that xj (tk ) > for all
k ∈ N.
Since there is an upper bound for |fj (x)| for all x ∈ Sp[V ],
there is a finite duration time τ > 0 such that
xj (tk + τ 0 ) ≥ /2,
∀τ 0 ∈ [0, τ ] and ∀k ∈ N.
Rescale τ , if necessary, so that τ ≤ 1. On the other hand, by
Corollary 4, there exists an instant T ≥ 0 such that x(t) ∈ Q
for all t ≥ T . This, in particular, implies that xj (t) ≤ 1/2 for
all t ≥ T . Without loss of generality, assume that tk ≥ T for
all k ∈ N, and hence
xj (t) ∈ [/2, 1/2]
for all t ∈ [tk , tk + τ ] and for all k ∈ N. It then follows that
cji (1 − xj (t))xj (t) ≥ cji (1 − /2)/2.
(17)
for all t ∈ [tk , tk + τ ] and for all k ∈ N.
For convenience, let δ := cji (1 − /2)/2. Since xi (t)
converges to 0, there is an instant T 0 such that
xi (t) ≤ τ δ/2,
∀t ≥ T 0 .
(18)
Using the fact that τ < 1 and (1 − xi (t)) < 1, we obtain
(1 − xi (t))xi (t) < δ/2,
∀t ≥ T 0 .
(19)
We recall that fi (x) (defined in (10)) is given by
X
fi (x) = −(1 − xi )xi +
cji (1 − xj )xj .
j∈Vi+
We first compute the time derivative of φr (x) along a trajectory x(t) of system (3): first, let E 0 be a subset of E defined
by
E 0 := {i → j ∈ E | i ∈ DVr (1), j ∈ Vr };
we then have
X
d
cij (1 − xi )xi ≥ 0.
φr (x) =
dt
0
Note that by its definition (21), we have that φr (x) ≤ 1
for all x ∈ Sp[V ]. Also, note that from (22), φr (x(t)) is
monotonically increasing in t. So, limt→∞ φr (x(t)) exists.
Next, we show that φ̇r (x(t)) converges to 0. The proof
will be similar to the proof of Lemma 3, and is carried out
by contradiction. Suppose that there exists an , and a time
sequence {tk }k∈N with limk→∞ tk = ∞, such that
φ̇r (x(tk )) ≥ ,
Combining (17) and (19), we then obtain
fi (x(t)) > δ/2,
as long as tk ≥ T 0 . It then follows that
Z tk +τ
xi (tk + τ ) = xi (tk ) +
fi (x(t))dt > δτ /2
φ̇r (x(t)) ≥ /2,
tk
Recall that Si ⊂ V is the supporting set of i, and Si (k) ⊂ V
is defined such that if j ∈ Si (k), then there is a path γ from
j to i, with l(γ) ≤ k. Also, recall that for a subset V 0 of
V , we have defined SV 0 = ∪i∈V 0 Si . In the remainder of this
subsection, we focus on the case where V 0 is Vr , the set of
roots of G.
First, note that by the definition of Vr , we have SVr = V ;
indeed, for any vertex j, there is a path from j to a root of
G. Now, for each k ≥ 0, we define
SVr (k) := ∪i∈Vr Si (k).
For k = 0, we have SVr (0) = Vr . It should be clear that
SVr (k − 1) ⊆ SVr (k), and moreover, there is an integer m
such that
SVr (m) = V.
(20)
We now assume that m is the least integer such that (20) holds.
Let
DVr (k) := SVr (k) − SVr (k − 1).
With Lemma 3 and the notations above, we prove Proposition 4.
Proof of Proposition 4. We prove the proposition by subsequently showing that
lim xi (t) = 0,
t→∞
∀i ∈ DVr (k),
for k = 1, . . . , m.
Base case. We prove for k = 1. Let φr be a function on
Sp[V ] defined as follows:
X
φr (x) :=
xi .
(21)
i∈Vr
∀k ∈ N.
Note that |φ̈r (x(t))| is bounded above for any x(t) ∈ Sp[V ],
and hence there is a duration time τ > 0 such that
∀t ∈ [tk , tk + τ ]
which contradicts (18). This completes the proof.
(22)
i→j∈E
∀t ∈ [tk , tk + τ ] and ∀k ∈ N.
By passing to a subsequence, we can assume that tk +τ < tk+1
for all k ∈ N. Thus, we have
X Z tk +τ
lim φr (x(t)) − φr (x(0)) ≥
φ̇r (x(s))ds.
t→∞
k∈N
tk
By construction, each summand on the right hand side is
bounded below by τ /2, and hence the summation is infinite
which contradicts the fact that limt→∞ φr (x(t)) exists.
Since φ̇r (x(t)) converges to 0 and each summand of
φ̇r (x(t)) in (22) is non-negative, we have
lim xi (t)(1 − xi (t)) = 0,
t→∞
∀i ∈ DVr (1).
Thus, xi (t) converges to either 0 or 1. Since x(0) is not a
vertex, by Corollary 4, the trajectory x(t) enters Q in a finite
amount of time. We thus conclude that xi (t) converges to 0
for all i ∈ DVr (1).
Inductive step. We assume that xj (t) converges to 0 for all
j ∈ DVr (k − 1) for k ≥ 2, and prove that xi (t) converges
to 0 for all i ∈ DVr (k). Fix a vertex i in DVr (k); then, from
Lemma 1, there is a vertex j ∈ Vi− ∩ DVr (k − 1). From
the induction hypothesis, we have that xj (t) converges to 0.
Since i ∈ Vj+ , by Lemma 3, xi (t) also converges to 0. This
completes the proof.
B. The non-vertex equilibrium point
In this subsection, we prove that there exists a unique nonvertex equilibrium point of system (3). First, recall that Q is
defined in (15) and Sp[Vr ] is the convex hull spanned by ei ,
for i ∈ Vr . We then define
Qr := Q ∩ Sp[Vr ].
(23)
From Corollaries 2 and 4, both Sp[Vr ] and Q are positive
invariant sets for system (3), and hence Qr is also a positive
invariant set. We establish in this subsection the following
result:
Proposition 5. There is a unique non-vertex equilibrium point
x∗ of system (3), which lies in Qr , and 0 < x∗i < 1/2 for all
i ∈ Vr . Moreover, x∗ is exponentially stable.
for all i ∈ V . Thus,
vi ≤
1
1X
vj = (1 − vi ),
2
2
j6=i
To prove Proposition 5, we re-write system (3) into a matrix
form. We first recall the definition of infinitesimally stochastic
matrices:
Definition 3 (Infinitesimally stochastic matrix). A square
matrix A is an infinitesimally stochastic matrix if each offdiagonal entry of A is non-negative, and the entries of each
row sum to zero.
We now define an infinitesimally stochastic matrix C by
specifying its off-diagonal entries: define the ij-th entry of C
to be cij if i → j ∈ E, and 0 otherwise. The diagonal entries
of C are then uniquely determined by the condition that the
rows of C sum to zero. We note here that the negative of
the matrix C is known as the Laplacian matrix of a weighted
graph G, with weights cij for i → j ∈ E.
For a vector x in Sp[V ], let X be a diagonal matrix defined
as follows:
X := diag(x1 , . . . , xn ).
Then, system (3) can be rewritten into the following matrix
form:
ẋ = C > (I − X)x.
(24)
For simplicity, but without loss of generality, we assume in
the rest of this subsection that
Vr = {1, . . . , k}.
(25)
Note that k ≥ 3, i.e., the cardinality of Vr is at least three;
indeed, by the assumption, each vertex i of G has at least two
outgoing neighbors, and moreover, an outgoing neighbor of a
root is also a root. Following (25), we have that C is a lower
block-triangular matrix:
C11
0
C=
(26)
C21 C22
with C11 a k-by-k infinitesimally stochastic matrix.
We next state some known facts about the matrix C: (i)
Since G is rooted and cij > 0 for all i → j ∈ E, the matrix C
has zero as a simple eigenvalue while all the other eigenvalues
of C have negative real parts (see, for example, [29]). Using
the fact that C11 is an infinitesimally stochastic matrix, we
have that C11 has zero as a simple eigenvalue, with all the
other eigenvalues of C11 having negative real parts, and that
C22 is a stable matrix, i.e., all the eigenvalues of C22 have
negative real parts. (ii) Let v ∈ Rn (resp. v 0 ∈ Rk ) be the
left-eigenvector of C (resp. C11 ) correspondingPto the zero
n
eigenvalue; then v = (v 0 , 0). Scale v such that i=1 vi = 1;
0
then all entries of v are positive, and they sum to one (see,
for example, [30]).
We next state another relevant property of v. Recall that
cij ≤ 1/2 for all i → j ∈ E. It then follows that
vi ≤ 1/3,
∀i ∈ V.
To see this, note that C > v = 0, and hence
X
vi =
cji vj
j∈Vi+
(27)
from which (27) holds.
We now return to the proof of Proposition 5. We establish
the proposition by first showing that the unique equilibrium
point x∗ lies in Qr , with 0 < x∗i < 1/2 for all i ∈ Vr , and
then showing that x∗ is exponentially stable. This is done in
Lemmas 4 and 5 below.
Lemma 4. There exists a unique non-vertex equilibrium point
x∗ of system (3). Moreover, x∗ ∈ Qr , and 0 < x∗i < 1/2 for
all i ∈ Vr .
Proof. Note that from (24), if x∗ is an equilibrium point of
system (3), then
C > (I − X ∗ )x∗ = 0.
Since the null space of C > is spanned by the single vector v,
we must have
(I − X ∗ )x∗ = µv
(28)
for some µ ∈ R. Note that the entries of (1 − X ∗ )x∗ and of
v are all nonnegative, and thus we have µ ≥ 0.
We first show that x∗i = 0 for all i > k. Since vi = 0 for
all i > k, we have (1 − x∗i )x∗i = 0 for all i > k. Since x∗ is
not a vertex, we have x∗i 6= 1, and hence x∗i = 0 for all i > k.
We now solve for x∗i for all i ≤ k. Following (28), we have
(1 − x∗i )x∗i = µvi ,
∀i ≤ k.
From Corollary 3, we have x∗i < 1/2 for all i ≤ k. Thus, for
a fixed µ, we can solve for x∗i (µ) as
p
1
1 − 1 − 4µvi .
x∗i (µ) =
2
It now
suffices
to
show
that
there is a unique positive µ such
Pk
that i=1 x∗i (µ) = 1. Define
ψ(µ) :=
k
X
x∗i (µ);
i=1
note that ψ(0) = 0, and moreover, ψ(µ) is strictly monotonically increasing in µ as long as (1 − 4µvi ) ≥ 0, for
all i = 1, . . . , k. Without loss of generality, we assume that
v1 ≥ vi for all i > 1. Let µ1 := 1/(4v1 ); it then suffices to
show that ψ(µ1 ) > 1.
Let wi := vi /v1 ; then, ψ(µ1 ) can be expressed as follows:
k
ψ(µ1 ) =
k 1 X√
−
1 − wi .
2 2 i=2
We then consider the following optimization problem:
Pk √
max i=2 1 − wi
Pk
s.t. 0 ≤ wi ≤ 1 and
i=2 wi ≥ 2.
Note that the first constraint 0 ≤ wi ≤ 1 comes
the
Pfrom
k
assumption that vi ≤ v1 , and the second constraint i=2 wi ≥
2 follows from (27); indeed, we have
k
X
i=2
wi = (1 − v1 )/v1 ≥ 2.
Pk √
Since the function i=2 1 − wi is strictly concave in wi ,
the maximum is achieved uniquely at
2
, ∀i = 2, . . . , k.
k−1
We recall that the cardinality of the root set Vr is at least three,
i.e., k ≥ 3, and hence
2
≤ 1,
wi =
k−1
which satisfies the constraint that 0 ≤ wi ≤ 1. With the above
values of wi , for i = 2, . . . , k, we have
r
k
X
√
2
max
1 − wi = (k − 1) 1 −
< (k − 2). (29)
k−1
i=2
wi =
The last inequality in (29) holds because
2
2
1
1−
=
<1−
+
k−1
k − 1 (k − 1)2
k−2
k−1
2
and hence J(x∗ )> is an infinitesimally stochastic matrix.
Moreover, the digraph induced by J(x∗ )> is the same as the
digraph induced by C, which is G. Since G is rooted, J(x∗ )
has zero as a simple eigenvalue and all the other eigenvalues
of J(x∗ ) have negative real parts. Let L be a linear subspace
of Rn perpendicular to the vector 1:
L := v ∈ Rn | hv, 1i = 0 .
Then, L can be viewed as the tangent space of Sp[V ] at x for
all x ∈ Sp[V ]. Note that L is invariant under J(x∗ ), i.e., for
any vector v ∈ L, we have J(x∗ )v ∈ L; indeed, we have
hJ(x∗ )v, 1i = hv, J(x∗ )> 1i = 0.
This, in particular, implies that the eigenvalues of J(x∗ )
have negative real parts when restricted to L, and hence the
equilibrium point x∗ of system (3) is exponentially stable.
Proposition 5 is then established by combining Lemmas 4
and 5.
.
Following (29), we then have
C. System convergence over Qr
k 1
− (k − 2) = 1.
2 2
which completes the proof.
ψ(µ1 ) >
Note that in the case when G is strongly connected, the set
of equilibrium points of system (3) coincides with the set of
equilibrium points of the opinion system studied in [18], [19],
though with a completely different dynamical system.
It now remains to show that x∗ is exponentially stable. We
first have some preliminaries.
n×n
Definition 4 (Induced graphs). Let A ∈ R
be an infinitesimally stochastic matrix. We say a digraph GA = (V, EA ) of
n vertices is induced by A if the edge set EA satisfies the
following condition: i → j ∈ EA if and only if the ij-th entry
of A is positive.
It is well known that if GA is rooted, then A has zero
as a simple eigenvalue and all the other eigenvalues have
negative real parts (see, for example, [26], [29]). With the
notions above, we establish the following result:
Lemma 5. Let x∗ be the non-vertex equilibrium point of
system (3). Then, x∗ is exponentially stable.
∗
Proof. Let J(x ) be the Jacobian matrix of the vector field
f (x) at x∗ , i.e.,
∂f (x∗ )
J(x∗ ) :=
.
∂x
Then, by a direct computation,
J(x∗ ) := C > (I − 2X ∗ ).
Since x∗ ∈ Qr , we have
1 − 2x∗i ≥ 1 − 2α = 1/(4n − 6),
∀i ∈ V,
with α defined in (14). In particular, each diagonal entry of
the diagonal matrix (I − 2X ∗ ) is positive. Let 1 ∈ Rn be a
vector of all ones; then, we have
∗ >
∗
J(x ) 1 = (I − 2X )C1 = 0,
In this subsection, we establish the convergence for a
special class of trajectories x(t) of system (3). These are the
trajectories whose initial conditions are in the positive invariant
set Qr defined in (23). Precisely, we have the following:
Proposition 6. Suppose that the initial condition x(0) of
system (3) is in Qr ; then, the trajectory x(t) converges to
the non-vertex equilibrium point x∗ .
We establish Proposition 6 below. First, recall that we have
labeled the vertices of G such that the root set Vr is comprised
of the first k vertices. Let
x0 := (x1 , . . . .xk )
and similarly, define
X 0 := diag(x1 , . . . , xk ).
Note that if x ∈ Qr , then x = (x0 , 0). Since Qr is positive
invariant, the dynamics of x0 is simply given by
>
ẋ0 = C11
(I − X 0 )x0
(30)
with C11 the k-by-k block matrix defined in (26). To prove
Proposition 6, it suffices to show that for any initial condition
x0 (0), the solution x0 (t) of system (30) converges to x0∗ :=
(x∗1 , . . . , x∗k ).
To do so, we first make a change of variables for system (30). Recall that Gr = (Vr , Er ) is the subgraph induced
by the root set Vr . We now define a set of new variables yi ’s
as follows:
(1 − xi )xi
, ∀i ∈ Vr .
yi :=
(1 − x∗i )x∗i
This is well defined because, from Proposition 5, we have x∗i ∈
(0, α], and hence the denominator is nonzero. We consider
below the dynamics of yi , for i ∈ Vr : First, we define a set
of new coefficients e
cji as follows:
e
cji := cji
(1 − x∗j )x∗j
,
(1 − x∗i )x∗i
∀i → j ∈ Er .
(31)
+
For a vertex i ∈ Vr , denote by Vr,i
the set of incoming
neighbors of i in Gr . It should be clear that
+
Vr,i
= Vi+ ∩ Vr .
Moreover, following (35), we obtain
aij (t) ≥ (1 − 2α) e
cji ,
+
j∈Vr,i
We further describe below some relevant properties of
system (32). First, note that x∗ ∈ Qr is an equilibrium point
of system (3), and hence
>
C11
(I − X 0∗ )x0∗ = 0,
e r = (Vr , E
er )
G
be defined by reversing directions of edges of Gr , i.e.,
er if and only if j → i ∈ Er .
i→j∈E
e r is the digraph induced by A(t).
Then, from (35) and (36), G
e r . In
Furthermore, since Gr is strongly connected, so is G
summary, we have that
e r for all t ≥ 0, and G
er
1. The induced digraph of A(t) is G
is strongly connected.
2. Following (39), there is a uniform lower bound δ > 0
such that
which implies that
aij (t) ≥ δ,
X
cji (1 − x∗j )x∗j ,
∀i ∈ Vr .
(33)
+
j∈Vr,i
+
j∈Vr,i
e
cji =
X
+
j∈Vr,i
(1 − x∗j )x∗j
cji
= 1,
(1 − x∗i )x∗i
∀i ∈ Vr .
(34)
er .
∀t ≥ 0 and ∀i → j ∈ E
Then, it is known from [26] that the transition matrix Φ(t) of
the linear system (38) converges to a rank one matrix:
lim Φ(t) = 1 · u>
Combining (31) and (33), we have that the coefficients e
cji , for
i → j ∈ Er , satisfy the following condition:
X
(39)
Now, consider the digraph induced by A(t): Let
With the notations above, we can express the evolution equations of yi , for i ∈ Vr , as follows:
X
ẏi = (1 − 2xi ) −yi +
e
cji yj , ∀i ∈ Vr .
(32)
(1 − x∗i )x∗i =
∀t ≥ 0 and ∀j → i ∈ Er .
t→∞
where 1 ∈ Rk is a vector of all ones, and u is a vector in the
(k − 1)-simplex. Hence, the solution y(t) also converges:
lim y(t) = hu, y(0)i · 1.
t→∞
Also, note that Qr is positive invariant for system (3). Thus,
if x(0) ∈ Qr , then x(t) ∈ Qr for all t ≥ 0; in particular, this
implies that
1 − 2xi (t) ≥ 1 − 2α > 0,
∀ t ≥ 0 and ∀ i ∈ Vr .
It suffices to show that the convergence of y(t) implies the
convergence of x0 (t) of system (30). To see this, first note
that Qr is positive invariant, and hence,
(35)
xi (t) ∈ [0, α],
∀t ≥ 0 and ∀i ∈ Vr .
for α defined in (14).
Combining (32) with (34) and (35), we recognize that the
dynamics of yi is a continuous-time consensus process scaled
by a time-varying, positive factor (1 − 2xi (t)). Now, define a
k-by-k infinitesimally stochastic matrix A(x0 ) = (aij (x0 )) as
follows: for i 6= j, let
(1 − 2xi )e
cji if j → i ∈ Er
0
aij (x ) :=
(36)
0
otherwise.
Since α < 1/2, the function (1 − x)x is invertible when
restricted to the closed interval [0, α]; indeed, (1 − x)x is
strictly monotonically increasing when restricted to [0, α].
Thus, if yi (t) = (1−xi (t))xi (t) converges, then so does xi (t),
for all i ∈ Vr . On the other hand, from Proposition 5, there is
only one equilibrium point x∗ in Qr ; we thus conclude that
x0 (t) converges to x0∗ , and hence x(t) converges to x∗ . This
completes the proof.
The diagonal entries of A(x0 ) are set to be −(1 − 2xi ), for
i ∈ Vr . With the matrix A(x0 ), we can rewrite system (32)
into the following matrix form:
D. Proof of Theorem 1
We provide in this subsection a complete proof of Theorem 1. We first recall that α is a scalar (defined in (14)) given
by
1
1
,
α= −
2 4(2n − 3)
ẏ = A(x0 )y.
(37)
With the definitions and notations above, we next prove
Proposition 6:
Proof of Proposition 6. We first establish the convergence of
system (37), and then show that it implies the convergence of
sysem (30).
Fix an initial condition x(0) ∈ Qr , and let x(t) be the
solution of system (3). Then, system (37) can be viewed
as a time-varying linear consensus process, with A(x(t))
(or simply written as A(t)) the time-varying infinitesimally
stochastic matrix, i.e.,
ẏ = A(t)y.
(38)
We also recall that Q, Qr are subsets of Sp[V ], given by
Q = {x ∈ Sp[V ] | xi ≤ α, ∀ i ∈ V }
and
Qr = Q ∩ Sp[Vr ].
To prove Theorem 1, we further need the following: for a
vector v = (v1 , . . . , vn ) ∈ Rn , let the one-norm of v be
defined as follows:
n
X
kvk1 :=
|vi |;
i=1
then, the following holds:
P
Lemma 6. Let x be in Q with i∈V
/ r xi ≤ , for ≤ 1/4;
then, there exists a point x0 in Qr such that
kx − x0 k1 ≤ 2.
Proof. We first show that
X
(α − xi ) ≥ 1/4
(40)
i∈Vr
This holds because
X
P
i∈Vr
xi ≤ 1, and
α ≥ 3α =
i∈Vr
3
5
3
−
≥ .
2 4(2n − 3)
4
i∈V
/ r
kx0 − xk1 = 2
xi ≤ 2,
i∈V
/ r
which completes the proof.
which implies that x ∈ U 0 ⊂ U .
We now show that any trajectory x(t) of system (3) intersects the open set U . Note that if this is the case, then by using
the arguments as before, we know that x(t) converges to x∗ .
To establish the statement, we first note that from Corollary 4,
there is an instant T1 such that x(t) ∈ Q for all t ≥ T1 . We
also note that from Proposition 4, there exists an instant T2
such that
X
xi (t) ≤ , ∀t ≥ T2 .
Thus,
if we let T := max{T1 , T2 }, then x(t) ∈ Q and
P
i∈V
/ r xi (t) ≤ for all t ≥ T , which implies that
x(t) ∈ U,
∀t ≥ T.
This completes the proof.
VI. C ONCLUSIONS
It then follows that
X
kx − x0 k ≤ 2 ≤ 0 ,
i∈V
/ r
The first (resp. last) inequality holds because |Vr | ≥ 3 (resp.
n ≥ 3). Since ≤ 1/4, from (40), there exists a vector x0 ∈ Qr
such that x0i ≥ xi for all i ∈ Vr , and moreover,
X
X
(x0i − xi ) =
xi .
i∈Vr
Then,
from Lemma 6, we have that for any x ∈ Q with
P
0
i∈V
/ r xi ≤ , there exists a x ∈ Qr such that
We are now in a position to prove Theorem 1:
Proof of Theorem 1. Parts 1 and 2 of Theorem 1 are established by Corollary 1 and Proposition 5, respectively. We prove
here Part 3.
First, we show that there is an open set U in Sp[V ]
containing Qr such that if x(0) ∈ U , then the solution x(t)
of system (3) converges to x∗ . By Proposition 5, the nonvertex equilibrium point x∗ is exponentially stable. Thus, there
exists an open neighborhood Ux∗ of x∗ in Sp[V ] such that if
x(0) ∈ Ux∗ , then x(t) converges exponentially fast to x∗ .
Now choose a point x ∈ Qr . From Proposition 6, we know
that there exists an instant T ≥ 0 such that if x(0) = x, then
x(t) ∈ Ux∗ for all t ≥ T . This, in particular, implies that
there exists an open neighborhood Ux of x in Sp[V ] such that
if x(0) ∈ Ux , then x(T ) ∈ Ux∗ . Now, define
U := ∪x∈Qr Ux .
Then, U is the desired open set of Qr in Sp[V ].
Next, we show that
P there exists an > 0 such that if x lies
in Q and satisfies i∈V
/ r xi ≤ , then x ∈ U . Since Qr is
a compact subset and U is an open set containing Qr , there
exists an 0 > 0 such that the 0 -neighborhood of Qr in Sp[V ]
is contained in U . Specifically, let
U 0 := {x ∈ Sp[V ] | kx − x0 k1 < 0 for some x0 ∈ Qr };
then, U 0 ⊂ U . Now, choose positive but sufficiently small
such that
≤ min{0 /2, 1/4}.
In this paper, we have introduced a continuous-time model
whereby a number of agents in a social network are able to
evaluate their self-appraisals over time in a distributed way. We
have shown that under a particular assumption (Assumption 1),
the solution of the system converges to the unique non-vertex
equilibrium point x∗ as long as the initial condition is not a
vertex, and moreover, x∗ is exponentially stable. This stable
equilibrium point can be interpreted as the steady state of the
self-appraisals of the agents, and we have related the value of
each x∗i to the values of cji , as described by Proposition 3 and
Corollary 3.
Future work may focus on the case where each cij is
time-variant. For example, we can assume that each cij also
depends on the self-appraisal of xj , e.g., how much opinion
agent i accepts from agent j now depends on how influential
agent j is in the social network. Also we note that this model
can be developed into many other interesting problems. For
example, a question related to sparse systems is that given
the underlying graph G, what is the collection of achievable
steady states x∗ by the choice of cij ’s? A similar question
has been asked and answered in the context of the consensus
process [31]. Another problem related to optimal control is
to assume that there is an agent i who is able to manipulate
her own weights cij , and we ask whether there is a choice of
these weights so that self-appraisal of agent i is maximized?
If further, we assume that there are two such players each of
whom is trying to maximize her own self-appraisal, then what
would be the strategy for each of the players to choose the
cij ’s? This list of examples indicates that the self-appraisal
model has a rich structure which can be investigated under
various assumptions and from different perspectives.
R EFERENCES
[1] M. H. DeGroot. Reaching a consensus. Journal of the American
Statistical Association, 69(345):118–121, 1974.
[2] E. M. Rogers. Diffusion of Innovations. The Free Press, New York, NY,
1995.
[3] T. Kuran. Chameleon voters and public choice. Public Choice, 53(1):53–
78, 1987.
[4] J. Liu, N. Hassanpour, S. Tatikonda, and A. S. Morse. Dynamic threshold
models of collective action in social networks. In Proceedings of the 51st
IEEE Conference on Decision and Control, pages 3991–3996, 2012.
[5] N. E. Friedkin and E. C. Johnsen. Social influence networks and opinion
change. Advances in Group Processes, 16(1):1–29, 1999.
[6] R. Hegselmann and U. Krause. Opinion dynamics and bounded confidence models, analysis, and simulation. Journal of Artificial Societies
and Social Simulation, 5(3), 2002.
[7] G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch. Mixing beliefs
among interacting agents. Advances in Complex Systems, 3(1-4):87–98,
2000.
[8] K. Sznajd-Weron and J. Sznajd. Opinion evolution in closed community.
International Journal of Modern Physics C, 11(6):1157–1165, 2000.
[9] P. Clifford and A. Sudbury. A model for spatial conflict. Biometrika,
60(3):581–588, 1973.
[10] R. A. Holley and T. M. Liggett. Ergodic theorems for weakly interacting
infinite systems and the voter model. The Annals of Probability,
3(4):643–663, 1975.
[11] D. Acemoglu, G. Como, F. Fagnani, and A. Ozdaglar. Opinion fluctuations and disagreement in social networks. Mathematics of Operations
Research, 38(1):1–27, 2013.
[12] E. Yildiz, A. Ozdaglar, D. Acemoglu, A. Saberi, and A. Scaglione.
Binary opinion dynamics with stubborn agents. ACM Transactions on
Economics and Computation, 1(4):19, 2013.
[13] J. Ghaderi and R. Srikant. Opinion dynamics in social networks: a local
interaction game with stubborn agents. In Proceedings of the 2013
American Control Conference, pages 1982–1987, 2013.
[14] V. D. Blondel, J. M. Hendrickx, and J. N. Tsitsiklis. On Krause’s
multi-agent consensus model with state-dependent connectivity. IEEE
Transactions on Automatic Control, 54(11):2586–2597, 2009.
[15] A. Mirtabatabaei and F. Bullo. Opinion dynamics in heterogeneous
networks: convergence conjectures and theorems. SIAM Journal on
Control and Optimization, 50(5):2763–2785, 2012.
[16] S. R. Etesami and T. Başar. Game-theoretic analysis of the HegselmannKrause model for opinion dynamics in finite dimensions. Automatic
Control, IEEE Transactions on, 60(7):1886–1897, July 2015.
[17] A. Mirtabatabaei, P. Jia, and F. Bullo. Eulerian opinion dynamics with
bounded confidence and exogenous inputs. SIAM Journal on Applied
Dynamical Systems, 13(1):425–446, 2014.
[18] P. Jia, A. Mirtabatabaei, N. E. Friedkin, and F. Bullo. Opinion dynamics
and the evolution of social power in influence networks. SIAM Review.
to be published.
[19] P. Jia, A. Mirtabatabaei, N. E. Friedkin, and F. Bullo. On the dynamics
of influence networks via reflected appraisal. In Proceedings of the 2013
American Control Conference, pages 1251–1256, 2013.
[20] V. D. Blondel, J. M. Hendrickx, and J. N. Tsitsiklis. Continuous-time
average-preserving opinion dynamics with opinion-dependent communications. SIAM Journal on Control and Optimization, 48(8):5214–5240,
2010.
[21] J. M. Hendrickx and A. Olshevsky. Symmetric continuum opinion dynamics: convergence, but sometimes only in distribution. In Proceedings
of the 52nd IEEE Conference on Decision and Control, pages 1989–
1994, 2013.
[22] N. E. Friedkin. A formal theory of reflected appraisals in the evolution
of power. Administrative Science Quarterly, 56(4):501–529, 2011.
[23] V. Gecas and M. L. Schwalbe. Beyond the looking-glass self: Social
structure and efficacy-based self-esteem. Social Psychology Quarterly,
pages 77–88, 1983.
[24] Z. Xu, J. Liu, and T. Başar. On a modified DeGroot-Friedkin model
of opinion dynamics. In Proceedings of the 2015 American Control
Conference, pages 1047–1052, 2015.
[25] X. Chen, J. Liu, Z. Xu, and T. Başar. Distributed evaluation and
convergence of self-appraisals in social networks. In Proceedings of
the 54th IEEE Conference on Decision and Control, 2015. to appear.
[26] L. Moreau. Stability of continuous-time distributed consensus algorithms. In Proceedings of the 43rd IEEE Conference on Decision and
Control, pages 3998–4003, 2004.
[27] Q. Zhu, H. Tembine, and T. Başar. Heterogeneous learning in zero-sum
stochastic games with incomplete information. In Proceedings of the
49th IEEE Conference on Decision and Control, pages 219–224. IEEE,
2010.
[28] I.M. Bomze. Lotka-volterra equation and replicator dynamics: new
issues in classification. Biological cybernetics, 72(5):447–453, 1995.
[29] W. Ren and R. W. Beard. Consensus seeking in multiagent systems
under dynamically changing interaction topologies. IEEE Transactions
on Automatic Control, 50(5):655–661, 2005.
[30] L. Farina and S. Rinaldi. Positive linear systems: theory and applications, volume 50. John Wiley & Sons, 2011.
[31] X. Chen, M.-A. Belabbas, and T. Başar. Consensus with linear objective
map. In Proceedings of the 54th IEEE Conference on Decision and
Control, 2015. to appear.
A PPENDIX
We prove here Proposition 1. We first have some definitions.
A polynomial m(x1 , . . . , xn ) is said to be a monomial if it
can be expressed as follows:
m(x1 , . . . , xn ) = Πni=1 xki i .
with ki ≥ 0 for all i = 1, . . . , n. The
Pndegree of the
monomial m(x1 , . . . , xn ) is defined to be i=1 ki . A polynomial p(x1 , . . . , xn ) can be uniquely expressed as a linear
combination of monomials. We say a monomial is contained in
p if the corresponding coefficient is nonzero. Furthermore, the
degree of p is defined to be the highest degree of a monomial
contained in p. With these definitions at hand, we have the
following result:
Lemma 7. Let x(0) be any initial condition of system (3)
(k)
in Sp[V ]. The time derivative xi (0) is a degree-(k + 1)
polynomial in xj (0), for j ∈ Si (k), and it does not contain
a constant term. Moreover, if Di (k) is nonempty for k ≥ 1,
then the following properties hold:
1) The only monomials in variables xj (0), for j ∈ Di (k),
(k)
contained in xi (t) are
xj (0) and x2j (0),
∀ j ∈ Di (k).
2) The coefficients of xj (0) and x2j (0), for j ∈ Di (k), in
(k)
xi (0) are αji and −αji , respectively.
(k)
Proof. In the proof, we will simply write xi by omitting the
argument. The proof will be carried out by induction on k.
Base case. We assume that k = 1; in this case, we have
X
(1)
xi = −(1 − xi )xi +
cji (1 − xj )xj .
j∈Vi+
We recall the fact that
Si (1) = Vi+ ∪ {i}
and
Di (1) = Vi+ ,
and the fact that
αji = cji ,
∀ j ∈ Vi+ .
It then follows that Lemma 7 holds for k = 1.
Inductive step. We assume that Lemma 7 holds for (k − 1),
(k)
and prove for k. We first show that xi is a degree-(k + 1)
polynomial in variables xj ’s, for j ∈ Si (k), and it does not
contain a constant term.
Label the elements of Si (k − 1) as i1 , . . . , is . By the
(k−1)
induction hypothesis, we can write xi
as follows:
(k−1)
xi
=
t
X
q=1
σq mq .
Each mq is a monomial in variables xi1 , . . . , xis , and σq is
the associated coefficient. The degree of mq is at most k. By
chain rule, we have
(k)
xi
=
t X
s
X
σq
q=1 p=1
∂mq
fi (x).
∂xip p
(41)
Note that each fip (x) (defined in (10)) is a degree-2 polynomial in variables xj , for j ∈ Sip (1), and it does not contain
a constant term. Also, note that since ip ∈ Si (k − 1), each
Sip (1) is a subset of Si (k). Thus, by following (41), we know
(k)
that (i) xi is a polynomial in variables xj , for j ∈ Si (k); (ii)
(k)
(k)
the degree of xi is at most (k + 1); and (iii) xi does not
contain a constant term. Further, a direct computation yields
(k)
that the monomial xk+1
is contained in xi with coefficient
i
(k)
k!. Thus, the degree of xi is (k + 1).
We now assume that Di (k) is nonempty, and establish
items 1 and 2 of the lemma. First, note that each ∂mq /∂xip
is a monomial in xj , for j ∈ Si (k − 1). Hence, if
(∂mq /∂xip )fip (x) contains a monomial in xj , for j ∈ Di (k);
then, we must have
m q = x ip
(42)
so that ∂mq /∂xip = 1, and moreover, fip (x) has to contain a
monomial in xj , for j ∈ Di (k). On the other hand, fip (x) is
a polynomial in variables xj , for j ∈ Sip (1). From Lemma 1,
if Sip (1) intersects Di (k), then
ip ∈ Di (k − 1).
(43)
(k)
Combining (42) and (43), we know that xi contains at most
xj and x2j , for j ∈ Di (k), as the monomials in xj , for j ∈
Di (k).
Now, fix a vertex j ∈ Di (k), and let σj0 and σj00 be the
(k)
coefficients of xj and x2j in xi . It suffices to show that
σj0 = αji
and
σj00 = −αji .
First, from Lemma 1, Vj− intersects Si (k − 1), and their
intersection lies in Di (k − 1). We thus define
W := Di (k − 1) ∩ Vj− .
Then, using the fact that (∂mq /∂xip )fip (x) contains the
monomials xj and x2j if and only if mq = xip and ip ∈ W ,
we obtain that
X
σj0 = −σj00 =
cjip σip ,
ip ∈W
(k−1)
where σip is the coefficient of xip in xi
. From the
induction hypothesis, we have σip = αip i , and hence
X
σj0 = −σj00 =
cjip αip i ,
ip ∈W
which is αji by (7). This completes the proof.
With Lemma 7 at hand, we now prove Proposition 1:
Proof of Proposition 1. We first prove for the case where
there does not exist a supporting path for i at x(0). The fact
that i ∈
/ Vr follows from Lemma 2. To see this, we assume
that i ∈ Vr . But then, Si = V , and hence, xj (0) = 0 for all
j ∈ V , which contradicts the fact that x(0) ∈ Sp[V ]. We now
show that
xj (t) = 0,
∀t ≥ 0 and ∀j ∈ Si .
(44)
Note that the dynamics of xj depends only on xk ’s for k ∈
Sj (1). Since j ∈ Si , we have Sj (1) ⊆ Si . This, in particular,
implies that the following system:
X
ẋj = −(1 − xj )xj +
ckj (1 − xk )xk , ∀j ∈ Si (45)
k∈Vj+
formed by xj ’s, for j ∈ Si , is an isolated system, independent
of xj 0 , for j 0 ∈
/ Si . Since there is no supporting path for i at
x(0), by Lemma 2, we have
xj (0) = 0,
∀j ∈ Si .
But then, following this initial condition, (44) is the unique
solution of system (45).
We now prove for the case when there exists a critical
supporting path for i at x(0) of length k. We first show that
(l)
(k)
xi (0) = 0 for all l < k, and then show that xi (0) > 0.
(l)
By Lemma 7, xi (0) is a polynomial in xj (0), for j ∈ Si (l),
and it does not contain a constant term. Since the length of a
critical supporting path for i at x(0) is k, we have xj (0) = 0
(l)
for all j ∈ Si (l). Thus, xi (0) = 0 for all l < k.
(k)
We now show that xi (0) > 0. Since there is a critical
supporting path for i at x(0), Di (k) is nonempty. Then, by
Lemma 7, we have
X
(k)
xi (0) =
αji (xj (0) − x2j (0)).
j∈Di (k)
Since x(0) is not a vertex of Sp[V ], we have xj (0) < 1 for
all j ∈ V . On the other hand, there exists at least a vertex
j ∈ Di (k) such that xj (0) > 0. This holds because there
exists a critical supporting path for i at x(0) of length k. It
(k)
then follows that xi (0) > 0, which completes the proof.
| 3cs.SY
|
arXiv:1303.4680v1 [math.AC] 19 Mar 2013
ON THE LINEARITY DEFECT OF THE RESIDUE FIELD
LIANA M. ŞEGA
Abstract. Given a commutative Noetherian local ring R, the linearity defect of a finitely generated R-module M , denoted ldR (M ), is an invariant that
measures how far M and its syzygies are from having a linear resolution. Motivated by a positive known answer in the graded case, we study the question
of whether ldR (k) < ∞ implies ldR (k) = 0. We give answers in special cases,
and we discuss several interpretations and refinements of the question.
Introduction
This paper is concerned with properties of the minimal free resolution of the
residue field k of a commutative Noetherian local ring R, and more precisely
with the observation that certain behavior of the tail of the resolution determines properties of its beginning. An instance of such behavior is displayed by
Castelnuovo-Mumford regularity: If A is a standard graded algebra over a field k
and regA (k) < ∞, then A is Koszul, cf. Avramov and Peeva [4].
We study an invariant related to regularity, defined for all local rings R: the
linearity defect of a finitely generated R-module M , denoted ldR (M ). This notion
was introduced by Herzog and Iyengar [8] and studied further by Iyengar and
g
Römer [9]. The definition is recalled in Section 1. We denote (−) the associated
graded objects with respect to the maximal ideal m of R, and we recall here that
ldR (M ) < ∞ if and only if there exists a syzygy N of M such that N g has a linear
resolution over Rg , and ldR (k) = 0 if and only if the algebra Rg is Koszul.
Linearity defect can also be defined for standard graded k-algebras. If A is a
standard graded k-algebra, then ldA (k) < ∞ implies regA (k) = 0, cf. [8, 1.12, 1.13],
hence A is Koszul. A natural question, raised in [8], is whether this property can
be extended to local rings (R, m, k):
Question 1. If ldR (k) < ∞, does it follow that ldR (k) = 0?
We give an interpretation of the linearity defect in terms of the maps
n
n+1
νin (M ) : TorR
) → TorR
i (M, R/m
i (M, R/m )
induced by the canonical surjection R/mn+1 → R/mn . More precisely, we show
n
in Theorem 2.2 that ldR (M ) < ∞ if and only if ν≫0
(M ) = 0 for all n > 0, and
n
n
ldR (M ) = 0 iff ν>0 (M ) = 0 for all n > 0. (The notation ν≫0
(M ) = 0 means
n
n
νi (M ) = 0 for all i ≫ 0 and the notation ν>0 (M ) = 0 means νin (M ) = 0 for all
i > 0.) Based on this interpretation, Question 1 can be refined as follows:
n
n
Question 2. Let n > 0. If ν≫0
(k) = 0, does it follow that ν>0
(k) = 0?
2000 Mathematics Subject Classification. Primary 13D07. Secondary 13D02.
Key words and phrases. linearity defect, Koszul local rings.
Research partly supported by NSF grant DMS-1101131 and Simons Foundation grant 20903.
1
2
L. M. ŞEGA
The maps ν 1 (k) are connected to the Yoneda algebra E = ExtR (k, k). Let R!
denote the subalgebra of E generated by Ext1R (k, k). A result of Roos shows that
1
1
ν>0
(k) = 0 if and only if E = R! and ν≫0
(k) = 0 if and only if E is finitely
!
generated as a module over R . When n = 1, Question 2 can be reformulated thus
as follows:
Question 3. If E is finitely generated as a module over R! , does it follow that
E = R! ?
As mentioned above, Question 1 has a a positive answer for standard graded
k-algebras. In this case, we show that Question 3 has a positive answer, as well;
see Corollary 5.7.
We further provide answers to these questions for a local ring (R, m, k), as indicated below:
(1) Question 1 has a positive answer under any of the following assumptions:
(a) m3 = 0 (Section 1);
(b) R is complete intersection and Rg is Cohen-Macaulay (Section 6);
(c) R is Golod and Rg is Cohen-Macaulay (Section 7).
(2) Question 2 has a positive answer under any of the following assumptions:
(a) R is complete intersection and n = 1 (Section 5);
(b) mn+2 = 0. (Section 7).
(3) Question 3 has a positive answer under any of the following assumptions:
(a) R is complete intersection (Section 5);
(b) E is generated over R! in degree 2 (Section 5).
The paper is organized as follows.
In Section 1 we mainly introduce terminology. In Section 2 we prove the interpretation of linearity defect ldR (M ) in terms of the maps νin (M ).
In Section 3 we note that ldR (k) = 0 implies ldR (mn ) = 0 for all n and we
provide a (possible) generalization: We show that ldR (mn ) ≤ ldR (k) for all n > 0.
We ask whether equality holds; note that a positive answer for n = 1 would give a
positive answer to Question 1.
1
In Section 4 we use a minimal Tate resolution of k to prove that ν≫0
(k) = 0
1
implies ν2 (k) = 0. The latter condition is shown in Section 5 to be equivalent to
b is a minimal Cohen presentation
the following quadratic property : If R ∼
= Q/a
with (Q, n, k) regular local and a ⊆ n2 , then n3 ∩ a ⊆ na.
Section 5 makes the connections with the Yoneda algebra. Section 6 is concerned
with complete intersection rings, and Section 7 with Artinian and Golod rings.
1. Linearity defect
In this section we establish notation, provide the definition of linearity defect,
and we introduce Question 1. Proposition 1.3 provides an easily available answer
in the radical cube zero case.
Throughout, (R, m, k) denotes a commutative Noetherian local ring; the notation
identifies m as the maximal ideal and k as the residue field. In addition, we assume
that m 6= 0. The embedding dimension of R, denoted edim R, is the minimal number
of generators of m, and the codimension of R, denoted codim R, is defined as the
number edim R − dim R.
LINEARITY DEFECT
3
If M is a finitely generated R-module, then the ith Betti number of M is the
number
βiR (M ) = rankk TorR
i (M, k) .
The Poincaré series of M over R is the formal power series
X
PR
βiR (M )ti ∈ Z[[t]] .
M (t) =
i≥0
The Hilbert series of M is the formal power series
X
HilbM (t) =
rankk (mn M/mn+1 M )tn ∈ Z[[t]] .
n≥0
We let R denote the associated graded ring and M g denote the associated
graded module with respect to m; that is,
M
M
Rg =
mn /mn+1 and M g =
mn M/mn+1 M .
g
n≥0
n≥0
1.1. Linearity defect. Let M be a finitely generated R-module and let
F =
∂
i
Fi → · · · → F0 → 0
· · · → Fi+1 −→
be a minimal free resolution of M , with differential ∂.
For any such F one constructs the complex
linR (F ) = · · · → Fi+1 g (−i − 1) → Fi g (−i) → · · · → F0 g → 0
with differentials induced from F . Herzog and Iyengar [8, 1.7] defined the linearity
defect of M to be the number:
ldR (M ) = sup{i ∈ Z | Hi (linR (F )) 6= 0} .
We make the convention that ldR (0) = 0.
If ldR (M ) = 0, we say, following [8], that M is a Koszul module. If k is Koszul,
we say that R is a Koszul ring. The connection with the classical Koszul algebra
notion, defined for standard graded rings, is as follows: R is a Koszul ring iff Rg is
a Koszul algebra. Also, M is Koszul iff M g has a linear resolution over Rg .
For an integer d, one has ldR (M ) ≤ d if and only if the dth syzygy module in a
minimal free resolution of M over R is Koszul.
If M is a Koszul R-module, then linR (F ) is a minimal free resolution of M g over
g
R . Since this resolution is minimal and Hilbert series are additive on short exact
sequences, we have:
(1.1.1)
Rg
PR
M (t) = PM g (t) =
HilbM (−t)
.
HilbR (−t)
Remark 1.2. Let A be a standard graded algebra over a field k, and let m denote
its maximal irrelevant ideal. When N is a finitely generated graded A-module, one
can define in the same manner the linearity defect ldA (N ) of N , by using a minimal
graded free resolution of N over A. All subsequent definitions and results can be
similarly adapted to the graded case.
As mentioned in the introduction, we are concerned with the following question:
Question 1. If ldR (k) < ∞ does it follow that R is Koszul?
4
L. M. ŞEGA
This question was posed in [8, 1.14], motivated by the fact that a positive answer
is available in the case of standard graded algebras; see [8, 1.13] for a proof of this
fact, and [8, 2.4] for a further discussion of the problem.
A positive answer in the case m3 = 0 is easily available.
Proposition 1.3. Assume m3 = 0. If there exists a finitely generated R-module
M such that ldR (M ) < ∞, then R is Koszul.
Proof. Since ldR (M ) < ∞, there exists a syzygy N in a minimal free resolution of
M such that ldR (N ) = 0. Since m3 = 0, one has that m2 N = 0. The Rg -module
N g then has a linear resolution, hence Rg is Koszul, according to Avramov et. al.
[3, Theorem 1.6]
2. The maps ν n
In this section we introduce the maps ν n and begin their study. The main result
is Theorem 2.2, which provides an interpretation of linearity defect in terms of
these maps, and motivates Question 2. A connection with Castelnouvo-Mumford
regularity is made in Proposition 2.7, based on the proof of results of Herzog and
Iyengar [8].
2.1. Let M be a finitely generated R-module and F a minimal free resolution of
M . For all integers n and i we consider the map
(2.1.1)
n+1
n
νin (M ) : TorR
) → TorR
i (M, R/m
i (M, R/m )
induced in homology by the canonical surjection R/mn+1 → R/mn .
When M = k, we drop the module argument: We set ν n = ν n (k).
Theorem 2.2. Let M be a finitely generated R-module. The following then hold:
(a) If i > 0 and F is a minimal free resolution of M , then Hi (linR (F )) = 0 if
n
and only if νi+1
(M ) = 0 = νin (M ) for all n > 0.
(b) ldR (M ) ≤ d if and only if νin (M ) = 0 for all i > d and all n ≥ 0.
(c) ldR (M ) = 0 if and only if νin (M ) = 0 for all i > 0 and all n ≥ 0.
Obviously, parts (b) and (c) are immediate consequences of part (a). The proof
of (a) will be given after some preliminaries.
n
If s is an integer such that νin = 0 for all i > s, we write ν>s
= 0. The notation
n
ν≫0 = 0 means that an integer s as above exists. Note that the statement of
Theorem 2.2 motivates Question 2 from the introduction.
We proceed now with some preparation for the proof of Theorem 2.2(a).
2.3. Let i > 0 and n ≥ 0. Computing homology in (2.1.1) by means of a minimal
free resolution F of M and using the natural isomorphism F ⊗R R/mk ∼
= F/mk F
n
for k = n, n + 1, the map νi (M ) can be realized as the map
Hi (F/mn+1 F ) → Hi (F/mn F )
induced in homology by the canonical surjection of complexes F/mn+1 F → F/mn F .
The following statements are thus equivalent:
(1) νin (M ) = 0;
(2) For every x ∈ Fi with ∂x ∈ mn+1 Fi−1 there exists y ∈ Fi+1 such that
x − ∂y ∈ mn Fi ;
(3) For every x ∈ Fi with ∂x ∈ mn+1 Fi−1 there exists u ∈ mn Fi such that
∂x = ∂u.
LINEARITY DEFECT
5
When n = 1, these conditions are also equivalent with
(2’) If x ∈ Fi satisfies ∂x ∈ m2 Fi−1 , then x ∈ mFi .
For the proof of the implication (2) =⇒ (3) take u = x − ∂y. For the proof of the
implication (3) =⇒ (2), take y such that ∂y = x − u, using the fact that Hi (F ) = 0.
Remark 2.4. Note that ν1n = 0 for all n ≥ 0. Indeed, let g1 , . . . , ge be a minimal
generating set for m and consider a minimal free resolution F of k with F0 = R
and F1 = Re . Let x ∈ F1 with ∂x ∈ mn+1 . Then there exist u1 , . . . , ue ∈ mn such
that ∂x = u1 g1 + · · · + ue ge . Let f1 , . . . , fe ∈ F1 such that ∂fi = gi for all i. If
u = u1 f1 + · · · + ue fe , then we have ∂x = ∂u and u ∈ mn F1 , hence condition (3) of
2.3 is satisfied.
Remark 2.5. Let ϕ : (R, m, k) → (R′ , m′ , k) be a flat homomorphism of local rings
such that ϕ(m) = m′ ; in particular, ϕ is faithfully flat. Let M be a finitely generated
R-module and set M ′ = M ⊗R R′ . When considering the maps ν n (M ′ ), we regard
M ′ as an R′ -module. Standard arguments show that for all integers n, i we have
νin (M ) = 0 iff νin (M ′ ) = 0. Also, if F is a minimal free resolution of M over R,
then F ′ = F ⊗R R′ is a minimal free resolution of F ′ over R′ and Hi (linR (F )) = 0
if and only if Hi (linR′ (F ′ )) = 0. In particular, ldR (M ) = ldR′ (M ′ ).
2.6. Notation. Set U = linR (F ). Note that U is a complex of graded Rg -modules,
and we will denote by Hi (U )j the jth graded component of the ith homology
module Hi (U ). For all integers j and s, one has that Hj (U )j+s is the homology of
the complex:
(2.6.1)
ms−1 Fj+1 /ms Fj+1 → ms Fj /ms+1 Fj → ms+1 Fj−1 /ms+2 Fj−1
with differentials induced by the differential ∂ of the complex F . We will use ∂
to denote the differential of this complex, as well. If x ∈ ms Fj we denote x the
image of x in ms Fj /ms+1 Fj . Note that ∂x ∈ ms+1 Fj−1 and, with the notational
convention above, one has ∂(x) = ∂x.
Proof of Theorem 2.2(a). In Remark 2.5, we may take R′ to be the m-adic completion of R. Consequently, we may assume that R is complete. Let U be as in
2.6.
Let i > 0. Assume that Hi (U ) = 0, hence Hi (U )i+s = 0 for all s. Let n ≥ 0. We
n
want to prove that νi+1
(M ) = 0 = νin (M ). To do so, we will verify condition (3)
of 2.3, with the appropriate indices.
n
To show νi+1
(M ) = 0, let x ∈ Fi+1 with ∂x ∈ mn+1 Fi . Note that ∂x is a cycle
in Ui,i+n+1 . Since Hi (U )i+n+1 = 0, we have that ∂x = ∂u0 for some u0 ∈ mn Fi+1 .
Then ∂(x− u0 ) ∈ mn+2 Fi . Using inductively this reasoning, we obtain that for each
k ≥ 0 there exists uk ∈ mn+k Fi such that ∂x − ∂u0 − ∂u1 − · · · − ∂uk ∈ mn+2+k Fi .
Since R is assumed complete, we can set u = u0 + u1 + · · · . Then u ∈ mn Fi+1 and
n
∂x − ∂u = 0. This shows that νi+1
(M ) = 0, using 2.3.
n
We now show νi (M ) = 0. We will prove this by induction on n ≥ 0. Obviously,
one has νi0 (M ) = 0. Assume now that n > 0 and νik (M ) = 0 for k < n. Let
x ∈ Fi with ∂x ∈ mn+1 Fi−1 . In particular, ∂x ∈ mn Fi−1 . Since we assumed
νin−1 (M ) = 0, condition (3) of 2.3 gives that there exists u′ ∈ mn−1 Fi such ∂x =
∂u′ . Considering u′ as an element in Ui,i+n−1 = mn−1 Fi /mn Fi , we have ∂u′ = 0.
Since Hi (U )i+n−1 = 0, there exists z ∈ mn−2 Fi+1 /mn−1 Fi+1 such that u′ = ∂z in
6
L. M. ŞEGA
mn−1 Fi /mn Fi , hence u′ − ∂z ∈ mn Fi . We take then u = u′ − ∂z and note that
u ∈ mn Fi and ∂x = ∂u. This shows νin (M ) = 0, using 2.3.
n
Assume now that νi+1
(M ) = νin (M ) = 0 for all n ≥ 0; we use again 2.3 to
translate below these conditions. Let s ≥ 0. We will show that Hi (U )i+s = 0. Let
x ∈ ms Fi /ms+1 Fi with ∂x = 0 in ms+1 Fi−1 /ms+2 Fi−1 . Thus ∂x ∈ ms+2 Fi−1 . Since
νis+1 (M ) = 0, there exists y ∈ ms+1 Fi and u ∈ Fi+1 such that x − y = ∂u. Then
s−1
∂u ∈ ms Fi and the fact that νi+1
(M ) = 0 shows that ∂u = ∂w with w ∈ ms−1 Fi+1 .
s+1
Thus x = y + ∂w, with y ∈ m Fi and w ∈ ms−1 Fi+1 , hence x = ∂w in U , showing
that Hi (U )i+s = 0.
In the graded case, the maps ν 1 are related to the Castelnuovo-Mumford regularity, denoted reg(−). Analyzing the proof [8, 1.12, 1.13], we identify below a
weaker hypothesis, in terms of the maps ν 1 . We repeat the main steps of the proofs
given in loc. cit., with the purpose of making sure that the weaker hypothesis is
indeed sufficient.
Proposition 2.7. Let A be a standard graded k-algebra, and N a graded finitely
1
(N ) = 0, then regA (N ) < ∞.
generated A-module. If ν≫0
1
In particular, if ν≫0 (k) = 0, then A is Koszul.
Proof. (Following [8, 1.12, 1.13].) We denote m the maximal irrelevant ideal of
A. Let F be a minimal free graded resolution of N over A. As described in [8,
1.11], U = linR (F ) decomposes as a direct sum of the linear strands complexes F r ,
defined in loc. cit., and regA (N ) is equal to sup{r ∈ Z | F r 6= 0}.
For each r, let n(r) denote the least integer such that (F r )n(r) 6= 0. If regA (N )
is infinite, then, as noted in the proof of [8, 1.12], there exists an infinite sequence
r1 < r2 < · · · with n(r1 ) < n(r2 ) < · · · . Since (F rs )n = 0 for n < n(rs ) and F rs is
minimal, it follows that there exists a nonzero element of bidegree (n(rs ), n(rs )+rs )
in F rs ⊗k. This element corresponds to a nonzero element of bidegree (n(rs ), n(rs ))
in H(U ). In particular, there exists an element xs in Fn(rs ) r mFn(rs ) such that
1
∂xs ∈ m2 Fn(rs )−1 , and this shows, using 2.3(2’), that νn(r
(N ) 6= 0 for each s.
s)
1
This contradicts the assumption that ν≫0 (N ) = 0.
For the last statement, one uses the result of Avramov and Peeva [4, (2)]: If
regA (k) < ∞, then A is Koszul.
Applying Theorem 2.2, one obtains the statement of [8, 1.13]:
Corollary 2.8. Let A be a standard graded algebra. If ldA (k) < ∞, then A is
Koszul.
3. The linearity defect of powers of the maximal ideal
The powers of the maximal ideal and, more generally, modules of the form mn M
with n > 0 and mn M 6= 0, are known to share some of the asymptotic properties of
the residue field; for example, they have the same complexity. With this thought in
mind, we proceed to investigate the connection between ldR (k) and ldR (mn ). We
prove in Proposition 3.2 an inequality between these numbers.
Remark 3.1. If ldR (k) = 0, then ldR (mn ) = 0 for all n.
Indeed, if ldR (k) = 0 then R is Koszul and Rg is a Koszul algebra. Then (mg )≥n
has a linear resolution over Rg by [9, 5.4], hence ldR (mn ) = 0.
LINEARITY DEFECT
7
The next proposition provides a possibly more general result.
Proposition 3.2. ldR (mn ) ≤ ldR (k) for all n.
Proof. Given a finite R-module M , recall that
n+1
n
νin (M ) : TorR
) → TorR
i (M, R/m
i (M, R/m )
is induced by the surjection R/mn+1 → R/mn . We (ab)use the same notation for
the avatar of this map which arises by switching the module entries:
n+1
n
νin (M ) : TorR
, M ) → TorR
i (R/m
i (R/m , M ) .
We set M = M/mM .
Let i > 1 and p, q ≥ 0. Consider the exact sequence
0 → mn → R/mn+1 → R/mn → 0
for n = p and n = q and the induced commutative diagram with exact rows and
columns.
νip (mq )
R
q
TorR
i (m , mp+1 )
/ TorR (mq , Rp )
i
m
γ
R
R
TorR
i ( mq+1 , mp+1 )
∂
νip (R/mq+1 )
R
R
/ TorR
i ( mq+1 , mp )
R
R
TorR
i ( mq , mp+1 )
α
R
p
/ TorR
i−1 ( mq+1 , m )
q
(mp )
νi−1
νiq (R/mp )
νiq (R/mp+1 )
/ TorR (mq , mp )
i−1
νip (R/mq )
R
R
/ TorR
i ( mq , mp )
β
R
p
/ TorR
i−1 ( mq , m )
n
Set ldR (k) = d and assume i > d + 1. Then νi−1
= νin = 0 for all n ≥ 0,
q
p
q
by Theorem 2.2. Since m and m are k-vector spaces, we also have νi−1
(mp ) =
p
q
νi (m ) = 0.
Claim 1. If νiq (R/mp+1 ) = 0, then νip (R/mq+1 ) = 0.
Claim 2. If νip (R/mq ) = 0, then νiq (R/mp ) = 0.
Claim 1 is obtained by analyzing the upper left square of the diagram: If
νiq (R/mp+1 ) = 0, then γ is surjective. Since νip (mq ) = 0, the commutativity of
the square gives that νip (R/mq+1 ) ◦ γ = 0, hence νip (R/mq+1 ) = 0.
Claim 2 is obtained by analyzing the lower right square of the diagram: If
q
νip (R/mq ) = 0, then the map β is injective. Since νi−1
(mp ) = 0, the commutativity
q
q
p
of the square yields that β ◦ νi (R/m ) = 0, hence νi (R/mp ) = 0.
Hence these claims hold for all p, q ≥ 0. Note that νip (R/mq+1 ) = 0 for q = 0,
since νip = 0. We then apply Claim 2 and we get νi1 (R/mp ) = 0. Then we apply
Claim 1 and we get νip−1 (R/m2 ) = 0. Then Claim 2 yields νi2 (R/mp−1 ) = 0. An
inductive repeated use of Claim 1 and Claim 2 then yields that νip (R/mq ) = 0 for
all p, q ≥ 0 and all i > d + 1. Using Theorem 2.2 we conclude ldR (R/mp ) ≤ d + 1
for all p ≥ 0.
If n > 0 and mn 6= 0, then mn is a first syzygy in a minimal free resolution of
R/mn and we conclude ldR (mn ) ≤ max{0, ldR (R/mn ) − 1} ≤ d = ldR (k).
Question 3.3. Is it true that ldR (mn ) = ldR (k) for all n > 0?
Note that for n = 1 this question subsumes Question 1 in the introduction, for
if ldR (k) < ∞ then ldR (m) = ldR (k) − 1, unless ldR (k) = 0.
8
L. M. ŞEGA
4. Tate resolutions and the map ν 1
In this section we are concerned with Question 2 in the introduction, for n = 1:
1
1
If ν≫0
= 0, does it follow that ν>0
= 0? Proposition 4.1, whose proof uses Tate
resolutions, gives a partial answer.
Recall that ν1n = 0 for all n ≥ 0, as noted in Remark 2.4. In particular, the
condition ν11 = 0 holds for all local rings (R, m, k). However, as we will see in
Section 5, the condition ν21 = 0 does not hold for all rings.
1
Proposition 4.1. If ν2n
= 0 for some n > 0, then ν21 = 0.
Some preliminaries are needed.
4.2. Divided powers and Tate resolutions. A system of divided powers on a graded
R-algebra A is an operation that for each j ≥ 1 and each i ≥ 0 assigns to every
a ∈ A2i an element a(j) ∈ A2ij , subject to certain axioms; cf. [7, 1.7.1]. A DGΓ Ralgebra is a DG R -algebra A with divided powers compatible with the differential ∂
of A: ∂(a(j) ) = ∂(a)a(j−1) . We denote by |x| the homological degree of an element
x.
Given a set x = {xi | |xi | ≥ 1}, we let Ahxi denote a DGΓ algebra with
M
M
R
Rx
Rx
⊗
Γ
A ⊗ R ΛR
R ∗
∗
x∈x
|x| even
x∈x
|x| odd
as underlying graded algebra and differential compatible with that of A and the
divided powers of x ∈ x.
A Tate resolution of a surjective ring homomorphism R → T is a quasi-isomorphism Rhxi → T , where x = {xi }i>1 and |xj | ≥ |xi | ≥ 1 holds for all j ≥ i ≥ 1.
Such a resolution always exists: see [7, 1.2.4]. Furthermore, such a resolution
can be chosen minimally, as described in [1, Construction 6.3.1]; for T = k this
construction yields a minimal free resolution of k, which we shall call a minimal
Tate resolution of k over R.
For each sequence of integers µ < · · · < ν and each sequence of integers iµ ≥
ν)
1, . . . , iν ≥ 1, the product xµ(iµ ) . . . x(i
is called a normal Γ-monomial; 1 is considν
ered to be a normal monomial. The normal monomials form the standard basis of
Rhxi, considered as a graded algebra over R.
For each µ > 0 we set
Iµ = {i ≥ 0 : |xi | = µ} .
Proof of Proposition 4.1. Let F = Rhxi be a minimal Tate resolution of k. Let
B be the standard basis of F over R, as described above. We will interpret the
vanishing of the maps νi1 in terms of condition (2’) in 2.3, which states that νi1 = 0
if and only if the following holds: If A ∈ Fi satisfies ∂A ∈ m2 Fi−1 , then A ∈ mFi .
1
Assume that ν2n
= 0 and let A ∈ F2 such that ∂A ∈ m2 F1 . Expressing this
element in terms of the basis B, we have:
X
X
A=
al xl +
bij xi xj
l∈I2
i,j∈I1 ,i<j
with al , bij ∈ R for l ∈ I2 and i, j ∈ I1 with i < j. Since A(n) ∈ F2n satisfies
1
∂A(n) = ∂A · A(n−1) ∈ m2 F2n−1 , the hypothesis that ν2n
= 0 yields A(n) ∈ mF2n .
LINEARITY DEFECT
9
(n)
The coefficient of xl in the expression of A(n) in terms of the basis B is anl , hence
anl ∈ m and thus al ∈ m. Consider now the element:
X
X
al xl =
bij xi xj
A′ = A −
i,j∈I1 ,i<j
l∈I2
′
To show A ∈ mF , it suffices to show A ∈ mF . Note that ∂A′ ∈ m2 F , since
∂A ∈ m2 F and al ∈ m.
If i > j, we set bij = bji . We compute next ∂A′ , and we note that for each xj
with j ∈ I1 , the coefficient of xj in ∂A′ is
X
cj =
±bij ∂xi
i∈I1 ,i6=j
Since {xj }j∈I1 is a basis for F1 and ∂A′ ∈ m2 F , we conclude cj ∈ m2 . Recall that
F is a minimal resolution of k. In particular, {∂xi }i∈I1 is a minimal generating set
for m. We conclude that bij ∈ m, hence A′ ∈ mF and thus A ∈ mF .
The next result will be needed later.
b = Q/a be a minimal Cohen
Lemma 4.3. Let R′ = R/I, with I ⊆ m2 . Let R
b′ = Q/b
presentation of R, with (Q, n, k) a regular local ring and a ⊆ n2 , and write R
2
for some b with a ⊆ b ⊆ n .
R
R′
Then the map Torϕ
2 (k, k) : Tor2 (k, k) → Tor2 (k, k) induced by the projection
ϕ : R → R′ is injective if and only if a ∩ nb ⊆ na.
Some preliminaries are needed for the proof.
4.4. Let A be a DGΓ algebra and let A>0 denote the ideal of elements of positive
degree. The module of indecomposables of A is the quotient of A>0 by the submodule generated by all elements of the form uv with u, v ∈ A>0 and w(n) with
w ∈ A2i , for n ≥ 0, i > 0. We denote by π∗ (R) the module of Γ-indecomposables
R
of TorR
∗ (k, k), where the DGΓ algebra structure on Tor∗ (k, k) is induced from a
minimal Tate resolution of k. A surjective homomorphism ϕ : R → R′ of local rings
induces canonically a map π∗ (ϕ) : π∗ (R) → π∗ (R′ ).
Proof of Lemma 4.3. The injectivity of the map Torϕ
2 (k, k) is invariant under completion, hence we may assume R = Q/a and R′ = Q/b.
R
Note that π1 (ϕ) = Torϕ
1 (k, k) is an isomorphism, since both π1 (R) = Tor1 (k, k)
′
2
and π1 (R′ ) = TorR
1 (k, k) can be canonically identified with n/n .
In view of [2, Corollary 1.3(b)-(c)], the kernel of the map Torϕ
2 (k, k) can be
identified with the kernel of the map π2 (ϕ). The proof of [7, Proposition 3.3.4]
canonically identifies π2 (R) with a/na and π2 (R′ ) with b/nb. Thus the map π2 (ϕ)
can be canonically indentified with the map a/na → b/nb induced by the inclusion
a ⊆ b. The kernel of this map is (a ∩ nb)/na, and the conclusion follows.
5. Finite generation of the Yoneda algebra over the Koszul dual
We consider now the graded algebra with Yoneda product E = ExtR (k, k). Let
R! denote the k-subalgebra of E = ExtR (k, k) generated by its elements of degree
1; this is sometimes referred to as the Koszul dual of R. Note that the Yoneda
product gives E a structure of right module over R! .
In this section we consider the following:
10
L. M. ŞEGA
Question 3. If E is finitely generated as a right module over R! , does it follow that
E = R! ?
As discussed after Theorem 5.5, Question 3 arises by interpreting Question 2 in
view of the following observation:
Remark 5.1. Let i > 0. Then νi1 = 0 if and only if the Yoneda multiplication map
E i−1 ⊗ E 1 → E i is surjective.
Indeed, note that νi1 = 0 if and only if HomR (νi1 , k) = 0 if and only if the map
ExtiR (k, k) → ExtiR (R/m2 , k) induced by the projection R → R/m2 is zero. The
statement then follows by applying a result of Roos [12, Corollary 1].
For i = 1, we recover the fact that ν11 = 0.
Let ρi : ExtiR/m2 (k, k) → ExtiR (k, k) denote the graded algebra map induced by
the canonical projection R → R/m2 .
Lemma 5.2. Let i ≥ 0. Consider the following statements:
(a) ρi is surjective;
(b) ρi+1 is surjective;
1
(c) νi+1
= 0.
Then: (b) implies (c); (a) and (c) imply (b).
In particular, ν21 = 0 if and only if ρ2 is surjective.
Proof. Consider the commutative diagram:
Exti+1
R/m2 (k, k)
O
ρi+1
/ Exti+1 (k, k)
R O
∼
=
ExtiR/m2 (k, k) ⊗ Ext1R/m2 (k, k)
ρi ⊗ρ1
/ ExtiR (k, k) ⊗ Ext1R (k, k)
where the vertical maps are Yoneda products. Note that the left vertical map is
an isomorphism, because ExtR/m2 (k, k) is the free tensor algebra on Ext1R/m2 (k, k),
cf. [12, Corollary 3]. Also, note that ρ1 is an isomorphism. The conclusion then
follows from the commutativity of the diagram and 5.1.
We can reformulate Question 2 in the introduction, for n = 1, in terms of the
maps ρ:
Question 5.3. If ρ≫0 is surjective, does it follow that ρ>0 is surjective?
b∼
Let R
= Q/a be a minimal Cohen presentation of R, with (Q, n) a regular local
ring and a ⊆ n2 . Noting that the left-hand side is independent on the choice of the
presentation, we set:
s(R) = inf{i ≥ 1 | a ∩ ni+2 ⊆ na}
Lemma 5.4. ν21 = 0 if and only if s(R) = 1.
Proof. As proved above, ν21 = 0 if and only if ρ2 is surjective. The induced map
HomR (ρ2 , k) can be indentified with the map
R/m2
R
Torϕ
2 (k, k) : Tor2 (k, k) → Tor2
2
(k, k)
induced by the surjection ϕ : R → R/m . Hence ρ2 is surjective if and only if
Torϕ
2 (k, k) is injective. Apply then Lemma 4.3.
LINEARITY DEFECT
11
We are now ready to translate information about the maps ν 1 in terms of the
Yoneda algebra. When we talk about finite generation of E over R! , we mean finite
generation as a right R! -module.
Theorem 5.5. Let r ≥ 1. The following implications hold:
(5)
1
ν>0
= 0 ey
(3)
(1)
1
ν>r
= 0 ey
(6)
E
is
generated
over R!
%9
by elements of degree r
(2)
ν21
%9 E = R!
(4)
= 0 ey
(7)
%9 s(R) = 1
Proof. The equivalences (5) and (6) are given by 5.1. The implication (2) is given
by Proposition 4.1. The equivalence (7) is given by Lemma 5.4.
In view of the implications (5) and (6) in the Theorem, observe that Question
2 in the introduction can be reformulated, for n = 1, as follows: Is the implication
(3) reversible? This is in fact Question 3. We give below some answers.
Corollary 5.6. If ExtR (k, k) is generated over R! by its elements of degree 2, then
ExtR (k, k) = R! .
Proof. Recall that ν11 = 0. The implication (2) in Theorem 5.5 shows then that
1
1
= 0. The conclusion is then given by the equivalences
ν>2
= 0 if and only if ν>0
(5) and (6).
Corollary 5.7. Let A be a standard graded k-algebra. If ExtA (k, k) is finitely
generated over A! , then ExtA (k, k) = A! and A is Koszul.
Proof. Use Proposition 2.7 and the graded version of the equivalence (6) in the
Theorem to conclude that ExtA (k, k) = A! . This is a known characterization of
Koszul algebras.
b ∼
We say that R is complete intersection if R
= Q/(f1 , · · · , fc ) with (Q, n, k) a
regular local ring and f = f1 , . . . , fc is a regular sequence. Note that c = codim R.
Corollary 5.8. Assume R is complete intersection. If ExtR (k, k) is finitely generated over R! , then ExtR (k, k) = R! .
Proof. Assume R is complete intersection. In this case, the equality s(R) = 1 is
equivalent to ExtR (k, k) = R! ; this can be seen from [13, 5.3, 5.4(2)]. In particular,
all implications in the diagram in Theorem 5.5 are reversible.
We now record some consequences in terms of the linearity defect, which follow
immediately from the results above and Theorem 2.2.
Corollary 5.9. If ldR (k) < ∞, then s(R) = 1.
Corollary 5.10. Assume that ldR (k) < ∞. If either ldR (k) ≤ 2 or R is complete
intersection, then ExtR (k, k) = R! .
12
L. M. ŞEGA
6. Complete intersection rings
In this section we answer positively Question 1 for complete intersection rings,
under the additional assumption that Rg is Cohen-Macaulay.
6.1. Set dim R = d. If M is a finitely generated R-module of dimension d, recall,
cf. [6, 4.1.8, 4.1.9], that
(6.1.1)
HilbM (t) =
QM (t)
(1 − t)d
for some polynomial QM (t) ∈ Z[t] with QM (1) 6= 0. We define the multiplicity
of M , denoted e(M ), by setting e(M ) = QM (1). We set e(M ) = 0 whenever
dim M < d. With this definition, e(−) is exact on short exact sequences, see [6,
4.6.7].
Lemma 6.2. Let R be a d-dimensional Cohen-Macaulay local ring. Assume that
PR
M (t) = A(t)/B(t) with A(t), B(t) relatively prime polynomials in Z[t].
If ldR (M ) < ∞, then B(−1) 6= 0 and e(M )/e(R) = A(−1)/B(−1).
Proof. Set βi = βiR (M ). Let K be a Koszul syzygy of M so that depth K ≥
depth R. Since R is Cohen-Macaulay, we have that dim K = d.
We have an exact sequence:
(6.2.1)
0 → K → Rβn → Rβn−1 → · · · → Rβ1 → Rβ0 → M → 0 .
Since K is Koszul, (1.1.1) gives
PR
K (t) =
(6.2.2)
HilbK (−t)
.
HilbR (−t)
As recalled in (6.1.1), we also have
(6.2.3)
HilbK (t) =
QK (t)
(1 − t)d
and
HilbR (t) =
QR (t)
(1 − t)d
for QR (t), QK (t) ∈ Z[t] with QK (1) = e(K) and QR (1) = e(R).
The hypothesis and the choice of K give that
A(t)
R
n
n+1
= PR
.
(6.2.4)
M (t) = β0 + β1 t + · · · + βn t + PK (t) · t
B(t)
plugging in (6.2.2) and (6.2.3) into (6.2.4), we have:
(6.2.5)
QK (−t) n+1
A(t)
= β 0 + β 1 t + · · · + β n tn +
·t
.
B(t)
QR (−t)
A multiplicity count in the exact sequence (6.2.1) gives:
(6.2.6)
(β0 − β1 + · · · + (−1)n βn )e(R) = e(M ) − (−1)n+1 e(K)
We then plug in t = −1 into (6.2.5). Note that QK (1) = e(K) and QR (1) = e(R).
We then use (6.2.6) in order to conclude:
e(M ) − (−1)n+1 e(K) QK (1)
A(−1)
=
+
· (−1)n+1
B(−1)
e(R)
QR (1)
e(M )
e(M ) − (−1)n+1 e(K) e(K)
+
· (−1)n+1 =
=
e(R)
e(R)
e(R)
LINEARITY DEFECT
13
A sequence x = x1 , . . . , xm is said to be strictly regular if the initial forms
x∗1 , . . . , x∗m form a regular sequence in Rg . A strictly regular sequence is, in particular, a regular sequence in R.
Lemma 6.3. If x ∈ m r m2 is a strictly regular sequence on R, then ldR (k) =
ldR/(x) (k).
Proof. We may assume that x consists of a single element x. Since x is strictly
regular, we have (mn+1 : x) ⊆ mn for all n. Then apply [13, 8.7] together with
Theorem 2.2.
Remark 6.4. Let R be a ring of dimension d such that Rg is Cohen-Macaulay.
Consider the ring R′ = R[t]m[t] . The residue field of R′ is infinite and the natural
map R → R′ is faithfully flat. We can then choose a strictly regular sequence
g = g1 , . . . , gd such that the length of R′′ = R′ /(g) is equal to e(R). Note that
edim R′′ = codim R.
Since the map of local rings ϕ : (R, m, k) → (R′ , m′ , k ′ ) is faithfully flat, with
ϕ(m) = m′ , we have that ldR (k) = ldR′ (k). Furthermore, Lemma 6.3 yields that
ldR (k) = ldR′′ (k).
If R is complete intersection of codimension c, then it is known e(R) ≥ 2c , cf. [5,
§7, Proposition 7]. We say that R has minimal multiplicity if e(R) = 2c .
Theorem 6.5. If R is a complete intersection, then the following statements are
equivalent:
(a) ldR (k) = 0;
(b) R has minimal multiplicity.
Furthermore, if Rg is Cohen-Macaulay, then they are also equivalent to
(c) ldR (k) < ∞.
Remark 6.6. If R is a d-dimensional complete intersection ring of embedding dimension e and codimension c, then d = e − c and a result of Tate and Zariski, see
for example [7, 3.4.3], gives
(6.6.1)
PR
k (t) =
(1 + t)d
(1 + t)e
=
.
(1 − t2 )c
(1 − t)c
Proof. (a) =⇒ (b): If ldR (k) = 0, then Rg is a Koszul ring. We use then (1.1.1)
and (6.1.1) to conclude
PR
k (t) =
1
HilbR (−t)
=
(1 + t)d
.
QR (−t)
Comparing this with the formula (6.6.1), it follows that QR (t) = (1 + t)c , hence
e(R) = QR (1) = 2c .
(b) =⇒ (a): If R is a complete intersection of minimal multiplicity, then Rg is
a complete interesection of quadrics, hence Rg is Koszul.
Obviously, (a) =⇒ (c).
Assume now that Rg is Cohen-Macaulay.
(c) =⇒ (b): Note that we may replace R with the ring R′′ of Remark 6.4. We
may assume thus that d = 0.
1
Then (6.6.1) gives PR
. Using Lemma 6.2 with M = k, we conclude
k (t) =
(1 − t)c
c
e(R) = 2 , hence R has minimal multiplicity.
14
L. M. ŞEGA
7. Artinian and Golod rings
In this section we use a result of Martsinkovsky to provide more evidence for
Question 2 in the case of Artinian rings (Theorem 7.1). We also settle Question 1
in the case of Golod rings R with Rg Cohen-Macaulay, using results of Avramov
and Levin (Theorem 7.2).
n−1
n−1
Theorem 7.1. Assume R is Artinian with mn+1 = 0. If ν≫0
= 0, then ν>0
= 0.
Proof. For integers i, s denote
γis : ExtiR (ms−1 , k) → ExtiR (ms , k)
the map induced in cohomology by the inclusion ms ֒→ ms−1 .
If i > 1 and s > 0, note that νis = 0 if and only if HomR (νis , k) = 0. Using the
a
natural isomorphisms ExtiR (R/ma , k) ∼
= Exti−1
R (m , k) for a = s, s + 1, we conclude
s+1
that νis = 0 if and only if γi−1 = 0.
n−1
n
Let j > 0 and assume that i large enough so that νi+j+1
= 0, thus γi+j
= 0.
Consider now the commutative diagram below, where where α = E i ⊗ γjn and
n
β = γi+j
= 0, the vertical arrows are given by the Yoneda product, and the
isomorphisms are due to the fact that mn is a k-vector space, since mn+1 = 0. The
rightmost square is induced by choosing a certain projection Hom(mn , k) → k; the
choice of this projection will be discussed later.
α
/ E i ⊗ Extj (mn , k)
R
β=0
/ Exti+j (mn , k)
R
E i ⊗ ExtjR (mn−1 , k)
n−1
Exti+j
(m
, k)
R
∼
=
/ E i ⊗ E j ⊗ Hom(mn , k)
/ Ei ⊗ Ej
/ E i+j ⊗ Hom(mn , k)
/ E i+j
∼
=
n−1
Assume that νj+1
6= 0, or, equivalently, that γjn 6= 0. There exists thus an element in ExtjR (mn−1 , k) whose image θ in ExtjR (mn , k) under γjn is non-zero. The
commutativity of the left square yields that the Yoneda product ϕθ is zero for all
ϕ ∈ Ei.
Let θe denote the image of θ in E j ⊗ Hom(mn , k). The commutativity of the
middle square shows that ϕθe = 0 for all ϕ ∈ E i . Since θe 6= 0, we can now construct
the right square in the diagram by choosing a projection Hom(mn , k) → k in such a
∼
=
way that the image of θe under the induced map E j ⊗ Hom(mn , k) → E j ⊗ k −
→ Ej
remains nonzero. Let θ denote this image.
The commutativity of the right square shows then that ϕθ = 0 for all ϕ ∈ E i .
The element θ of E is thus annihilated by all elements of E of sufficiently large
degree. This is a contradiction, according to the proof of [11, Theorem 6].
n−1
The contradiction shows that νj+1
= 0 for all j > 0. Recalling that ν1n−1 = 0
n−1
by 2.4, we conclude that ν>0
= 0.
If R is a Cohen-Macaulay ring, then one has an inequality codim R ≤ e(R) − 1.
If equality holds, we say that R has minimal multiplicity (as a Cohen-Macaulay
ring). We talked earlier about minimal multiplicity for complete intersections, and
one should distinguish between the two notions. In particular, note that a complete
intersection ring R has minimal multiplicity as a Cohen-Macaulay ring only when
codim R ≤ 1.
LINEARITY DEFECT
15
Besides complete intersections, Golod rings constitute another major class of
rings for which the homological behavior of modules is fairly well understood. Since
the definition of such rings is somewhat technical, we refer to [1, §5] for the definition
and properties.
Theorem 7.2. If R is Golod and Rg is Cohen-Macaulay, then the following statements are equivalent:
(a) ldR (k) = 0;
(b) ldR (k) < ∞;
(c) R has minimal multiplicity.
Remark 7.3. Assume that R is as in the hypothesis of the Corollary, and let R′
and R′′ as in Remark 6.4. Since ϕ : (R, m, k) → (R′ , m′ , k ′ ) is faithfully flat, with
g
ϕ(m) = m′ , note that R′ is Golod and R′ Cohen-Macaulay. Also, since the strictly
regular sequence g is contained in m r m2 , note that R′′ is Golod, as well, see [1,
5.2.4]
Proof. Using Remarks 6.4 and 7.3, we may assume that R is Artinian.
(c) =⇒ (a): Since R is Artinian of minimal multiplicity, we have m2 = 0. One
has R ∼
= Rg and Rg is obviously Koszul.
(b) =⇒ (c): Let n be such that mn+1 = 0 and mn 6= 0. By Theorem 7.1 and
n−1
Theorem 2.2, we have that ν>0
= 0. If n ≥ 2, then the natural map R → R/mn
is Golod (cf. Levin [10, 3.15]) and thus small (cf. Avramov [2, 3.5]). Since the only
small ideal of a Golod Artinian ring is (0), see [2, 4.7], we must have mn = 0, a
contradiction.We conclude n ≤ 1, hence m2 = 0.
Since (a) obviously implies (b), the proof is completed.
References
[1] L. L. Avramov, Infinite free resolutions, Six lectures on commutative algebra (Bellaterra,
1996), Progress in Math. 166, Birkhäuser, Basel, 1998; 1–118.
[2] L. L. Avramov, Small homomorphisms of local rings, J. Algebra 50 (1978), 400–453.
[3] L. L. Avramov, S. B. Iyengar, L. M. Şega, Short Koszul modules, J. Commut. Algebra 2
(2010), 249–279.
[4] L. L. Avramov, I. Peeva, Finite regularity and Koszul algebra, Amer. J. Math. (2001), 275–
281.
[5] N. Bourbaki, Algèbre commutative. Chapitre VIII: Dimension, Masson, Paris, 1983.
[6] W. Bruns, J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics
39, Cambridge University Press, Cambridge, 1993.
[7] T. H. Gulliksen, G. Levin, Homology of local rings, Queen’s Papers Pure Appl. Math. 20,
Queen’s Univ., Kingston, ON (1969).
[8] J. Herzog, S. Iyengar, Koszul modules, J. Pure Appl. Algebra 201 (2005), 154–188.
[9] S. Iyengar, T. Römer, Linearity defects of modules over commutative rings, J. Algebra 322
(2009), 3212–3237.
[10] G. Levin, Local rings and Golod homomorphisms, J. Algebra 37 (1975), 266-289.
[11] A. Martsinkovsky, A remarkable property of the (co)syzygy modules of the residue field of a
nonregular local ring, J. Pure Appl. Algebra 110 (1996), 9–13.
[12] J.-E. Roos, Relations between Poincaré-Betti series of loop spaces and of local rings,
Séminaire d’Algèbre Paul Dubreil 31ème année (Paris, 1977–1978), Lecture Notes in Math.
740, Springer, Berlin, 1979; 285–322.
[13] L. M. Şega, Homological properties of powers of the maximal ideal of a local ring 241 (2001),
827–858.
Liana M. Şega, Department of Mathematics and Statistics, University of Missouri,
Kansas City, MO 64110, U.S.A.
E-mail address: [email protected]
| 0math.AC
|
Minimax Estimation of Kernel Mean Embeddings
Minimax Estimation of Kernel Mean Embeddings
Ilya Tolstikhin
[email protected]
†
arXiv:1602.04361v2 [math.ST] 31 Jul 2017
Department of Empirical Inference
Max Planck Institute for Intelligent Systems
Spemanstraße 38, Tübingen 72076, Germany
Bharath K. Sriperumbudur
[email protected]
Department of Statistics
Pennsylvania State University
University Park, PA 16802, USA
Krikamol Muandet†
[email protected]
Department of Mathematics
Faculty of Science, Mahidol University
272 Rama VI Rd. Rajchathevi, Bangkok 10400, Thailand
Editor:
Abstract
In this paper, we study the minimax estimation of the Bochner integral
Z
k(·, x) dP (x),
µk (P ) :=
X
also called as the kernel mean embedding, based on random samples drawn i.i.d. from P ,
where k : X × X → R is a positive definite kernel. Various estimators (including the
empirical estimator), θ̂n of µk (P ) are studied in the literature wherein all of them satisfy
kθ̂n − µk (P )kHk = OP (n−1/2 ) with Hk being the reproducing kernel Hilbert space induced
by k. The main contribution of the paper is in showing that the above mentioned rate of
n−1/2 is minimax in k · kHk and k · kL2 (Rd ) -norms over the class of discrete measures and
the class of measures that has an infinitely differentiable density, with k being a continuous
translation-invariant kernel on Rd . The interesting aspect of this result is that the minimax
rate is independent of the smoothness of the kernel and the density of P (if it exists).
Keywords: Bochner integral, Bochner’s theorem, kernel mean embeddings, minimax
lower bounds, reproducing kernel Hilbert space, translation invariant kernel
1. Introduction
Over the last few years, kernel embedding of distributions (Smola et al., 2007), (Sriperumbudur et al.,
2010) has gained a lot of attention in the machine learning community due to the wide variety of applications it has been employed in. Some of these applications include kernel
two-sample testing (Gretton et al., 2007, 2012), kernel independence and conditional independence tests (Gretton et al., 2008; Fukumizu et al., 2008), covariate-shift (Smola et al.,
2007), density estimation (Sriperumbudur, 2011), feature selection (Song et al., 2012), causal
inference (Lopez-Paz et al., 2015), kernel Bayes’ rule (Fukumizu et al., 2013) and distribution regression (Szabó et al., 2015).
1
Tolstikhin, Sriperumbudur, and Muandet
Formally, let Hk be a separable reproducing kernel Hilbert space (RKHS) (Aronszajn,
1950) with a continuous reproducing kernel k : X × X → R defined on a separable
space X . Given a Borel probability measure P defined over X such that
p
Rtopological
k(x,
x)
dP (x) < ∞, the kernel mean or the mean element is defined as the Bochner
X
integral
Z
µP :=
X
k(·, x) dP (x) ∈ Hk .
(1)
We refer the reader to Diestel and Uhl (1977, Chapter 2) and Dinculeanu (2000, Chapter 1)
for the definition of a Bochner integral. The mean element in (1) can be viewed as an
embedding of P in Hk ,
µk : M+1 (X ) → Hk ,
µk (P ) = µP ,
where M+1 (X ) denotes the set of all Borel probability measures on X . Hence, we also refer to µk as the kernel mean embedding (KME). The mean embedding can be seen as a
generalization of the classical kernel feature map that embeds points of an input space X
as elements in Hk . The mean embedding µk can also be seen as a generalization of the
classical notions of characteristic function, moment generation function (if it exists), and
Weierstrass transform of P (all
defined on Rd ) to an arbitrary topological space X as the
√
2
choice of k(·, x) as (2π)−d/2 e− −1h·,xi , eh·,xi , and (4π)−d/2 e−k·−xk2 , x ∈ Rd respectively reduces µk to these notions. The mean embedding µk is closely related to the maximum mean
discrepancy (MMD) (Gretton et al., 2007), which is the RKHS distance between the mean
embeddings of two probability measures. We refer the reader to (Sriperumbudur et al.,
2010; Sriperumbudur, 2016) for more details on the properties of µk and the corresponding
MMD.
In all the above mentioned statistical and machine learning applications, since the underlying distribution P is known only through random samples X1 , . . . , Xn drawn i.i.d. from
it, an estimator of µP is employed. The goal of this paper is to study the minimax optimal
estimation of µP . In the literature, various estimators of µP have been proposed. The
simplest and most popular is the empirical estimator
µPn , which is constructed by replac1 Pn
ing P by its empirical counterpart, Pn := n i=1 δXi , where δx denotes a Dirac measure
at x ∈ X . In fact, all the above mentioned applications deal with the empirical estimator of µP because of its simplicity. Using Bernstein’s inequality in separable Hilbert
spaces (Yurinsky, 1995, Theorem 3.3.4), it follows that for bounded continuous kernels,
√
kµPn − µP kHk = OP (n−1/2 ) for any P , i.e., the empirical estimator is a n-consistent estimator of µP in Hk -norm. This result is also proved in Smola et al. (2007, Theorem 2),
Gretton et al. (2012), and Lopez-Paz et al. (2015) using McDiarmid’s inequality, which we
improve in Proposition A.1 (also see Remark A.2 in Appendix A) by providing a better constant. Assuming X = Rd and
R P to have a density p, Sriperumbudur (2016, Theorem 4.1)
proposed to estimate µP = Rd k(·, x)p(x) dx by replacing p with a kernel density estimator,
√
which is then shown to be n-consistent in Hk -norm if k is a bounded continuous translation
invariant kernel—see Section 2 for its definition—on Rd . Recently, Muandet et al. (2016,
Section 2.4, Theorem 7) proposed a non-parametric shrinkage estimator of µP and estab√
lished its n-consistency in Hk -norm for bounded continuous kernels on X . Muandet et al.
(2016, Section 3, Theorem 10) also proposed a penalized M-estimator for µP where the
2
Minimax Estimation of Kernel Mean Embeddings
penalization parameter is computed in a completely data-driven manner using leave-one√
out cross validation and showed that it is also n-consistent in Hk -norm. In fact, the
√
n-consistency of all these estimators is established by showing that they are all within a
k · kHk -ball of size oP (n−1/2 ) around the empirical estimator µPn .
In the above discussion, it is important to note that the convergence rate of µPn (and
also other estimators) to µP in Hk -norm does not depend on the smoothness of k or the
density, p (if it exists). Under some mild conditions on the kernel (defined on Rd ), it can be
shown (see Section 4) that Hk is continuously included in L2 (Rd ) and kf kL2 (Rd ) ≤ ck kf kHk
for all f ∈ Hk , where ck is a constant that depends only on the kernel. This means, k·kL2 (Rd )
is a weaker norm than k · kHk and therefore it could be possible that µPn converges to µP in
L2 (Rd ) at a rate faster than n−1/2 (depending on the smoothness of k). In Proposition A.1
(also see Remark A.3 in Appendix A) we show that kµPn − µP kL2 (Rd ) = OP (n−1/2 ). Now
given these results, it is of interest to understand whether these rates are optimal in a
minimax sense, i.e., whether the above mentioned estimators are minimax rate optimal or
can they be improved upon? Therefore the goal of this work is to obtain minimax rates for
the estimation of µP in k · kHk and k · kL2 (Rd ) .
Formally, we would like to find the minimax rate rn,k (F, P) and a positive constant
ck (F, P) (independent of n) such that
n
o
−1
(F, P)kθ̂n − µP kF ≥ ck (F, P) > 0,
(2)
inf sup P n rn,k
θ̂n P ∈P
where F is either Hk or L2 (Rd ), P is a suitable subset of Borel probability measures on X ,
and the infimum is taken over all estimators θ̂n mapping the i.i.d. sample X1 , . . . , Xn to F.
Suppose k(x, y) = hx, yi, x, y ∈ Rd . Norms k · kHk and k · kL2 (Rd ) match for this choice of k
and the corresponding
RKHS is finite dimensional, i.e., Hk = Rd . For
R a distribution P on
R
d
R satisfying Rd kxk2 dP (x) < ∞, this choice of kernel yields µP = x dP (x) as the mean
embedding of P which simply is the mean of P . It is well-known (Lehmann and Casella,
2008, Chapter 5, Example 1.14) that the minimax rate of estimating µP ∈ Rd based on
(Xi )ni=1 is rn,k (F, P) = n−1/2 for the class P of GaussianP
distributions on Rd . In fact,
this rate is attained by the empirical estimator µPn = n1 ni=1 Xi , which is the sample
mean. Based on this observation, while one can intuitively argue that the minimax rate of
estimating µP is n−1/2 even if Hk is an infinite dimensional RKHS, it is difficult to extend
the finite dimensional argument in a rigorous manner to the estimation of the infinite
dimensional object, µP . In this paper, through a key inequality—see (3)— we rigorously
show that it is indeed the case.
The main result of the paper is that if k is translation invariant on X = Rd (see
Theorems 1 and 9 for precise conditions on the kernel) and P is the set of all Borel discrete
probability measures on Rd , then the minimax rate rn,k (F, P) is n−1/2 for both F = Hk and
F = L2 (Rd ). Next, we show in Theorems 6 and 12 that the minimax rate for the estimation
of µP in both k · kHk and k · kL2 (Rd ) still remains n−1/2 even when P is restricted to the
class of Borel probability measures which have densities, p that are continuously infinitely
differentiable. The reason for considering such a class of distributions with smooth densities
is that µP , which is the convolution of k and p, is smoother than k. Therefore one might
wonder if it could be possible to estimate µP at a rate faster than n−1/2 that depends on
3
Tolstikhin, Sriperumbudur, and Muandet
the smoothness of k and p. Our result establishes that even for the class of distributions
with very smooth densities, the minimax rate is independent of the smoothness of k and
the density of P . The key ingredient in the proofs of Theorems 6 and 12 is the non-trivial
inequality (see Proposition 3)
kµG0 − µG1 kF ≥ c′k,σ2 kτ0 − τ1 k2 ,
(3)
which relates the F-distance between the mean embeddings of the Gaussian distributions,
G0 = N (τ0 , σ 2 I) and G1 = N (τ1 , σ 2 I) to the Euclidean distance between the means of these
Gaussians, where c′k,σ2 is a constant that depends only on σ 2 and the translation invariant
characteristic kernel k. Combining (3) with Le Cam’s method (see Appendix B) implies
that the estimation of an infinite dimensional object µP is as hard as the estimation of finite
dimensional mean of a Gaussian distribution, thereby establishing the minimax rate to be
n−1/2 . These results show that the empirical estimator—and other estimators we discussed
above—of µP is minimax rate optimal.
Ramdas et al. (2015, Corollary 1) derived a special case of (3) for the Gaussian kernel k by ignoring small terms in the Taylor series expansion of kµG0 − µG1 kHk (refer to
Remark 4). They used this result to show that the MMD between G0 and G1 decreases
to zero exponentially/polynomially fast in d even when the Kullback-Leibler divergence between the two is kept constant, which in turn sheds some light on the decaying power of
MMD-based hypothesis tests in high dimensions. Proposition 3 is more general, as it holds
for any translation-invariant kernel k and does not require a truncation of small reminder
terms.
The paper is organized as follows. Various notations used throughout the paper and
definitions are collected in Section 2. The main results on minimax estimation of µP in
k · kHk and k · kL2 (Rd ) for translation invariant kernels (and also radial kernels) on Rd are
presented in Sections 3 and 4 respectively. The proofs of the results are provided in Section 5
while some supplementary results needed in the proofs are collected in appendices.
2. Definitions & Notation
qP
Pd
d
d
2
Define kak2 :=
i=1 ai and ha, bi :=
i=1 ai bi , where a := (a1 , . . . , ad ) ∈ R and b :=
(b1 , . . . , bd ) ∈ Rd . C(Rd ) (resp. Cb (Rd )) denotes the space of all continuous (resp. bounded
continuous) functions on Rd . f ∈ C(Rd ) is said to vanish at infinity if for every ǫ > 0 the
set {x : |f (x)| ≥ ǫ} is compact. The class of all continuous f on Rd which vanish at infinity
is denoted as C0 (Rd ). For f ∈ Cb (Rd ), kf k∞ := supx∈Rd |f (x)| denotes the supremum
norm of f . Mb (Rd ) (resp. M+b (Rd )) denotes the set of all finite (resp. finite non-negative)
Borel measures on Rd . supp(µ) denotes the support of µ ∈ Mb (Rd ) which is defined as
supp(µ) = {x ∈ Rd | for any open set U such that x ∈ U, |µ|(U ) 6= 0}, where |µ| is the
total-variation of µ. M+1 (Rd ) denotes the set of Borel probability measures on Rd . For
µ ∈ M+b (Rd ), Lr (Rd , µ) denotes the Banach space of r-power (r ≥ 1) µ-integrable functions
and we will use Lr (Rd ) for Lr (Rd , µ) if µ is a Lebesgue measure on Rd . For f ∈ Lr (Rd , µ),
1/r
R
denotes the Lr -norm of f for 1 ≤ r < ∞ and we denote it as
kf kLr (Rd ,µ) := Rd |f |r dµ
k · kLr (Rd ) if µ is the Lebesgue measure. The convolution f ∗ g of two measurable functions
4
Minimax Estimation of Kernel Mean Embeddings
f and g on Rd is defined as
(f ∗ g)(x) :=
Z
f (y)g(x − y) dy,
Rd
provided the integral exists for all x ∈ Rd . The Fourier transforms of f ∈ L1 (Rd ) and
µ ∈ Mb (Rd ) are defined as
Z
∧
−d/2
f (y) := F [f ](y) = (2π)
f (x) e−ihy,xi dx,
y ∈ Rd
Rd
and
∧
−d/2
µ (y) := F [µ](y) = (2π)
Z
Rd
respectively, where i denotes the imaginary unit
e−ihy,xi dµ(x),
√
y ∈ Rd
−1.
A kernel k : Rd × Rd → R is called translation invariant if there exists a symmetric
positive definite function, ψ such that k(x, y) = ψ(x − y) for all x, y ∈ Rd . Bochner’s
theorem (see Wendland, 2005, Theorem 6.6) provides a complete characterization for a
positive definite function ψ: A continuous function ψ : Rd → R is positive definite if and
only if it is the Fourier transform of Λψ ∈ M+b (Rd ), i.e.,
Z
1
e−ihx,wi dΛψ (w), x ∈ Rd .
(4)
ψ(x) =
(2π)d/2 Rd
A kernel k is called radial if there exists φ : R+ → R such that k(x, y) = φ(kx − yk22 ) for all
x, y ∈ Rd . From Schönberg’s representation (Schoenberg, 1938; Wendland, 2005, Theorems
7.13 & 7.14) it is known that a kernel k is radial on every Rd if and only if there exists
ν ∈ M+b ([0, ∞)) such that the following holds for all x, y ∈ Rd :
Z ∞
2
e−tkx−yk dν(t).
(5)
k(x, y) = φ(kx − yk2 ) =
0
Some examples of reproducing kernels on Rd (in fact all these are radial) that appear
throughout the paper are:
kx−yk2
1. Gaussian: k(x, y) = exp − 2η2 2 , η > 0;
P
kx−yk22
2. Mixture of Gaussians: k(x, y) = M
β
exp
−
, where M ≥ 2, η12 ≥ η22 ≥
i
2
i=1
2ηi
PM
2 > 0, and positive constants β , . . . , β
· · · ≥ ηM
1
M such that
i=1 βi = CM < ∞;
3. Inverse Multiquadrics: k(x, y) = (c2 + kx − yk22 )−γ , c, γ > 0;
d
kx−yk2 τ − 2
c2τ −d
4. Matérn: k(x, y) = Γ(τ − d )2τ −1−d/2
K d −τ (ckx − yk2 ), τ > d/2, c > 0,
c
2
2
where Kα is the modified Bessel function of the third kind of order α and Γ is the
Gamma function.
A kernel k is said to be characteristic if the mean embedding, µk : P → µP is injective, where µP is defined in (1). Various characterizations for the injectivity of µk (or k
5
Tolstikhin, Sriperumbudur, and Muandet
being characteristic) are known in literature (for details, see Sriperumbudur et al., 2011
and references therein). If k is a bounded continuous translation invariant positive definite kernel on Rd , a simple characterization can be obtained for it to be characteristic
(Sriperumbudur et al., 2010, Theorem 9): k is characteristic if and only if supp(Λψ ) = Rd
where Λψ is defined in (4). This characterization implies that the above mentioned examples
are characteristic kernels. Examples of non-characteristic kernels of translation invariant
type include k(x, y) = sin(x−y)
x−y , x, y ∈ R and k(x, y) = cos(x − y), x, y ∈ R. More generally,
polynomial kernels of any finite order are non-characteristic.
3. Minimax Estimation of µP in the RKHS Norm
In this section, we present our main results related to the minimax estimation of kernel mean
embeddings (KMEs) in the RKHS norm. As discussed in Section 1, various estimators of
µk (P ) are known in literature and all these have a convergence rate of n−1/2 if the kernel is
bounded. The main goal of this section is to show that the rate n−1/2 is actually minimax
optimal for different choices of P (see (2)) under some mild conditions on k.
First, choosing P to be the set of all discrete probability measures on Rd , in Section 3.1
(see Theorem 1 and Corollary 2), we present the minimax lower bounds of order Ω(n−1/2 )
with constant factors depending only on the properties of the kernel for translation invariant
and radial kernels respectively. Next we will show in Section 3.2 that the rate n−1/2 remains
minimax optimal for translation invariant and radial kernels even if we choose the class P
to contain only probability distributions with infinitely continuously differentiable densities.
For translation invariant kernels the result (see Theorem 6) is based on a key inequality,
which relates the RKHS distance between embeddings of Gaussian distributions to the
Euclidean distance between the mean vectors of these distributions (see Proposition 3). The
minimax lower bound for radial kernels (see Theorem 8) is derived using a slightly different
argument. Instead of applying the bound of Theorem 6 to the particular case of radial
kernels, we will present a direct analysis based on the special properties of radial kernels.
This will lead us to the lower bound with almost optimal constant factors, depending only
on the shape of Borel measure ν corresponding to the kernel.
Our analysis is based on the following simple idea: if a kernel k is characteristic, there
is a one-to-one correspondence between any given set of Borel probability measures P
defined over Rd and a set µk (P) of their embeddings into the RKHS Hk . This means
that distributions in P are indexed by their embeddings Θ := µk (P) and so (2) can be
equivalently written as
o
n
−1
inf sup Pθ rn,k
(Hk , P)kθ̂n − θkHk ≥ ck (Hk , P) > 0,
(6)
θ̂n θ∈Θ
where the goal is to find the minimax rate rn,k (Hk , P) and a positive constant ck (Hk , P)
(independent of n) such that (6) holds and Pθ = P n when θ = µk (P ). Using this equivalence, we obtain the minimax rates by employing Le Cam’s method (Tsybakov, 2008)—see
Theorems B.1 and B.2 for a reference.
6
Minimax Estimation of Kernel Mean Embeddings
3.1 Lower Bounds for Discrete Probability Measures
The following result (proved in Section 5.1) presents a minimax rate of n−1/2 for estimating
µk (P ), where k is assumed to be translation invariant on Rd .
Theorem 1 (Translation invariant kernels) Let P be the set of all Borel discrete probability measures on Rd . Suppose k(x, y) = ψ(x − y), where ψ ∈ Cb (Rd ) is positive definite
and k is characteristic. Assume there exists z ∈ Rd and β > 0, such that ψ(0) − ψ(z) ≥ β.
Then the following holds:
(
r )
1
1 2β
n
inf sup P
≥ .
kθ̂n − µk (P )kHk ≥
6
n
4
θ̂n P ∈P
The result is based on Le Cam’s method involving two hypotheses (see Theorem B.1), where
we choose them to be KMEs of discrete measures, both supported on the same pair of points
separated by z in Rd .
Remark (Choosing z and β) As discussed in Sriperumbudur et al. (2010, Section 3.4),
if k is translation invariant and characteristic on Rd , then it is also strictly positive definite. This means that ψ(0) > 0. Moreover, the following hold: (a) Since ψ is positive
definite, we have |ψ(x)| ≤ ψ(0) for all x ∈ Rd and (b) since ψ is characteristic, it cannot be a constant function. Together these facts show that there always exist z ∈ Rd
and β > 0 satisfying the assumptions
of Theorem 1. For instance, a Gaussian kernel
k(x, v) = exp −kx − vk22 /(2η 2 ) satisfies ψ(0) − ψ(z) ≥ kzk22 /(4η 2 ) if kzk22 ≤ 2η 2 , where we
used a simple fact that 1 − e−x ≥ x/2 for 0 ≤ x ≤ 1.
While Theorem 1 dealt with general translation invariant kernels, the following result
(proved in Section 5.2) specializes it to radial kernels, i.e., kernels of the form in (5), by
providing a simple condition on ν under which Theorem 1 holds.
Corollary 2 (Radial kernels) Let P be the set of all BorelR discrete probability measures
2
∞
on Rd and k be radial on Rd , i.e., k(x, y) = ψν (x − y) := 0 e−tkx−yk2 dν(t), where ν ∈
M+b ([0, ∞)) such that supp(ν) 6= {0}. Assume there exist 0 < t1 < ∞ and α > 0 satisfying
ν([t1 , ∞)) ≥ α. Then the following holds:
r
1
1 α
n
≥ .
inf sup P
kθ̂n − µk (P )kHk ≥
6 n
4
θ̂n P ∈P
Remark (Choosing t1 and α) Since supp(ν) 6= {0} the assumption of ν[t1 , ∞) ≥ α is
always satisfied. For instance, if ν is a probability measure with positive median η then we
can set t1 = η and α = 12 . Based on this, it is easy to verify (see Appendix D.1) that α = 1
−2γ
for Gaussian, α = CM for mixture of Gaussian kernels, α = c 2 for inverse multiquadrics
and α = 12 for Matérn kernels.
3.2 Lower Bounds for Probability Measures with Smooth Densities
So far, we have shown that the rate n−1/2 is minimax optimal for the problem of KME
estimation (both for translation invariant and radial kernels). As discussed in Section 1,
7
Tolstikhin, Sriperumbudur, and Muandet
since this rate is independent of the smoothness of the estimand (which is determined by the
smoothness of the kernel), one might wonder whether the minimax rate can be improved by
restricting P to distributions with smooth densities. We show in this section (see Theorems 6
and 8) that this is not the case by restricting P to contain only distributions with infinitely
continuously differentiable densities and proving the minimax lower bound of order n−1/2 .
We will start the analysis with translation invariant kernels and present a corresponding
lower bound in Theorem 6. The proof of this result is again based on an application
of Le Cam’s method involving two hypotheses (see Theorem B.1), where this time these
hypotheses are chosen to be embeddings of the d-dimensional Gaussian distributions. One
of the main steps, when applying Theorem B.1, is to lower bound the distance between these
embeddings. This is done in the following result (proved in Section 5.3), which essentially
shows that if we take two Gaussian distributions G(µ0 , σ 2 I) and G(µ1 , σ 2 I) with the mean
vectors µ0 , µ1 ∈ Rd which are close enough to each other, then the RKHS distance between
the corresponding embeddings can be lower bounded by the Euclidean distance kµ0 − µ1 k2 .
Proposition 3 Let σ > 0. Suppose k(x, y) = ψ(x − y), where ψ ∈ Cb (Rd ) is positive
definite and k is characteristic. Then there exist constants ǫψ,σ2 , cψ,σ2 > 0 depending only
on ψ and σ 2 , such that the following condition holds for any a ∈ Rd with kak22 ≤ ǫψ,σ2 :
cψ,σ2
2
≤ min
ez ∈S d−1 (2π)d/2
Z
e−σ
2 kwk2
2
Rd
hez , wi2 cos (ha, wi) dΛψ (w) < ∞,
(7)
where S d−1 is a unit sphere in Rd and Λψ ∈ M+b (Rd ) is defined in (4). Moreover, for all
vectors µ0 , µ1 ∈ Rd satisfying kµ0 − µ1 k22 ≤ ǫψ,σ2 , the following holds:
kθ0 − θ1 kHk ≥
r
cψ,σ2
kµ0 − µ1 k2 ,
2
(8)
where θ0 and θ1 are KMEs of Gaussian measures G(µ0 , σ 2 I) and G(µ1 , σ 2 I) respectively.
Remark 4 (KME expands small distances) For a Gaussian kernel, it is possible to
show (Sriperumbudur et al., 2012, Example 3; Ramdas et al., 2015, Proposition 1) that
kθ0 − θ1 k2Hk = C1 1 − exp(−C2 kµ0 − µ1 k22 ) , where C1 and C2 are positive constants that
depend only on σ 2 and η 2 . This
shows that (8) holds for kµ0 − µ1 k2 ∈ [0, D], where D
satisfies C1 1 − exp(−C2 D 2 ) = 12 D 2 cψ,σ2 . In other words,
Proposition 3 states that the
2
d
mapping fσ2 : R → Hk defined by fσ2 (x) := µk G(x, σ I) expands small distances.
Remark 5 (Computing cψ,σ2 and ǫψ,σ2 ) Generally it may be very hard to compute (or
bound) the constants cψ,σ2 and ǫψ,σ2 appearing in the statement of Proposition 3. However,
in some cases this may be still possible. In Appendix E we will provide an extensive analysis
for the case of radial kernels.
Based on Proposition 3, the following result shows that the rate of n−1/2 remains minimax
optimal for the problem of KME estimation with translation invariant kernels, even if we
restrict the class of distributions P to contain only measures with smooth densities.
8
Minimax Estimation of Kernel Mean Embeddings
Theorem 6 (Translation invariant kernels) Let P be the set of distributions over Rd
whose densities are continuously infinitely differentiable. Suppose k(x, y) = ψ(x − y), where
ψ ∈ Cb (Rd ) is positive definite and k is characteristic. Define cψ := cψ,1 and ǫψ := ǫψ,1
where cψ,1 and ǫψ,1 are positive constants that satisfy (7) in Proposition 3. Then for any
n ≥ ǫ1ψ , the following holds:
r
1
1 cψ
≥ .
inf sup P
kθ̂n − µk (P )kHk ≥
2 2n
4
θ̂n P ∈P
n
Proof The proof will be based on Theorem B.1. For this we need to find two probability
measures P0 and P1 on Rd and corresponding KMEs θ0 and θ1 , such that kθ0 − θ1 kHk is
of the order Ω(n−1/2 ), while KL(P0n kP1n ) is upper bounded by a constant independent of
n. Here KL(P0 kP1 ) denotes
Kullback-Leibler divergence between P0 and P1 , which is
R the
dP0
defined as KL(P0 kP1 ) = log dP1 dP0 where P0 is absolutely continuous w.r.t. P1 .
Pick two Gaussian distributions G0 := G(µ0 , σ 2 I) and G1 := G(µ1 , σ 2 I) for µ0 , µ1 ∈ Rd ,
and σ 2 > 0. It is known that (Tsybakov, 2008, Section 2.4)
KL(Gn0 kGn1 ) = n ·
kµ0 − µ1 k22
,
2σ 2
(9)
where Gn0 and Gn1 are n-fold product distributions. Choose µ0 and µ1 such that
kµ0 − µ1 k22 =
1
.
n
Denote KMEs of G0 and G1 using θ0 and θ1 respectively. Next we will take σ 2 = 1 and apply
Proposition 3. Since cψ and ǫψ satisfy (7) in Proposition 3, it follows from Proposition 3
that for 1/n ≤ ǫψ ,
cψ
cψ
kθ0 − θ1 k2Hk ≥
kµ0 − µ1 k22 =
.
2
2n
This
p shows that the first condition of Theorem B.1 is satisfied for θ0 and θ1 with s :=
1
cψ /(2n). Moreover, using (9) we can show that the second condition of Theorem B.1 is
2
satisfied with α = 21 . We conclude the proof with an application of Theorem B.1.
Remark 7 (Lower bound on the sample size n) Note that Theorem 6 holds only for
large enough sample size n (i.e., n ≥ 1/ǫψ ). This assumption on p
n can be dropped if we set
kµ0 − µ1 k22 = ǫψ /n in the proof. In this case, the lower bound 21 cψ /(2n) will be replaced
p
with 21 cψ ǫψ /(2n), while the lower bound on the minimax probability 1/4 will be replaced
with
!
p
1 − ǫψ 1 − ǫψ /4
.
e 2,
max
4
2
The latter is generally undesirable, especially if ǫψ grows with d → ∞, since we want the
minimax probability to be lower bounded by some universal non-zero constant that does not
depend on the properties of the problem at hand.
9
Tolstikhin, Sriperumbudur, and Muandet
Since radial kernels are particular instances of translation invariant kernels, Theorem 6
can be specialized by explicitly computing the constants cψ and ǫψ to derive a minimax
lower bound of order Ω(n−1/2 ). Unfortunately, the resulting lower bound will depend on
the dimensionality d in a rather bad way and, as a consequence, is suboptimal in some
situations. For instance, if we consider a Gaussian kernel k(x, y) = exp − 2η12 kx − yk22 ,
then a straightforward
computation of cψ shows that the lower bound in Theorem 6 has
p
2
−d/2
the form (1 + 2/η )
/n which shrinks to zero as d → ∞, while Proposition A.1 (also
see Remark A.2) provides a dimension independent upper bound of the order Op (n−1/2 ).
Therefore, instead of specializing Theorem 6 to radial kernels, we obtain the following result
for radial kernels by using a refined analysis which yields a minimax rate of Ω(n−1/2 ) that
matches the upper bound of Proposition A.1 up to constant factors that depend only on
the shape of Borel measure ν. In particular, when specialized to the Gaussian kernel, the
result matches the upper bound up to a constant factor independent of d.
R∞
2
Theorem 8 (Radial kernels) Let k be radial on Rd , i.e., k(x, y) = 0 e−tkx−yk2 dν(t),
where ν ∈ M+b ([0, ∞)) and P be the set of distributions over Rd whose densities are continuously infinitely differentiable. Assume that supp(ν) 6= {0} and there exist 0 < t0 ≤ t1 < ∞,
0 < β < ∞ such that ν([t0 , t1 ]) ≥ β. Then the following holds:
s
(
)
1
βt
2
1
1
0
≥ .
·
1−
inf sup P n kθ̂n − µk (P )kHk ≥
50 n t1 e
2+d
5
θ̂n P ∈P
Proof The proof, which is presented in Section 5.4, is based on an application of Le Cam’s
method involving multiple hypotheses (see Theorem B.2), where we use exponential (in d)
number of Gaussian distributions with variances decaying as d1 .
Remark (Non-trivial lower bound as d → ∞) The proof of Theorem 8 is based on
Gaussian distributions with variances decaying as 1/d. As d → ∞, it is obvious the densities of these distributions do not have uniformly bounded Lipschitz constants, i.e., they
are arbitrarily “peaky”. Hence, if we choose P to be class of distributions with infinitely
differentiable densities that have uniformly bounded Lipschitz constants, then as d → ∞, the
densities considered in the proof of Theorem 8 do not belong to P. On the other hand, the
densities considered in the proof of Theorem 6 still belong to P but yielding an uninteresting
result since cψ → 0 when d → ∞. Therefore, it is an open question whether a non-trivial
lower bound can be obtained for radial kernels (or any other translation invariant kernels) if
we choose P to contain only distributions with densities having uniformly bounded Lipschitz
constants.
Remark (Alternative Proof ) For completeness, we also present an alternative proof of
Theorem 8 in Appendix E. It is based on Proposition 3, which holds for any translation invariant kernel. As a result, this proof leads to slightly worse constants compared to Theorem
8 (where we used an analysis specific to radial kernels), as well as a superfluous condition
on the minimal sample size n.
0
In Appendix D.2, we compute the positive constant Bk := βt
t1 that appears in the lower
bound in Theorem 8 in a closed form for Gaussian, mixture of Gaussian, inverse multiquadric and Matérn kernels.
10
Minimax Estimation of Kernel Mean Embeddings
4. Minimax Estimation of µP in the L2 (Rd ) Norm
So far, we have discussed the minimax estimation of the kernel mean embedding (KME) in
the RKHS norm. In this section, we investigate the minimax estimation of KME in L2 (Rd )
norm. The reason for this investigation is as follows. Let k(x, y) = ψ(x − y), x, y ∈ Rd ,
where ψ ∈ L1 (Rd ) ∩ C(Rd ) is strictly positive definite. The corresponding RKHS is given
by (see Wendland, 2005, Theorem 10.12)
)
(
Z
∧ (ω)|2
|f
dω < ∞ ,
(10)
Hk = f ∈ L2 (Rd ) ∩ C(Rd ) :
∧
Rd ψ (ω)
R
∧
∧ (ω)
dω with f ∧ being the
which is endowed with the inner product hf, giHk = Rd f (ω)g
ψ∧ (ω)
Fourier transform of f in the L2 -sense. It follows from (10) that for any f ∈ Hk ,
(†)
(‡)
|f ∧ (ω)|2 ∧
ψ
(ω)
dω
≤ kψ ∧ k∞ kf k2Hk < ∞,
∧
Rd ψ (ω)
Rd
(11)
where (⋆) follows from Plancherel theorem (Wendland, 2005, Corollary 5.25), kf kHk is
defined in (10), (†) follows from Hölder’s inequality, and (‡) holds since ψ ∧ ∈ C0 (Rd ) (by
Riemann-Lebesgue lemma, Folland, 1999, Theorem 8.22). Note that ψ ∧ is non-negative
(Wendland, 2005, Theorem 6.11) and so the inequality in (†) is valid. It therefore follows
from (11) that Hk is continuously included in L2 (Rd ) and k · kL2 (Rd ) is a weaker norm than
k · kHk .1 This means it is possible that the minimax rate of estimating µP in k · kL2 (Rd ) could
be faster than its RKHS counterpart with the rate possibly depending on the smoothness
of k. Hence, it is of interest to analyze the minimax rates of estimating µP in k · kL2 (Rd ) .
Interestingly, we show in this section that the minimax rate in the L2 setting is still n−1/2 .
(⋆)
kf k2L2 (Rd ) =
kf ∧ k2L2 (Rd )
=
Z
∧
f (ω)
2
dω =
Z
The analysis in the L2 setting follows ideas similar to those of the RKHS setting wherein,
first, in Section 4.1, we consider the minimax rate of estimating µP for translation invariant
and radial kernels when P is the set of all Borel discrete probability measures on Rd (see
Theorem 9 and Corollary 10). Next, in Section 4.2, we choose P to be the set of all
probability distributions that have infinitely continuously differentiable densities and study
the question of minimax rates for translation invariant (see Theorem 12) and radial kernels
(see Theorem 13). For both these choices of P, we show that the rate is n−1/2 irrespective
of the smoothness of k. Exploiting the injectivity of mean embedding for characteristic
kernels (see the paragraph below and the paragraph around (6)), these results are derived
using Le Cam’s method (see Theorems B.1 and B.2). Combined with Proposition A.1 (also
see Remark A.3), these results show that the empirical estimator, µPn is minimax optimal.
Finally, in Section 4.3 we discuss the relation between our results and some classical results
of nonparametric density estimation, particularly, those of the kernel density estimator.
1. The continuous inclusion of Hk in L2 (Rd ) is known for Gaussian kernels on Rd (e.g., see Vert and Vert,
2006, Lemma 11). Similar result is classical for Sobolev spaces in general (e.g., see Folland, 1999,
Section 9.3, p. 302) and particularly for those induced by Matérn kernels. Steinwart and Christmann
2
p Theorem 4.26) provides a general result for continuous inclusion of Hk in L (µ) assuming
R(2008,
k(x, x) dµ(x) < ∞ where µ is a σ-finite measure. However, the result does not hold for transX
lation invariant kernels on Rd as the integrability condition is violated.
11
Tolstikhin, Sriperumbudur, and Muandet
Before we proceed to the main results of this section, we briefly discuss the difference
between estimation in RKHS and L2 (Rd ) norms. Suppose k(x, y) = ψ(x − y), x, y ∈ Rd
where ψ ∈ L1 (Rd ) ∩ L2 (Rd ) ∩ C(Rd ) is positive definite and characteristic. It is easy to
verify that µP ∈ L1 (Rd ) ∩ L2 (Rd ). Since µP = ψ ∗ P , (10) implies
kµP k2Hk
=
Z
Rd
|(ψ ∗ P )∧ |2
dω =
ψ ∧ (ω)
Z
Rd
|φP (ω)|2 ψ ∧ (ω) dω = kφP k2L2 (Rd ,ψ∧ )
(12)
whereas
(⋆)
kµP k2L2 (Rd ) =
Z
Rd
2
|µ∧
P (ω)| dω
=
Z
Rd
|φP (ω)|2 (ψ ∧ )2 (ω) dω = kφP k2L2 (Rd ,(ψ∧ )2 ) ,
(13)
R
T
where φP (ω) := e−iω x dP (x) is the characteristic function of P and (⋆) follows from
Plancherel’s theorem. It follows from (12) and (13) that the RKHS norm emphasizes
the high frequencies of φP compared to that of the L2 -norm. Since ψ is characteristic,
i.e., P 7→ µk (P ) ∈ Hk is injective, which is guaranteed if and only if supp(ψ ∧ ) = Rd
(Sriperumbudur et al., 2010, Theorem 9), it follows from (13) that P 7→ µk (P ) ∈ L1 (Rd ) ∩
L2 (Rd ) is injective. Therefore (2) can be equivalently written as (6) by replacing k · kHk
with k · kL2 (Rd ) (see the discussion around (6)) and we obtain minimax rates by employing
Le Cam’s method as we did in the previous section.
4.1 Lower Bounds for Discrete Probability Measures
The following result (proved in Section 5.5) for translation invariant kernels is based on
an application of Le Cam’s method involving two hypotheses (see Theorem B.1), where we
choose them to be KMEs of discrete measures, both supported on the same pair of points
separated by a vector z in Rd .
Theorem 9 (Translation invariant kernels) Let P be the set of all Borel discrete probability measures on Rd . Suppose k(x, y) = ψ(x − y), x, y ∈ Rd where ψ ∈ L2 (Rd ) ∩ C(Rd )
is positive definite and k is characteristic. Define
Z
ψ
2
ψ(y)ψ(y + z)dy
(14)
Cz := 2 kψkL2 (Rd ) −
Rd
for some z ∈ Rd \ {0}. Then Czψ > 0 and
s
ψ
1
1 Cz
≥ .
inf sup P n kθ̂n − µk (P )kL2 (Rd ) ≥
6
n
4
θ̂n P ∈P
Using Cauchy-Schwartz inequality, the constant Czψ in Theorem 9 can be shown (see the
proof of Lemma 15 in Section 5.5) to be positive for every z ∈ Rd \{0} if k is characteristic,
i.e., supp(Λψ ) = Rd (see (4) for Λψ ). The following result (proved in Section 5.6) specializes
Theorem 9 to radial kernels.
12
Minimax Estimation of Kernel Mean Embeddings
Corollary 10 (Radial kernels) Let P be the set of all BorelR discrete probability measures
2
∞
on Rd and k be radial on Rd , i.e., k(x, y) = ψν (x − y) := 0 e−tkx−yk2 dν(t), where ν ∈
M+b ([0, ∞)) such that supp(ν) 6= {0} and
Z
0
∞
t−d/2 dν(t) < ∞.
(15)
Assume that there exist 0 < δ0 ≤ δ1 < ∞ and β > 0 such that ν([δ0 , δ1 ]) ≥ β. Then the
following holds:
s
d/2
1
π
β
1
≥ .
inf sup P n kθ̂n − µk (P )kL2 (Rd ) ≥
4
6 n 2δ1
θ̂n P ∈P
In Corollary 10, since supp(ν) 6= {0}, the assumption of ν([δ0 , δ1 ]) ≥ β is always satisfied.
In addition, the condition (15) on ν is satisfied by Gaussian, mixture of Gaussians, inverse
multiquadric (while (15) is satisfied for γ > d/2, the result in Corollary 10 holds for γ > d/4)
and Matérn kernels—refer to Remark A.3 for more details. Also, for these examples of
−d/2
kernels, the positive constant Ak := β 2 δ1
in the lower bound in Corollary 10 can be
computed in a closed form (see Appendix D.3 for details).
4.2 Lower Bounds for Probability Measures with Smooth Densities
Next, as we did in Section 3.2, we choose P to be the set of all probability measures
that have infinitely continuously differentiable densities and show that the minimax rate
of estimating µP in L2 -norm for translation invariant (see Theorem 12) and radial kernels
(see Theorem 13) is n−1/2 . The proof of these results are again based on an application of
Le Cam’s method involving two (see Theorem B.1) and multiple hypotheses (see Theorem
B.2), where these hypotheses are chosen to be embeddings of the d-dimensional Gaussian
distributions. As in Section 3.2, the results of this section are based on the following result
(proved in Section 5.7), which is conceptually similar to that of Proposition 3.
Proposition 11 Let σ > 0. Suppose k(x, y) = ψ(x − y), where ψ ∈ L1 (Rd ) ∩ Cb (Rd ) is
positive definite and k is characteristic. Then there exist constants ǫψ,σ2 , cψ,σ2 > 0 depending
only on ψ and σ 2 , such that the following condition holds for any a ∈ Rd with kak22 ≤ ǫψ,σ2 :
cψ,σ2 ≤ min 2
ez ∈S d−1
Z
e−σ
Rd
2 kwk2
2
2
hez , wi2 cos (ha, wi) ψ ∧ (w) dw < ∞,
(16)
where S d−1 is a unit sphere in Rd . Moreover, for all vectors µ0 , µ1 ∈ Rd satisfying kµ0 −
µ1 k22 ≤ ǫψ,σ2 , the following holds:
kθ0 − θ1 kL2 (Rd ) ≥
r
cψ,σ2
kµ0 − µ1 k2 ,
2
where θ0 and θ1 are KMEs of the Gaussian measures G(µ0 , σ 2 I) and G(µ1 , σ 2 I) respectively.
13
Tolstikhin, Sriperumbudur, and Muandet
The following result for translation invariant kernels is established using the above result
wherein the proof is exactly the same as that of Theorem 6 except for an application of
Proposition 11 in place of Proposition 3.
Theorem 12 (Translation invariant kernels) Let P be the set of distributions over Rd
whose densities are continuously infinitely differentiable. Suppose k(x, y) = ψ(x − y), where
ψ ∈ L1 (Rd ) ∩ Cb (Rd ) is positive definite and k is characteristic. Define cψ := cψ,1 and
ǫψ := ǫψ,1 where cψ,1 and ǫψ,1 are positive constants that satisfy (16) in Proposition 11.
Then for any n ≥ ǫ1ψ , the following holds:
r
1 cψ
1
inf sup P
kθ̂n − µk (P )kL2 (Rd ) ≥
≥ .
2
2n
4
θ̂n P ∈P
n
As discussed in Remark 7, it is possible to remove the requirement of minimal sample size in
Theorem 12. Also, as discussed in Remark 5 and in the paragraph following Remark 7, the
constants cψ and ǫψ appearing in the bound in Theorem 12 are not only difficult to compute
but also may depend on the dimensionality d in a sup-optimal manner, particularly as
d → ∞. Therefore, similar to what was done in Section 3.2, we will not specialize Theorem
12 to radial kernels but instead present the following result (proved in Section 5.8 and the
proof closely follows that of Theorem 8), which is based on a direct analysis involving the
properties of radial kernels. For the particular case of a Gaussian kernel, this lower bound
matches the upper bound of Proposition A.1 (also see Remark A.3) up to a constant factor
independent of d.
R∞
2
Theorem 13 (Radial kernels) Let k be radial on Rd , i.e., k(x, y) = 0 e−tkx−yk2 dν(t),
where ν ∈ M+b ([0, ∞)) and P be the set of distributions over Rd whose densities are continuously infinitely differentiable. Assume that (15) holds, supp(ν) 6= {0} and there exist
0 < δ0 ≤ δ1 < ∞, 0 < β < ∞ such that ν([δ0 , δ1 ]) ≥ β. Then the following holds:
s
d/2 2
β δ0
π
1
2
1
1
≥ .
1−
inf sup P n kθ̂n − µk (P )kL2 (Rd ) ≥
50
n
2δ
δ
e
2
+
d
5
1
1
θ̂n P ∈P
− d+2
The constant Bk := β 2 δ0 δ1 2 in the lower bound in the above result can be computed in
a closed form for Gaussian, mixture of Gaussian, inverse multiquadric, and Matérn kernels
(see Appendix D.4 for details). The factor (π/2)d/4 can be eliminated from the lower bound
by considering a rescaled kernel (π/2)−d/4 ψ(x−y). Nevertheless, the bound will still depend
on d exponentially as captured by the constant Bk . This can be further overcome by using
the normalized kernel k(x, y)/kψkL2 (Rd ) . In the particular case of normalized Gaussian
kernels (πη 2 )−d/2 exp − 2η12 kx − yk22 this will lead to dimension-free lower bounds.
4.3 Relation to Kernel Density Estimation
In this section, we discuss the relation between the estimation of µP and density estimation.
The problem of density estimation deals with estimating an unknown density, p based on
14
Minimax Estimation of Kernel Mean Embeddings
random samples (Xi )ni=1 drawn i.i.d. from it. One of the popular non-parametric methods
for density estimation is kernel density estimation (KDE), where the estimator is of the
form (Tsybakov, 2008, Section 1.2)
n
X
Xi,d − xd
Xi,1 − x1
1
p̂n (x1 , . . . , xd ) = Qd
K
,··· ,
.
h1
hd
n i=1 hi
i=1
Rd
Here K :
→ R is the smoothing kernel (this kernel should not be confused with the reproducing kernel k which we used throughout the paper), h1 , . . . , hd > 0 are bandwidths, and
Xi,j is the j-th coordinate of the i-th sample point. Assuming p ∈ L2 (Rd ), the consistency of
p̂n is usually studied in the sense of mean integrated squared error (MISE) Ekp̂n − pk2L2 (Rd ) ,
which can be decomposed into variance and bias terms as:
2
L2 (Rd )
Ekp̂n − pk2L2 (Rd ) = E p̂n − E[p̂n ]
+ p − E[p̂n ]
2
.
L2 (Rd )
(17)
Assume K to be bounded and h1 = · · · = hd = h. Define Kh := h−d K(·/h). Then for any
fixed x ∈ Rd ,
Z
n
n
Xi − x
1X
1 X
Kh (z − x) dPn (z)
K
Kh (Xi − x) =
=
p̂n (x) =
nhd
h
n
Rd
i=1
i=1
and
E[p̂n (x)] =
1
hd
Z
K
Rd
z−x
h
p(z)dz = (Kh ∗ p)(x).
This shows that p̂n = µKh (Pn ) and E[p̂n ] = µKh (P ) where P is the distribution with p as its
density w.r.t. the Lebesgue measure and Pn is the empirical measure constructed based on
samples (Xi )ni=1 drawn from p. Therefore the results of Section 4 (and more generally of this
paper) are about the minimax rates for E[p̂n ]. However, note that Kh need not be positive
definite (and therefore need not be the reproducing kernel of some RKHS).
R On the other
d
hand, K has to be positive, i.e., K(x) ≥ 0, ∀ x ∈ R and normalized, i.e., Rd K(x) dx = 1
to yield an estimator that is a valid density, unlike in kernel mean estimation where k need
not be positive nor normalized. The minimax rate of n−1/2 for estimating E[p̂n ] is achieved
by the kernel density estimator p̂n (which is nothing but the empirical estimator of µKh (P ))
as it is known (based on a straightforward generalization of Tsybakov, 2008, Proposition
1.4 for multiple dimensions) that
2
L2 (Rd )
kKk2L2 (Rd )
,
nhd
where we assume K ∈ L2 (Rd ). The bandwidth parameter h is immaterial in the estimation
of µKh (P ) and can be treated as a constant (independent of n) unlike in the problem
of estimating p where h should decay to zero at an appropriate rate for the bias p −
E[p̂n ] L2 (Rd ) to converge to zero as n → ∞. In particular, if p lies in a Sobolev space of
smoothness index s, then the bias-squared term in (17) behaves as h2s , which combined
1
2s
with the above bound on the variance yields a rate of n− 2s+1 for h = n− 2s+1 . This rate is
known to be minimax optimal for the problem of estimating p while our rates are minimax
optimal for the problem of smoothed density estimation where the smoothing is carried out
by the kernel.
E p̂n − E[p̂n ]
15
≤
Tolstikhin, Sriperumbudur, and Muandet
5. Proofs
In this section we present all the missing proofs of results of Sections 3 and 4.
5.1 Proof of Theorem 1
Pick two discrete distributions P0 = p0 δx + (1 − p0 )δv and P1 = p1 δx + (1 − p1 )δv , where
x, v ∈ Rd , 0 < p0 < 1, 0 < p1 < 1 and δx denotes
R R a Dirac measure supported at x. Define
2
k(x, y) dP0 (x) dP0 (y), which follows from
θ0 = µk (P0 ) and θ1 = µk (P1 ). Since kθ0 kHk =
the reproducing property of k, it is easy to verify that
kθ0 − θ1 k2Hk = E[k(ξ, ξ ′ )] + E[k(η, η ′ )] − 2 E[k(ξ, η)],
where ξ and η are random variables distributed according to P0 and P1 respectively, and
ξ ′ and η ′ are independent copies of ξ and η. Since k is translation invariant, we have
k(v, v) = k(x, x) = ψ(0) and k(x, v) = k(v, x) = ψ(x − v), which imply
kθ0 − θ1 k2Hk = 2(p0 − p1 )2 ψ(0) − ψ(x − v) .
(18)
Also note that
1 − p0
p0
+ (1 − p0 ) log
p1
1 − p1
p0 − p1
p1 − p0
+ (1 − p0 ) log 1 +
= p0 log 1 +
p1
1 − p1
(∗)
p1 − p0
p0 − p1
+ (1 − p0 ) 1 +
≤ log p0 1 +
p1
1 − p1
p0 1 − p0
= log 1 + (p0 − p1 )
−
,
p1 1 − p1
KL(P0 kP1 ) = p0 log
where we used Jensen’s inequality in (∗) for the logarithmic function, which is concave.
Next, using a simple inequality log(1 + x) ≤ x, which holds for all x > −1, we get
p0 1 − p0
(p0 − p1 )2
KL(P0 kP1 ) ≤ (p0 − p1 )
−
.
=
p1 1 − p1
p1 (1 − p1 )
Note that a maximal value of denominator is achieved when p1 = 21 . Setting p1 = 12 we get
2
the following upper bound: KL(P0 kP1 ) ≤ 4 p0 − 21 , which when used in the chain rule
of KL-divergence yields
1 2
n
n
.
KL(P0 kP1 ) ≤ 4n p0 −
2
2
1
Choosing p0 such that (p0 − 12 )2 = 9n
yields KL(P0n kP1n ) ≤ 49 and kθ0 − θ1 k2Hk = 9n
ψ(0) −
d
ψ(x − v) . Choose x and v in such a way that x − v = z, where z ∈ R is a point for which
ψ(0) − ψ(z) ≥ β and β > 0. This yields
kθ0 − θ1 k2Hk ≥
16
2β
,
9n
Minimax Estimation of Kernel Mean Embeddings
which shows that the assumptions of Theorem B.1 are satisfied with s :=
The result follows from an application of Theorem B.1 by noticing that
q
2β
1
and α := 49 .
6
√n
1− α/2
> 1/4.
2
Remark (Measures with bounded support) It is evident from the above proof that
exactly the same lower bound holds if we restrict P to contain only probability measures
with bounded support. We can proceed further and assume that for each P ∈ P the radius
of supp(P ) is upper bounded by some positive constant R. In this case the same reasoning
will work as long as ψ is not “flat” on the ball of radius R centered around origin.
5.2 Proof of Corollary 2
The proof is based on application of Theorem 1. Since supp(ν) 6= {0}, it follows from
(Sriperumbudur et al., 2011, Proposition 5) that k is characteristic. We now show that
there exist z ∈ Rd and β > 0, such that ψν (0) − ψν (z) ≥ β. Note that for any x ∈ Rd
Z ∞
Z ∞
2
−tkxk22
dν(t) ≥
1 − e−tkxk2 dν(t)
1−e
ψν (0) − ψν (x) =
t1
Z0 ∞
2
2
≥
1 − e−t1 kxk2 dν(t) = ν([t1 , ∞)) 1 − e−t1 kxk2
t1
αt
2
1
≥ α 1 − e−t1 kxk2 ≥
kxk22 ,
2
where the last inequality holds whenever kxk22 ≤ t11 . Choosing z such that kzk22 = t11 yields
ψν (0) − ψν (z) ≥ α2 . The result therefore follows from Theorem 1 by choosing β = α2 .
5.3 Proof of Proposition 3
Before we prove Proposition 3, first we will derive a closed form expression for the RKHS
distance between KMEs of two d-dimensional Gaussian distributions with the kernel being
translation invariant, i.e., k(x, y) = ψ(x − y). Throughout this section Λψ will denote a
finite non-negative Borel measure corresponding to the positive-definite function ψ from
(4).
Lemma 14 Let θ0 and θ1 be KME of Gaussian measures G(µ0 , σ 2 I) and G(µ1 , σ 2 I) for
µ0 , µ1 ∈ Rd and σ 2 > 0. Suppose k(x, y) = ψ(x − y), where ψ ∈ Cb (Rd ) is positive definite.
Then
Z
2
2
2
(19)
e−σ kwk2 (1 − cos (hµ0 − µ1 , wi)) dΛψ (w).
kθ0 − θ1 k2Hk =
(2π)d/2 Rd
Proof Note that
kθ0 − θ1 k2Hk = kθ0 k2Hk + kθ1 k2Hk − 2hθ0 , θ1 iHk ,
(20)
where hθ0 , θ1 iHk = EX EY [k(X, Y )] with X ∼ G(µ0 , σ 2 I) and Y ∼ G(µ1 , σ 2 I). We will now
derive the closed form for the inner product:
Z Z
1
1
2
2
1
e− 2σ2 kx−µ0 k2 − 2σ2 ky−µ1 k2 dxdy
ψ(x − y)
hθ0 , θ1 iHk =
2
d
(2πσ )
Rd Rd
17
Tolstikhin, Sriperumbudur, and Muandet
=
Z
Rd
Z
Rd
1
(2π)d/2
Z
Rd
e−ihx−y,wi dΛψ (w)
1
1
2
2
1
e− 2σ2 kx−µ0 k2 − 2σ2 ky−µ1 k2 dx dy,
2
d
(2πσ )
where we used (4). The function appearing under the integral is absolutely integrable and
so by Tonelli-Fubini theorem (Dudley, 2002, Theorem 4.4.5) we obtain
)
(Z
1
2
Z Z
e−ihx,wi − 12 kx−µ0 k22
eihy,wi e− 2σ2 ky−µ1 k2
dx dy dΛψ (w)
e 2σ
hθ0 , θ1 iHk =
2 d/2 (2π)d/2
2 d/2
Rd (2πσ )
Rd Rd (2πσ )
Z Z
σ 2 kwk2
1
1
1
2
ihy,wi − 2σ 2 ky−µ1 k22 −ihµ0 ,wi−
2
=
e
e
e
dy dΛψ (w)
2 )d/2 (2π)d/2
d
d
(2πσ
R
R
Z Z
σ 2 kwk2
1
1
1
2
−ihµ0 ,wi−
ihy,wi − 2σ 2 ky−µ1 k22
2
dΛψ (w)
=
dy
e
e
e
d/2
2
d/2
(2π)
Rd
Rd (2πσ )
Z
2
2
σ 2 kwk2
1
2 −ihµ ,wi− σ kwk2
ihµ1 ,wi−
0
2
2
e
dΛψ (w),
=
e
(2π)d/2 Rd
where we used Lemma C.1 to compute the Fourier transform for a Gaussian density. Using
Euler’s formula and the fact that Λψ is symmetric according to Lemma C.2, while sin(x) is
an odd function, we get
Z
1
2
2
hθ0 , θ1 iHk =
(21)
cos (hµ0 − µ1 , wi) e−σ kwk2 dΛψ (w).
d/2
(2π)
Rd
The result in (19) follows by using (21) in (20).
Proof of Proposition 3: Define a := µ0 − µ1 and
Z
2
2
2
G(a) :=
e−σ kwk2 (1 − cosha, wi) dΛψ (w).
d/2
(2π)
Rd
Note that G(0) = 0. Next, since for any i = 1, . . . , d
∂ −σ2 kwk22
2
2
2
2
(1 − cosha, wi) = e−σ kwk2 wi sinha, wi ≤ e−σ kwk2 wi ∈ L1 (Λψ ),
e
∂ai
we can differentiate G under the integral sign (Folland, 1999, Theorem 2.27) and get
∇G(0) = 0.
If a function f : Rd → R is strongly convex with parameter m > 0 on some set A ⊆ Rd ,
then for all x, y ∈ A:
f (x) ≥ f (y) + h∇f (y), x − yi +
m
kx − yk22 .
2
If we can show that G is strongly convex on Bǫ := {b ∈ Rd : kbk22 ≤ ǫ} for some ǫ > 0, then
we can apply previous inequality with y = 0 and x = a to obtain
G(a) ≥
m
kak22 , ∀a ∈ Bǫ .
2
It is known that a twice continuously differentiable function f is strongly convex on A ⊆ Rd
with parameter m > 0 if the matrix ∇2 f (x) − m · I is positive definite for all x ∈ A, where
18
Minimax Estimation of Kernel Mean Embeddings
∇2 f is the Hessian and I ∈ Rd×d is an identity matrix. Next we compute the Hessian of G
by once again employing differentiation under the integral sign (justified in the similar way
as above) to obtain
2
∂ 2 G(a)
=
∂ai ∂aj
(2π)d/2
Z
e−σ
2 kwk2
2
Rd
wi wj cos (ha, wi) dΛψ (w),
0 ≤ i, j ≤ d.
Thus
2
∇ G(a) =
(2π)d/2
2
Z
e−σ
2 kwk2
2
Rd
wwT cos (ha, wi) dΛψ (w).
In order to prove that G is strongly convex on Bǫ ⊆ Rd we need to show that ∇2 G(a) −
m · I is positive definite for each a ∈ Bǫ and some m > 0. In other words, we need to show
that there is m > 0 such that for each z ∈ Rd \ {0} and a ∈ Bǫ the following holds:
hz, ∇2 G(a)zi ≥ mkzk22 ,
or, equivalently,
2
(2π)d/2
Z
Rd
e−σ
2 kwk2
2
hez , wi2 e−iha,wi dΛψ (w) ≥ m,
(22)
where ez := z/kzk2 ∈ Rd is a vector of unit length pointed in the direction of z. Note that
l.h.s. of (22) is the Fourier transform of a measure Tz on Rd , which is absolutely continuous
2
2
with respect to Λψ with Radon-Nikodym derivative 2e−σ kwk2 hez , wi2 .
Fix any z ∈ Rd . We will first show that we can apply Bochner’s Theorem (see (4)) for
the measure Tz . For this we need to check that it is (a) non-negative and (b) finite. Part
(a) is apparent from the facts that Λψ is non-negative and Tz has a non-negative density
with respect to Λψ . To check (b) we write
Z
Z
2
2
2
dTz (x) =
e−σ kwk2 hez , wi2 dΛψ (w) < ∞,
d/2
(2π)
Rd
Rd
2
2
as e−σ kwk2 hez , wi2 is positive and bounded for any z ∈ Rd , while Λψ is finite. We conclude
from Bochner’s Theorem, that function ψ̃z (x) defined in the following way:
Z
e−ihx,wi dTz (w),
ψ̃z (x) =
Rd
is positive-definite. Moreover it is well known (Dudley, 2002, Theorem 9.4.4) that ψ̃z ∈
Cb (Rd ) as ψ̃z is the characteristic function of Tz . Finally, it follows from the discussion
in (Sriperumbudur et al., 2011, Section 3.3), that if supp(Tz ) = Rd , then a bounded and
continuous function ψ̃z (x) is strictly positive definite. To check the condition supp(Tz ) = Rd
we note that supp(Λψ ) = Rd since ψ is characteristic (Sriperumbudur et al., 2010, Theorem
2
2
9), and no open sets of Rd are contained in the region where e−σ kwk2 hez , wi2 = 0.
Summarizing, we have established that the l.h.s. of (22) is equal to ψ̃z (a), where
ψ̃z : Rd → R is a bounded, continuous and strictly positive definite function for each
19
Tolstikhin, Sriperumbudur, and Muandet
z ∈ Rd \ {0}. In particular, we have ψ̃z (0) > 0 for all z ∈ Rd . Note that ψ̃z (0) depends on
z only through its direction. Next we want to show that
inf ψ̃z (0) > 0,
z∈Sd
(23)
where the infimum is over the unit sphere Sd := {b ∈ Rd : kbk22 = 1}. Note that the function
F : z → ψ̃z (0) defined on Sd is continuous. Since Sd is closed and bounded, we know that
F attains its minimum on it. In other words, there is z ∗ ∈ Sd , such that
inf ψ̃z (0) = ψ̃z ∗ (0).
z∈Sd
Thus, if inf z∈Sd ψ̃z (0) = 0, we will also get ψ̃z ∗ (0) = 0, which will contradict the fact that
ψ̃z (0) > 0 for each z ∈ Rd \{0}. This proves (23). Using Lemma C.3 we also conclude that
inf z∈Sd ψ̃z : Rd → R is a continuous function. Now we may finally conclude that there are
constants cψ,σ2 , ǫψ,σ2 > 0 such that
inf ψ̃z (a) ≥ cψ,σ2
z∈Sd
for all a ∈ Bǫψ,σ2 . Finally, we take m = cψ,σ2 and this concludes the proof.
5.4 Proof of Theorem 8
The proof is based on application of Theorem B.2 where we choose θ0 , . . . , θM to be KMEs
of d-dimensional Gaussian measures with variances decaying to zero as d → ∞.
Let G(µ0 , σ 2 I) and G(µ1 , σ 2 I) be two d-dimensional Gaussian distributions with mean
vectors µ0 , µ1 ∈ Rd and variance σ 2 > 0. Define θ0 and θ1 to be the embeddings of
G(µ0 , σ 2 I) and G(µ1 , σ 2 I) respectively.
(A) Deriving a closed form expression for kθ0 − θ1 k2Hk .
Using Lemma 14, presented in Section 5.3, we have
Z
2
2
2
2
e−σ kwk2 (1 − cos (hµ0 − µ1 , wi)) dΛψ (w),
kθ0 − θ1 kHk =
d/2
(2π)
Rd
(24)
where Λψ is a finite non-negative Borel measure from the Bochner’s Theorem corresponding
to the kernel k. We now show that Λψ is absolutely continuous with respect to the Lebesgue
measure on Rd and has the following density:
Z ∞
kwk2
1
− 4t 2
e
λψ (w) =
dν(t), w ∈ Rd .
(2t)d/2
0
Indeed, by noticing that
Z ∞ Z
0
−ihw,xi
e
Rd
kwk2
1
− 4t 2
dw dν(t) < ∞
e
(2t)d/2
20
Minimax Estimation of Kernel Mean Embeddings
we may apply Tonelli-Fubini theorem (Dudley, 2002, Theorem 4.4.5) to interchange the
order of integration and get
!
Z ∞
Z ∞ Z
Z
−ihw,xi
kwk2
kwk2
e
e−ihw,xi
1
1
2
2
e− 4t dν(t) dw =
e− 4t dw dν(t)
d/2
d/2 (2t)d/2
(2t)d/2
Rd (2π)
0
Rd (2π)
0
Z ∞
2
e−tkxk2 dν(t).
=
0
Substituting the form of λψ into (24) we can write
kθ0 − θ1 k2Hk =
2
(2π)d/2
Z
Rd
Z
∞
e−σ
2 kwk2
2
0
(1 − cos (hµ0 − µ1 , wi))
kwk2
1
− 4t 2
dν(t) dw.
e
(2t)d/2
Applying Tonelli-Fubini theorem once again and using Lemma C.1 together with Euler’s
formula we obtain
Z ∞
Z
kwk2
1
2
1
)
− 2 2 (2σ2 + 2t
(1 − cos (hµ0 − µ1 , wi)) dw dν(t)
kθ0 − θ1 k2Hk =
e
d/2
d/2
d
(2π)
(2t)
R
0
Z ∞
Z ∞
tkµ0 −µ1 k2
2
2
2
−
4σ 2 t+1 dν(t)
=
dν(t)
−
e
(4σ 2 t + 1)d/2
(4σ 2 t + 1)d/2
0
0
d/2
Z ∞
tkµ0 − µ1 k22
1
1 − exp −
dν(t).
(25)
2
=
1 + 4tσ 2
1 + 4tσ 2
0
(B) Lower bounding kθ0 − θ1 k2Hk in terms of kµ0 − µ1 k22 .
It follows from (25) that
kθ0 −
θ1 k2Hk
≥
Z
t1
2
t0
1
1 + 4tσ 2
d/2
where 0 < t0 ≤ t1 < ∞. Note that 1 − e−x ≥
tkµ0 − µ1 k22
1 − exp −
1 + 4tσ 2
x
2
≥
tkµ0 − µ1 k22
1 − exp −
1 + 4tσ 2
dν(t),
for 0 ≤ x ≤ 1. Using this we get
tkµ0 − µ1 k22
,
2(1 + 4tσ 2 )
∀t ∈ [t0 , t1 ]
as long as
t1 kµ0 − µ1 k22 ≤ 1 + 4t1 σ 2 .
(26)
Thus, as long as (26) holds, we can lower bound the RKHS distance as:
kθ0 −
θ1 k2Hk
d/2
tkµ0 − µ1 k22
1
dν(t)
≥
1 + 4tσ 2
1 + 4tσ 2
t0
Z t1
t
dν(t).
= kµ0 − µ1 k22
2 )(d+2)/2
(1
+
4tσ
t0
Z
t1
21
(27)
Tolstikhin, Sriperumbudur, and Muandet
Note that the function t 7→
global maximum at t =
Z
t1
t0
1
2dσ2
t
(1+4tσ2 )(d+2)/2
1
monotonically increases on [0, 2dσ
2 ], reaches its
1
and then decreases on [ 2dσ
2 , ∞). Thus we have
t
dν(t) ≥ β min
(1 + 4tσ 2 )(d+2)/2
t0
t1
,
(1 + 4t0 σ 2 )(d+2)/2 (1 + 4t1 σ 2 )(d+2)/2
Setting
σ2 =
1
2t1 d
yields that t = t1 is the global maximum of the function t 7→
.
(28)
1+ t2td
1
t
(d+2)/2 ,
in which case
t0
t1
≤
.
2
(d+2)/2
(1 + 4t0 σ )
(1 + 4t1 σ 2 )(d+2)/2
With this choice of σ 2 we get
Z t1
βt0
t
βt0
dν(t) ≥
(d+2)/2 ≥
(d+2)/2
2
(d+2)/2
t0 (1 + 4tσ )
0
1 + d2
1 + 2t
t1 d
(d+2)/2
βt0
2
2
= βt0 1 −
≥
1−
,
2+d
e
2+d
(29)
where we used the fact that (1 − x1 )x−1 monotonically decreases to 1e . Using (29) in (27),
we obtain
βt0
2
2
kθ0 − θ1 kHk ≥
1−
kµ0 − µ1 k22 .
(30)
e
2+d
(C.1) Application of Theorem B.2: Choosing θ0 , . . . , θM .
Now we are going to apply Theorem B.2. First of all, we need to choose M + 1 embeddings. Recall that Theorem B.2 requires these embeddings to be sufficiently distant from
each other, while the corresponding distributions should be close. We will choose the embeddings {θ0 , . . . , θM } to be KMEs of Gaussian distributions G(µi , σ 2 I) for specific choice
d
M
of σ 2 > 0 and µi ∈ R
, i =d 0, . . . 2, M . Mean vectors {µi }i=0 will be constrained to live in
the ball B(cν , n) := x ∈ R : kxk2 ≤ cν /n , where cν is a positive constant to be specified
later. This guarantees that KL-divergences between the Gaussian distributions will remain
small. At the same time, it was shown in (30) that the RKHS distance between embeddings
θi and θj is lower bounded by the Euclidean distance between µi and µj . In other words,
in order for the embeddings θ0 , . . . , θM to be sufficiently separated we need to make sure
that the mean vectors µ0 , . . . , µM are not too close to each other. Summarizing, we face
the problem of choosing a finite collection of pairwise distant points in the Euclidean ball.
This question is closely related to the concepts of packing and covering numbers.
For any set A ∈ Rd its ǫ-packing number M (A, ǫ) is the largest number of points in A
separated from each other by at least a distance of ǫ. An ǫ-covering number N (A, ǫ) of A
22
Minimax Estimation of Kernel Mean Embeddings
is the minimal number of balls of radius ǫ needed to cover A. Packing numbers are lower
bounded by the covering numbers (Dudley, 1999, Theorem 1.2.1):
N (A, ǫ) ≤ M (A, ǫ).
Also, it is well known that the ǫ-covering number of a unit d-dimensional Euclidean ball is
lower bounded by ⌊ǫ−d ⌋. Together, these facts state that we can find at least ⌊ǫ−d ⌋ points
in the d-dimensional unit ball, which are at least ǫ away from each other. Similarly (just by
a simple scaling) we can find at least ⌊ǫ−d ⌋ points in the d-dimensional ball of radius R > 0,
which are at least R · ǫ away from each other. Applying this fact to B(cν p
, n) we can finally
d
−1
argue that there are at least N points in B(cν , n) which are at least N
cν /n away from
each other, where N ≥ 3 is an integer to be specified later. Now, take M = N d − 1 ≥ 2
(which explains the lower bound N ≥ 3) and fix µ0 , . . . , µM to be these M + 1 points.
(C.2) Application of Theorem B.2: Lower bounding kθi − θj kHk .
With this choice of parameters µ0 , . . . , µM , for any 0 ≤ i < j ≤ M , we have
2
cν βt0
2
1−
,
kθi − θj kHk ≥ 2
N n e
2+d
where we used (30) and the lower bound on kµi − µj k2 . Setting
cν =
C
βt0
(31)
for some C > 0 we obtain
kθi −
θj k2Hk
C
≥ 2
N en
2
1−
2+d
This satisfies the first assumption of Theorem B.2 with s :=
.
1
2N
r
C
en
1−
2
2+d
.
(C.3) Application of Theorem B.2: Upper bounding KL(Pθi kPθj ).
Note that for any 0 ≤ i < j ≤ M we have
2cν
t1 d
kµi − µj k22
≤ 2 = 4C
,
KL Gn (µi , σ 2 I)kGn (µj , σ 2 I) = n ·
2
2σ
σ
βt0
where the inequality holds since µi ∈ B(cν , n) and the equality follows from (28) and (31).
Here we used the fact that for any points x and y contained in a ball of radius R we obviously
have kx − yk ≤ 2R. Also note that we chose M = N d − 1 ≥ 2 and thus
log(M ) = d log(N ) + log(1 − N −d ) ≥ d log(N ) + 1 −
23
1
Nd
≥ d log(N ) −
Nd − 1
N −1
Tolstikhin, Sriperumbudur, and Muandet
for d ≥ 1, where we used the inequality log(x) ≥ 1 − x1 which holds for x ≥ 0. Taking
βt0
1
log N −
C=
(32)
32t1
N −1
we get
1
t1 d
=
4C
βt0
8
d
d log N −
N −1
for any d ≥ 1. Concluding, we get
1
≤
8
1
d log N −
N −1
≤
1
log(M )
8
1
KL Gn (µi , σ 2 I)kGn (µj kσ 2 I) ≤ log(M )
8
and thus the second assumption of Theorem B.2 is satisfied with α = 18 . Finally, it is easy
to check that if we take N = 5 then condition (26) will be satisfied. Indeed,
4t1 βt0
1
4t1 cν
1
1
2
=
log N −
t1 kµ0 − µ1 k2 ≤
=
log N −
,
n
βt0 n 32t1
N −1
8n
N −1
while
1 + 4t1 σ 2 ≥ 1.
Thus (26) holds whenever
1
≤ 8n,
N −1
which obviously holds for N = 5 and any n ≥ 1.
r
1
C
1−
To conclude the proof we insert (32) into s := 2N en
4
, and notice that
using 81 log N − N 1−1 ≥ 25
log N −
√
M
√
1+ M
1 − 2α −
s
2α
log(M )
!
≥
2
2+d
, lower bound this value
1
.
5
5.5 Proof of Theorem 9
The proof is based on the following result, which gives a closed form expression for the
L2 (Rd ) distance between embeddings of two discrete distributions supported on the same
pair of points in Rd .
Lemma 15 Suppose k(x, y) = ψ(x − y), where ψ ∈ L2 (Rd ) ∩ Cb (Rd ) is positive definite
and k is characteristic. Define P0 = p0 δx + (1 − p0 )δv and P1 = p1 δx + (1 − p1 )δv , where
0 < p0 < 1, 0 < p1 < 1, x, v ∈ Rd , and x 6= v. Then
ψ
kµk (P0 ) − µk (P1 )k2L2 (Rd ) = Cx,v
(p0 − p1 )2 ,
where
ψ
Cx,v
:= 2
kψk2L2 (Rd )
−
Z
Rd
ψ(y)ψ(y + x − v)dy
24
> 0.
Minimax Estimation of Kernel Mean Embeddings
Proof We have
kµk (P0 ) − µk (P1 )k2L2 (Rd ) = (p0 − p1 )2 kk(x, ·) − k(v, ·)k2L2 (Rd )
Z
2
ψ(x − y) − ψ(v − y) dy
= (p0 − p1 )2
d
ZR
2
(⋆)
= (p0 − p1 )2
ψ(y) − ψ(y + x − v) dy
d
ZR
ψ(y) ψ(y) − ψ(y + x − v) dy
= 2(p0 − p1 )2
d
R
Z
2
2
= 2(p0 − p1 ) kψkL2 (Rd ) −
ψ(y)ψ(y + x − v)dy ,
Rd
where we used the symmetry of ψ in (⋆). Cauchy-Schwartz inequality states that
sZ
sZ
Z
(⋆)
ψ 2 (y)dy
ψ 2 (y + x − v)dy = kψk2L2 (Rd ) ,
ψ(y)ψ(y + x − v)dy ≤
Rd
Rd
Rd
where the equality in (⋆) holds if and only if ψ(y) = λ ψ(y + x − v) for some constant λ
and all y ∈ Rd . If we take y = v − x the above condition implies ψ(v − x) = λ ψ(0) and
with y = 0 we get ψ(0) = λ ψ(x − v). Together these identities show that ψ(0) = λ2 ψ(0).
Since the kernel is characteristic and translation invariant, ψ is strictly positive definite,
which means ψ(0) > 0. We conclude that λ = ±1. Assume that λ = −1. In this case
ψ(x − v) = −ψ(0) < 0. Repeating the argument we can show that ψ 2(x − v) = ψ(0) and
generally ψ m(x − v) = (−1)m ψ(0) for all m ∈ N. Since ψ ∈ L2 (Rd ) we need ψ 2 to be
integrable on Rd . Summarizing, we showed that a non-negative, integrable, and continuous
function takes the same strictly positive value ψ 2 (0) > 0 infinitely many times, leading to
a contradiction. Arguing similarly for λ = 1 will result in a contradiction. This means the
equality in (⋆) is never attained which concludes the proof.
The proof of Theorem 9 is carried out by simply repeating the proof of Theorem 1 but
replacing (18) with the result in Lemma 15 and using x − v := z.
5.6 Proof of Corollary 10
The proof will be based on Theorem 9. The moment condition (15) on ν is sufficient for
ψν ∈ L2 (Rd ) Rto hold (see Remark A.3). Thus we only need to compute the expression
kψν k2L2 (Rd ) − Rd ψν (y)ψν (y + z)dy appearing in (14). Note that
Z Z ∞Z ∞
Z
2
2
(33)
e−t1 kyk2 −t2 ky+zk2 dν(t1 )dν(t2 )dy.
ψν (y)ψν (y + z)dy =
Rd
Rd
Since
Z ∞Z ∞Z
0
0
Rd
2
2
0
0
∞
d/2 t t kzk2
2
π
− 1 2
e t1 +t2 dν(t1 )dν(t2 )
t1 + t2
0
0
Z ∞ d/2
π
dν(t1 ) < ∞,
≤ ν([0, ∞))
t
1
0
e−t1 kyk2 −t2 ky+zk2 dy dν(t1 )dν(t2 ) =
25
Z ∞Z
Tolstikhin, Sriperumbudur, and Muandet
we may apply Tonelli-Fubini theorem to switch the order of integration in (33) and get
d/2 t t kzk2
Z ∞Z ∞
Z
π
− 1 2 2
e t1 +t2 dν(t1 )dν(t2 ).
ψν (y)ψν (y + z)dy =
t1 + t2
0
0
Rd
Using this we get
!
d/2
t1 t2 kzk2
2
π
−
dν(t1 )dν(t2 )
1 − e t1 +t2
ψν (y)ψν (y + z)dy =
kψν k2L2 (Rd ) −
t1 + t2
0
0
Rd
!
Z Z
t t kzk2
π d/2 δ1 δ1
− 1t 2+t 2
≥
1−e 1 2
dν(t1 )dν(t2 ).
2δ1
δ0 δ0
Z ∞Z
Z
∞
Since t1 and t2 are bounded below by δ0 > 0 we may take kzk2 large enough so that the
following will hold:
Z
β2
π d/2
2
.
ψν (y)ψν (y + z)dy ≥
kψν kL2 (Rd ) −
2 2δ1
Rd
5.7 Proof of Proposition 11
The proof is based on the following result, which provides a closed form expression for the
L2 (Rd ) distance between the embeddings of Gaussian measures.
Lemma 16 Let θ0 and θ1 be KME of Gaussian measures G(µ0 , σ 2 I) and G(µ1 , σ 2 I) for
µ0 , µ1 ∈ Rd and σ 2 > 0. Suppose k(x, y) = ψ(x − y), where ψ ∈ L1 (Rd ) ∩ Cb (Rd ) is positive
definite and and k is characteristic. Then
Z
2
2
2
(34)
1 − e−ihw,µ0 −µ1 i ψ ∧ (w) e−σ kwk dw.
kθ0 − θ1 k2L2 (Rd ) = 2
Rd
Proof First of all, note that ψ ∈ L2 (Rd ) since ψ ∈ L1 (Rd ) and ψ is bounded. This
shows that θ0 , θ1 ∈ L2 (Rd ). We will use P0 and P1 to denote the corresponding Gaussian
distributions G(µ0 , σ 2 I) and G(µ1 , σ 2 I). By definition we have
Z
Z Z
k(z, y) dP1 (z) dy
k(x, y) dP0 (x)
hθ0 , θ1 iL2 (Rd ) =
Rd
Rd
Rd
Z
Z Z
k(x, y)k(z, y) dP0 (x) dP1 (z) dy.
=
Rd
Rd
Rd
Using the fact that ψ is bounded (Wendland, 2005, Theorem 6.2) we get
ZZZ
Z Z Z
k(x, y) dy dP0 (x) dP1 (z)
k(x, y)k(z, y) dy dP0 (x) dP1 (z) ≤ ψ(0)
Rd
Rd Rd Rd
ZZZ
ψ(x − y) dy dP0 (x)dP1 (z)
= ψ(0)
Rd Z Z
dP0 (x)dP1 (z) < ∞.
= ψ(0)kψkL1 (Rd )
Rd
26
Minimax Estimation of Kernel Mean Embeddings
This allows us to use Tonelli-Fubini theorem (Dudley, 2002, Theorem 4.4.5) and get
Z Z Z
k(x, y)k(z, y)dy dP0 (x)dP1 (z)
hθ0 , θ1 iL2 (Rd ) =
Rd Rd
Rd
Z Z Z
ψ(y)ψ(y + z − x)dy dP0 (x)dP1 (z)
=
Rd Rd
Rd
Z
Z Z Z
1
(⋆)
−ihy+z−x,wi
e
dΛψ (w)dy dP0 (x)dP1 (z)
ψ(y)
=
(2π)d/2 Rd Rd
Rd
Rd
Z Z Z Z
1
−ihy+z−x,wi
=
ψ(y)e
dΛψ (w)dy dP0 (x)dP1 (z),
(2π)d/2 Rd Rd
Rd Rd
where we used (4) in (⋆). Since ψ ∈ L1 (Rd ) we have
Z
Z Z
−ihy,wi
−ihz−x,wi
e
ψ(y)e
dy dΛψ (w) =
Rd
Rd
Rd
Z
Rd
|ψ(y)| dy dΛψ (w) < ∞
and thus we can use Tonelli-Fubini theorem to switch the order of integration:
Z Z Z Z
1
−ihy+z−x,wi
hθ0 , θ1 iL2 (Rd ) =
ψ(y)e
dy dΛψ (w) dP0 (x)dP1 (z)
(2π)d/2 Rd Rd
Rd Rd
Z
Z Z
ψ ∧ (w)e−ihz−x,wi dΛψ (w) dP0 (x)dP1 (z).
(35)
=
Rd
Rd
Rd
Next we are going to argue that if both ψ and ψ ∧ belong to L1 (Rd ) (the latter is true as it
follows from Wendland, 2005, Corollary 6.12) then ψ ∧ is the Radon-Nikodym derivative of
Λψ with respect to the Lebesgue measure. To this end, since ψ ∧ ∈ L1 (Rd ), Fourier inversion
theorem (Wendland, 2005, Corollary 5.24) yields that for all x ∈ Rd , the following holds:
Z
1
ψ(x) =
eihw,xi ψ ∧ (w)dw.
(2π)d/2 Rd
On the other hand, using (4) and Lemma C.2, we also have
Z
1
eihw,xi dΛψ (w).
ψ(x) =
(2π)d/2 Rd
These two identities show that for all x ∈ Rd
Z
Z
ihw,xi ∧
e
ψ (w)dw =
Rd
Rd
eihw,xi dΛψ (w).
(36)
Note that since k is translation invariant and characteristic, ψ is a strictly positive definite
function (Sriperumbudur et al., 2010, Section 3.4) and therefore it follows from (Wendland,
2005, Theorem 6.11) that ψ ∧ is non-negative (and nonvanishing). Since ψ ∧ ∈ L1 (Rd ) we
conclude that ψ ∧ is the Radon-Nikodym derivative of a finite non-negative measure Tψ on
Rd , which is absolutely continuous with respect to the Lebesgue measure. (36) (after proper
normalization) shows that the characteristic functions of measures Λψ and Tψ coincide. We
finally conclude from (Dudley, 2002, Theorem 9.5.1) that Λψ = Tψ , which means that Λψ
is absolutely continuous with respect to the Lebesgue measure and has a density ψ ∧ .
27
Tolstikhin, Sriperumbudur, and Muandet
Returning to (35) we can write it as
Z Z Z
2 −ihz−x,wi
∧
hθ0 , θ1 iL2 (Rd ) =
ψ (w) e
dw dP0 (x)dP1 (z).
Rd
Rd
Rd
We already showed that ψ ∈ L2 (Rd ). From Plancherel’s theorem (Wendland, 2005, Corollary 5.25), we have kψkL2 (Rd ) = kψ ∧ kL2 (Rd ) and thus ψ ∧ ∈ L2 (Rd ). Another application of
Tonelli-Fubini theorem yields
Z Z
Z
2
−ihz−x,wi
∧
e
dP0 (x)dP1 (z) dw
hθ0 , θ1 iL2 (Rd ) =
ψ (w)
d
Rd Rd
ZR
2
2
2
ψ ∧ (w) eihw,µ0 −µ1 i e−σ kwk2 dw.
=
Rd
Noticing that the Fourier transform of a real and even function is also even, we conclude
that ψ ∧ is also even. This finishes the proof since kθ0 −θ1 k2L2 (Rd ) = kθ1 k2L2 (Rd ) +kθ0 k2L2 (Rd ) −
2hθ0 , θ1 iL2 (Rd ) .
Now we turn to the proof of Theorem 11. We will write Λ̃ψ to denote a non-negative
finite measure, absolutely continuous with respect to the Lebesgue measure with density
(2π)d/2 (ψ ∧ )2 . Then (34) in Lemma 16 can be written as
Z
2
2
2
2
ihw,µ0 −µ1 i
1
−
e
e−σ kwk2 dΛ̃ψ (w).
kθ0 − θ1 kL2 (Rd ) =
d/2
d
(2π)
ZR
2
2
2
=
(1 − cos(hw, µ0 − µ1 i)) e−σ kwk2 dΛ̃ψ (w),
(2π)d/2 Rd
which is exactly of the form in Lemma 14 but with Λψ replaced by Λ̃ψ . From the proof of
Lemma 16, since ψ ∧ is even and non-vanishing, the corresponding measure Λ̃ψ is symmetric
and supp(Λ̃ψ ) = Rd . The result therefore follows by carrying out the proof of Proposition 3
verbatim but for replacing Λψ with Λ̃ψ .
5.8 Proof of Theorem 13
The proof will closely follow that of Theorem 8. Let G(µ0 , σ 2 I) and G(µ1 , σ 2 I) be two
d-dimensional Gaussian distributions with mean vectors µ0 , µ1 ∈ Rd and variance σ 2 > 0.
Let θ0 and θ1 denote the kernel mean embeddings of G(µ0 , σ 2 I) and G(µ1 , σ 2 I) respectively.
(A) Deriving a closed form expression for kθ0 − θ1 k2L2 (Rd ) .
The condition in (15) ensures that ψν ∈ L2 (Rd ) (see Remark A.3). In fact, using a similar
argument it can be shown that ψν ∈ L1 (Rd ). Also it is easy to verify that ψν ∈ Cb (Rd ).
Next, under this moment condition we may apply Tonelli-Fubini theorem to compute the
Fourier transform of ψν :
Z ∞
Z ∞
Z
kwk2
1
1
−ihw,xi
−tkxk22
− 4t 2
∧
e
dν(t).
(37)
dν(t)dx
=
e
e
ψν (w) =
(2π)d/2 Rd
(2t)d/2
0
0
28
Minimax Estimation of Kernel Mean Embeddings
It is immediate to see that ψν∧ ∈ L1 (Rd ). Therefore Lemma 16 yields
Z
2
2
2
2
1 − e−ihw,µ0 −µ1 i ψν∧ (w) e−σ kwk2 dw.
kθ0 − θ1 kL2 (Rd ) = 2
(38)
Rd
σ 2 kwk2
2
Denoting G(w) := ψν∧ (w)e− 2 and using a well-known property of the Fourier transform
we get
Z
Z
1
1
−ihw,τ i 2
2 ∧
e
G (w)dw =
G∧ (x)G∧ (τ − x)dx.
(39)
(G ) (τ ) =
(2π)d/2 Rd
(2π)d/2 Rd
Next we compute the Fourier transform of G using (37):
Z ∞
Z
kwk2
σ 2 kwk2
1
1
2
∧
−ihw,xi
− 4t 2
−
2
G (x) =
e
e
dν(t)
e
dw.
d/2
d/2
(2π)
(2t)
Rd
0
Using the moment condition on ν we have
Z
0
∞
Z
1
(4πt)d/2
−
e
kxk2
2
4t
Rd
−
e
σ 2 kxk2
2
2
Z ∞
1 −d/2
1
t+ 2
dν(t)
dxdν(t) =
2σ
(2σ 2 )d/2 0
Z ∞
1
≤
t−d/2 dν(t) < ∞.
(2σ 2 )d/2 0
(40)
This allows us to use Tonelli-Fubini theorem and write
Z ∞
Z
σ 2 kwk2
kwk2
1
1
2
−ihw,xi − 4t 2 −
2
e
dw dν(t)
e
e
G∧ (x) =
(2π)d/2 0 (2t)d/2 Rd
Z ∞
1
tkxk22
dν(t).
=
exp −
2tσ 2 + 1
(2tσ 2 + 1)d/2
0
Returning to (39) and denoting ∆1 := 2t1 σ 2 + 1, ∆2 := 2t2 σ 2 + 1 we obtain
2 ∧
(G ) (τ ) =
Z
Rd
Z
0
∞Z ∞
0
1
t1 kxk22 t2 kτ − xk22
exp −
−
dν(t1 ) dν(t2 ) dx.
∆1
∆2
(2π∆1 ∆2 )d/2
Using a simple identity
akxk22 + bkx − yk22 = (a + b) x −
b
y
a+b
2
+
2
ab
kyk22 ,
a+b
which holds for any x, y ∈ Rd and a, b ∈ R with a + b 6= 0, we obtain
(G2 )∧ (τ ) =
Z
Rd
Z
0
∞Z ∞
0
1
t2
t2 ∆ 1 τ
t1
+
x−
exp −
d/2
∆1 ∆2
t1 ∆ 2 + t2 ∆ 1
(2π∆1 ∆2 )
t1 t2 kτ k22
× exp −
dν(t1 ) dν(t2 ) dx.
t1 ∆ 2 + t2 ∆ 1
29
2
2
!
Tolstikhin, Sriperumbudur, and Muandet
Using an argument similar to (40) we can show that Tonelli-Fubini theorem is applicable
to the r.h.s. of the above equation. Therefore, changing the order of integration we get
d/2
Z ∞Z ∞
1
t1 t2 kτ k22
2 ∧
(G ) (τ ) =
exp −
dν(t1 ) dν(t2 ).
2(t1 ∆2 + t2 ∆1 )
t1 ∆ 2 + t2 ∆ 1
0
0
R
Noticing that (G2 )∧ (0) = (2π)1d/2 Rd G2 (w) dw and returning to (38) we get
kθ0 − θ1 k2L2 (Rd ) = 2(2π)d/2 (G2 )∧ (0) − (G2 )∧ (µ0 − µ1 )
d/2
Z ∞Z ∞
t1 t2 kµ0 − µ1 k22
π
1 − exp −
dν(t1 )dν(t2 ).
=2
t1 ∆ 2 + t2 ∆ 1
t1 ∆ 2 + t2 ∆ 1
0
0
(B) Lower bounding kθ0 − θ1 k2L2 (Rd ) in terms of kµ0 − µ1 k22 .
Consider
kθ0 − θ1 k2L2 (Rd )
d/2
Z δ1Z δ1
π
t1 t2 kµ0 − µ1 k22
≥2
1 − exp −
dν(t1 )dν(t2 )
t1 ∆ 2 + t2 ∆ 1
t1 ∆ 2 + t2 ∆ 1
δ0 δ0
d/2
Z δ1Z δ1
π
t1 t2 kµ0 − µ1 k22
=2
1 − exp −
dν(t1 )dν(t2 ).
4t1 t2 σ 2 + t1 + t2
4t1 t2 σ 2 + t1 + t2
δ0 δ0
Using the fact that 1 − e−x ≥ x2 for 0 ≤ x ≤ 1, we obtain
d/2
Z δ1Z δ1
π
t1 t2 kµ0 − µ1 k22
2
dν(t1 ) dν(t2 )
kθ0 − θ1 kL2 (Rd ) ≥
4t1 t2 σ 2 + t1 + t2
4t1 t2 σ 2 + t1 + t2
δ0 δ0
(41)
whenever
t1 t2 kµ0 − µ1 k22
≤ 1.
4t1 t2 σ 2 + t1 + t2
Note that the expression on the left hand side of the previous inequality is increasing both
in t1 and t2 . This means that for t1 , t2 ∈ [δ0 , δ1 ] we have:
t1 t2 kµ0 − µ1 k22
δ1 kµ0 − µ1 k22
≤
4t1 t2 σ 2 + t1 + t2
4δ1 σ 2 + 2
and thus (41) holds whenever
δ1 kµ0 − µ1 k22 ≤ 4δ1 σ 2 + 2
which will be satisfied later. (41) can be rewritten as
Z δ1Z δ1
(π)d/2 t1 t2 kµ0 − µ1 k22
2
kθ0 − θ1 kL2 (Rd ) ≥
dν(t1 ) dν(t2 )
2
d/2+1
δ0 δ0 (4t1 t2 σ + t1 + t2 )
d/2 Z δ1
Z δ1
t1 t2
2
π
t1 +t2 kµ0 − µ1 k2
=
dν(t1 ) dν(t2 )
d/2+1
t2 2
t1 + t2
δ0
δ0
4 t1t1+t
σ
+
1
2
d/2 Z δ1Z δ1
S(t1 , t2 )kµ0 − µ1 k22
π
≥
d/2+1 dν(t1 ) dν(t2 ),
2δ1
δ0 δ0
4S(t1 , t2 )σ 2 + 1
30
(42)
Minimax Estimation of Kernel Mean Embeddings
t2
. Note that S(t1 , t2 ) takes values in [ δ20 , δ21 ] as t1 and t2 varies
where S(t1 , t2 ) := t1t1+t
2
in [δ0 , δ1 ]. We can now repeat part of the proof of Theorem 8 where we showed that the
1
function t 7→ (1+4tσ2t)(d+2)/2 monotonically increases on [0, 2dσ
2 ], reaches its global maximum
at t =
Z δZ1
1
2dσ2 ,
1
and then decreases on [ 2dσ
2 , ∞). Using this fact we have
)
(
β2
δ1
S(t1 , t2 )
δ0
,
.
d
d
d +1 dν(t1 )dν(t2 ) ≥ 2 min
(1 + 2δ0 σ 2 ) 2 +1 (1 + 2δ1 σ 2 ) 2 +1
4S(t , t )σ 2 + 1 2
δ1
δ0 δ0
1
2
By setting σ 2 := δ11d we ensure that t = δ21 is the global maximum of the function t 7→
δ0
δ1
t
d +1 ≤
d +1 . Combining this with (42) we have
d +1 and thus
1+ δ4td
(1+2δ0 σ2 ) 2
2
(1+2δ1 σ2 ) 2
1
kθ0 −
≥
θ1 k2L2 (Rd )
≥
β 2 δ0
2(1 +
2 d2 +1
d)
β 2 δ0
2
π
2δ1
1 + δ2δ1 d0
d/2
d +1
2
kµ0 −
π
2δ1
µ1 k22
d/2
kµ0 − µ1 k22
β 2 δ0
≥
2e
2
1−
2+d
π
2δ1
d/2
kµ0 − µ1 k22 ,
where we used an analysis similar to (29).
(C) Application of Theorem B.2.
We finish by repeating all the remaining steps carried out in the proof of Theorem 8
(steps C.1, C.2, and C.3), where we set
1
1
1
2
and cν :=
log N −
.
σ :=
δ1 d
16δ1
N −1
Acknowledgements
The authors thank the action editor and two anonymous reviewers for their detailed comments, which helped to improve the presentation. The authors would also like to thank
David Lopez-Paz, Jonas Peters, Bernhard Schölkopf, and Carl-Johann Simon-Gabriel for
useful discussions.
Appendix A.
√
n-consistency of µk (Pn )
In the following, we present a general result whose special cases establishes the convergence
rate of n−1/2 for kµk (Pn ) − µk (P )kF when F = Hk and F = L2 (Rd ).
Proposition A.1 Let (Xi )ni=1 be random samples drawn i.i.d. from P defined on a separable topological space X . Suppose r : X → H is continuous and
sup kr(x)k2H ≤ Ck < ∞,
x∈X
31
(43)
Tolstikhin, Sriperumbudur, and Muandet
where H is a separable Hilbert space of real-valued functions. Then for any 0 < δ ≤ 1 with
probability at least 1 − δ we have
Z
X
r(x) dPn (x) −
Z
r(x) dP (x)
X
≤
H
r
Ck
+
n
r
2Ck log(1/δ)
.
n
Proof Note that r : X → H is a H-valued measurable function as r is continuous and
H is separable
2008, Lemma A.5.18). The condition
in (43)
R
R (Steinwart and Christmann,
√
1
ensures that kr(x)kH dQ(x) ≤ Ck < ∞ for any Q ∈ M+ (X ) and therefore r(x) dQ(x)
is well defined as a Bochner integral for any Q ∈ M+1 (X ) (Diestel and Uhl, 1977, Theorem
2, p.45). By McDiarmid’s inequality, it is easy to verify that with probability at least 1 − δ,
Z
X
r(x) dPn (x) −
Z
r(x) dP (x)
X
H
Z
≤E
X
Z
r(x) dP (x)
r(x) dPn (x) −
X
r
2Ck log(1/δ)
,
+
n
H
(44)
where
E
=
Z
X
s
r(x) dPn (x) −
E
Z
Z
r(x) dP (x)
X
H
2
r(x) dPn (x)
X
+
+
Z
X
2
H
Z
E
X
r(x) dPn (x) −
2
Z
H
v
u Z
u
r(x) dPn (x)
= tE
≤
s
r(x) dP (x)
X
H
2
r(x) dP (x)
X
H
− 2E
Z
Z
2
r(x) dP (x)
X
H
r(x) dPn (x),
X
Z
X
r(x) dP (x)
H
Z
n
2X
−
r(x) dP (x)
. (45)
E r(Xi ),
n
X
H
i=1
To simplify the r.h.s. of (45), we make the following observation. Note that for
R any g ∈ H,
Tg : H → R, f 7→ hg, f iH is a bounded linear functional on H. Choose f = X r(y) dP (y).
It follows from (Diestel and Uhl, 1977, Theorem 6, p.47) that
Z
r(y) dP (y)
g,
X
= Tg
Z
X
H
Z
Z
hg, r(y)iH dP (y).
Tg (r(y)) dP (y) =
r(y) dP (y) =
X
X
Applying (46) to the third term in the r.h.s. of (45) with g =
Z
r(x) dP (x)
E r(Xi ),
X
=
H
Z
X
hr(xi ), giH dP (xi ) =
Z
R
X
(46)
r(x) dP (x), we obtain
r(xi ) dP (xi ), g
X
H
= kgk2H
and so (45) reduces to
E
Z
X
r(x) dPn (x) −
Z
r(x) dP (x)
X
H
≤
s
E
32
Z
X
2
r(x) dPn (x)
H
−
Z
2
.
r(x) dP (x)
X
H
(47)
Minimax Estimation of Kernel Mean Embeddings
Consider
E
Z
2
r(x) dPn (x)
X
H
2
n
n
1 X
1X
Ehr(Xi ), r(Xj )iH
r(Xi )
= 2
=E
n
n
i,j=1
i=1
H
1 X
1 X
Ehr(Xi ), r(Xj )iH + 2
= 2
Ehr(Xi ), r(Xj )iH
n
n
i=j
=
i6=j
1
n−1
EX∼P kr(X)k2H +
EX∼P,Y ∼P hr(X), r(Y )iH .
n
n
Using (46), the second term in (48) can be equivalently written as
Z Z
hr(x), r(y)iH dP (y) dP (x)
EX∼P,Y ∼P hr(X), r(Y )iH =
ZX X Z
(⋆)
r(y) dP (y)
dP (x)
=
r(x),
X
(⋆)
=
Z
X
(48)
H
r(x) dP (x),
X
Z
X
Z
r(y) dP (y)
x=
r(x) dP (x)
H
X
2
,
H
where we invoked (46) in (⋆). Combining the above with (48) and using the result in (47)
yields
s
r
R
Z
Z
EX∼P kr(X)k2H − k X r(x) dP (x)k2H
Ck
r(x) dP (x)
≤
r(x) dPn (x) −
E
≤
n
n
X
X
H
and the result follows.
Remark A.2 Suppose H is an RKHS with a reproducing kernel k that is continuous and
satisfies supx∈X k(x, x) < ∞. Choosing r(x) = k(·, x), x ∈ X in Proposition A.1 yields a
concentration inequality for kµk (Pn ) − µk (P )kH with Ck := supx∈X k(x, x), thereby establishing a convergence rate of n−1/2 for kµk (Pn ) − µk (P )kH . While such a result has already
appeared in Smola et al. (2007, Theorem 2), Gretton et al. (2012) and Lopez-Paz et al.
(2015), the result derived from Proposition A.1 improves upon them by providing better
constants. While all these works including Proposition A.1 are based on McDiarmid’s inequality (see (44)), the latter obtains better constants by carefully bounding the expectation
term in (44). It is easy to verify that Ck = 1 for Gaussian and Ck = CM for mixture of
Gaussian kernels, Ck = c−2γ for inverse multiquadrics, and Ck = 1 for Matérn kernels.
Remark A.3 Assuming X = Rd , H = L2 (Rd ) and r(x) = k(·, x), x ∈ Rd , where k is a
continuous positive definite kernel on Rd , Proposition A.1 establishes a convergence rate of
n−1/2 for kµk (Pn )−µk (P )kL2 (Rd ) under the condition that supx∈Rd kk(x, ·)k2L2 (Rd ) < ∞. If k
is translation invariant on Rd , i.e., k(x, y) = ψ(x−y), x, y ∈ Rd where ψ ∈ C(Rd ) is positive
definite, then ψ ∈ L2 (Rd ) ensures that supx∈Rd kk(x, ·)k2L2 (Rd ) = supx∈Rd kψ(x − ·)k2L2 (Rd ) =
kψk2L2 (Rd ) and therefore Propositions A.1 holds with Ck := kψk2L2 (Rd ) . On the other hand,
for radial kernels on Rd , i.e., kernels of the form in (5), the condition in (43) is ensured if
Z ∞
t−d/2 dν(t) < ∞
(49)
0
33
Tolstikhin, Sriperumbudur, and Muandet
since
sup kk(x, ·)k2L2 (Rd ) = sup
x∈Rd
x∈Rd
(†)
Z
Rd
Z
∞
0
≤ ν([0, ∞)) sup
x∈Rd
(‡)
= ν([0, ∞)) sup
x∈Rd
2
2
e−tkx−yk dν(t) dy
Z
Z
∞
Rd 0
∞Z
Z
0
2
e−2tkx−yk dν(t)dy
−2tkx−yk2
e
Rd
ν([0, ∞))
dy dν(t) =
(2/π)d/2
Z
0
∞
dν(t)
,
td/2
where we used Jensen’s inequality in (†) and Fubini’s theorem in (‡). Therefore the bound in
R ∞ π d/2
Proposition A.1 holds with Ck := ν([0, ∞)) 0 2t
dν(t). (49) is satisfied by Gaussian,
mixture of Gaussian and Matérn (see Sriperumbudur, 2016, Equation 6.17) kernels. For
inverse multiquadrics, while (49) holds for γ > d/2 since ν = c−2γ Gamma(γ, c2 ) (see
Wendland, 2005, Theorem 7.15), in fact the condition in (43) holds for γ > d/4 (see
Lemma C.4).
Appendix B. Minimax Lower Bounds and Le Cam’s Method
Let Θ be a set of parameters (or functions) containing the element θ which we want to
estimate. Assume there is a class P = {Pθ : θ ∈ Θ} of probability measures on Rd indexed
by Θ. Suppose d : Θ × Θ → [0, ∞) is a metric on Θ. Le Cam’s method provides a lower
bound on the minimax probability, inf θ̂n supθ∈Θ Pθn (d(θ̂n , θ) ≥ s) for s > 0, where the
infimum is taken over all possible estimators θ̂n : Rd → Θ that are constructed from an
i.i.d. sample (Xi )ni=1 drawn from Pθ . The following two results which we used throughout
this work are based on Le Cam’s method and they provide a lower bound on the minimax
probability. The first one follows from Theorem 2.2 and Equation (2.9) of Tsybakov (2008).
It requires a construction of two sufficiently distant elements of the set Θ corresponding to
the probability distributions similar in the Kullback-Leibler (KL) divergence sense, where
the KL divergence betweenR two distributions P and Q with P absolutely continuous w.r.t. Q
dP
dP .
is defined as KL(P kQ) = log dQ
Theorem B.1 (Lower bound based on two hypotheses) Assume Θ contains θ0 and
θ1 such that d(θ0 , θ1 ) ≥ 2s and KL(Pθn0 kPθn1 ) ≤ α for some s > 0 and 0 < α < ∞. Then
n
o
inf sup Pθn d(θ̂n , θ) ≥ s ≥ max
θ̂n θ∈Θ
!
p
1 −α 1 − α/2
.
e ,
4
2
Note that the second condition of the theorem bounds the distance between the n-fold
product distributions by a constant independent of n. Recalling the chain rule of the
KL-divergence, which states that KL(Pθn0 kPθn1 ) = n · KL(Pθ0 kPθ1 ), we can see that this condition is rather restrictive and requires the marginal distributions to satisfy KL(Pθ0 kPθ1 ) =
O(n−1 ). This condition is slightly relaxed in the following result, which follows from Theorem 2.5 of Tsybakov (2008).
34
Minimax Estimation of Kernel Mean Embeddings
Theorem B.2 (Lower bound based on many hypotheses) Assume M ≥ 2 and suppose that there exist θ0 , . . . , θM ∈ Θ such that (i) d(θi , θj ) ≥ 2s > 0, ∀ 0 ≤ i < j ≤ M ; (ii)
1 PM
n
n
Pθj is absolutely continuous w. r. t. Pθ0 for all j = 1, . . . , M , and M
i=1 KL(Pθj kPθ0 ) ≤
α log M with 0 < α < 1/8. Then
√
r
n
o
M
2α
n
√
> 0.
1 − 2α −
inf sup Pθ d(θ̂n , θ) ≥ s ≥
log M
1+ M
θ̂n θ∈Θ
The above result is commonly used with M tending to infinity as n → ∞. In this case
the second condition on the KL-divergence indeed becomes less restrictive than the one
of Theorem B.1, since the upper bound α log M may now grow with the sample size n.
At the same time, Theorem
√ B.2 still
√ provides a lower bound on the minimax probability
independent of n, since M /(1 + M ) and log M can be lower bounded by 1/2 and log 2
respectively.
Appendix C. Technical Lemmas
The following technical results are used to prove the main results of Sections 3 and 4.
Lemma C.1 (Theorem 5.18, Wendland, 2005) For any µ ∈ Rd and σ 2 > 0 the following holds:
∧
kx−µk2
2
σ 2 kwk22
1
1
−
2σ 2
e
exp
−ihµ,
wi
−
, w ∈ Rd .
(w)
=
2
(2πσ 2 )d/2
(2π)d/2
Lemma C.2 Let ψ : Rd → R be a symmetric and positive definite function. Let Λψ be
the corresponding finite non-negative Borel measure from (4). Then Λψ is symmetric, i.e.,
Λψ (A) = Λψ (−A) for all A ⊂ Rd .
Proof From the definition of Λψ we know that it is finite, non-negative, and
Z
Z
Z
sin(hw, xi)Λψ (dw).
cos(hw, xi)Λψ (dw) − i ·
e−ihw,xi Λψ (dw) =
ψ(x) =
Rd
Rd
Rd
R
Since ψ(−x) = ψ(x) for all x ∈ Rd , we get Rd sin(hw, xi)Λψ (dw) = 0. Note that ψ(−x) is
by definition a characteristic function of measure Λψ , and we have just proved that it is
real-valued. It is known (Bogachev, 2007, Corollary 3.8.7) that in this case the measure Λψ
is invariant under the mapping x → −x.
Lemma C.3 Assume X, Y ⊆ Rd . If f : X × Y → R is a continuous function and Y is a
compact set then g(x) := inf y∈Y f (x, y) is continuous.
Proof First, the map g : X → R is well defined since fx (y) := f (x, y) is a continuous
function for any x ∈ X and thus fx achieves its infimum since Y is a compact set. We will
prove that the map g : X → R is continuous by showing that g−1 (−∞, a) and g−1 (a, ∞)
are open sets for all a ∈ R (Dudley, 2002, Corollary 2.2.7 (a)).
35
Tolstikhin, Sriperumbudur, and Muandet
Now we will show that g −1 (−∞, a) is open for any a ∈ R. It suffices to show that
for any x ∈ g −1 (−∞, a) there is an open neighborhood Ux of x which also belongs to
g−1 (−∞, a). The set g−1 (−∞, a) consists of elements x ∈ X for which g(x) < a. In other
words, it consists of such elements x ∈ X for which there is corresponding yx ∈ Y satisfying
f (x, yx ) < a. Take any x ∈ g−1 (−∞, a). Since f is continuous, f −1 (−∞, a) is open and
contains (x, yx ). Moreover f −1 (−∞, a) contains Ux × Vy , where Ux and Vy are open sets
with x ∈ Ux and yx ∈ Vy . Now suppose x′ ∈ Ux . Then for any y ∈ Vy we have f (x′ , y) < a.
In particular, f (x′ , yx ) < a, which means that g(x′ ) < a and x′ ∈ g−1 (−∞, a). This shows
that g −1 (−∞, a) is open.
Next we will show that g −1 (a, ∞) is also an open set for any a ∈ R. Assume this is not
the case. Then there is x ∈ g−1 (a, ∞) such that for any neighborhood Ux of x there is a point
x′ ∈ Ux such that x′ 6∈ g−1 (a, ∞). This means that for any such x′ there is yx′ satisfying
f (x′ , yx′ ) ≤ a. Using this we can construct a sequence {xn , yn } from X × Y , such that
xn ∈
/ g−1 (a, ∞) for every n, limn→∞ xn = x and for any n it holds that f (xn , yn ) ≤ a. Since
Y is compact we conclude that {yn } has a converging subsequence {yn(k) } (Dudley, 2002,
Theorem 2.3.1) with limit y ∗ ∈ Y . We just showed that there is a sequence {xn(k) , yn(k) }
in X × Y , which converges to (x, y ∗ ), such that limk→∞ f (xn(k) , yn(k) ) ≤ a. Since f is
continuous, this also means that limk→∞ f (xn(k) , yn(k) ) = f (x, y ∗ ) ≤ a. This means that
inf y∈Y f (x, y) ≤ f (x, y ∗ ) ≤ a. In other words, this shows that x 6∈ g −1 (a, ∞) leading to a
contradiction and therefore g −1 (a, ∞) is open.
Lemma C.4 (L2 norm of inverse multiquadrics kernels) For any c > 0 and γ > d4 ,
Z
Proof
Z
Rd
2
(c +
Rd
kxk22 )−2γ dx
(c2 + kxk22 )−2γ dx = cd−4γ π d/2
Γ(2γ − d2 )
.
Γ(2γ)
Z
−2γ
x 2 −2γ
d−4γ
=c
1+
dx = c
1 + kxk22
dx
c 2
Rd
Rd
d/2 Z ∞
−2γ d−1
d−4γ 2π
1 + r2
r dr
=c
Γ(d/2) 0
d/2 Z ∞
d−4γ π
(1 + x)−2γ xd/2−1 dx
=c
Γ(d/2) 0
−4γ
Z
= cd−4γ
π d/2 Γ(d/2)Γ(2γ − d/2)
,
Γ(d/2)
Γ(2γ)
where last identity can be found in (Gradshteyn and Ryzhik, 2000, 3.194.3).
Appendix D. Bounds on Constants for Various Radial Kernels
In this appendix, we present bounds on the constants that appear in Corollaries 2, 10 and
Theorems 8, 13. for various radial kernels.
36
Minimax Estimation of Kernel Mean Embeddings
D.1 α in Corollary 2
In Corollary 2, we assumed that there exist 0 < t1 < ∞ and α > 0 such that ν([t1 , ∞)) ≥ α.
In the following, we present the values of t1 and α for various radial kernels.
and so for any t1 < 2η12 , we obtain α = 1.
P
(ii) Mixture of Gaussians: ν = M
i=1 βi δ 1 and so α = CM for any t1 <
(i) Gaussian kernel: ν = δ
1
2η 2
2η 2
i
1
.
2η12
(iii) Inverse multiquadric kernel: It follows from (Wendland, 2005, Theorem 7.15) that
k(x, y) =
Z
∞
−tkx−yk22 t
e
0
γ−1 e−c2 t
Γ(γ)
dt,
(50)
and so
ν = c−2γ Gamma(γ, c2 )
(51)
where the density of a Gamma distribution with parameters a, b > 0 is defined as
Gamma(t; a, b) =
ba a−1 −tb
t e , t ≥ 0.
Γ(a)
Therefore choosing t1 to be the median of ν, we obtain α =
c−2γ
2 .
(iv) Matérn kernel: We know from (Wendland, 2005, Theorem 6.13) that Matérn kernel is
related to the Fourier transform of the inverse multiquadric kernel as
Z
kwk2 τ −d/2
1
21−τ
−ihv,wi 2
2 −τ
,
K
(ckwk2 )
e
c + kvk2
dv =
Γ(τ ) d/2−τ
c
(2π)d/2 Rd
where c > 0 and τ > d/2. Using this together with the representation (50) of an inverse multiquadrics kernel we obtain the following identity, which already appeared in
(Sriperumbudur, 2016, Equation (72)):
"Z
#
Z
∞
τ −1 e−c2 t
Γ(τ )c2τ −d 2d/2
1
−tkvk22 t
−ihv,x−yi
k(x, y) =
e
e
dt dv
Γ(τ − d/2) (2π)d/2 Rd
Γ(τ )
0
Z
2τ −d 2d/2 Z ∞
1
2
2
(⋆) c
tτ −1 e−c t
e−ihv,x−yi e−tkvk2 dv dt
=
d/2
Γ(τ − d/2) 0
(2π)
Rd
Z ∞
2τ
−d
d/2
kx−yk2
2
1
c
2
− 4t 2
e
tτ −1 e−c t
=
dt
Γ(τ − d/2) 0
(2t)d/2
Z ∞
kx−yk2
c2τ −d
2
2
=
tτ −d/2−1 e−c t e− 4t dt,
Γ(τ − d/2) 0
where we invoked Tonelli-Fubini theorem (Dudley, 2002, Theorem 4.4.5) in (⋆) since c2 +
−τ
∈ L1 (Rd ) for τ > d/2. After change of variables we finally obtain
k · k22
1
k(x, y) =
Γ τ − d2
c2
4
τ − d2 Z
37
∞
0
2
c2
e−tkx−yk td/2−τ −1 e− 4t dt,
Tolstikhin, Sriperumbudur, and Muandet
which shows that Matérn kernel is a particular instance of radial kernels with
d c2
,
ν = InvGamma τ − ,
2 4
where the density of an inverse-Gamma distribution with parameters a, b > 0 has the form
InvGamma(t; a, b) =
Therefore α =
D.2
βt0
t1
1
2
ba −a−1 −b/t
t
e
.
Γ(a)
for the choice of t1 to be the median of ν.
in Theorem 8
In Theorem 8, we assumed that there exist 0 < t0 ≤ t1 < ∞ and 0 < β < ∞ such that
0
ν([t0 , t1 ]) ≥ β. Define Bk := βt
t1 . In the following, we present the values of Bk for various
radial kernels.
(i) Gaussian kernel: Choose t0 = t1 =
(ii) Mixture of Gaussians: Set t0 =
1
2η2
1
, t1
2η12
so that β = 1 and Bk = 1.
=
1
2
2ηM
so that β = CM implying Bk =
2
CM ηM
2
η1
.
(iii) Inverse multiquadric kernel: From (51), we have ν = c−2γ Gamma(γ, c2 ). Therefore
ν
Z γ/c2
h γ γ i
1
2
,
=
tγ−1 e−tc dt
2
2
2c c
Γ(γ) γ/(2c2 )
(
γ γ−1
1
exp − cγ2 c2 2cγ2 ,
Γ(γ) 2c2
≥
γ γ−1
1
exp − cγ2 c2 2cγ2 ,
Γ(γ) c2
Therefore with t0 =
γ
, t1
2c2
=
γ
c2
Bk =
and β =
(
c−2γ
2Γ(γ)
c−2γ
4Γ(γ)
(
c−2γ γ γ
Γ(γ) 2e
c−2γ γ γ
2Γ(γ) e
γ γ
,
2e
γ γ
,
e
,
,
for γ ≥ 1;
for γ ∈ (0, 1).
for γ ≥ 1;
for γ ∈ (0, 1)
for γ ≥ 1;
for γ ∈ (0, 1)
, we obtain
.
(iv) Matérn kernel: It is easy to check that if X ∼ Gamma(a, b) and Y ∼ InvGamma(a, b)
for a, b > 0 then for any 0 < x ≤ y < ∞ the following holds:
P{x ≤ X ≤ y} = P{1/y ≤ Y ≤ 1/x}.
This means, the above calculations for inverse multiquadrics can be used to obtain the
following for the Matérn kernel:
d
1 d 2τ4e−d τ − 2 , for τ − d2 ≥ 1;
2Γ(τ − 2 )
.
Bk =
τ − d2
1 d 2τ2e−d
, for τ − d2 ∈ (0, 1)
4Γ(τ − 2 )
38
Minimax Estimation of Kernel Mean Embeddings
−d/2
D.3 β 2 δ1
in Corollary 10
In Corollary 10, we assumed that there exist 0 < δ0 ≤ δ1 < ∞ and 0 < β < ∞ such that
−d/2
ν([δ0 , δ1 ]) ≥ β. Define Ak := β 2 δ1 . Based on the analysis carried out in Appendix D.2,
in the following, we present the values of Ak for various radial kernels.
(i) Gaussian kernel: Choose δ0 = δ1 =
(ii) Mixture of Gaussians: Set δ0 =
1
so that β =
2η2
1
, δ1 = 2η12
2η12
M
1 and Ak = (2η 2 )d/2 .
so that β = CM implying Ak =
2 (2η 2 )d/2 .
CM
M
(iii) Inverse multiquadric kernels: Choosing δ0 = t0 and δ1 = t1 as in Appendix D.2, we
obtain
d
cd−4γ γ 2γ− 2 ,
for γ ≥ 1;
2
2γ
(γ) (2e)
Ak = Γd−4γ
.
2γ− d
2
c
γ
,
for
γ
∈
(0,
1)
2
2γ
4Γ (γ) e
(iv) Matérn kernel: Define γ̃ := τ − d2 and c̃ := 2c . Choosing δ0 =
β=
1
Γ(γ̃)
γ̃
1
2Γ(γ̃)
γ̃
2e
γ̃
γ̃
e
,
for γ̃ ≥ 1;
,
for γ̃ ∈ (0, 1)
c̃2
γ̃
and δ1 =
2c̃2
γ̃ ,
we obtain
,
using the analysis in Appendix D.2. Therefore,
Ak =
2γ̃+ d
2
c−d2e−2γ̃ γ̃
,
2
Γ (γ̃)
c−d2e−2γ̃
Γ (γ̃)
− d+2
2
D.4 β 2 δ0 δ1
γ̃
2γ̃+ d
2
d
22+ 2
,
for γ̃ ≥ 1;
.
for γ̃ ∈ (0, 1)
in Theorem 13
In Theorem 13, we assumed that there exist 0 < δ0 ≤ δ1 < ∞ and 0 < β < ∞ such that
− d+2
ν([δ0 , δ1 ]) ≥ β. Define Bk := β 2 δ0 δ1 2 . Based on the analysis carried out in Appendix D.2,
in the following, we present the values of Bk for various radial kernels.
(i) Gaussian kernel: Choose δ0 = δ1 =
(ii) Mixture of Gaussians: Set δ0 =
2 2d/2 η d+2
CM
M
η12
1
2η2 so that β =
1
, δ1 = 2η12
2η12
M
1 and Bk = (2η 2 )d/2 .
so that β = CM implying Bk =
.
(iii) Inverse multiquadric kernels: Choosing δ0 = t0 and δ1 = t1 as in Appendix D.2, we
obtain
d
cd−4γ γ 2γ− 2 , for γ ≥ 1;
2
2γ
(γ) (2e)
Bk = 2Γd−4γ
.
d
c
γ 2γ− 2
,
for
γ
∈
(0,
1)
8Γ2 (γ) e2γ
39
Tolstikhin, Sriperumbudur, and Muandet
(iv) Matérn kernel: With the choice of δ0 and δ1 as in Appendix D.3, we obtain
2γ̃+ d
2
c−d2e−2γ̃ γ̃
, for γ̃ ≥ 1;
2
2Γ (γ̃)
.
Bk =
d
2γ̃+
c−d2e−2γ̃ γ̃ d2 ,
for γ̃ ∈ (0, 1)
Γ (γ̃)
3+
2
2
Appendix E. Alternate Proof of Theorem 8
In Theorem 8 we presented a minimax lower bound for radial kernels based on an appropriate
construction of d-dimensional Gaussian distributions. By a clever choice of the variance
σ 2 , which decays to zero as d → ∞, we obtained a lower bound of the order Ω(n−1/2 )
independent of d. This result was based on the direct analysis and special properties of
radial kernels. In this appendix we will show that we can recover almost the same result
using only Proposition 3, which holds for any translation invariant kernel. As we will see,
this leads to slightly worse constant factors and an additional lower bound on the sample
size n in terms of the properties of distribution ν, which specifies the kernel. Essentially we
will repeat the main steps of the proof of Theorem 8. However, we will use Proposition 3
instead of direct computations (based on the form of radial kernels) to lower bound the
RKHS distance between embeddings of Gaussian distributions with the Euclidean distance
between their mean vectors.
Theorem E.1 Let P be the set of distributions over Rd whose densities are continuously
infinitely differentiable and k be radial on Rd , i.e.,
Z ∞
2
e−tkx−yk2 dν(t),
k(x, y) =
0
where ν ∈ M+b ([0, ∞)) such that supp(ν) 6= {0}. Assume that there exist 0 < t0 ≤ t1 < ∞
1 Zν
where Zν := ν([0, ∞)). Then
and 0 < β < ∞ such that ν([t0 , t1 ]) ≥ β. Suppose n ≥ 24 tβt
0
s
(
)
1
βt
2
1
1
0
≥ .
·
1−
inf sup P n kθ̂n − µk (P )kHk ≥
50 2n t1 e
2+d
5
θ̂n P ∈P
Proof We apply Proposition 3 to the radial kernel k. In order to do so, we need to lower
bound the quantity appearing in r.h.s. of Condition (7), which we do as follows. We already
saw in the proof of Theorem 8 that in our case Λψ is absolutely continuous with respect to
the Lebesgue measure on Rd and has the following density:
Z ∞
kwk2
1
− 4t 2
dν(t), w ∈ Rd .
λψ (w) =
e
d/2
(2t)
0
Therefore the r.h.s. of (7) reduces to
Z
2
2
2
e−σ kwk2 hez , wi2 cos (ha, wi) dΛψ (w)
d/2
(2π)
Rd
Z ∞
Z
kwk2
1
2
−σ2 kwk22
2
− 4t 2
=
e
hez , wi cos (ha, wi)
e
dν(t) dw
(2π)d/2 Rd
(2t)d/2
0
40
Minimax Estimation of Kernel Mean Embeddings
2
=
(2π)d/2
Z
∞
0
1
(2t)d/2
Z
1
2σ2 + 2t
kwk22
1
e− 2
Rd
|
hez , wi2 e−iha,wi dw dν(t),
{z
}
(52)
♣
where we used Euler’s formula and Tonelli-Fubini theorem in the last equality. Denoting
1
, we have
δ := 2σ 2 + 2t
! d
!
Z
d
d
X
X
X
X
δ
exp −
♣=
wℓ2 (ez )2j wj2 +
(ez )j (ez )ℓ wj wℓ exp −i
aℓ wℓ dw
2
Rd
j=1
ℓ=1
j6=ℓ
ℓ=1
= ⋆ + ♠,
where
⋆ :=
d Z
X
j=1
and
♠ :=
d Z
X
j6=l
1
Rd
1
Rd
2
e− 2 δkwk2 (ez )2j wj2 e−iha,wi dw
2
e− 2 δkwk2 (ez )j (ez )l wj wl e−iha,wi dw.
Note that
Z
YZ
2
− 12 δkwk22 2 −iha,wi
2
wj e
dw = (ez )j
e
(ez )j
Rd
ℓ6=j
∞
−∞
×
=
(ez )2j
Y
ℓ6=j
r
δ
2
e− 2 wℓ e−iaℓ wℓ dwℓ
Z
∞
−∞
− 2δ wj2
e
2π − a2ℓ
e 2δ ·
δ
wj2 e−iaj wj dwj
Z
∞
− 2δ wj2
e
−∞
wj2 e−iaj wj dwj
,
where we used Lemma C.1. It follows from (Folland, 1999, Theorem 8.22(d)) that if g =
x2 f ∈ L1 (R), then f ∧ is twice differentiable and
g∧ (y) = −
∂ 2 f ∧ (y)
,
∂2y
which together with Lemma C.1 shows that
r
Z ∞
2
1 2π − aj
− 2δ wj2 2 −iaj wj
wj e
dwj =
e
e 2δ
δ
δ
−∞
a2j
1−
δ
!
.
Therefore, we get
(ez )2j
Z
− 21 δkwk22
e
Rd
wj2 e−iha,wi dw = (ez )2j
(ez )2j
=
δ
Y
ℓ6=j
41
2π
δ
r
a2
− 2δℓ
2π
e
δ
d/2
kak2
2
e− 2δ
r
2
aj
2π 1 e− 2δ
δ δ
!
a2j
.
1−
δ
a2j
1−
δ
!
Tolstikhin, Sriperumbudur, and Muandet
Summing over j = 1, . . . , d we get
Z
d
X
2
(ez )j
⋆=
j=1
− 21 δkwk22
e
Rd
1
=
δ
wj2 e−iha,wi dw
2π
δ
d/2
d
X
kak2
(ez )2j a2j 2π d/2− kak22
− 2δ 2
e
e 2δ .
−
δ2
δ
j=1
Next, for any j 6= ℓ we compute
!
!
d
d
X
δX 2
exp −
wℓ (ez )j (ez )ℓ wj wℓ exp −i
aℓ wℓ dw
2
Rd
ℓ=1
ℓ=1
Z ∞
Y
Y Z ∞
δ 2
δ 2
e− 2 wq wq e−iaq wq dwq
= (ez )j (ez )ℓ
e− 2 wq e−iaq wq dwq
Z
= (ez )j (ez )ℓ
= (ez )j (ez )ℓ
q∈{j,ℓ} −∞
q6∈{j,ℓ} −∞
Y
q6∈{j,ℓ}
2π
δ
r
d/2
a2
− 2δq
2π
e
δ
e−
kak2
2
2δ
Y
q∈{j,ℓ}
r
a2
− 2δq
2π iaq
e
δ δ
a a
j ℓ
− 2 .
δ
Summing over j 6= ℓ we get
hez , ai2
♠=−
δ2
2π
δ
d/2
−
e
kak2
2
2δ
d
X
(ez )2j a2j 2π d/2 − kak22
e 2δ .
+
δ2
δ
j=1
Returning to (52), we get
Z
2
2
2
e−σ kwk2 hez , wi2 cos (ha, wi) dΛψ (w)
d/2
(2π)
Rd
d/2
Z ∞
kak2
2
2π
1
1
hez , ai2
− 2δ 2
=
e
1−
dν(t)
δ
δ
δ
(2π)d/2 0 (2t)d/2
Z ∞
2thez , ai2
t
1 2tkak22
1−
dν(t).
exp −
=4
2 4σ 2 t + 1 (4σ 2 t + 1)1+d/2
4σ 2 t + 1
0
In order to apply Proposition 3 we need to lower bound the following value, appearing in
Condition (7):
Z
2
2
2
e−σ kwk2 hez , wi2 cos (ha, wi) dΛψ (w)
∆(a) := min
d/2
d
z∈R \{0} (2π)
Rd
Z ∞
1 2tkak22
2tkak22
t
exp −
=4
1− 2
dν(t).
2 4σ 2 t + 1 (4σ 2 t + 1)1+d/2
4σ t + 1
0
Next we will separately treat two different cases.
Case 1: d > 2. Note that the function ρ(t) = t(4σ 2 t + 1)−(d+2)/2 is positive and
bounded on [0, ∞) for any d > 0. Thus, we can define a non-negative and finite measure
42
Minimax Estimation of Kernel Mean Embeddings
R∞
τ̃ , absolutely continuous with respect to ν with density ρ(t). If we denote Zτ := 0 1 dτ̃ (t)
and write τ for the normalized version of τ̃ , then we can rewrite
Z ∞
1 2tkak22
2tkak22
exp −
∆(a) = 4
1− 2
dτ̃ (t)
2 4σ 2 t + 1
4σ t + 1
0
1 2tkak22
2tkak22
= 4 Zτ Et∼τ exp −
1− 2
2 4σ 2 t + 1
4σ t + 1
2
1 2tkak22
2tkak22
1 2tkak2
− 4 Zτ Et∼τ exp −
.
= 4 Zτ Et∼τ exp −
2 4σ 2 t + 1
2 4σ 2 t + 1 4σ 2 t + 1
t2
(d+2)/2 is bounded and
(4σ2 t+1)
2tkak22
is a convex
7→ exp − 12 4σ2 t+1
Note that for d > 2, Et∼τ [|t|] is finite, since in this case t 7→
ν is a finite measure. Denote µτ := Et∼τ [t] and note that t
function on [0, ∞). Thus, for d > 2 we can use Jensen’s inequality to get
Et∼τ
1 2tkak22
exp −
2 4σ 2 t + 1
1 2µτ kak22
≥ exp −
.
2 4σ 2 µτ + 1
Also note that
2tkak22
2µτ kak22
1 2tkak22
2tkak22
,
−4 Zτ Et∼τ exp −
≥
4
Z
E
≥
−4
Z
−
τ t∼τ
τ
2 4σ 2 t + 1 4σ 2 t + 1
4σ 2 t + 1
4σ 2 µτ + 1
where we used inequality e−x ≤ 1, which holds for x ≥ 0, together with Jensen’s inequality
2tkak22
and the fact that t 7→ − 4σ2 t+1
is concave on [0, ∞). Summarizing, we have
2µτ kak22
1 2µτ kak22
− 2
∆(a) ≥ 4Zτ exp −
2 4σ 2 µτ + 1
4σ µτ + 1
2
1 2µτ kak2
2µτ kak22
≥ 4Zτ 1 −
−
2 4σ 2 µτ + 1 4σ 2 µτ + 1
2µτ kak22
,
= 2Zτ 2 − 3 2
4σ µτ + 1
where we used a simple inequality ex ≥ 1 + x. If the following condition is satisfied:
2µτ kak22
1
≤ ,
4σ 2 µτ + 1
3
then we get
∆(a) ≥ 2Zτ = 2
Z
∞
0
(4σ 2 t
t
dν(t).
+ 1)1+d/2
(53)
(54)
Together with Proposition 3 this leads to the following lower bound, which holds for any
µ0 , µ1 ∈ Rd and σ 2 > 0 satisfying (53) with a := µ0 − µ1 :
kθ0 −
θ1 k2Hk
≥
Z
∞
0
tkµ0 − µ1 k22
dν(t),
(4σ 2 t + 1)1+d/2
43
Tolstikhin, Sriperumbudur, and Muandet
where θ0 and θ1 are KME’s of Gaussian measures G(µ0 , σ 2 I) and G(µ1 , σ 2 I) respectively.
Note that this lower bound is identical to the one in (27), which we obtained using direct
analysis for the radial kernels. However, condition (26) is now replaced with the stronger
one in (53). We can now repeat the proof of Theorem 8 starting from inequality (27)
and making sure that condition (53) is satisfied when we choose constants appearing in
definitions of µ0 , µ1 and σ 2 .
In order to check condition (53) we need to upper bound the expectation µτ . It is easily
1
t2
∗ =
seen that for d > 2, t 7→ (4σ2 t+1)
(d+2)/2 achieves its maximum on [0, ∞) for t
σ2 (d−2) .
R∞
Using this fact, denoting Zν = 0 1 dν(t), and setting σ 2 = 2t11 d we get
Z ∞
Zν
t2
1
(t∗ )2
dν(t)
≤
µτ =
Zτ 0 (4σ 2 t + 1)1+d/2
Zτ (4σ 2 t∗ + 1)1+d/2
4t2 Zν
d2
1
= 1
4
2
Zτ (d − 2) ( d−2 + 1)1+d/2
2
(d+2)/4 !2
2
2
2
4
4t21 Zν
4t21 Zν
.
1+
1+
1−
≤
=
Zτ
d−2
d+2
Zτ e2
d−2
We may finally use (29) to get
βt0
Zτ ≥
e
2
1−
d+2
,
which leads to the following upper bound on the expectation µτ :
µτ ≤
4t21 Zν d(d + 2)
.
βt0 e (d − 2)2
This upper bound shows that the condition (53) is satisfied if the following holds:
2
βt0 e (d − 2)2
t1 kak22 ≤ t1 σ 2 +
.
3
24t1 Zν d(d + 2)
(55)
We conclude the proof by repeating the remaining steps of the proof of Theorem 8 and
replacing condition (26) on the value kµ0 − µ1 k22 with (55) specified to a = µ0 − µ1 .
Case 2: d ≤ 2. We can use a simple inequality e−x/2 (1 − x) ≥ 1 − 3x/2 which holds for
any x and get the following lower bound:
∆(a) ≥ 4
Z
∞
0
t
(4σ 2 t + 1)1+d/2
3tkak2
1− 2 2
4σ t + 1
dν(t).
Assuming kak22 ≤ σ 2 we further get
3tkak2
1− 2 2
4σ t + 1
≥
3tσ 2
1− 2
4σ t + 1
44
≥
3tσ 2
1− 2
4σ t
=
1
4
Minimax Estimation of Kernel Mean Embeddings
and as a consequence, we also get
∆(a) ≥
Z
∞
0
(4σ 2 t
t
dν(t),
+ 1)1+d/2
which coincides with (54) up to an additional factor of 2. We can now repeat all the steps
for the previous case, and it is also easy to check that in this case kµ0 − µ1 k22 ≤ σ 2 will be
indeed satisfied. This concludes the proof.
Remark E.2 This result should be compared to Theorem 8, which was based on the direct
analysis for radial kernels. We see that apart from an extra factor 2 appearing under the
square root in the lower bound, Theorem E.1 also requires a superfluous condition on the
minimal sample size n, which depends on properties of ν. For instance, for Gaussian kernel
with ν concentrated on a single point 2η12 for some η 2 > 0, the result holds as long as n ≥ 24,
because in this case we can take t0 = t1 = 2η12 and β = 1. However, other choices of ν may
lead to quite restrictive lower bounds on n.
Remark E.3 Conceptually, the main difference between the proofs of Theorems 8 and E.1
lies in the way we lower bound the RKHS distance between embeddings of Gaussian measures
with the Euclidean distance between their mean vectors. In Theorem 8 we derived a closedform expression for the RKHS distance in (25) and then lower bounded it directly using the
properties specific to its form. On the other hand, in Theorem E.1 we resorted to the lower
bound of Lemma 3, which holds for any translation invariant kernel and hence is less tight.
References
N. Aronszajn. Theory of reproducing kernels. Trans. Amer. Math. Soc., 68:337–404, 1950.
V. I. Bogachev. Measure Theory, volume 1. Springer, 2007.
J. Diestel and J. J. Uhl. Vector Measures. American Mathematical Society, Providence,
1977.
N. Dinculeanu. Vector Integration and Stochastic Integration in Banach Spaces. Wiley,
2000.
R. M. Dudley. Uniform Central Limit Theorems. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 1999.
R. M. Dudley. Real analysis and probability. Cambridge University Press, 2002.
G. B. Folland. Real Analysis: Modern Techniques and Their Applications. Wiley, 1999.
K. Fukumizu, A. Gretton, X. Sun, and B. Schölkopf. Kernel measures of conditional dependence. In J.C. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural
Information Processing Systems 20, pages 489–496, Cambridge, MA, 2008. MIT Press.
K. Fukumizu, L. Song, and A. Gretton. Kernel Bayes’ rule: Bayesian inference with positive
definite kernels. J. Mach. Learn. Res., 14:3753–3783, 2013.
45
Tolstikhin, Sriperumbudur, and Muandet
I. S. Gradshteyn and I. M. Ryzhik. Table of Integrals, Series, and Products. Academic
Press, San Diego, USA, 2000.
A. Gretton, K. M. Borgwardt, M. Rasch, B. Schölkopf, and A. Smola. A kernel method for
the two sample problem. In B. Schölkopf, J. Platt, and T. Hoffman, editors, Advances in
Neural Information Processing Systems 19, pages 513–520, Cambridge, MA, 2007. MIT
Press.
A. Gretton, K. Fukumizu, C. H. Teo, L. Song, B. Schölkopf, and A. J. Smola. A kernel
statistical test of independence. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors,
Advances in Neural Information Processing Systems 20, pages 585–592. MIT Press, 2008.
A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. J. Smola. A kernel
two-sample test. Journal of Machine Learning Research, 13:723–773, 2012.
E. L. Lehmann and G. Casella. Theory of Point Estimation. Springer-Verlag, New York,
2008.
D. Lopez-Paz, K. Muandet, B. Schölkopf, and I. Tolstikhin. Towards a learning theory of
cause-effect inference. In Proceedings of the 32nd International Conference on Machine
Learning, ICML 2015, Lille, France, 6-11 July 2015, 2015.
K. Muandet, B. Sriperumbudur, K. Fukumizu, A. Gretton, and B. Schölkopf. Kernel mean
shrinkage estimators. Journal of Machine Learning Research, 2016. To appear.
A. Ramdas, S. Reddi, B. Poczos, A. Singh, and L. Wasserman. On the decreasing power of
kernel and distance based nonparametric hypothesis tests in high dimensions. In AAAI
Conference on Artificial Intelligence, 2015.
I. J. Schoenberg. Metric spaces and completely monotone functions. The Annals of Mathematics, 39(4):811–841, 1938.
A. J. Smola, A. Gretton, L. Song, and B. Schölkopf. A Hilbert space embedding for distributions. In Proceedings of the 18th International Conference on Algorithmic Learning
Theory (ALT), pages 13–31. Springer-Verlag, 2007.
L. Song, A. Smola, A. Gretton, J. Bedo, and K. Borgwardt. Feature selection via dependence
maximization. Journal of Machine Learning Research, 13:1393–1434, 2012.
B. K. Sriperumbudur. Mixture density estimation via Hilbert space embedding of measures.
In Proceedings of International Symposium on Information Theory, pages 1027–1030,
2011.
B. K. Sriperumbudur. On the optimal estimation of probability measures in weak and
strong topologies. Bernoulli, 22(3):1839–1893, 2016.
B. K. Sriperumbudur, A. Gretton, K. Fukumizu, B. Schölkopf, and G. R. G. Lanckriet.
Hilbert space embeddings and metrics on probability measures. J. Mach. Learn. Res.,
11:1517–1561, 2010.
46
Minimax Estimation of Kernel Mean Embeddings
B. K. Sriperumbudur, K. Fukumizu, and G. R. G. Lanckriet. Universality, characteristic
kernels and rkhs embedding of measures. J. Mach. Learn. Res., 12:2389–2410, 2011.
B. K. Sriperumbudur, K. Fukumizu, A. Gretton, B. Schölkopf, and G. R. G. Lanckriet. On
the empirical estimation of integral probability metrics. Electronic Journal of Statistics,
6:1550–1599, 2012.
I. Steinwart and A. Christmann. Support Vector Machines. Springer, 2008.
Z. Szabó, A. Gretton, B. Póczos, and B. K. Sriperumbudur. Two-stage sampled learning
theory on distributions. In Proceedings of the Eighteenth International Conference on
Artificial Intelligence and Statistics, volume 38, pages 948–957. JMLR Workshop and
Conference Proceedings, 2015.
A. B. Tsybakov. Introduction to Nonparametric Estimation. Springer, NY, 2008.
R. Vert and J-P. Vert. Consistency and convergence rates of one-class SVMs and related
algorithms. Journal of Machine Learning Research, 7:817–854, 2006.
H. Wendland. Scattered Data Approximation. Cambridge Monographs on Applied and
Computational Mathematics. Cambridge University Press, 2005.
V. Yurinsky. Sums and Gaussian Vectors, volume 1617 of Lecture Notes in Mathematics.
Springer-Verlag, Berlin, 1995.
47
| 10math.ST
|
Framework for state and unknown input estimation of linear
time-varying systems ?
Peng Lu a , Erik-Jan van Kampen a , Cornelis C. de Visser a , Qiping Chu a
arXiv:1606.08090v1 [cs.SY] 26 Jun 2016
a
Delft University of Technology, Kluyverweg 1, 2629HS Delft, The Netherlands
Abstract
The design of unknown-input decoupled observers and filters requires the assumption of an existence condition in the literature.
This paper addresses an unknown input filtering problem where the existence condition is not satisfied. Instead of designing a
traditional unknown input decoupled filter, a Double-Model Adaptive Estimation approach is extended to solve the unknown
input filtering problem. It is proved that the state and the unknown inputs can be estimated and decoupled using the extended
Double-Model Adaptive Estimation approach without satisfying the existence condition. Numerical examples are presented
in which the performance of the proposed approach is compared to methods from literature.
Key words: Kalman filtering; state estimation; unknown input filtering; fault estimation; Double-Model Adaptive Estimation.
1
Introduction
Faults and model uncertainties such as disturbances can
be represented as unknown inputs. The problem of filtering in the presence of unknown inputs has received
intensive attention in the past three decades.
It is common to treat the unknown inputs as part of
the system state and then estimate the unknown inputs
as well as the system state [18]. This is an augmented
Kalman filter, whose computational load may become
excessive when the number of the unknown inputs is
comparable to the states of the original system [10].
Friedland [10] derived a two-stage Kalman filter which
decomposes the augmented filter into two reduced-order
filters. However, Friedland’s approach is only optimal in
the presence of a constant bias [18]. Hsieh and Chen derived an optimal two-stage Kalman filter which performance is also optimal for the case of a random bias [18].
On the other hand, unknown input filtering can be
achieved by making use of unbiased minimum-variance
? This paper was not presented at any IFAC meeting. Corresponding author Peng Lu. Tel. +31 152783466. Fax +31
152786480.
Email addresses: [email protected] (Peng Lu),
[email protected] (Erik-Jan van Kampen),
[email protected] (Cornelis C. de Visser),
[email protected] (Qiping Chu).
Preprint submitted to Automatica
estimation [16,21,5,14,15,3]. Kitanidis [21] first developed an unbiased recursive filter based on the assumption that no prior information about the unknown input
is available [12]. Hou and Patton [14] used an unknowninput decoupling technique and the innovation filtering
technique to derive a general form of unknown-input
decoupled filters [14,15]. Darouach, Zasadzinski and
Boutayeb [7] extended Kitanidis’ method using a parameterizing technique to derive an optimal estimator
filter. The problem of joint input and state estimation,
when the unknown inputs only appear in the system
equation, was addressed by Hsieh [15] and Gillijns and
De Moor [11]. Gillijns and De Moor [12] further proposed a recursive three-step filter for the case when the
unknown inputs also appear in the measurement equation. However, their approach requires the assumption
that the distribution matrix of the unknown inputs in
the measurement equation is of full rank. Cheng et al.
[4] proposed a global optimal filter which removed this
assumption, but this filter is limited to state estimation
[1]. Later, Hsieh [17] presented a unified approach to
design a specific globally optimal state estimator which
is based on the desired form of the distribution matrix
of the unknown input in the measurement equation [17].
However, all the above-mentioned filters require the assumption that an existence condition is satisfied. This
necessary condition is given by Hou and Patton [14] and
Darouach, Zasadzinski and Boutayeb [7], in the form
of rank condition (5). Hsieh [17] presents different decoupling approaches for different special cases. However,
28 June 2016
these approaches also have to satisfy the existence condition (5). In some applications, such as that presented
in the current paper, the existence condition is not satisfied. Therefore, a traditional unknown input decoupled
filter can not be designed.
2.1
Consider the following linear time-varying system:
xk+1 = Ak xk + Bk uk + Ek dk + wk
yk = Hk xk + Fk fk + vk
Recently, particle filters are also applied to unknown input estimation [13,8,28]. These filters can cope with systems with non-Gaussian noise and have a number of applications such as for robot fault detection [2,9,30]. In
this paper, the performance of unknown input estimation using particle filters will be compared with that of
our approach.
(1)
(2)
where xk ∈ Rn represents the system states, yk ∈ Rm
the measurements, dk and fk are the unknown inputs.
Specifically, dk ∈ Rnd the disturbances, fk ∈ Rnf are
the output faults. wk and vk are assumed to be uncorrelated zero-mean white noise sequences with covariance Qk and Rk respectively. uk , the known inputs,
is omitted in the following discussion because it does
not affect the filter design [14]. Without loss of generality, we consider the case: n = m = nd = nf and
rank Hk = rank Ek = rank Fk = m, which implies all
the states are influenced by dk and fk . It should be noted
that the approach proposed in this paper can be readily
extended to the case when n 6= m or rank Hk 6= rank Ek .
This paper proposes an extended Double-Model Adaptive Estimation (DMAE) approach, which can cope with
the unknown input filtering problem when a traditional
unknown input filter can not be designed. The original
DMAE approach, which was proposed by Lu et al. [22]
for the estimation of unknown inputs in the measurement equation, is extended to allow estimation of the
unknown inputs which appear both in the system equation and the measurement equation. The unknown inputs are augmented as system states and are modeled
as random walk processes. The unknown inputs in the
system equation are assumed to be Gaussian random
processes of which covariances are estimated on-line. It
is proved that the state and unknown inputs can be
estimated and decoupled while not requiring the existence condition. Two illustrative examples are given to
demonstrate the effectiveness of the proposed approach
with comparison to other methods from literature such
as the Robust Three-Step Kalman Filter (RTSKF) [12],
the Optimal Two-Stage Kalman Filter (OTSKF) [18]
and the particle filters [13,8].
The unknown inputs are denoted as d0k , i.e., d0k =
"
dk
#
∈
fk
Rnd0 . Then, model (1) and (2) can be reformulated into
the general form as given in Hou and Patton [14] and
Darouach, Zasadzinski and Boutayeb [7]:
xk+1 = Ak xk + Ek0 d0k + wk
yk = Hk xk + Fk0 d0k + vk
(3)
(4)
In this paper, Ek0 = [Ek 0], Fk0 = [0 Fk ]. The existence
of an unknown-input decoupled filter must satisfy the
following existence condition [14,7]:
The structure of the paper is as follows: the preliminaries of the paper are given in Section 2, formulating the
filtering problem when the existence condition for a traditional unknown input decoupled filter is not satisfied
and generalizing the DMAE approach. In Section 3, the
extension of the DMAE approach to the filtering problem when the unknown inputs appear both in the system
equation and the measurement equation is presented.
Furthermore, the on-line estimation of the covariance
matrix of the unknown inputs is introduced. It is proved
that the state and the unknown inputs can still be estimated and decoupled in Section 4. In Section 5, two illustrative examples are given to show the performance of
the proposed approach with comparison to some existing unknown-input decoupled filters. Finally, Section 6
concludes the paper.
2
Problem formulation
rank
"
Fk0 Hk Ek0
0
Fk0
#
= rank
[Fk0 ]
+ rank
" #
Ek0
Fk0
(5)
In our case, since rank Hk = m, the left-hand side of condition (5) is 2m while the right-hand side is 3m. Therefore, the above existence condition does not hold, which
means that all the unknown-input filters mentioned in
the introduction can not be directly implemented.
In this paper, we consider the consecutive bias fault estimation of a system subjected to disturbances, as described in Eqs. (1) and (2). Although the existence condition of designing a traditional unknown input decoupled filter is not satisfied, it will be shown that the unknown inputs can still be decoupled using an extended
DMAE approach.
Remark 1. The model described by Eqs. (1) and (2) is
useful for applications where the disturbances appear in
the system equation and the faults appear in the measurement equation, such as bias fault estimation in aircraft air data sensors [22].
The DMAE approach
This section presents the problem formulation and the
DMAE approach.
2
x̂nf (k)
0
Filter based on x̂nf
no fault
Y
u
x̂af (k)
then the conditional probability of the two filters can be
updated recursively using the following equation:
x̂nf (k + 1)
γnf
Selective
Reinitialization x̂af (k + 1)
Filter based on
x̂0af
fault
γaf
pi (k) =
j=1
Probability evaluator
paf
Fig. 1. Block diagram for the DMAE approach
2.2
,
i = 1, 2
fyk |a,Yk−1 (yk |aj , Yk−1 )pj (k − 1)
(8)
where Yk−1 is the measurement history vector which is
defined as Yk−1 = {y(1), y(2), .., y(k − 1)}.
fyk |a,Yk−1 (yk |ai , Yk−1 ) is the probability density function
which is given by the following Gaussian form [24]:
pnf
Hypothesis Conditional
•
fyk |a,Yk−1 (yk |ai , Yk−1 )pi (k − 1)
2
P
The DMAE approach
fy(k)|a,Yk−1 (y(k)|ai , Yk−1 )
=βi (k) exp{−γiT (k)Ci−1 (k)γi (k)/2}
(9)
The DMAE1 approach proposed in Lu et al. [22] considers the model (1) and (2) for dk = 0 (nd = 0). It is
referred to as the DMAE approach in this paper, which
is generalized in the following.
where
The DMAE [22], which is a modified approach of
multiple-model-based approach [23,24], is composed of
two Kalman Filters (KFs) operating in parallel: a nofault (or fault-free) filter and an augmented fault filter.
These two filters are based on two modes of the system:
fault-free (fk = 0) and faulty (fk 6= 0). The two filters
use the same vector of measurements Y and vector of
input u, and are based on the same equations of motion,
while each hypothesizes a different fault scenario. The
state vector of the no-fault filter xnf and that of the
augmented fault filter xaf are as follows:
In Eq. (10), |•| denotes the determinant of the covariance
matrix Ci (k) which is computed by the KF at time step
k. The filter which matches the fault scenario produces
the smallest innovation which is the difference between
the estimated measurement and the true measurement.
Therefore, the conditional probability of the filter which
matches the true fault scenario is the highest between
the two filters. After the computation of the conditional
probability, the state estimate of the nonlinear system
x̂(k) can be generated by the weighted state estimate
x̂i (k) of the two filters:
xnf,k = xk , xaf,k =
"
xnf,k
fk
#
βi (k) =
x̂(k) =
(6)
(10)
x̂i (k)pi (k)
= x̂nf (k)pnf (k) + x̂af (k)paf (k).
(11)
The fault is only estimated by the augmented fault filter
and the estimate is denoted as fˆ(k). The probabilityweighted fault estimate of the DMAE approach f¯(k) is
calculated as follows:
At time step k, each of the filters produces a state estimate x̂0i (k) and a vector of innovations γi (k). The
principle is that the KF which produces the most wellbehaved innovations, contains the model which matches
the true faulty model best [23,24]. The block diagram of
the DMAE is given in Fig. 1.
f¯(k) = fˆ(k)paf (k)
(12)
The core of the DMAE approach is selective reinitialization. The flow chart of the selective reinitialization
algorithm is presented in Fig. 2.
In the algorithm, x̂0nf (x̂0af ) and x̂nf (x̂af ) denote the state
estimate of the no-fault (augmented fault) filter before
0
0
and after the reinitialization, respectively. Pnf
(Paf
) and
Pnf (Paf ) denote the covariance of state estimate error of
the no-fault (augmented fault) filter before and after the
reinitialization, respectively. x̂t , pt and Pt are the vectors
which contain the state estimate, model probability and
the covariance matrix of state estimation error of the nofault filter and the fault filter respectively. imax,k is the
index of the model with the maximum model probability
A hypothesis test uses the innovation γi (k) and the innovation covariance matrix Ci (k) of the filters in order
to assign a conditional probability to each of the filters.
Let a denote the fault scenarios of the system. If we define the hypothesis conditional probability pi (k) as the
probability that a is assigned ai for i = 1, 2 (a1 = nf ,
a2 = af ), conditioned on the measurement history up to
time step k:
i = 1, 2
1/2
i (k)|
i=1
where “nf ” means no fault and “af ” means augmented
fault. It can be noted that the state vector of the augmented fault filter is the state vector of the no-fault filter
with augmentation of the fault vector fk .
pi (k) = Pr[a = ai |Y (k) = Yk ],
2
X
1
(2π)m/2 |C
(7)
3
where wd,k is a white noise sequence with covariance:
E{wd,k (wd,l )T } = Qdk δkl . fk is also modeled as a random
walk process as:
pt,1 = pnf ; pt,2 = paf
x̂t,1 = x̂0nf,j ; x̂t,2 = x̂0af,j , j = 1 : n
0
0
Pt,1 = Pnf,jj
; Pt,2 = Paf,jj
,j = 1 : n
fk+1 = fk + wf,k ,
(15)
where wf,k is a white noise sequence with covariance:
E{wf,k (wf,l )T } = Qfk δkl . Then, the system model and
measurement model of the no-fault filter can be described as follows:
Find imax,k with
pt,imax,k = max (pt )
Yes
x̄nf,k+1 = Ānf,k x̄nf,k + w̄nf,k
yk = H̄nf,k x̄nf,k + vk
No
imax,k = 1?
(16)
(17)
where
x̂af =
#
"
x̂t,imax,k
xf0
, Paf =
#
"
Pt,imax,k 0
0
Ānf,k =
x̂nf = x̂t,imax,k
P0f
Pnf = Pt,imax,k
Āaf,k =
The DMAE approach can achieve an unbiased estimation of xk and fk when dk = 0 [22]. However, when
dk 6= 0, the unknown-input filtering problem becomes
more challenging. Since the existence condition (5) is
no longer satisfied, traditional unknown-input decoupled
filters can not be designed.
x̄nf,k =
dk
, x̄af,k =
fk
#
(19)
(20)
"
Ānf,k 0
0
I
#
, H̄af,k = [H̄nf,k Fk ], w̄af,k =
"
w̄nf,k
#
wf,k
(21)
This paper proposes a method to adapt Qdk by making
use of the augmented fault filter of the DMAE approach.
To compensate for the effect of a bad choice of Qdk on
the estimation of xk , the system noise vector w̄nf,k in
Eqs.(16), (18) and (21) is modified to:
(13)
where x̄nf,k ∈ Rn+nd and x̄af,k ∈ Rn+nd +nf . The state
vector of the augmented fault filter is that of the nofault filter augmented with the fault vector. Therefore,
the state vector of the no-fault filter can be inferred from
that of the augmented fault filter and vice versa.
w̄nf,k =
"
#
wk + wk0
wd,k
(22)
where wk0 is the noise used to compensate for the effect of
a bad choice of Qdk on the estimation of xk . In this paper,
we approximate wk0 by Ek wd,k . Therefore, w̄nf,k is
The random walk process provides a useful and general
tool for the modeling of unknown time-varying processes
[10,27,15]. dk can be modeled by a random walk process
[27,15] as:
dk+1 = dk + wd,k ,
wd,k
#
Since the difference from the DMAE in Lu et al. [22] is
the augmentation of dk , only the covariance related to
wd,k , i.e., Qdk is discussed below. It should be noted that
Qdk is usually unknown, the optimality of the filter can
be compromised by a poor choice of Qdk [21,15]. If Qdk is
not properly chosen, it can influence the estimation of
dk as well as xk .
In this section, the DMAE is extended to the case when
dk 6= 0. In order to achieve this, the state vectors of the
no-fault filter and augmented fault filter are changed to:
x̄nf,k
, H̄nf,k = [Hk 0], w̄nf,k =
wk
where
Extension of the DMAE approach
"
I
"
x̄af,k+1 = Āaf,k x̄af,k + w̄af,k
yk = H̄af,k x̄af,k + vk
P0f
" #
xk
0
#
The model of the augmented fault filter is as follows:
at time step k.
and
are the parameters which are
used for the initialization of the fault filter.
3
Ak Ek
(18)
Fig. 2. Flow chart of the Selective Reinitialization algorithm.
Note n refers to the dimension of x̂nf .
xf0
"
w̄nf,k =
(14)
4
"
#
wk + Ek wd,k
wd,k
(23)
ˆaf,k−1|k−1 denote the unbiased estimate of x̄af,k−1
Let x̄
given measurements up to time k −1. x̂k−1|k−1 , dˆk−1|k−1
and fˆk−1|k−1 denote the estimates of xk−1 , dk−1 and
fk−1 , respectively. The innovation of the augmented
fault filter is:
where Q̂k,jj , j = 1, 2, ..., m is the jth diagonal element
of Q̂k which is denoted as:
ˆaf,k|k−1
γaf,k = yk − H̄af,k x̄
The restriction Q̃k,jj ≥ 0, j = 1, 2, ..., m in Eq. (31) is to
preserve the properties of a variance [19].
Q̂k = (Ĉaf,k − Hk Qk−1 HkT − Fk Qfk FkT − Rk )
= Hk Ak−1 x̃k−1|k−1 + Hk Ek−1 d˜k−1|k−1 + Fk f˜k−1|k−1
+ Hk wk−1 + Hk Ek−1 wd,k−1 + Fk wf,k−1 + vk
(24)
with
x̃k−1|k−1 := xk−1 − x̂k−1|k−1
d˜k−1|k−1 := dk−1 − dˆk−1|k−1
f˜k−1|k−1 := fk−1 − fˆk−1|k−1
4
(32)
Unknown input decoupled filtering
This section proves that the unknown input decoupled
filtering can be achieved using the extended DMAE approach which does not need to satisfy the existence condition (5). Let l (l ≥ 1) denote the time step when the
first fault occurs and le denote the time step when the
first fault is removed, which means fk = 0 when k < l
and fk 6= 0 when l ≤ k ≤ le . Without loss of generality,
it will be proven that fk can be estimated when k ≤ le .
(25)
(26)
(27)
Therefore, the innovation covariance of the augmented
fault filter is:
T
Caf,k = E{γaf,k γaf,k
}
4.1
x
= Hk Ak−1 Pk−1|k−1
ATk−1 HkT
f
d
T
+ Hk Ek−1 Pk−1|k−1
Ek−1
HkT + Fk Pk−1|k−1
FkT
xf
xd
T
+ Hk Ak−1 Pk−1|k−1
Ek−1
HkT + Hk Ak−1 Pk−1|k−1
FkT
df
dx
T
+ Hk Ek−1 Pk−1|k−1
ATk−1 HkT + Hk Ek−1 Pk−1|k−1
Fk−1
Unknown input estimation during k < l
Theorem 1 During k < l, an unbiased estimate of dk
can be achieved by the fault-free filter of the extended
DMAE approach.
fx
fd
T
+ Fk−1 Pk−1|k−1
ATk−1 HkT + Fk Pk−1|k−1
Ek−1
HkT + Rk
T
+ Hk Qk−1 HkT + Hk Ek−1 Qdk−1 Ek−1
HkT + Fk Qfk FkT
(28)
PROOF. When k < l, fk = 0. The fault-free model
matches the true fault scenario while the augmented
fault filter does not. Therefore, according to the DMAE
approach, imax,k = 1 during this time period.
where the covariance matrices are defined as follows:
x
d
Pk|k
:= E[x̃k|k x̃Tk|k ], Pk|k
:= E[d˜k|k d˜Tk|k ]
f
T
xd
Pk|k
:= E[f˜k|k f˜k|k
], Pk|k
:= E[x̃k|k d˜Tk|k ]
The system model during this period is as follows:
xk+1 = Ak xk + Ek dk + wk
yk = Hk xk + vk
xf
dx
T
Pk|k
:= E[d˜k|k x̃Tk|k ], Pk|k
:= E[x̃k|k f˜k|k
]
fx
df
T
Pk|k
:= E[f˜k|k x̃Tk|k ], Pk|k
:= E[d˜k|k f˜k|k
]
Under this situation, dk can be estimated using the faultfree filter whose convergence condition will be discussed
later.
2
fd
Pk|k
:= E[f˜k|k d˜Tk|k ].
The actual Caf,k is approximated as follows [26,29]:
Ĉaf,k =
1
N
k
X
T
γaf,j γaf,j
(33)
(34)
(29)
The estimation of dk and fk when l ≤ k ≤ le will be
discussed in the following.
j=k−N +1
Qdk can be approximated by the main diagonal of
−1
T
Ek−1
Hk−1 Q̃k (HkT )−1 (Ek−1
)−1
4.2
(30)
For the sake of readability, the subscript “af ” will be
discarded for the remainder of the section. All the variables with a bar on top in the remainder of this section
refer to the augmented fault filter.
with Q̃k is a diagonal matrix defined as:
Q̃k := diag(max{0, Q̂k,11 }, ..., max{0, Q̂k,mm })
Unknown input estimation at k = l
(31)
5
Consequently, the expectation of fˆl|l
Using the DMAE approach, the Kalman gain K̄l can be
partitioned as follows:
Klx
E[fˆl|l ] = E[Klf Fl fl ].
d
K̄l =
Kl
(35)
Therefore, it can concluded that fl can be estimated if
and only if Klf satisfies
Klf
where Klx , Kld and Klf are the Kalman gains associated
with xk , dk and fk , respectively.
Klf Fl = I.
(36)
˜l−1|l−1 x̄
˜Tl−1|l−1 ]
P̄l−1|l−1 := E[x̄
˜l−1|l−1 = x̄l−1 − x̄
ˆl−1|l−1 .
where x̄
(37)
Due to the selective reinitialization algorithm given in
f
Fig. 2, Pl−1|l−1
= P0f . Therefore, the covariance of the
state prediction error P̄l|l−1 can be computed and partitioned as follows:
x
xd
Pl−1|l−1
Pl−1|l−1
0
T
dx
d
P̄l|l−1 = Āl−1
Pl−1|l−1 Pl−1|l−1 0 Āl−1
0
0
P0f
T
Ql−1 + El−1 Qdl−1 El−1
El−1 Qdl−1 0
d
T
d
+
Q
E
Q
0
l−1
l−1
l−1
0
0
Qfl−1
(45)
Px
P xd
0
l|l−1 l|l−1
dx
d
= Pl|l−1 Pl|l−1 0
(46)
f
0
0 Pl|l−1
where el is defined as
el := Hl Al−1 x̃l−1|l−1 + Hl El−1 d˜l−1|l−1
+ Hl wl−1 + Hl El−1 wd,l−1 + vl
(38)
Since x̂l−1|l−1 and dˆl−1|l−1 are unbiased (this can be
achieved by the DMAE1 in Lu et. al [22] since fk = 0
when k < l), E[el ] = 0.
Consequently, the expectation of γ̄l is:
E[γ̄l ] = Fl fl .
(39)
The estimation of the fault can be given by
fˆl|l = fˆl|l−1 + Klf γ̄l
= fˆl−1|l−1 + K f γ̄l
l
(44)
PROOF. Define the following covariance matrix:
PROOF. The innovation of the augmented filter is
γ̄l = el + Fl fl
2
Theorem 3 Let x̂l−1|l−1 and dˆl−1|l−1 be unbiased, then
fl can be estimated by the augmented fault filter of the
DMAE approach by choosing a sufficiently large P0f and
a sufficiently small xf0 .
Lemma 2 Let x̂l−1|l−1 and dˆl−1|l−1 be unbiased, if xf0
is chosen to be 0 or sufficiently small, then fl can be
estimated by the augmented fault filter if and only if Klf
satisfies
Klf Fl = I.
(43)
(40)
where
Since imax,k = 1 when k < l, according to the flow chart
of the selective reinitialization algorithm given in Fig. 2,
Eq. (40) can be further written into
fˆl|l = xf0 + Klf γ̄l
x
x
d
T
ATl−1 + El−1 Pl−1|l−1
El−1
Pl|l−1
:= Al−1 Pl−1|l−1
xd
T
dx
+ Al−1 Pl−1|l−1
El−1
+ El−1 Pl−1|l−1
ATl−1
T
+ Ql−1 + El−1 Qdl−1 El−1
(41)
d
d
Pl|l−1
:= Pl−1|l−1
+ Qdl−1
xd
xd
d
Pl|l−1
:= Al−1 Pl−1|l−1
+ El−1 Pl−1|l−1
+ El−1 Qdl−1
Substituting (37) into (41), yields
fˆl|l = Klf Fl fl + Klf el
dx
dx
d
T
T
Pl|l−1
:= Pl−1|l−1
ATl−1 + Pl−1|l−1
El−1
+ Qdl−1 El−1
f
Pl|l−1
:= P0f + Qfl−1
(42)
6
where
Define
C̄l := H̄l P̄l|l−1 H̄lT + Rl .
(47)
x̂∗k|k−1
Substituting Eqs. (21) and (46) into the above equation,
it follows that
f
x
HlT + Fl Pl|l−1
FlT + Rl
C̄l = Hl Pl|l−1
x̂∗k|k
k|k
(48)
f
f
fˆk|k = fˆl|l , Pk|k
= Pl|l
, l < k ≤ le
Fl−1 .
Therefore,
=
can be estimated.
xk+1 = Ak xk + Ek dk + wk
yk = Hk xk + Fk fˆl|l + vk
(50)
It follows from Lemma 2 that fl
2
Error analysis
In the previous sections, it is assumed that x̂l−1|l−1 and
dˆl−1|l−1 are unbiased. We analyze the estimation error
of fl when x̂l−1|l−1 and dˆl−1|l−1 are biased.
Theorem 4 Provided that fk has been estimated at k =
l, dk can be estimated by the augmented fault filter of the
extended DMAE approach.
Through Eq. (44), Eq. (42) can be further rewritten into
fˆl|l = fl + Fl−1 el
PROOF. During this period, the augmented fault
model matches the true fault scenario. Therefore,
imax,k = 2, which means that the fault-free filter is reinitialized by the fault filter during this period. Since this
paper considers bias fault, fk is constant for l < k ≤ le .
Therefore, during this period, we can set:
ˆk|k
x̄
(55)
Therefore, dk can be estimated using the augmented
fault filter under the same condition as for the model
described by Eqs. (33) and (34).
2
Unknown input estimation during l < k ≤ le
ˆk|k−1
x̄
(54)
As can be seen, the only unknown input is dk since the
fault filter treats fk as a known input during this period.
Since a known input does not affect the design of a filter
[14], the convergence condition of this fault filter is the
same as that of the fault-free filter based on Eqs. (33)
and (34).
4.4
4.3
(53)
It can be inferred that the model of the fault filter is
equivalent to:
(49)
C̄l ≈ Fl P0f FlT . It follows that
Klf
k
are updated by the normal Kalman filtering procedure.
It can be seen that during this period, the estimation of
the fault and the covariance are:
f
If P0f is chosen sufficiently large, then Pl|l−1
≈ P0f and
x
Pl|l−1
HlT C̄l−1
dx
T −1
K̄l =
Pl|l−1 Hl C̄l
Fl−1
k|k
(52)
Consequently, the Kalman gain of the augmented filter
can be calculated and partitioned as follows:
K̄l = P̄l|l−1 H̄lT C̄l−1
x
Pl|l−1
HlT
dx
T −1
=
Pl|l−1 Hl C̄l
f
Pl|l−1
FlT
"
"
#
#
x
xd
P
P
x̂k|k−1
k|k−1
k|k−1
∗
,
:=
:=
, Pk|k−1
dx
d
Pk|k−1
Pk|k−1
dˆk|k−1
"
"
#
#
" #
x
xd
P
P
x̂k|k
Kkx
k|k
k|k
∗
:=
:=
, Pk|k
, Kk∗ :=
dˆk|k
P dx P d
Kd
(56)
Substitute Eq. (38) into Eq. (56), it follows
fˆl|l = fl + Fl−1 (Hl Al−1 x̃l−1|l−1 + Hl El−1 d˜l−1|l−1
+ Hl wl−1 + Hl El−1 wd,l−1 + vl )
(57)
"
#
"
#
∗
Pk|k−1
0
x̂∗k|k−1
=
, P̄k|k−1 =
f
0
Pl|l
fˆl|l
"
#
"
#
" #
∗
Pk|k
0
x̂∗k|k
Kk∗
=
, K̄k =
,
, P̄k|k =
f
0 P
0
fˆl|l
The estimation error of fl as a function of x̃l−1|l−1 and
d˜l−1|l−1 can be obtained as follows:
f˜l|l = fl − fˆl|l
(58)
˜
=
+ Hl El−1 dl−1|l−1
+ Hl wl−1 + Hl El−1 wd,l−1 + vl )
(59)
Fl−1 (Hl Al−1 x̃l−1|l−1
l|l
(51)
7
If x̃l−1|l−1 and d˜l−1|l−1 are unbiased, the expectation of
f˜l|l is zero, which means the fault estimate is unbiased.
If x̃l−1|l−1 and d˜l−1|l−1 are biased, assume
The system is described by model (1) and (2) where
A=
≤ f¯I,
aI ≤ Al−1 ≤ āI, f I ≤ Fl−1
¯
¯
hI ≤ Hl−1 ≤ h̄I, eI ≤ El−1 ≤ ēI,
¯
¯
ex I ≤ x̃l−1|l−1 ≤ ēx I, ed I ≤ d˜l−1|l−1 ≤ ēd I,
¯
¯
wI ≤ wl−1 ≤ w̄I, wd I ≤ wd,l−1 ≤ w̄d I,
¯
¯
vI ≤ vl−1 ≤ v̄I.
¯
"
#
−0.0005 −0.0084
,
(67)
1.7902
"
" #
" #
0.629
1 0
1 0
E=
,H =
,F =
, (68)
0 −0.52504
0 1
0 1
"
#
"
#
0.0022
0
0.012 0
Q=
,R =
(69)
0
0.0022
0 0.012
(60)
(61)
(62)
(63)
(64)
0.0517
0.8069
#
0
,B =
"
#
0.1815
Then it follows that the fault estimation error is bounded
by the following:
The input uk is: uk = −0.5 when 200 < k ≤ 300, otherwise uk = 0.5. fk is given by the red solid lines in
Fig. 3(c). It can be noted that the number of unknown
h(aex + eed + w + ewd ) + v h̄(āēx + ēēd + w̄ + ēw̄d ) + v̄ inputs in [27], [7] and [16] is n (n = 2) while this paper
d
d
¯ ¯¯
¯¯ ¯¯ ¯ ¯
¯,
f
deals with 2nd unknown inputs.
f
¯
(65)
4.5
In both examples, since Ek0 = [Ek 0], Fk0 = [0 Fk ], condition (5) is not satisfied. In addition, rank yk < rank
d0k . Consequently, all the unknown input decoupled filters in the introduction are not applicable to solve the
problem, except for special cases when dk = 0 or fk = 0.
N in Eq. (29) is set to be 10. In both examples, Qfk = 0,
Qdk is updated by the main diagonal of the matrix given
in (30), xf0 = [10−3 , 10−3 ]T , P0f = 102 I.
Discussion
For the model given in Eqs. (33) and (34), the convergence condition for time-invariant case has been given
by Darouach et al. [6], which is given as follows:
rank
"
#
zI − A −E
H
0
= n + nd , ∀z ∈ C, |z| ≥ 1
Example 1. In this example, dk is a constant bias vector, which is shown by the red solid lines in Fig. 3(b).
The condition (5) is not satisfied. Therefore, traditional
unknown input filters, which require the satisfaction of
condition (5), can not be implemented.
(66)
This convergence condition is also required by traditional unknown input filters such as those in Darouach,
Zasadzinski and Boutayeb [7] and Cheng et al. [4].
The extended DMAE approach is implemented. The
true and estimated pnf and paf using the extended
DMAE approach are well matched. The probabilityweighted estimates of xk , dk , which are calculated using
Eq. (11), are shown in Fig. 3(a) and 3(b), respectively.
The probability-weighted estimate of fk (calculated using Eq. (12)) is shown in Fig. 3(c). As can be seen, xk ,
dk and fk can all be estimated.
The system considered in this paper is linear and the
noise is assumed to be Gaussian. If the system is nonlinear, the DMAE should be extended using Unscented
Kalman Filters [20,22] or particle filters [13,8,28]. If the
system noise is non-Gaussian, then it should be extended
by making use of particle filters [13,8,28]. However, this
is out of the scope of the present paper.
5
Example 2. In this example 1 , the disturbances,
"
# which
d1,k
are taken from [25], are stochastic. dk =
is gend2,k
Illustrative examples with comparison to existing methods
In this section, two examples similar to that in [27], [7]
and [16] are provided to demonstrate the performance
of the extended DMAE approach. Note that both E and
F are of full rank in this example.
1
The implementation of this work is available at:
https://www.researchgate.net/profile/Peng_Lu15/
publications?pubType=dataset
8
True
DMAE
0.9
0.02
∆ f1
1
0.8
x
0.7
0
−0.02
0.6
0.5
0
RTSKF
DMAE
0.04
100
200
300
400
−0.04
0
500
100
200
100
200
300
400
500
300
400
500
0.04
0
∆ f2
2
0.02
x
−5
0
−0.02
−10
0
100
200
time (s)
300
400
−0.04
0
500
(a) True and estimated states, example 1
1
d
Fig. 4. Errors of estimation of f1 and f2 using the RTSKF
and the DMAE approach, case 1, example 2
True
DMAE
1.5
erated using the following model [25]:
#
"
di,k
1
0
di,k
0.5
0
100
200
300
400
500
2.5
d
2
2
1
100
200
time (s)
300
400
500
(b) True and estimated disturbances, example 1
"
#
di,k−1
0
= V2
− L2 −2 LVgi
di,k−1
gi
q
σi L3Vgi
0
wd,k
q
+
, i = 1, 2
√
(1 − 2 3)σi ( LVgi )3
0
1
(70)
Three cases are considered for this example. The first
two cases are special cases. In these two cases, the existence condition (5) is satisfied. Therefore, some of the
approaches mentioned in the introduction can still be
used.
True
DMAE
4
where V = 35, σ1 = 0.5, σ2 = 0.8, Lg1 = 2500, Lg2 =
0
1500 and wd,k
∼ N (0, 1). The generated dk is shown by
the red solid lines in Fig. 5(b). It should be noted that
the DMAE approach still models dk as a random walk
process since dk is treated as an unknown input.
1.5
0.5
0
time (s)
f1
2
Case 1 dk = 0, fk 6= 0
0
−2
0
100
200
100
200
300
400
500
300
400
500
In this case, Ek is a zero matrix. Therefore, condition (5)
is satisfied. The probability-weighted estimate of fk using the extended DMAE is the same as in Fig. 3(c). The
RTSKF in Gillijns and De Moor [12] is also applied and
the errors of estimation of fk compared to the DMAE
are shown in Fig. 4. In addition, particle filters [13,8] are
also applied. The model used for estimation of fk is also
the random walk. 100 particles are used. The root mean
square errors (RMSEs) of estimation of f1 and f2 using
the RTSKF, the particle filter [13,8] and the extended
DMAE are shown in Table 1.
3
f2
2
1
0
−1
0
time (s)
(c) True and estimated faults, example 1
Case 2 dk 6= 0, fk = 0
Fig. 3. Results of the DMAE approach, example 1
9
Table 1
RMSEs of the fault and disturbance estimation for Example
2
Case 2
Case 3
d1
d2
f1
f2
RTSKF [12]
-
-
0.0103
0.0102
PF [13,8]
-
-
0.1549
0.1496
DMAE
-
-
0.0060
0.0047
OTSKF [18]
0.0697
0.1442
-
-
PF [13,8]
0.1088
0.2035
-
-
DMAE
0.0709
0.1459
-
-
[12,11,18,13,8,15]
N/A
N/A
N/A
N/A
DMAE
0.0845
0.1655
0.0230
0.0283
pnf
Methods
100
200
100
200
300
400
500
300
400
500
1
0.5
0
0
In this case, Fk is a zero matrix. Therefore, condition (5)
is also satisfied. The true and estimated pnf and paf using
the extended DMAE approach are shown in Fig. 5(a).
The probability-weighted estimate of dk is presented in
Fig. 5(b). The results using the methods in Heish [15],
Heish and Chen [18], and Gillijns and De Moor [11], are
similar to that of the DMAE. Particle filter is also applied. The model used for estimation of dk is the random
walk. The RMSEs of estimation of d1 and d2 using the
OTSKF in Heish [18], the particle filter [13,8] and the
extended DMAE are shown in Table 1.
True
DMAE
0.5
0
0
paf
Case 1
1
time (s)
(a) True and estimated model probabilities, case 2, example 2
True
DMAE
1
d1
0
−1
−2
0
Case 3 dk 6= 0, fk 6= 0
100
200
100
200
300
400
500
300
400
500
4
In this case, condition (5) is not satisfied. Thus, all the
conventional filters mentioned in the introduction are
not applicable.
d2
2
0
The true and estimated pnf and paf using the extended DMAE approach are also well matched. The
probability-weighted estimates of xk , is shown in Fig. 6.
The probability-weighted estimates of dk and fk are the
same as in Figs. 5(b) and 3(c) respectively. It can be
seen that despite the fact that the existence condition
for traditional unknown-input decoupled filters is not
satisfied, xk , dk and fk can all be estimated using the
extended DMAE approach. The RMSEs of the estimation of dk and fk using the extended DMAE approach
are shown in Table 1.
−2
0
time (s)
(b) True and estimated disturbances, case 2, example 2
Fig. 5. Results of the DMAE approach, case 2, example 2
However, it is also noted that the extended DMAE approach is more sensitive to Rk errors. The RMSE of the
fault estimation increases to 0.063 when Qk is multiplied
with 103 and increases to 1.79 when Rk is multiplied
with 103 . This is expected since in section 3, the process
noise w̄nf,k is adapted while the output noise vk is not
adapted. Therefore, selection of Rk should be performed
with more caution.
Finally, the sensitivity of the DMAE with respect to
errors in Qk and Rk is discussed. To demonstrate the
sensitivity with respect to errors in Qk , Rk is fixed and
Qk is multiplied with a coefficient kQ . The sensitivity
result of the RMSE of fault estimation with kQ ranging
from 10−3 to 103 is shown in Fig. 7(a). To show the
sensitivity with respect to Rk errors, Qk is fixed and Rk
is multiplied with a coefficient kR . The sensitivity result
of the RMSE of fault estimation with kR ranging from
10−3 to 103 is shown in Fig. 7(b).
6
Conclusion
In this paper, the unknown input decoupling problem
is extended to the case when the existence condition of
traditional unknown input filters is not satisfied. It is
proved that the states, disturbances and faults can be estimated using an extended DMAE approach which does
not require the existence condition. Therefore, it can
It can be seen from Fig. 7(a) and 7(b) that the minimum RMSEs are obtained when kQ = 1 or kR = 1.
10
True
DMAE
0.1
0.5
0.09
0
0.08
−0.5
0.07
−1
0
0.06
100
200
300
400
RMSE
x
1
1
500
x
2
10
0.05
0.04
5
0.03
0
0.02
−5
0.01
−10
0
0
100
200
time (s)
300
400
500
−2
10
0
10
2
kQ
10
(a) Sensitivity with respect to Qk errors, case 3, example 2
Fig. 6. True and estimated states, case 3, example 2
be applied to a wider class of systems and applications.
Two illustrative examples demonstrate the effectiveness
of the extended DMAE approach. Future work would
consider extending the DMAE to deal with systems with
non-Gaussian noise.
2
1.8
1.6
1.4
RMSE
1.2
References
1
0.8
[1] Fayçal Ben Hmida, Karim Khémiri, José Ragot, and
Moncef Gossa.
Unbiased Minimum-Variance Filter for
State and Fault Estimation of Linear Time-Varying Systems
with Unknown Disturbances. Mathematical Problems in
Engineering, 2010:1–17, 2010.
0.6
0.4
0.2
0
[2] François Caron, Manuel Davy, Emmanuel Duflos, and
Philippe Vanheeghe. Particle Filtering for Multisensor Data
Fusion With Switching Observation Models: Application to
Land Vehicle Positioning. IEEE Transactions on Signal
Processing, 55(6):2703–2719, 2007.
−2
10
0
10
2
kR
10
(b) Sensitivity with respect to Rk errors, case 3, example 2
Fig. 7. Sensitivity of the fault estimation using the DMAE
approach with respect to Qk and Rk errors, case 3, example
2
[3] Jie Chen and Ron J. Patton. Optimal Filtering and
Robust Fault Diagnosis of Stochastic Systems with Unknown
Disturbances.
IEEE Proceedings Control Theory and
Applications, 143:31–36, 1996.
[9] Nando De Freitas. Rao-Blackwellised Particle Filtering for
Fault Diagnosis. In Proceedings IEEE Aerospace Conference,
pages 4–1767–4–1772, 2002.
[4] Yue Cheng, Hao Ye, Yongqiang Wang, and Donghua Zhou.
Unbiased Minimum-Variance State Estimation for Linear
Systems with Unknown Input. Automatica, 45(2):485–491,
February 2009.
[10] Bernard Friedland. Treatment of Bias in Recursive Filtering.
IEEE Transactions on Automatic Control, 14(4):359–367,
1969.
[5] M. Darouach and M. Zasadzinski. Unbiased Minimum
Variance Estimation for Systems with Unknown Exogenous
Inputs. Automatica, 33(4):717–719, 1997.
[11] Steven Gillijns and Bart De Moor. Unbiased MinimumVariance Input and State Estimation for Linear DiscreteTime Systems. Automatica, 43(1):111–116, January 2007.
[6] M. Darouach, M. Zasadzinski, O.A. Bassong, and
S. Nowakowski. Kalman Filtering with Unknown Inputs via
Optimal State Estimation of Singular Systems. International
Journal of Systems Science, 26(10):2015–2028, October 1995.
[12] Steven Gillijns and Bart De Moor. Unbiased MinimumVariance Input and State Estimation for Linear DiscreteTime Systems with Direct Feedthrough.
Automatica,
43(5):111–116, May 2007.
[7] M. Darouach, M. Zasadzinski, and M. Boutayeb. Extension
of Minimum Variance Estimation for Systems with Unknown
Inputs. Automatica, 39(5):867–876, May 2003.
[13] N.J. Gordan, D.J. Salmond, and A.F.M. Smith. Novel
Approach to Nonlinear/non-Gaussian Bayesian State
Estimation. In Proc. Inst. Elect. Eng., F, volume 140, pages
107–113, 1993.
[8] Arnaud Doucet, Simon Godsill, and Christophe Andrieu.
On Sequential Monte Carlo Sampling Methods for Bayesian
Filtering. Statistics and Computing, 10:197–208, 2000.
[14] M. Hou and R. J. Patton. Optimal Filtering for Systems with
11
Unknown Inputs. IEEE Transactions on Automatic Control,
43(3):445–449, 1998.
[15] Chien-Shu Hsieh. Robust Two-Stage Kalman Filters for
Systems with Unknown Inputs. IEEE Transactions on
Automatic Control, 45(12):2374–2378, 2000.
[16] Chien-Shu Hsieh. Extension of unbiased minimum-variance
input and state estimation for systems with unknown inputs.
Automatica, 45(9):2149–2153, September 2009.
[17] Chien-Shu Hsieh. On the Global Optimality of Unbiased
Minimum-variance State Estimation for Systems with
Unknown Inputs. Automatica, 46(4):708–715, April 2010.
[18] Chien-Shu Hsieh and Fu-Guang Chen. Optimal solution
of the two-stage Kalman filter. IEEE Transactions on
Automatic Control, 44(1):194–199, 1999.
[19] A. H. Jazwinski. Adaptive Filtering. Automatica, 5:475–485,
1969.
[20] Simon J. Julier and Jeffrey K. Uhlmann. A New Extension of
the Kalman Filter to Nonlinear Systems. in Proc. AeroSense:
11th Int. Symp. Aerospace/Defense Sensing, Simulation and
Controls, pages 182–193, 1997.
[21] Peter K. Kitanidis. Unbiased Minimum-variance Linear State
Estimation. Automatica, 23(6):775–778, 1987.
[22] Peng Lu, Laurens Van Eykeren, E. van Kampen,
Cornelis Coen de Visser, and Qiping Chu. Double-Model
Adaptive Fault Detection and Diagnosis Applied to Real
Flight Data. Control Engineering Practice, 36:39–57, March
2015.
[23] D. T. Magill. Optimal Adaptive Estimation of Sampled
Stochastic Processes. IEEE Transactions on Automatic
Control, 10(4):434–439, 1965.
[24] Peter S. Maybeck. Multiple Model Adaptive Algorithms for
Detecting and Compensating Sensor and Actuator/Surface
Failures in Aircraft Flight Control Systems. International
Journal of Robust and Nonlinear Control, 9(14):1051–1070,
December 1999.
[25] Donald Mclean.
Automatic Flight Control Systems.
Englewood Cliffs, NJ: Prentice-Hall, 1990.
[26] Raman K. Mehra. On the Identification of Variances
and Adaptive Kalman Filtering. IEEE Transactions on
Automatic Control, 15(2):175–184, April 1970.
[27] Sang Hwan Park, Pyung Soo Kim, Oh-kyu Kwon, and
Wook Hyun Kwon. Estimation and Detection of Unknown
Inputs Using Optimal FIR Filter. Automatica, 36:1481–1488,
2000.
[28] Vandi Verma, Geoff Gordon, Reid Simmons, and Sebastian
Thrun. Real-Time Fault Diagnosis. IEEE Robotics &
Automation Magzine, 11(1):56–66, 2004.
[29] Qijun Xia, Ming Rao, Yiqun Ying, and Xuemin Shen.
Adaptive Fading Kaiman Filter with an Application.
Automatica, 30(8):1333–1338, 1994.
[30] Bo Zhao, Roger Skjetne, Mogens Blanke, and Fredrik Dukan.
Particle Filter for Fault Diagnosis and Robust Navigation of
Underwater Robot. IEEE Transactions on Control Systems
Technology, 22(6):2399–2407, 2014.
12
| 10math.ST
|
arXiv:1702.00810v2 [math.AC] 25 Mar 2018
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
PEYMAN NASEHPOUR
Abstract. In Section 1 of the paper, we prove McCoy’s property for the
zero-divisors of polynomials in semirings. We also investigate zero-divisors
of semimodules and prove that under suitable conditions, the monoid semimodule M [G] has very few zero-divisors if and only if the S-semimodule M
does so. The concept of Auslander semimodules are introduced in this section
as well. In Section 2, we introduce Ohm-Rush and McCoy semialgebras and
prove some interesting results for prime ideals of monoid semirings. In Section 3, we investigate the set of zero-divisors of McCoy semialgebras. We also
introduce strong Krull primes for semirings and investigate their extension in
semialgebras.
0. Introduction
The concept of zero-divisors in ring theory was one of the main concepts that
Fraenkel introduced in his paper [11] in 1914 [23, p. 59]. Vandiver introduced the
concept of zero-divisors in semirings in his 1934 paper [39], where he introduced the
algebraic structure of semirings itself as well [13]. The main purpose of the current
paper is to focus on zero-divisors in semimodules and semialgebras and continue
our study of these elements in our 2016 paper [30]. Before proceeding to explain
what we do in this paper, it is important to clarify, from the beginning, what we
mean by a semiring in this paper.
In this paper, by a semiring, we understand an algebraic structure, consisting
of a nonempty set S with two operations of addition and multiplication such that
(S, +) is a commutative monoid with the identity element 0, (S, ·) is a commutative
monoid with identity element 1 6= 0, multiplication distributes over addition, i.e.
a · (b + c) = a · b + a · c for all a, b, c ∈ S and finally the element 0 is the absorbing
element of the multiplication, i.e. s·0 = 0 for all s ∈ S. Semimodules over semirings
are defined similar to the concept of modules over rings in module theory [13,
Chap. 14]. For semirings and semimodules and their applications, refer to the
books [13–15, 19].
Let us recall that for a semiring S and a nonzero S-semimodule M , an element
s ∈ S is called a zero-divisor on M , if there is a nonzero m ∈ M such that sm = 0.
Section 1 of the paper is devoted to the zero-divisors of semimodules. A classical
result in commutative algebra states that if R is a commutative ring with a nonzero
identity, f is a zero-divisor on R[X], then f can be annihilated by a nonzero constant
b ∈ R [25, Theorem 2]. In Theorem 1.1, we show the following:
2010 Mathematics Subject Classification. 16Y60, 13B25, 13F25, 06D75.
Key words and phrases. Zero-divisors of Semimodules, Semiring polynomials, Monoid semirings, Auslander semimodules, McCoy semialgebras.
1
2
PEYMAN NASEHPOUR
Let M be an arbitrary S-semimodule and G a cancellative torsion-free commutative monoid. If f ∈ S[G] is a zero-divisor on M [G], then f can be annihilated by
a nonzero constant b ∈ M (McCoy’s Theorem for Semimodules).
This useful statement helps us to obtain some interesting results related to the
set of zero-divisors ZS (M ) of an S-semimodule M . In order to explain some of
the results that we obtain in Section 1, we need to recall a couple of concepts in
semiring theory. A nonempty subset I of a semiring S is said to be an ideal of S,
if a + b ∈ I for all a, b ∈ I and sa ∈ I for all s ∈ S and a ∈ I [1]. An ideal I of
a semiring S is called a proper ideal of the semiring S, if I 6= S. A proper ideal
p of a semiring S is called a prime ideal of S, if ab ∈ p implies either a ∈ p or
b ∈ p. Similar to commutative algebra, if M is an S-semimodule, we define an
ideal p of a semiring S an associated prime ideal of M , if p = Ann(m) for some
m ∈ M . Note that Ann(m), for each m ∈ M , is the set of all elements s ∈ S such
that s · m = 0. We denote the set of all associated prime ideals of M by AssS (M )
or simply Ass(M ) if there is no fear of ambiguity. In Theorem 1.18, we prove that
if ZS (M ) = p1 ∪ p2 ∪ · · · ∪ pn , where pi ∈ AssS (M ) for all 1 ≤ i ≤ n, then
ZS[G] (M [G]) = p1 [G] ∪ p2 [G] ∪ · · · ∪ pn [G]
and pi [G] ∈ AssR[G] (M [G]) for all 1 ≤ i ≤ n.
We define an S-semimodule M to be primal if ZS (M ) is an ideal of S (see
Definition 1.21). Note that an S-semimodule M has Property (A) if each finitely
generated ideal I ⊆ ZS (M ) has a nonzero annihilator in M (check Definition 1.14).
In Corollary 1.23, we show that if S is a weak Gaussian semiring and G is a
cancellative torsion-free commutative monoid, then the S[G]-semimodule M [G] is
primal if and only if the S-semimodule M is primal and has Property (A). We
recall that a semiring S is weak Gaussian if and only if each of its prime ideals is
subtractive [30, Theorem 19]. An ideal I of a semiring S is subtractive if a + b ∈
I and a ∈ I imply that b ∈ for all a, b ∈ S [13, p. 66]. In this section, we
also introduce Auslander semimodules. We define an S-semimodule M to be an
Auslander semimodule, if Z(S) ⊆ ZS (M ) (see Definition 1.24). In Theorem 1.26,
we show that if M is an Auslander S-semimodule and has Property (A) and G is
a cancellative torsion-free commutative monoid, then M [G] is an Auslander S[G]semimodule.
The definition of Auslander semimodules, inspired by the definition of Auslander
modules in [27], is related to Auslander’s Zero-Divisor Theorem in commutative
algebra, which says that if R is a Noetherian local ring, M is an R-module of finite
type and finite projective dimension, and r ∈ R is not a zero-divisor on M , then r
is not a zero-divisor on R (cf. [17, p. 8] and [3, Remark 9.4.8]).
At the end of Section 1, we define an S-semimodule M to be torsion-free if
ZS (M ) ⊆ Z(S) (See Definition 1.27). After that, in Theorem 1.31, we prove
that if a semiring S has property (A) and G is a cancellative torsion-free commutative monoid, then the S[G]-semimodule M [G] is torsion-free if and only if the
S-semimodule M is torsion-free.
A semiring B is called to be an S-semialgebra, if there is a semiring morphism
λ : S → B. For the definition of semiring morphisms, one can refer to [13, Chap.
9].
In Section 2 of the paper, we introduce Ohm-Rush and McCoy semialgebras.
We define an S-semialgebra B to be an Ohm-Rush S-semialgebra, if f ∈ c(f )B
for each f ∈ B, where by c(f ), we mean the intersection of all ideals I of S such
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
3
that f ∈ IB (see Definition 2.1). Note that if R is a commutative ring the term
Ohm-Rush algebra has been used for an R-algebra that is a content R-module by
Epstein and Shapiro in [9]. For more on content modules and algebras, refer to [35].
Now let B be an S-semialgebra. We define B to be a McCoy S-semialgebra if
B is an Ohm-Rush S-semialgebra and g · f = 0 for g, f ∈ B with g 6= 0 implies
s · c(f ) = 0 for a nonzero s ∈ S (check Definition 2.4).
After giving these definitions and some examples for them, we generalize some
theorems related to weak Gaussian semirings [30, Definition 18] and weak content
semialgebras [30, Definition 36] (for example, see Theorems 2.11 and Theorem 2.13).
These are useful for the main purpose of the last section of the paper. In fact Section
2, beside its interesting results, can be considered as a preparatory section for the
last section of the paper which discusses the zero-divisors of McCoy semialgebras.
In Section 3 of the present paper, we prove a couple of theorems for the relationship between the set of zero-divisors of a semiring S and a McCoy S-semialgebra
B. For example, in Theorem 3.5, we show that if S is a weak Gaussian semiring,
B is a McCoy and weak content S-semialgebra with homogeneous c, and the corresponding homomorphism λ of the S-semialgebra B is injective, then S has very
few zero-divisors if and only if B does so. Note that in Definition 3.2, we define
the content function from an Ohm-Rush S-semialgebra B to the set of ideals of S
to be homogeneous if c(s · f ) = s · c(f ) for each s ∈ S and f ∈ B.
We also show that if S is a weak Gaussian semiring and B is a McCoy and weak
content S-semialgebra such that the corresponding homomorphism λ is injective
and c : B → Fid(S) is homogeneous and onto, then B has few zero-divisors if and
only if S has few zero-divisors and property (A) (see Theorem 3.10). Note that by
Fid(S), we mean the set of all finitely generated ideals of S.
We finalize the last section by giving the definition of strong Krull prime ideals
for semirings. In Definition 3.12, we define a prime ideal p of a semiring S to be
a strong Krull prime of S, if for any finitely generated ideal I of S, there exists a
z ∈ S such that I ⊆ Ann(z) ⊆ p, whenever I ⊆ p. After giving this definition,
in Lemma 3.13, we prove that if B is a weak content S-semialgebra such that the
corresponding homomorphism λ is injective and c is homogeneous and p is a strong
Krull prime of S, then either pB = B or pB is a strong Krull prime of B.
Finally, in Theorem 3.14, we show that if B is a McCoy and weak content
S-semialgebra such that the corresponding homomorphism λ is injective, c is homogeneous, and Z(S) is a finite union of strong Krull primes of S, then Z(B) is a
finite union of strong Krull primes of B. We emphasize that one of the corollaries of
these results (see Corollary 3.17) is that if R is a commutative ring with a nonzero
identity and B is a content R-algebra and Z(R) is a finite union of strong Krull
primes of R, then Z(B) is a finite union of strong Krull primes of B. Now we pass
to the next section to investigate zero-divisors of semimodules.
1. Zero-Divisors of Semimodules
Let us recall that for a semiring S and a nonzero S-semimodule M , an element
s ∈ S is called a zero-divisor on M , if there is a nonzero m ∈ M such that sm =
0. The set of all zero-divisors on M is denoted by ZS (M ), or simply by Z(M ),
whenever there is no fear of ambiguity.
From monoid theory, we know that a cancellative torsion-free commutative
monoid can be embedded into a totally ordered Abelian group (cf. [12, Corollary
4
PEYMAN NASEHPOUR
15.7]). We use this to prove Theorem 1.1, which is a generalization of a classical
result for zero-divisors of polynomial rings showed by Neal Henry McCoy [25, Theorem 2].
Also, let us recall that any nonzero finitely generated torsion-free Abelian group
is isomorphic to the sum of n copies of Z, i.e., ⊕ni=1 Z [38, Theorem 25.23]. This
means that arguments on monoid semiring S[G] (semimodule M [G]) - where S
(M ) is a semiring (semimodule) and G is a cancellative torsion-free commutative
monoid - can be reduced to arguments on finite-variable Laurent polynomial semiring (semimodule) over S (M ), as for instance, we will see in the proof of Theorem
1.3.
Theorem 1.1 (McCoy’s Theorem for Semimodules). Let M be an arbitrary Ssemimodule and G a cancellative torsion-free commutative monoid. If f ∈ S[G] is
a zero-divisor on M [G], then it can be annihilated by a nonzero constant b ∈ M .
Proof. For the proof, we take f ∈ S[G] to be of the form f = a0 X un + · · · + an X u0
(n ≥ 0, ai ∈ S, ui > ui−1 and a0 6= 0) and let g ∈ M [G] − {0} with f · g = 0. If g
is a monomial, then the unique nonzero coefficient of g annihilates f . So, we can
assume that g = b0 X tm + · · · + bm X t0 (m ≥ 1, bi ∈ M , ti > ti−1 and b0 6= 0). Now
consider the elements a0 · g, . . . , an · g of M [G]. If all of these elements are zero,
then b0 annihilates any coefficient of f and the theorem is proved. Otherwise, there
is an r ∈ {0, 1, . . . , n} such that a0 · g = · · · = ar−1 · g = 0, while ar · g 6= 0. This
causes (ar X un−r + · · · + an X u0 ) · g = 0, since f · g = 0. Clearly, this implies that
ar ·b0 = 0. Therefore, h = ar ·g 6= 0 has less monomials than g and f ·h = f ·ar ·g = 0.
Therefore, by mathematical induction on the numbers of the monomials of g, there
is a nonzero constant b ∈ M such that f · b = 0 and the proof is complete.
Corollary 1.2 (McCoy’s Theorem for Semirings). Let S be a semiring and G a
cancellative torsion-free commutative monoid. If f ∈ S[G] is a zero-divisor on
S[G], then it can be annihilated by a nonzero constant s ∈ S.
Let us recall that a nonempty subset I of a semiring S is said to be an ideal of
S, if a + b ∈ I for all a, b ∈ I and sa ∈ I for all s ∈ S and a ∈ I [1]. We denote the
set of all ideals of S by Id(S). An ideal I of a semiring S is called a proper ideal
of the semiring S if I 6= S. An ideal I of a semiring S is said to be subtractive, if
a + b ∈ I and a ∈ I imply that b ∈ I for all a, b ∈ S. We say that a semiring S is
subtractive if each ideal of the semiring S is subtractive. Finally, we note that a
proper ideal p of a semiring S is called a prime ideal of S, if ab ∈ p implies either
a ∈ p or b ∈ p.
Now we proceed to prove a theorem for prime ideals of monoid semirings that is
a generalization of a theorem for prime ideals of polynomial semirings due to Susan
LaGrassa. In fact, LaGrassa in Theorem 2.6 of her dissertation [24] shows that if p
is an ideal of a semiring S and X is an indeterminate over S, then p[X] is a prime
ideal of S[X] if and only if p is a subtractive prime ideal of S.
Theorem 1.3. Let S be a semiring and G a cancellative torsion-free commutative
monoid. Then p[G] is a prime ideal of S[G] if and only if p is a subtractive and
prime ideal of S.
Proof. (⇒): Suppose that p[G] is a prime ideal of S[G] and a, b ∈ S such that
ab ∈ p. Then either a ∈ p[G] or b ∈ p[G]. Since p[G] ∩ S = p, we have already
proved that p is a prime ideal of S. Now suppose that a, b ∈ S such that a+b, a ∈ p
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
5
and take a nonzero v ∈ G. Set f = a + bX v and g = b + (a + b)X v . Clearly,
f g = ab + (a2 + ab + b2 )X v + (ab + b2 )X 2v and so, f g ∈ p[G]. On the other hand,
since p[G] is prime, either f ∈ p[G] or g ∈ p[G] and in each case, b ∈ p. So, we
have showed that p is subtractive.
(⇐): Suppose that p is a subtractive prime ideal of S and f g ∈ p[G] for some
f, g ∈ S[G]. Imagine f ∈ S[G] is of the form f = a0 X un + · · · + an X u0 (n ≥ 0,
ai ∈ S, and ui > ui−1 ) and g ∈ S[G] of the form g = b0 X tm + · · · + bm X t0 (m ≥ 0,
bi ∈ S and ti > ti−1 ). Since G is a cancellative torsion-free commutative monoid,
it can be embedded into a torsion-free Abelian group gp(G) [2, p. 50]. Therefore,
the subgroup G0 generated by un , . . . , u0 , tm , . . . , t0 is isomorphic to a finite copies
of Z and this means that f g ∈ p[G0 ], where S[G0 ] is isomorphic to a finite-variable
Laurent polynomial semiring. Now note that by Theorem 39 in [30], p[G0 ] is a
prime ideal and therefore, either f ∈ p[G0 ] or g ∈ p[G0 ] and this means that either
f ∈ p[G] or g ∈ p[G] and the proof is complete.
Let us recall that a semiring E is called entire if ab = 0 implies either a = 0 or
b = 0 for all a, b ∈ E [13, p. 4].
Corollary 1.4. Let S be a semiring and G a cancellative torsion-free commutative
monoid. Then S is an entire semiring if and only if S[G] is an entire semiring. In
particular, if k is a semifield and G a cancellative torsion-free commutative monoid,
then k[G] is an entire semiring.
Remark 1.5.
(1) By considering Corollary 1.4, one may ask if it is possible for
S[G] to be entire if S is a semiring, while G is not necessarily a cancellative
torsion-free commutative monoid. In the following, we give an affirmative
answer to this question:
Let us recall that a semiring S is called zerosumfree if s1 + s2 = 0 implies
that s1 = s2 = 0, for all s1 and s2 in S. Also, a semiring S is called an
information algebra if it is both zerosumfree and entire [13, p. 4]. Now if S
is an information algebra and G is a commutative monid with at least two
elements, then the monoid semiring S[G] is entire and here is the proof:
Proof. It is clear that if S is an information algebra and a and b are nonzero
elements of S, then ab + s is also nonzero, for any s ∈ S. Now if f and g are
nonzero elements of S[G], then f and g, respectively, have monomials of the
form aX g and bX h , where a and b are both nonzero. Clearly, (ab + s)X g+h
is a nonzero monimial of f g. Therefore, f g is nonzero and the proof is
complete.
(2) If k is a field and G is a torsion-free and non-Abelian group, Kaplansky’s
zero-divisor conjecture states that the group ring k[G] has no non-trivial
zero-divisors (cf. [22, Problem 6] and [36, Chap. 13]). This longstanding
algebra conjecture is related to the linear independence of time-frequency
shifts (also known as HRT) conjecture in wavelet theory [20].
Corollary 1.6. Let M be an S-semimodule and G a cancellative torsion-free commutative monoid. If p = Ann(m) is a prime ideal of S for some m ∈ M , then
p[G] = Ann(m) is a prime ideal of S[G].
Proof. Since p = Ann(m) is subtractive, by Theorem 1.3, p[G] is a prime ideal of
S[G]. The proof of p[G] = Ann(m) is straightforward.
6
PEYMAN NASEHPOUR
In [5], it has been defined that a ring R has few zero-divisors, if Z(R) is a
finite union of prime ideals. Rings having very few zero-divisors were investigated
in [31]. Modules having few and very few zero-divisors were defined and investigated
in [28,29,32,33]. We give the following definition and prove some interesting results
for zero-divisors of monoid semimodules. Now let M be an S-semimodule. Similar
to commutative algebra, we define an ideal p of a semiring S to be an associated
prime ideal of M if p is a prime ideal of S and p = Ann(m) for some m ∈ M . The
set of all associated prime ideals of M is denoted by AssS (M ) or simply Ass(M ).
Definition 1.7. We define an S-semimodule M to have very few zero-divisors, if
ZS (M ) is a finite union of prime ideals in AssS (M ).
In order to give some suitable examples for semimodules having very few zerodivisors, we prove the following theorem that is a semiring version of a theorem in
commutative algebra due to I. N. Herstein (1923-1988):
Theorem 1.8. Let M be a nonzero S-semimodule. If p is a maximal element of
all ideals of the form Ann(m) of S, where m is a nonzero element of M , then p is
prime.
Proof. Take p = Ann(m) and let ab ∈ Ann(m). Let a ∈
/ p. So, am 6= 0 and
Ann(am) ⊇ p. Since p is maximal among ideals of S in the form of Ann(x) with
x ∈ M − {0}, we have that Ann(am) = Ann(m). On the other hand, abm = 0. So,
b ∈ Ann(m) and this means that p is prime and the proof is complete.
Similar to commutative algebra, an S-semimodule M is called Noetherian if any
S-subsemimodule of M is finitely generated. A semiring S is Noetherian if it is
Noetherian as an S-semimodule [13, p. 69].
Corollary 1.9. If S is a Noetherian semiring and M is an S-semimodule, then
Z(M ) is a union of subtractive prime ideals of S.
S
Proof. Note that if M is an S-semimodule, then Z(M ) = m∈M−{0} Ann(m). Set
C to be the collection of all maximal elements of {Ann(m)
: m ∈ M − {0}}. Since
S
S is Noetherian, C is nonempty. Therefore, Z(M ) = I∈C I. By Theorem 1.8, any
element of C is prime of the form Ann(m) for some m 6= 0, which is a subtractive
ideal. Hence, Z(M ) is a union of subtractive prime ideals of S and the proof is
complete.
Corollary 1.10. If S is Noetherian and M is an S-semimodule, then Ass(M ) 6= ∅
and Z(M ) is the union of all prime ideals in Ass(M ).
Now we prove the following interesting theorem:
Theorem 1.11. Let S be a Noetherian semiring and M a Noetherian S-semimodule.
Then 0 <| AssS (M ) |< ∞.
Proof. Let S be Noetherian. So by Corollary 1.10, Ass(M ) 6= ∅. Assume that
{pi = Ann(mi )}i is the family of maximal primes of Z(M ). Take N to be the Ssubsemimodule of M generated by the elements {mi }i , where pi = Ann(mi ). Since
M is Noetherian, N is generated by a finite number of elements in {mi }i , say by
m1 , m2 , . . . , mk . If any further mi ’s exists, it can be written as a linear combination
of m1 , m2 , . . . , mk , i.e., for example we have mk+1 = t1 m1 +t1 m1 +t2 m2 +· · ·+tk mk ,
where ti ∈ S. This implies that p1 ∩ p2 ∩ · · · ∩ pk ⊆ pk+1 and therefore, pl ⊆ pk+1
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
7
for some 1 ≤ l ≤ k, contradicting the maximality of pl . Hence, there are no further
mi ’s (pi ’s) and the proof is complete.
Example 1.12. A family of examples for Theorem 1.11: Let n ≥ 2 be a non-prime
natural number. Clearly, the Noetherian semiring Id(Zn ) possesses non-trivial zerodivisors. Now if we take M = S n , then M is also Noetherian. Therefore, by
Theorem 1.11, M has very few zero-divisors.
The following statement is a semiring version of Remark 2.3.3 in [28]:
Proposition 1.13. Let S be a semiring and consider the following three conditions
on S:
(1) The semiring S is Noetherian.
(2) The semiring S has very few zero-divisors.
(3) The semiring S has few zero-divisors.
Then (1) ⇒ (2) ⇒ (3) and none of the implications are reversible.
Proof. For a proof of the implication (1) ⇒ (2), use Corollary 1.10 and Theorem
1.11. It is obvious that (2) implies (3).
(2) ; (1): Let x1 , x2 , x3 , . . . , xn , . . . be indeterminates over the semifield F and
put E = F [x1 , x2 , x3 , . . . , xn , . . .]. Clearly, E is an entire semiring and so, it has
very few zero-divisors, while it is not Noetherian.
(3) ; (2): Let k be a field and imagine D = k[x1 , x2 , x3 , . . . , xn , . . .]. Also set
m = (x1 , x2 , x3 , . . . , xn , . . .) and a = (x21 , x22 , x23 , . . . , x2n , . . .), where xi s are indeterminates over k. Now put R = D/a. It is easy to check that R is a local ring
with the only prime ideal m/a and Z(R) = m/a, while m/a ∈
/ AssR (R). In fact,
AssR (R) = ∅.
In [18], it has been defined that a ring R has Property (A), if each finitely
generated ideal I ⊆ Z(R) has a nonzero annihilator. In [32], a generalization of
this definition was given for modules. We define semimodules having Property (A)
as follows:
Definition 1.14. We define an S-semimodule M to have Property (A) if each
finitely generated ideal I ⊆ ZS (M ) has a nonzero annihilator in M .
The Prime Avoidance Lemma is one of the most important theorems in commutative ring theory [21, Theorem 81]. A more general result called Prime Avoidance
Lemma for Semirings has been given in [16, Lemma 3.11]. Therefore, we bring the
following result without proving it:
Theorem 1.15 (Prime Avoidance Theorem for Semirings). Let S be a semiring,
I an ideal, and pi (1 ≤ i ≤ n) subtractive prime ideals of S. If I ⊆ ∪ni=1 pi , then
I ⊆ pi for some i.
Proof. The proof is exactly the same as the proof given for the Prime Avoidance
Theorem in commutative ring theory [21, Theorem 81]. Therefore, its proof is
omitted here.
Proposition 1.16. If an S-semimodule M has very few zero-divisors, then M has
Property (A).
Proof. Use Theorem 1.15.
8
PEYMAN NASEHPOUR
Let M be an S-semimodule. We say an element s ∈ S is M -regular if it is
not a zero-divisor on M , i.e. s ∈
/ ZS (M ). We say an ideal of S is M -regular if
it contains an M -regular element. Note that if G is a commutative monoid and
f = s1 X g1 +· · ·+sn X gn is an element of the monoid semiring S[G], then the content
of f , denoted by c(f ), is defined to be the finitely generated ideal (s1 , . . . , sn ) of S.
Theorem 1.17. Let G be a cancellative torsion-free commutative monoid and M
be an S-semimodule. Then the following statements are equivalent:
(1) The S-semimodule M has Property (A).
(2) For all f ∈ S[G], f is M [G]-regular if and only if c(f ) is M -regular.
Proof. (1) ⇒ (2): Let the S-semimodule M have Property (A). If f ∈ S[G] is
M [G]-regular, then f · m 6= 0 for all nonzero m ∈ M and so, c(f ) · m 6= (0) for all
nonzero m ∈ M and according to the definition of Property (A), c(f ) 6⊆ ZS (M ).
This means that c(f ) is M -regular.
Now let c(f ) be M -regular. So c(f ) 6⊆ ZS (M ) and this means that c(f ) · m 6= (0)
for all nonzero m ∈ M and hence f · m 6= 0 for all nonzero m ∈ M . Since G is
a cancellative torsion-free commutative monoid, by Theorem 1.1, f is not a zerodivisor on M [G], i.e. f is M [G]-regular.
(2) ⇒ (1): Let I be a finitely generated ideal of S such that I ⊆ ZS (M ). Then
there exists an f ∈ S[G] such that c(f ) = I. But c(f ) is not M -regular, therefore,
according to our assumption, f is not M [G]-regular. Therefore, there exists a
nonzero m ∈ M such that f · m = 0 and this means that I · m = (0), i.e. I has a
nonzero annihilator in M and the proof is complete.
Theorem 1.18. Let M be an S-semimodule and G a cancellative torsion-free commutative monoid. Then the S[G]-semimodule M [G] has very few zero-divisors if
and only if the S-semimodule M has very few zero-divisors.
Proof. (⇐): Let ZS (M ) = p1 ∪ p2 ∪ · · · ∪ pn , where pi ∈ AssS (M ) for all 1 ≤ i ≤ n.
First, we show that ZS[G] (M [G]) = p1 [G]∪p2 [G]∪· · ·∪pn [G]. Let f ∈ ZS[G] (M [G]).
So, there exists an m ∈ M −{0} such that f ·m = 0 and so, c(f )·m = (0). Therefore,
c(f ) ⊆ ZS (M ) and this means that c(f ) ⊆ p1 ∪ p2 ∪ · · · ∪ pn and according to the
Prime Avoidance Theorem for Semirings (Theorem 1.15), we have c(f ) ⊆ pi , for
some 1 ≤ i ≤ n and therefore, f ∈ pi [G]. Now let f ∈ p1 [G] ∪ p2 [G] ∪ · · · ∪ pn [G].
Therefore, there exists an i such that f ∈ pi [G]. So, c(f ) ⊆ pi and c(f ) has a
nonzero annihilator in M and this means that f is a zero-divisor on M [G]. Note
that by Corollary 1.6, pi [G] ∈ AssR[G] (M [G]) for all 1 ≤ i ≤ n.
(⇒): Let ZS[G] (M [G]) = ∪ni=1 qi , where qi ∈ AssS[G] (M [G]) for all 1 ≤ i ≤ n.
Therefore, ZG (M ) = ∪ni=1 (qi ∩ S). Without loss of generality, we can assume that
qi ∩ S * qj ∩ S for all i 6= j. Now we prove that qi ∩ S ∈ AssS (M ) for all 1 ≤ i ≤ n.
Consider h ∈ M [G] such that qi = Ann(h) and h = m1 X g1 +m2 X g2 +· · ·+mn X gn ,
where m1 , . . . , mn ∈ M and g1 , . . . , gn ∈ G. It is easy to see that qi ∩ S =
Ann(c(h)) ⊆ Ann(m1 ) ⊆ ZS (M ) and by Theorem 1.15, q1 ∩ S = Ann(m1 ).
p Let us recall that a semiring S is called a weak Gaussian semiring, if c(f )c(g) ⊆
c(f g) for all f, g ∈ S[X] [30, Definition 18]. Also note that a semiring S is weak
Gaussian if and only if each of its prime ideals is subtractive [30, Theorem 19]. Now
let, for the moment, S be a weak Gaussian semiring and M an S-semimodule such
that the set ZS (M ) of zero-divisors of M is a finite union of prime ideals. One can
consider ZS (M ) = ∪ni=1 pi such that pi * ∪nj=1∧j6=i pj for all 1 ≤ i ≤ n. Obviously,
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
9
we have pi * pj for all i 6= j. Also, it is easy to check that, if ZS (M ) = ∪ni=1 pi
and ZS (M ) = ∪m
k=1 qk such that pi * pj for all i 6= j and qk * ql for all k 6= l,
then m = n and {p1 , . . . , pn } = {q1 , . . . , qn }, i.e., these prime ideals are uniquely
determined (For the proof, we have the permission to use Theorem 1.15, since each
prime ideal of a weak Gaussian semiring is subtractive). Now we give the following
definition:
Definition 1.19. An S-semimodule M is said to have few zero-divisors of degree
n, if ZS (M ) is a finite union of n prime ideals p1 , . . . , pn of S such that pi * pj
for all i 6= j.
Theorem 1.20. Let S be a weak Gaussian semiring, M an S-semimodule, and G
a cancellative torsion-free commutative monoid. Then the S[G]-semimodule M [G]
has few zero-divisors of degree n if and only if the S-semimodule M has few zerodivisors of degree n and Property (A).
Proof. (⇐): By taking into consideration of this assumption that the S-semimodule
M has Property (A), similar to the proof of Theorem 1.18, if ZS (M ) = ∪ni=1 pi ,
then ZS[G] (M [G]) = ∪ni=1 pi [G]. Also, it is obvious that pi [G] ⊆ pj [G] if and only if
pi ⊆ pj , for all 1 ≤ i, j ≤ n. So, these two imply that the S[G]-semimodule M [G]
has few zero-divisors of degree n.
(⇒): Note that ZS (M ) ⊆ ZS[G] (M [G]). It is easy to check that if ZR[G] (M [G]) =
∪ni=1 qi , where qi is a prime ideal of S[G] for each 1 ≤ i ≤ n, then ZS (M ) =
∪ni=1 (qi ∩ S). Now we prove that the S-semimodule M has Property (A). Let
I ⊆ ZS (M ) be a finitely generated ideal of S. Choose f ∈ S[G] such that I = c(f ).
So, c(f ) ⊆ ZS (M ) and obviously, f ∈ ZS[G] (M [G]) and according to Theorem 1.1,
there exists a nonzero m ∈ M such that f · m = 0. This means that I · m = 0 and I
has a nonzero annihilator in M . Consider that by a similar discussion in (⇐), the
S-semimodule M has few zero-divisors obviously not less than degree n and this
completes the proof.
Let us recall that an R-module M is said to be primal, if ZR (M ) is an ideal of
R [4]. Similarly, we define primal semimodules as follows:
Definition 1.21. We define an S-semimodule M to be primal if ZS (M ) is an ideal
of S.
It is easy to check that if ZS (M ) is an ideal of S, then it is a prime ideal and
therefore, the S-semimodule M is primal if and only if M has few zero-divisors of
degree one.
Example 1.22. Let (P, +, 0) be an idempotent commutative monoid and set S =
P ∪ {1}. Now extend addition on S as a + 1 = 1 + a = 1 for all a ∈ S and define
multiplication over S as ab = 0 for all a, b ∈ P and a · 1 = 1 · a = a for all a ∈ S. It
is, then, easy to check that (S, +, ·) is a semiring and ZS (S) = P is a prime ideal
of S and therefore, S is a primal S-semimodule [30, Proposition 20].
Corollary 1.23. Let S be a weak Gaussian semiring, M an S-semimodule, and G
a cancellative torsion-free commutative monoid. Then the S[G]-semimodule M [G]
is primal if and only if the S-semimodule M is primal and has Property (A).
Auslander’s Zero-Divisor Theorem in module theory states that if R is a Noetherian local ring, M is an R-module of finite type and finite projective dimension
10
PEYMAN NASEHPOUR
and r ∈ R is not a zero-divisor on M , then r is not a zero-divisor on R (cf. [17, p.
8], [3, Remark 9.4.8]). In [27], we have defined an R-module M to be Auslander, if
Z(R) ⊆ ZR (M ). This inspires us to give the following definition:
Definition 1.24. We define an S-semimodule M to be an Auslander semimodule,
if Z(S) ⊆ ZS (M ).
Examples 1.25. Here we give some examples for Auslander semimodules:
(1) If S is an entire semiring, then obviously, any R-semimodule M is Auslander.
(2) If for any nonzero s ∈ S, there is an x ∈ M such that s · x 6= 0, then
HomS (M, M ) is an Auslander S-semimodule, where by HomS (M, M ), we
mean the set of all S-endomorphisms on M [13, p. 159]. Here is the proof:
Proof. Let s ∈ ZS (S). So, there is a nonzero t ∈ S such that s · t = 0.
Define ft : M −→ M by ft (x) = t · x. By assumption, ft is a nonzero
element of HomR (M, M ). Clearly, sft = 0. So, s ∈ ZS (Hom(M, M )) and
this completes the proof.
(3) If N is an S-subsemimodule of an S-semimodule M and N is Auslander,
then clearly, M is also Auslander. Therefore, if M is an Auslander Ssemimodule, then M ⊕ M ′ is also an Auslander S-semimodule for any Ssemimodule M ′ . In particular, if {Mi }i∈Λ is a family of S-semimodules
and there
Q say i0 , such that Mi0 is an Auslander S-semimodule,
L is an i ∈ Λ,
then i∈Λ Mi and i∈Λ Mi are both Auslander S-semimodules.
Theorem 1.26. Let M be an Auslander S-semimodule, have Property (A) and
G a cancellative torsion-free commutative monoid. Then M [G] is an Auslander
S[G]-semimodule.
Proof. Let f ∈ Z(S[G]). By Corollary 1.2, there is a nonzero element s ∈ S
such that f · s = 0. This implies that c(f ) ⊆ Z(S). But M is an Auslander
semimodule, so Z(S) ⊆ ZS (M ), which implies that c(f ) ⊆ ZS (M ). On the other
hand, M has Property (A). So, c(f ) has a nonzero annihilator, which implies that
f ∈ ZS[G] (M [G]) and the proof is complete.
Let us recall that if R is a commutative ring with a nonzero identity, M a unital
R-module, and Q the total ring of fractions of R, then M is torsion-free if the
natural map M → M ⊗ Q is injective [3, p. 19]. It is starightforward to see that
M is a torsion-free R-module if and only if ZR (M ) ⊆ ZR (R). Therefore, the notion
of Auslander modules defined in [27] is a kind of dual to the notion of torsion-free
modules. Inspired by this, we define torsion-free semimodules as follows:
Definition 1.27. We define an S-semimodule M to be torsion-free, if ZS (M ) ⊆
Z(S).
Examples 1.28. Here we bring some examples for torsion-free semimodules:
L
(1) Clearly, any free S-semimodule F =
S is torsion-free.
(2) Let us recall that if M is an S-semimodule, then the content of m ∈ M ,
denoted by c(m), is defined to be the intersection of all ideals I of S such
that m ∈ IM . An S-semimdoule M is called to be a content S-semimodule
if m ∈ c(m)M , for all m ∈ M [30, Definition 24]. Now let M be a content
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
11
S-semimodule such that c(sm) = sc(m), for all s ∈ S and m ∈ M . Then
M is torsion-free and here is its proof:
Proof. Let s ∈ Z(M ). So, by definition, there is a nonzero m ∈ M such that
sm = 0. Clearly, this implies that c(sm) = 0, and so, we have sc(m) = 0.
Note that since M is a content semimodule, c(m) = 0 if and only if m = 0.
This already implies that the ideal c(m) is nonzero and so, s ∈ Z(S),
Q.E.D.
(3) It is straightforward to seeL
that if Pi isQa family of torsion-free S-semimodules,
then the S-semimodules i Pi and i Pi are also torsion-free.
Definition 1.29. Let M be an S-semimodule. We define the dual semimodule
of M , denoted by M ∗ , to be the S-semimodule M ∗ = HomS (M, S), where by
HomS (M, S), it is meant the set of all semimodule morphisms from M into S.
Proposition 1.30. Let M be an S-semimdoule. If the dual semimodule M ∗ is
nonzero, then it is torsion-free.
Proof. Clearly, if s is an element of Z(M ∗ ), then there is a nonzero semiring morphism f : M → S such that sf = 0. This means that there is an m ∈ M such that
f (m) 6= 0, while sf (m) = 0. But f (m) ∈ S. Therefore, s ∈ Z(S) and the proof is
complete.
Theorem 1.31. Let the semiring S have property (A) and G be a cancellative
torsion-free commutative monoid. Then the S[G]-semimodule M [G] is torsion-free
if and only if the S-semimodule M is torsion-free.
Proof. (⇒): Let s ∈ ZS (M ). Clearly, this implies that s ∈ ZS[G] (M [G]). But the
S[G]-semimodule M [G] is torsion-free. Therefore, ZS[G] (M [G]) ⊆ ZS[G] (S[G]). So,
s ∈ ZS (S).
(⇐): Let f ∈ ZS[G] (M [G]). By Theorem 1.1, there is a nonzero m ∈ M such
that c(f ) · m = 0, which means that c(f ) ⊆ ZS (M ). Since M is torsion-free,
c(f ) ⊆ ZS (S), and since S has property (A), f ∈ ZS[G] (S[G]) and the proof is
complete.
2. McCoy Semialgebras
The main purpose of this section is to introduce McCoy semialgebras. Our
definition for McCoy semialgebras is inspired by a classical result in commutative
algebra which states that if R is a commutative ring and f ∈ R[X] is a zero-divisor
on R[X], then there is a nonzero r ∈ R such that r · f = 0 [25, Theorem 2]. In
order to define McCoy semialgebras, we need to recall the definition of content
functions and to define Ohm-Rush semialgebras. We recall that an R-algebra B is
an Ohm-Rush algebra, if R as an R-module is a content module [9, Definition 2.1].
The concepts of content modules and algebras were introduced and investigated
in [35], [6], and [37].
Definition 2.1. Let S be a semiring and B an S-semialgebra.
(1) The function c from B into ideals of S is called the content function if c(f )
is the intersection of all ideals I of S such that f ∈ IB for each f ∈ B.
(2) We define an S-semialgebra B to be an Ohm-Rush S-semialgebra, if f ∈
c(f )B for each f ∈ B.
12
PEYMAN NASEHPOUR
The proof of the following proposition is straightforward, but we bring it here
only for the sake of reference.
Proposition 2.2. If B is an Ohm-Rush S-semialgebra, then the following statements hold:
(1) For any f ∈ B, c(f ) = (0) if and only if f = 0
(2) For any f ∈ B, c(f ) is a finitely generated ideal of S;
(3) For any ideal I of S, c(f ) ⊆ I if and only if f ∈ IB;
(4) For each f, g ∈ B, c(f g) ⊆ c(f )c(g).
Imagine B is an Ohm-Rush S-semialgebra and take f ∈ B and s ∈ S. By
definition, f ∈ c(f )B and this implies that s · f ∈ s · c(f )B. So, c(s · f ) ⊆ s · c(f )
and therefore, if s · c(f ) = (0), then c(s · f ) = (0), which implies that s · f = 0. On
the other hand, we know that if B is an Ohm-Rush R-algebra, then c(r ·f ) = r ·c(f )
for any r ∈ R and f ∈ B if and only if B is a flat R-algebra [35, Corollary 1.6].
Therefore, for flat Ohm-Rush R-algebra B, r · f = 0 implies r · c(f ) = (0) for any
r ∈ R and f ∈ B. So, the question arises if for any S-semialgebra B, s · f = 0
implies s · c(f ) = 0. The following example shows that this is not the case even for
some Ohm-Rush algebras.
Example 2.3. Let (R, m) be a discrete valuation ring, where m is its unique
maximal ideal. Let B = R/m2 . By [35, Proposition 2.1], B is an Ohm-Rush Ralgebra. Now let g and f be in B − {0} such that gf = 0. If we take f = r + m2
and g = s + m2 , we have r, s ∈
/ m2 , while rs ∈ m2 and in the language of valuation
rings, we get 0 ≤ v(r), v(s) ≤ 1, while v(r) + v(s) = v(rs) ≥ 2. These two imply
that v(r) = v(s) = 1. Obviously, in the R-algebra B, we have sf = 0, since
s(r + m2 ) T
= rs + m2 = m2 . Now let usTcalculate c(f ):
= T {I ∈ Id(R) : (r + m2 ) ∈ I · (R/m2 )} =
T c(f ) = {I ∈ Id(R)2 : f ∈ IB}
2
{I ∈ Id(R) : (r + m ) ∈ I/m } = {I ∈ Id(R) : r ∈ I} = (r) and obviously
sc(f ) = s(r) = (sr) 6= (0), since r and s are nonzero and R is an integral domain.
This argument is the base for our definition for McCoy semialgebras:
Definition 2.4. Let B be an S-semialgebra. We define B to be a McCoy Ssemialgebra, if B is an Ohm-Rush S-semialgebra and g · f = 0 for g, f ∈ B with
g 6= 0 implies s · c(f ) = 0 for some nonzero s ∈ S.
Examples 2.5. Here we give some examples for McCoy semialgebras:
(1) Let X be an indeterminate on a semiring S. It is clear that by Corollary
1.2, S[X] is a McCoy S-semialgebra.
(2) Let B be a content S-semialgebra [30, Definition 30], then by [30, Proposition 31], B is a McCoy S-semialgebra.
Remark 2.6. Let us recall that David Fields, a student of Robert Gilmer, has
proved that if R is a commutative Noetherian ring and a nonzero power series
f ∈ R[[X]] is a zero-divisor in power series ring R[[X]], then there exists a nonzero
element r ∈ R such that r · f = 0 [10, Theorem 5]. We also recall that if B is a
content R-algebra and a nonzero element f ∈ B is a zero-divisor in B, then there
exists a nonzero element r ∈ R such that r·f = 0 [35, Statement 6.1]. In fact, David
Field’s Theorem for zero-divisors of formal power series is a corollary of this fact
that formal power series R[[X]] over a Noetherian ring R is a content R-algebra, an
important theorem that was recently proved by Epstein and Shapiro [7, Theorem
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
13
2.6]. On the other hand, if S is a Noetherian semiring, X is an indeterminate over
S, and f = s0 + s1 X + s2 X 2 + · · · + sn X n + · · · is an element of S[[X]], then c(f )
is the ideal generated by the coefficients of f and f ∈ c(f )S[[X]] [30, Proposition
42]. Therefore, S[[X]] is an Ohm-Rush S-semialgebra. Also, one can easily check
that c(sf ) = sc(f ) for all s ∈ S and f ∈ S[[X]] if S is Noetherian. This motivates
us to propose the following questions:
Questions 2.7. Let S be a Noetherian semiring.
(1) Is S[[X]] a McCoy S-semialgebra?
(2) Is S[[X]] is a content S-semialgebra?
Now we proceed to give more examples for McCoy semialgebras. First we recall
that an S-semialgebra B is called a weak contentp
semialgebra if B is an Ohm-Rush
semialgebra and the content formula c(f )c(g) ⊆ c(f g) holds, for all f, g ∈ B [30,
Definition 36]. Also, a semiring S is said to be nilpotent-free, if sn = 0 implies s = 0
for all s ∈ S and n ∈ N. Now we give an interesting family of McCoy semialgebras
in the following:
Proposition 2.8. Let S be a nilpotent-free semiring and B a weak content Ssemialgebra. Then B is a McCoy S-semialgebra.
p
c(gf ) =
Proof.
Let
gf
=
0,
where
g,
f
∈
B
and
g
=
6
0.
Consequently
c(g)c(f
)
⊆
√
0 = 0. Now c(g) 6= 0, since g 6= 0. Take a nonzero element s ∈ c(g). Obviously,
s · c(f ) = (0). Q.E.D.
Remark 2.9. Let S be a semiring and B be an Ohm-Rush S-semialgebra. In
Examples 2.5, we have already mentioned that if B is a content S-semialgebra,
then it is a McCoy S-semialgebra. On the other hand, in Proposition 2.8, we have
shown that if S is nilpotent-free and B is a weak content S-semialgebra, then B is
a McCoy S-semialgebra.
In the following, we show that there exists a McCoy semialgebra that is not a
weak content semialgebra:
Let S be a semiring such that one of its prime ideals is not subtractive. Clearly,
by Corollary 1.2, S[X] is a McCoy S-semialgebra, while by Theorem 19 in [30], it is
not a weak content S-semialgebra. For example, consider the idempotent semiring
S = {0, u, 1}, where 1 + u = u + 1 = u [24]. It is clear that the ideal {0, u} is prime
but not subtractive. Now imagine f = 1 + uX and g = u + X. It is easy to see that
fpg = (1 + p
uX)(u + X) = u + uX + uX 2 , c(f g) = {0, u} andpc(f )c(g) = S while
c(f g) = {0, u} = {0, u} and this means that c(f )c(g) * c(f g), i.e., S[X] is
not a weak content S-semialgebra, while it is a McCoy S-semialgebra. From this
discussion, we propose the following question:
Question 2.10. Is there any weak content semialgebra that is not a McCoy semialgebra?
Since content semialgebras in general and weak content semialgebras over nilpotentfree semirings are good examples for McCoy semialgebras, we devote the rest of this
section to these semialgebras.
Let us note that if R is a commutative ring with a nonzero identity
and G
p
a cancellative torsion-free commutative monoid, then c(f )c(g) ⊆ c(f g), for all
f, g ∈ R[X]. Still, there are some semirings that this content formula does not
hold [30, Example 17]. In fact, a semiring S is called weak Gaussian if c(f )c(g) ⊆
14
PEYMAN NASEHPOUR
p
c(f g), for all f, g ∈ R[X]. And it has been proved in Theorem 19 in [30] that
a semiring S is weak Gaussian if and only if each prime ideal of S is subtractive.
Now we give the following theorem which is a generalization of Theorem 19 in [30]:
Theorem 2.11. Let S be a semiring and G a cancellative torsion-free commutative
monoid. Then the following statements are equivalent:
p
(1) c(f
√ g) ⊆ c(f )c(g) ⊆ c(f g), for all f, g ∈ S[G],
(2) I is subtractive for each ideal I of the semiring S,
(3) Each prime ideal p of S is subtractive.
√
Proof. (1) ⇒ (2): Suppose I √
is an ideal of S and a, b ∈ S such that a + b, a ∈ I.
We need to show that b ∈ I. Let v ∈ G − {0} and put f = a + bX v √and
g = b + (a + b)X v . Then just like the proof of [30, Theorem 19], we have b ∈ I.
(2) ⇒ (3): Obvious.
(3) ⇒ (1):
p Let each prime ideal p of S be subtractive. We need to show that
c(f )c(g) ⊆ c(f g), for all f, g ∈ S[G]. Let f, g ∈ S[G] and suppose that p is a
prime ideal of S and c(f g) ⊆ p. Obviously, f g ∈ p[G]. Now by Theorem 1.3, the
ideal p[G] is a prime ideal of S[G] and so, either f ∈ p[G] or g ∈ p[G] and this
means that either c(f ) ⊆ p or c(g) ⊆ p and in any case c(f )c(g)√⊆ p. Consequently,
T
by [13, Proposition 7.28 (Krull’s Theorem)] - that says that I = p∈SpecI (S) p,
where by SpecI (S) we mean the
p set of all prime ideals of S containing I, we have
T
c(f )c(g) ⊆ p∈Specc(f g) (S) p = c(f g) and the proof is complete.
Corollary 2.12. Let S be a semiring and G a cancellative torsion-free commutative
monoid. Then the following statements hold:
(1) If S is weak Gaussian, then S[G] is a weak content S-semialgebra.
(2) If S is weak Gaussian and nilpotent-free, then S[G] is a McCoy S-semialgebra.
In Theorem 2 of the paper [34], Douglas Geoffrey Northcott (1916–2005) proves
that if R is a commutative ring with a nonzero identity and G is a cancellative
torsion-free commutative monoid, then R[G] is a content R-algebra. In the following, we generalize this great result for subtractive semirings:
Theorem 2.13. Let S be a semiring and G a cancellative torsion-free commutative
monoid. Then S[G] is a content S-semialgebra if and only if S is a subtractive
semiring.
Proof. (⇒): Let I be an ideal of S. Take v ∈ G − {0} and a, b ∈ S such that
a + b, a ∈ I. Define f = 1 + X v and g = a + bX v + aX 2v . It is easy to see that
f g = a+(a+b)X v +(a+b)X 2v +aX 3v , c(f ) = S, c(g) = (a, b) and c(f g) = (a+b, a).
But according to our assumption, Dedekind-Mertens content formula holds and
therefore, there exists an m ∈ N0 such that c(f )m+1 c(g) = c(f )m c(f g). This
means that (a, b) = (a + b, a), which implies b ∈ (a + b, a) ⊆ I and I is subtractive.
(⇐): Take f, g ∈ S[G]. The same discussion, in the proof of Theorem 1.3,
shows that there exists a finitely generated torsion-free Abelian group G0 such that
f, g ∈ S[G0 ], which means that f, g can be considered to be elements of a Laurent
polynomial semiring with finite number of indeterminates. So, by Theorem 6 in [30],
Dedekind-Mertens content formula holds for f, g and this finishes the proof.
Remark 2.14. In Theorem 3 in [32], it has been shown that if R is a ring and M is
a commutative monoid and R[M ] is a content R-algebra, then M is a cancellative
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
15
torsion-free monoid. This is not necessarily the case if R is a proper semiring. For
example, let B = {0, 1} be the Boolean semifield. Now let M be a monoid with at
least two elements. It is, then, easy yo see that the monoid semiring B[M ] is an
entire semiring. Also note that if f ∈ B[M ] is nonzero, then c(f ) = B. Therefore,
from all we said, we see that if f, g ∈ B[M ] are both nonzero, then c(f g) = B =
c(f )c(g). On the other hand, if either f = 0 or g = 0, then c(f g) = c(f )c(g) = (0),
which means the B[M ] is a content (in fact, Gaussian) B-semialgebra, while M is
quite arbitrary.
We end this section with the following statement for McCoy semialgebras:
Proposition 2.15. Let B be a McCoy S-semialgebra. If S is an entire semiring,
then so is B.
Proof. Let g · f = 0, where f, g ∈ B and g 6= 0. By definition, there exists a nonzero
s ∈ S such that s · c(f ) = 0. Since S is entire, c(f ) = 0 and finally f = 0.
3. Zero-Divisors of McCoy Semialgebras
In this section, we explore some properties of the set of zero-divisors of McCoy
semialgebras. We also introduce the concept of strong Krull primes of semirings.
We start our investigation with the following lemma:
Lemma 3.1. Let S be a semiring and pi a subtractive prime
S ideal of S for any
1 ≤ i ≤Sn. If B is a McCoy S-semialgebra, then Z(S) ⊆ ni=1 pi implies that
Z(B) ⊆ ni=1 (pi B).
Proof. Let f ∈ Z(B). Since B is a McCoy S-semialgebra, there is some
Sn s ∈ S − {0}
such that s · c(f ) = 0. This means that c(f ) ⊆ Z(S). But Z(S) ⊆ i=1 pi , so by
Theorem 1.15, there is an 1 ≤ i ≤ n such that c(f ) ⊆ pi and finally f ∈ pi B.
Let us recall that if R is a ring and B is an Ohm-Rush R-algebra, then B is
flat if and only if c(r · f ) = r · c(f ) for any r ∈ R and f ∈ B [35, Corollary 1.6].
Since the corresponding property for the content function on semialgebras is useful
as we will see in this section very soon, we believe it is a good idea to give a name
to this property. It is good to mention that the use of the term “homogeneous”
in the following definition stems from the definition of “homogeneous functions” in
mathematical analysis (cf. [40, Chap. 1, 4.1]).
Definition 3.2. Let B be an Ohm-Rush S-semialgebra. We say the content function c is homogeneous (of degree 1), if c(s · f ) = s · c(f ) for each s ∈ S and f ∈ B.
Proposition 3.3. Let S be a semiring. Then the following statements hold:
(1) If B = S[G], where G is a commutative monboid, then the content function
of the S-semialgebra B is homogeneous.
(2) If S is Noetherian, then the content function of the S-semialgebra S[[X]] is
homogeneous.
Proof. (1): Let f = s1 X g1 + · · · + sn X gn be an element of the monoid semiring
S[G]. Then by using Proposition 23 in [30], one can easily see that the content of
f is the ideal (s1 , . . . , sn ). Therefore, c(sf ) = sc(f ), for all s ∈ S.
(2): Let f = s0 + s1 X + · · · + sn X n + · · · be an element of S[[X]] and set
Af to be the ideal generated by the coefficients of f . If S is Noetherian, then
Af = c(f ) [30, Proposition 42]. Therefore, for any s ∈ S, we have the following:
16
PEYMAN NASEHPOUR
c(sf ) = Asf = sAf = sc(f ).
This finishes the proof.
Lemma 3.4. Let B be an Ohm-Rush S-semialgebra with homogeneous c and the
corresponding homomorphism λ of the S-semialgebra B be injective. Then the
following statements hold:
(1) AnnS (s)B S
= AnnB (λ(s)) for any
S s ∈ S;
(2) If Z(B) = qi , then Z(S) = (qi ∩ S).
Proof. (1): f ∈ AnnB (λ(s)) ⇔ s · f = 0 ⇔ s · c(f ) = 0 ⇔ c(f ) ⊆ AnnS (s) ⇔ f ∈
AnnS (s)B.
(2): Let s ∈ Z(S) be nonzero. So, there is a nonzero s′ ∈ Z(S) such that
s · s′ = 0. This implies that λ(s) · λ(s′ ) = 0. Since λ is injective, λ(s) and λ(s′ )
are both nonzero in B. So, λ(s) ∈ Z(B) and therefore, by our assumption, there
is an i such that λ(s) ∈ qi , i.e., s ∈ qi ∩ S. Now let s ∈ qi ∩ S. So, λ(s) ∈ qi and
this implies that λ(s) is a zero-divisor on Z(B). This means that there is a nonzero
g ∈ B such that s · g = 0. Since c is homogeneous, s · c(g) = (0) and if we choose
a nonzero element t of c(g), we have s · t = 0, which means that s ∈ Z(S) and the
proof is complete.
Let us recall that a semiring S has very few zero-divisors if the set of zero-divisors
Z(S) of S is a finite union of primes in Ass(S) [30, Definition 48]. For instance,
by Theorem 49 in [30], any Noetherian semiring has very few zero-divisors. Rings
and modules having very few zero-divisors were introduced and studied in [31], [33]
and [32].
Theorem 3.5. Let S be a weak Gaussian semiring and B a McCoy and weak
content S-semialgebra with homogeneous c and the corresponding homomorphism λ
of the S-semialgebra B be injective. Then S has very few zero-divisors if and only
if so does B.
Proof. (⇒): Let S have very few zero-divisors. So by definition, Z(S) is a finite
union of primes pi in Ass(S). Our claim is that Z(B) is unionSof the ideals pi B.
n
Since pi = Ann(si ) is subtractive, by Lemma 3.1, Z(B) ⊆ i=1 (pi B). Now let
f ∈ pi B. So c(f ) ⊆ pi . But c is homogeneous, so si · f = 0 and this means that
f ∈ Z(B). Also, note that since B is a weak content S-semialgebra, by Lemma 3.4,
pi B ∈ Ass(B), in which
Sn B has very few zero-divisors.
(⇐): Let Z(B) =S i=1 qi , where qi ∈ Ass(B) for all 1 ≤ i ≤ n. Therefore, by
n
Lemma 3.4, Z(S) = i=1 (qi ∩ S). Without loss of generality, we can assume that
qi ∩ S * qj ∩ S for all i 6= j. Now we prove that qi ∩ S ∈ Ass(S) for all 1 ≤ i ≤ n.
Consider f ∈ B such that qi = Ann(f ) and c(f ) = (s1 , s2 , . . . , sm ). It is easy to
see that qi ∩ S = Ann(c(f )) ⊆ Ann(s1 ) ⊆ Z(S) and since every prime ideal of S is
subtractive, by Theorem 1.15, qi ∩ S = Ann(s1 ).
Corollary 3.6. Let S be a weak Gaussian semiring. Then the following statements
hold:
(1) If X is an indeterminate over S, then S has very few zero-divisors if and
only if S[X] does so.
(2) If X1 , X2 , . . . , Xn are distinct indeterminates over S, then S has very few
zero-divisors if and only if S[X1 , X2 , . . . , Xn ] does so.
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
17
(3) If G is a cancellative torsion-free commutative monoid, then S has very few
zero-divisors if and only if S[G] does so.
Remark 3.7.
(1) Let S be a semiring. Then it is clear that any content Ssemialgebra is a weak content and a McCoy S-semialgebra. Now we show
that there are weak content and McCoy S-semialgebras that are not content
S-semialgebras:
(2) Let S be a weak Gaussian and non-subtractive semiring, that is, all its
prime ideals are subtractive, while S possesses an ideal that is not subtractive. Note that there are such semirings. For example, refer to Proposition
21 in [30]. Then the S-semialgebra S[X] is a weak content semialgebra,
which obviously satisfies McCoy property (See Corollary 1.2 of the current
paper), but still it is not a content S-semialgebra. So we have already obtained a weak content and a McCoy S-semialgebra that is not a content
S-semialgebra.
Theorem 3.8. Let B be a McCoy S-semialgebra, which the corresponding homomorphism λ is injective. Let the content function c : B −→ Fid(S) be homogeneous
and onto, where by Fid(S), we mean the set of finitely generated ideals of S. Then
the following statements are equivalent:
(1) S has Property (A),
(2) For all f ∈ B, f is a regular element of B if and only if c(f ) is a regular
ideal of S.
Proof. (1) ⇒ (2): Let S have Property (A). If f ∈ B is regular, then for all nonzero
s ∈ S, s · f 6= 0 and so for all nonzero s ∈ S, s · c(f ) 6= (0), i.e. Ann(c(f )) = (0) and
according to the definition of Property (A), c(f ) 6⊆ Z(S). This means that c(f ) is
a regular ideal of S. Now let c(f ) be a regular ideal of S. So, c(f ) 6⊆ Z(S) and
therefore, Ann(c(f )) = (0). This means that for all nonzero s ∈ S, s · c(f ) 6= (0)
and so, for all nonzero s ∈ S, s · f 6= 0. Since B is a McCoy S-semialgebra, f is not
a zero-divisor of B.
(2) ⇒ (1): Let I be a finitely generated ideal of S such that I ⊆ Z(S). Since the
content function c : B → Fid(S) is onto, there exists an f ∈ B such that c(f ) = I.
But c(f ) is not a regular ideal of S, therefore, according to our assumption, f is not
a regular element of B. Since B is a McCoy S-semialgebra, there exists a nonzero
s ∈ S such that s · c(f ) = 0 and this means that s · I = (0), i.e. I has a nonzero
annihilator and the proof is complete.
Remark 3.9. In the above theorem the surjectivity condition for the content function c is necessary, because obviously S is a content S-semialgebra and the condition (2) is satisfied, while one can choose the semiring S such that it does not have
Property (A) (Cf. [21, Exercise 7, p. 63]).
Theorem 3.10. Let S be a weak Gaussian semiring and B a McCoy and weak
content S-semialgebra such that the corresponding homomorphism λ is injective
and c : B → Fid(S) is a homogeneous and an onto function. Then B has few
zero-divisors if and only if S has few zero-divisors and property (A).
S
Proof. (⇐): Let Z(S) = pi . Take f ∈ pi B, for some i. So, c(f ) ⊆ pi ⊆ Z(S).
Since S has property (A), there is some nonzero s ∈ S such that s · c(f ) = (0).
Since c is homogeneous and λ is injective, f ∈ Z(B). Now by lemma 3.1, the proof
of this part is complete.
18
PEYMAN NASEHPOUR
(⇒): By considering Lemma 3.4, we only need to prove that S has property (A).
Let I ⊆ Z(S) be a finitely generated ideal of S. Since c : B → Fid(S) is onto, there
is an f ∈ B such that c(f ) =SI and by Prime Avoidance Theorem for Semirings,
c(f ) ⊆ qi ∩ S, where Z(B) = qi . Now we have f ∈ (qi ∩ S)B ⊆ qi ⊆ Z(B). But
B is a McCoy S-semialgebra. So, there is some nonzero s ∈ S such that s · I = (0).
Q.E.D.
Corollary 3.11. Let S be a weak Gaussian semiring. Then S[X1 , X2 , . . . , Xn ] has
few zero-divisors if and only if S has few zero-divisors and property (A).
A prime ideal p of a commutative ring R is said to be a strong Krull prime
of R, if for any finitely generated ideal I of R, there exists a z ∈ R such that
I ⊆ Ann(z) ⊆ p, whenever I ⊆ p [26]. For a nice introduction to strong Krull
primes and their properties, one can refer to a recent paper by Epstein and Shapiro
[8].
Definition 3.12. We define a prime ideal p of a semiring S to be a strong Krull
prime of S, if for any finitely generated ideal I of S, there exists a z ∈ S such that
I ⊆ Ann(z) ⊆ p, whenever I ⊆ p.
Lemma 3.13. Let B be a weak content S-semialgebra such that the corresponding
homomorphism λ is injective and c is homogeneous. If p is a strong Krull prime
of S, then either pB = B or pB is a strong Krull prime of B.
Proof. Let pB 6= B and J be a finitely generated ideal of B such that J ⊆ pB,
where p is a strong Krull prime of S. Assume that J = (f1 , . . . , fk ) for f1 , . . . , fk ∈
B. This means that fi ∈ pB for all 1 ≤ i ≤ n. This implies that c(fi ) ⊆ p for all
1 ≤ i ≤ n. Since p is a strong Krull prime of S and c(f1 )+· · ·+c(fk ) ⊆ p is a finitely
generated ideal of S, there exists an s ∈ S such that c(f1 )+· · ·+c(fk ) ⊆ Ann(s) ⊆ p.
Obviously, J is annihilated by s, i.e. J ⊆ Ann(λ(s)). The final phase of the proof
is to show that Ann(λ(s)) ⊆ p. Let f ∈ Ann(λ(s)). So s · f = 0 and therefore,
s · c(f ) = 0. This means that c(f ) ⊆ Ann(s) ⊆ p and finally f ∈ pB.
Theorem 3.14. Let B be a McCoy and weak content S-semialgebra such that the
corresponding homomorphism λ is injective and c is homogeneous. If Z(S) is a
finite union of strong Krull primes of S, then Z(B) is a finite union of strong Krull
primes of B.
Proof. Let f ∈ pi B for some i. So, c(f ) ⊆ pi . But pi is a strong Krull prime of
S and c(f ) is a finitely generated ideal of S. So, there exists a z ∈ S such that
c(f ) ⊆ Ann(z).
This implies that f ∈ Z(B). On the other hand, by Lemma 3.1,
Sn
Z(B) = i=1 (pi B) and at least one of those pi Bs is a proper ideal of B. Therefore,
by Lemma 3.13, if pi B 6= B, then pi B must be a strong Krull prime of B and this
completes the proof.
Corollary 3.15. Let S be a semiring and B a content S-semialgebra such that c
is homogeneous. If Z(S) is a finite union of strong Krull primes of S, then Z(B)
is a finite union of strong Krull primes of B.
Proof. Since any content semialgebra is a McCoy [30, Proposition 31] and weak
content semialgebra [30, Proposition 37], by Theorem 3.14, the statement holds
and the proof is complete.
ON ZERO-DIVISORS OF SEMIMODULES AND SEMIALGEBRAS
19
Corollary 3.16. Let S be a weak Gaussian semiring. If Z(S) is a finite union
of strong Krull primes of S, then Z(S[X1 , X2 , . . . , Xn ]) is a finite union of strong
Krull primes of S[X1 , X2 , . . . , Xn ].
Proof. Since S is a weak Gaussian semiring, each prime ideal of S is subtractive [30,
Theorem 19]. Therefore, each ideal pS[X1 , X2 , . . . , Xn ] is prime if p is prime
(See Theorem 1.3). So, S[X1 , X2 , . . . , Xn ] is a weak content S-semialgebra [30,
Proposition 37]. On the other hand, by Corollary 1.2, S[X1 , X2 , . . . , Xn ] is a McCoy semialgebra. So, by Theorem 3.14, if Z(S) is a finite union of strong Krull
primes of S, then Z(S[X1 , X2 , . . . , Xn ]) is a finite union of strong Krull primes of
S[X1 , X2 , . . . , Xn ].
Corollary 3.17. Let R be a commutative ring with a nonzero identity and B be a
content R-algebra. If Z(R) is a finite union of strong Krull primes of the ring R,
then Z(B) is a finite union of strong Krull primes of B.
Acknowledgements
This work is supported by Golpayegan University of Technology. Our special
thanks go to the Department of Engineering Science in Golpayegan University of
Technology for providing all the necessary facilities available to us for successfully
conducting this research. We also like to thank Prof. Dara Moazzami, the President
of Golpayegan University of Technology, for his help, motivation, and encouragement.
References
[1] Bourne, S.: The Jacobson radical of a semiring, Proc. Nat. Acad. Sci. 37 (1951), 163–170.
[2] Bruns, W., Gubeladze, J.: Polytopes, Rings, and K-Theory, Springer, Dordrecht, 2009.
[3] Bruns, W., Herzog, J.: Cohen-Macaulay Rings, revised edn., Cambridge University Press,
Cambridge, 1998.
[4] Dauns, J.: Primal modules, Comm. Algebra 25 (8) (1997), 2409–2435.
[5] Davis, E.: Overrings of commutative rings II. Integrally closed overrings, Trans. Amer.
Math. Soc, 110 (1964), 196–212.
[6] Eakin, P., Silver, J.: Rings which are almost polynomial rings, Trans. Amer. Math. Soc. 174
(1974), 425–449.
[7] Epstein, N., Shapiro, J.: A Dedekind-Mertens theorem for power series rings, Proc. Amer.
Math. Soc. 144 (2016), 917–924.
[8] Epstein, N., Shapiro, J.: Strong krull primes and flat modules, J. Pure Appl. Algebra 218
(2014), No. 9, 1712–1729.
[9] Epstein, N., Shapiro, J.: The Ohm-Rush content function, J. Algebra Appl., 15, No. 1
(2016), 1650009 (14 pages).
[10] Fields, D. E.: Zero divisors and nilpotent elements in power series rings, Proc. Amer. Math.
Soc. 27 (1971), no. 3, 427–433.
[11] Fraenkel, A.: Über die Teiler der Null und die Zerlegung von Ringen, J. Reine Angew.
Math., 145 (1914), 139–176.
[12] Gilmer, R.: Multiplicative Ideal Theory, Marcel Dekker, New York, 1972.
[13] Golan, J. S.: Semirings and Their Applications, Kluwer Academic Publishers, Dordrecht,
1999.
[14] Golan, J. S.: Semirings and Affine Equations over Them: Theory and Applications, Kluwer
Academic Publishers, Dordrecht, 2003.
[15] Gondran, M., Minoux, M.: Graphs, Dioids and Semirings, Springer, New York, 2008.
[16] Hetzel, A. J., Lewis Lufi, R. V.: On the relationship between zero-sums and zero-divisors of
semirings, Kyungpook Math. J., 49 (2009), 221–233.
[17] Hochster, M.: Topics in the homological theory of modules over commutative rings, CBMS
Regional Conf. Ser. in Math., 24, Amer. Math. Soc, Providence, RI, 1975.
20
PEYMAN NASEHPOUR
[18] Huckaba, J. A., Keller, J. M.: Annihilation of ideals in commutative rings, Pac. J. Math.
83 (1979), 375–379.
[19] Hebisch, U., Weinert, H. J.: Semirings - Algebraic Theory and Applications in Computer
Science, World Scientific, Singapore, (1998).
[20] Heil, C., Speegle, D.: The HRT conjecture and the zero divisor conjecture for the Heisenberg
group, In: Balan R., Begué M., Benedetto J., Czaja W., Okoudjou K. (eds) Excursions in
Harmonic Analysis, Vol. 3. Applied and Numerical Harmonic Analysis. Birkhäuser, Cham,
2015.
[21] Kaplansky, I.: Commutative Rings, Allyn and Bacon, Boston, 1970.
[22] Kaplansky, I.: “Problems in the theory of rings” revisited, Am. Math. Mon., 77, (5) (1970),
445–454.
[23] Kleiner, I.: A History of Abstract Algebra, Birkhäuser, Boston, 2007.
[24] LaGrassa, S.: Semirings: Ideals and Polynomials, PhD. Thesis, University of Iowa, 1995.
[25] McCoy, N. H.: Remarks on divisors of zero, Amer. Math. Monthly, 49 (1942), 286–295.
[26] McDowell, K.: Associated prime divisors and finitely presented modules, unpublished manuscript, Wilfrid Laurier University, Waterloo, Ontario, 1975.
[27] Nasehpour, P.: Auslander modules, Beitr Algebra Geom (2018), arXiv:1705.03980.
[28] Nasehpour, P.: Content Algebras and Zero-Divisors, dissertation, University of Osnabrück,
Osnabrück, 2011.
[29] Nasehpour, P.: Content algebras over commutative rings with zero-divisors, preprint (2008),
arXiv:0807.1835
[30] Nasehpour, P.: On the content of polynomials over semirings and its applications, J. Algebra
Appl., 15, No. 5 (2016), 1650088 (32 pages).
[31] Nasehpour, P.: Zero-divisors of content algebras, Arch. Math. (Brno), 46 (4) (2010), 237–
249.
[32] Nasehpour, P.: Zero-divisors of semigroup modules, Kyungpook Math. J., 51 (1) (2011),
37–42.
[33] Nasehpour, P., Payrovi, Sh.: Modules having very few zero-divisors, Comm. Algebra 38
(2010), 3154–3162.
[34] Northcott, D. G.: A generalization of a theorem on the content of polynomials, Proc. Cambridge Phil. Soc. 55 (1959), 282–288.
[35] Ohm, J., Rush, D. E.: Content modules and algebras, Math. Scand. 31 (1972), 49–68.
[36] Passman D. S.: The Algebraic Structure of Group Rings, Wiley-Interscience, New York,
1977.
[37] Rush, D. E.: Content algebras, Canad. Math. Bull. Vol. 21 (3) (1978), 329–334.
[38] Spindler, K.: Abtract Algebra with Applications, Vol. 1, Dekker, New York, 1994.
[39] Vandiver, H. S.: Note on a simple type of algebra in which cancellation law of addition does
not hold, Bull. Amer . Math. Soc., 40 (1934), 914–920.
[40] Widder, D. V.: Advanced Calculus, 2nd ed, Dover Publications, New York, 1989.
Peyman Nasehpour, Department of Engineering Science, Golpayegan University of
Technology, Golpayegan, Iran
E-mail address: [email protected], [email protected]
| 0math.AC
|
A Domain Specific Approach to
Heterogeneous Computing:
From Availability to Accessibility
Gordon Inggs, David Thomas
Wayne Luk
Department of Electrical and Electronic Engineering
Imperial College London
London, United Kingdom
g.inggs11;[email protected]
Department of Computing
Imperial College London
London, United Kingdom
[email protected]
Our position is that the solution to the heterogeneous
programming challenge is the use of domain specific abstraction. Software developers are already familiar with domain
specific approaches, through the popularity of programming
environments such as Matlab and software libraries such as
OpenCV. We argue that through the use of domain specific
languages and programming frameworks, three benefits may
be realised:
1) Portable, Efficient Execution: A well-established property of domain specific approaches is that the relationships between computations may be captured in greater
detail [2], [3], the exploitation of which allows for safe,
parallel scaling across different architectures.
2) Domain Specific Metrics: the application domain provides unique measures of performance, which we call
metrics. These metrics allow for the performance of
tasks upon platforms to be characterised within the
context of the domain.
3) Automatic Partitioning: Through the performance predictions provided by domain specific metric models as
well as the portable execution capability enabled by the
extraction, domain tasks could then be shared automatically across the computational platforms available to the
user in an optimal manner.
We describe our domain specific approach to heterogeneous
computing, illustrated by an example from the domain of
computational finance. We then provide further details on our
case study in computational finance to evaluate the viability
of this approach. Finally, we conclude by outlining the possibilities offered by domain specific heterogeneous computing
to software developers.
Abstract—We advocate a domain specific software development
methodology for heterogeneous computing platforms such as
Multicore CPUs, GPUs and FPGAs. We argue that three specific
benefits are realised from adopting such an approach: portable,
efficient implementations across heterogeneous platforms; domain specific metrics of quality that characterise platforms in
a form software developers will understand; automatic, optimal
partitioning across the available computing resources. These
three benefits allow a development methodology for software
developers where they describe their computational problems
in a single, easy to understand form, and after a modeling
procedure on the available resources, select how they would like
to trade between various domain specific metrics. Our work
on the Forward Financial Framework (F 3 ) demonstrates this
methodology in practise. We are able to execute a range of
computational finance option pricing tasks efficiently upon a
wide range of CPU, GPU and FPGA computing platforms. We
can also create accurate financial domain metric models of walltime latency and statistical confidence. Furthermore, we believe
that we can support automatic, optimal partitioning using this
execution and modelling capability.
I. O UR P OSITION
The increasing availability of heterogeneous computing
platforms such as Multicore CPUs, GPUs and especially
FPGAs represents both an opportunity and challenge to high
performance computing software developers. As the number
of software developers working in fields such as scientific
computing, data analytics and computational finance increase,
the need for resolution to this dilemma is pressing.
These application-focused software developers would benefit from the performance and flexibility offered by heterogeneous platforms, however often lack the detailed architectural
knowledge (or design ability in the case of FPGAs) to realise
such benefits. We are encouraged by the growing portability
offered by standards such as OpenCL [1], however the often
orthogonal paradigms through which modern computing platforms need to be programmed hinders the cooperative use of
platforms. For example, code optimised to take advantage of
the extra control logic available in multicore CPUs woefully
under-performs upon the data parallel-oriented architecture of
GPUs and vice-versa.
Copyright is held by the author/owner(s).
1st International Workshop on FPGAs for Software Programmers
(FSP 2014), September 1, 2014, Munich, Germany.
II. O UR A PPROACH
We believe that through a domain specific approach that
harnesses the three benefits outlined above, a software development model for heterogeneous computing that incorporates
FPGAs such as illustrated in Figure 1 becomes possible [4].
We have illustrated this approach using an example from the
financial engineering domain.
55
Fig. 1. Our proposed approach for Heterogeneous Computing Software Developers
import ForwardFinancialFramework as F3
U_II = F3.Heston(0.05,100,0.09,1,-0.3,2,0.09)
Task Description
O_2 = F3.Barrier(U_II,True,100,5,4096,True,120)
O_2.get_price(interactive=True)
FPGA
Latency
Generation of Design
Space
Latency
Multicore CPU
Confidence
Interval
Confidence
Interval
Heterogeneous System
Latency
F^3 System Dialog
Trade-off
Selection
Confidence
Interval
Computation
print(O_2.price)
Result
>>> $0.10 +- $0.01
56
1) The software developer specifies their task in a high
level, domain specific form. In our example, we have
used a computational financial application framework
written in the Python programming language: An asset
such as a stock or a unit of foreign currency is described
using a Heston model underlying object; a knock-out
barrier option which depends on this asset is described
using barrier option object; finally a method of the
option object is called to initiate the option pricing task.
2) The design space of the task is generated by modelling
the relationship between the domain specific metrics
of the specified task upon the computational resources
available. In our example, the platforms are a multicore
CPU and a FPGA while the two financial domain
metrics are wall-time latency of the computation and
the size of the 95% confidence interval.
3) The software developer is able to make a selection of a
particular combination of metrics. This selection allows
the developer to balance their objectives while making
optimal use of the computing resources at their disposal.
Furthermore, this objective balancing does not require
detailed understanding of the computational resources
available. In our example, the user makes their selection
on a Pareto tradeoff curve of latency and the size of the
statistical confidence interval, a commonly used metric
of quality from the computational finance domain.
4) The task is then executed upon the resources available
so as to achieve the specified combination of metrics. In
our example, this would be to perform the pricing task
to the required degree of statistical confidence.
5) The result is then available to the user. In our example,
this is the option price, which is an attribute of the
Barrier Option object in the application framework.
TABLE I
L ATENCY S PEEDUP OF F 3 I MPLEMENTATIONS AND R EFERENCES
OVER S EQUENTIAL CPU I MPLEMENTATION
Target Platform
F3
Reference
Xilinx 7Z045
Altera Stratix V
GXA7
Xilinx Virtex 6
SX475T
AMD
Opteron
6272
AMD
Firepro
W5000
Intel Xeon Phi
3120P
See [5], [6]
Platform
Type
Kaiserslatuarn BlackScholes
Heston
Asian
Option
Option
Benchmark
FPGA
9.53
9.77
FPGA
274.87
194.85
FPGA
223.93
353.59
CPU
28.99
25.36
GPU
58.40
85.67
Co-processor
156.42
421.63
FPGA
CPU
GPU
26.51
11.53
64.22
248.64
115.21
103.06
library include both the underlyings that are used to model
assets such as stocks or commodities, such as the BlackScholes or Heston Models, as well as the derivative product
contracts such as futures or options that derive value from
these underlyings.
When the software developer calls the methods associated
with obtaining the price of the derivative objects, the framework is capable of generating, compiling and executing the
necessary implementations of pricing algorithms, such as the
Monte Carlo algorithm, for a range of platforms, including
Multicore CPUs, GPUs and FPGA. This heterogeneous implementation is capable of pricing the derivative in question,
and returning the result to the software developer within the
framework as an attribute of the derivative object.
III. O UR C ASE S TUDY
To evaluate this proposed approach to software development
for heterogeneous computing, we have undertaken a case
study in the application domain of Computational Finance,
focusing on the sub-domain of forward looking derivatives
pricing. We have created the Forward Financial Framework
(F 3 ) 1 , a computational finance application framework for
heterogeneous computing that makes use of a variety of
implementation technologies such as OpenCL for GPUs and
FPGAs, Maxeler Tools and Xilinx’s Vivado for FPGAs and
multithreaded C for multicore CPUs.
We have used F 3 to prove the first two benefits outlined in
section 1, portable heterogeneous execution as well as metric
modelling, and work on proving the third benefit, automatic
partitioning is in progress.
Table 1 provides the relative latency performance of F 3 ’s
implementations of the Kaiserslatuarn Option Pricing Benchmark2 , as well as a Black-Scholes model-based Asian Option.
We have compared F 3 ’s implementations to two hand-written,
manual implementations [5]–[7].
The results illustrate that from a single high level software
task description, software developers using F 3 could execute
tasks across a wide range of heterogeneous platforms including
several different FPGA toolflows, and achieve performance
within the same order of magnitude as those implementations created by embedded computing experts. The efficiency
that a hypothetical software developer would harness results
from the expert architectural knowledge that is built into the
framework’s implementation generation capability. However,
this approach allows for the architectural knowledge of the
framework developer to be shared with many software developers.
A. Heterogeneous Execution
In F 3 , computational finance tasks are described in a high
level, domain specific manner using a library of objects in
the Python programming language. The classes within the F 3
2 http://www.uni-kl.de/en/benchmarking/option-pricing/
1 https://github.com/Gordonei/ForwardFinancialFramework
57
Fig. 2. Financial Domain Metric Modelling within F 3
650
Metric Model produced by F3
Actual Performance
500
400
300
Latency (S)
200
100
Server CPU
Maxeler FPGA
50
40
Desktop CPU
Nvidia GPU
30
20
AMD GPU
0.03
0.04
0.1
0.2
0.3
0.4
0.5
matrix as well as a matrix of domain specific metrics values,
C.
C is a matrix of domain specific metric values that has the
same shape as A, where each element represents the value
required to achieve a vector of a targeted domain specific
metric values, ~t.
To generate a computational finance design space, we would
select a ~t of statistical confidence values and then use our
metric models to find the latency values to achieve this
across the platforms achieve this, i.e. our C. Our automatic
partitioning tool would then seek A such that the value of
F (A, C) is minimised.
We believe that the methods utilised in operations research,
particularly the bottleneck assignment problem could offer
provably optimal solutions to this problems.
IV. O UR C ONCLUSION
95% Confidence Interval ($)
In this paper we have argued that domain specific abstractions provide a means to make heterogeneous computing
accessible to software developers. Our approach requires no
knowledge of the computing platforms in questions, a crucial
advantage in the case of FPGAs. Our case study in computational finance, as illustrated by F 3 shows that portable
heterogeneous execution and domain specific metric modelling
is achievable. Furthermore, it suggests the possibility of automatic task partitioning across a range of platforms.
Our approach does however require expert knowledge of
the application domain, the computational architectures being
targeted, as well as the mapping of the former onto the latter.
However, given that popular application domains have often
spawned specialised software frameworks, extending these
to support multiple platforms is not inconceivable. Although
the scope of problems that can be addressed by a particular
instance of our approach is narrow, this approach enables the
use of heterogeneous computing platforms when none could
be used before.
B. Performance Modelling
F 3 also provides the capability to model the computational
finance metrics of wall-time latency and statistical confidence
interval size for the range of platforms that execution is
supported upon. This capability builds upon the heterogeneous
execution supported by the framework. The framework’s modelling capability requires a small amount of online benchmarking, executing a subset of the computational task on the
targeted platform. The model predictions are then generated
using knowledge from the computational finance domain embedded in the framework, exploiting a priori knowledge of the
structure of the pricing tasks.
Figure 2 illustrates the performance of the predictive modelling capability of F 3 for a portfolio of option pricing tasks.
This portfolio was comprised of the Kasiserslaturn option
pricing benchmark as well as the Black Scholes Asian Option
described in the previous section. We used benchmarking runs
orders of magnitude shorter than the predication target.
Software developers in computational finance are familiar
with metrics such as latency and confidence interval size, and
so by presenting platforms in terms of trade-offs between these
metrics we make a diverse range of computing platforms including FPGAs tractable. Furthermore, as the next subsection
will describe, this feature enables automatic task partitioning.
R EFERENCES
[1] J. Stone, D. Gohara, and G. Shi, “OpenCL: A Parallel Programming
Standard for Heterogeneous Computing Systems,” Computing in Science
Engineering, vol. 12, no. 3, pp. 66 –73, May-June 2010.
[2] H. Chafi, A. K. Sujeeth, K. J. Brown, H. Lee, A. R. Atreya, and K. Olukotun, “A Domain-Specific Approach to Heterogeneous Parallelism,” in
PPOPP, 2011, pp. 35–46.
[3] D. B. Thomas and W. Luk, “A Domain Specific Language for Reconfigurable Path-based Monte Carlo Simulations,” in Proc. Int. Conf. on
Field-Programmable Technology, 2007, pp. 97–104.
[4] T. D. Braun, H. J. Siegel, and A. A. Maciejewski, “Heterogeneous
computing: Goals, methods, and open problems,” in High Performance
Computing—HiPC 2001. Springer, 2001, pp. 307–318.
[5] C. de Schryver, I. Shcherbakov, F. Kienle, N. Wehn, H. Marxen,
A. Kostiuk, and R. Korn, “An Energy Efficient FPGA Accelerator for
Monte Carlo Option Pricing with the Heston Model,” in 2011 International Conference on Reconfigurable Computing and FPGAs (ReConFig),
30 2011-dec. 2 2011, pp. 468 –474.
[6] A. H. Tse, D. B. Thomas, K. H. Tsoi, and W. Luk, “Efficient
reconfigurable design for pricing asian options,” SIGARCH Comput.
Archit. News, vol. 38, no. 4, pp. 14–20, Jan. 2011. [Online]. Available:
http://doi.acm.org/10.1145/1926367.1926371
[7] G. Inggs, D. Thomas, and W. Luk, “A heterogeneous computing framework for computational finance,” in Proceedings of the 42nd International
Conference on Parallel Processing (ICPP), 2013, pp. 688–697.
C. Automatic Partitioning
We are currently investigating how we may use the heterogeneous execution and metric modelling capabilities of F 3 to
partition tasks across the platforms available automatically.
To express the task allocation problem more formally, we
seek an automatic means to find a task allocation matrix, A,
where each row represents a platform available and each column a computational task, thus each each element represents
the proportion of a specific task to a particular platform.
We seek values for A such that we minimise the value of
F (A, C), which has a value based upon the task allocation
58
| 5cs.CE
|
Sensor Fusion for Robot Control through Deep Reinforcement Learning
arXiv:1703.04550v1 [cs.RO] 13 Mar 2017
Steven Bohez, Tim Verbelen, Elias De Coninck, Bert Vankeirsbilck, Pieter Simoens and Bart Dhoedt
Abstract— Deep reinforcement learning is becoming increasingly popular for robot control algorithms, with the aim for a
robot to self-learn useful feature representations from unstructured sensory input leading to the optimal actuation policy. In
addition to sensors mounted on the robot, sensors might also be
deployed in the environment, although these might need to be
accessed via an unreliable wireless connection. In this paper, we
demonstrate deep neural network architectures that are able to
fuse information coming from multiple sensors and are robust
to sensor failures at runtime. We evaluate our method on a
search and pick task for a robot both in simulation and the
real world.
I. INTRODUCTION
Deep reinforcement learning has recently been applied to
transform raw sensory input to control signals for various
robotic tasks such as locomotion [20], grasping [12], manipulation [6], autonomous driving [7], etc. Generally the
sensors and actuators are treated as a single, static system and
the policy is trained on a fixed input size. With the current
trend towards the Internet-of-Things, more and more sensor
inputs can be provided by the environment the (mobile) robot
operates in. However, off-board sensors are accessed over an
unreliable wireless link, and it might not be known upfront
which sensors are actually operational.
In this paper, we propose a deep learning approach to fuse
information coming from multiple, possibly remote sensors,
while being robust to sensors being unavailable at runtime. A
naive way of combining multiple inputs is by concatenating
either the raw inputs or features calculated from these inputs
in the neural network architecture [17], [5], [12], [22]. This
approach however does not scale well in the number of
inputs, and none define the behavior in the case of a missing
input. Our contribution is twofold. First, we explore different
neural network architectures for fusing multiple inputs that
can be trained end-to-end with deep reinforcement learning
methods. Second, we use a DropPath regularization method
during training to make the neural network robust to missing
inputs at runtime.
We test our approach on a search and pick scenario, in
which a robot consisting of an omnidirectional platform with
a 5-DOF manipulator has to drive towards a can and pick
it up. As input, the robot receives lidar data from a sensor
mounted on the robot, as well as sensors mounted in the
environment. Our neural networks are trained and evaluated
in simulation, and we also present preliminary results transferring the trained policies to a real Kuka YouBot.
The remainder of the paper is structured as follows. We
discuss related work in Section II. In Section III we present
1
the detailed setup of our experiments. Section IV presents
our neural network architectures for fusing multiple inputs
and DropPath regularization. In Section V we present our
experimental results and finally Section VI concludes this
paper.
II. RELATED WORK
In this section, we will first discuss some important related
work on both sensor fusion and deep reinforcement learning.
A. Sensor Fusion
Combining information from several sources in order
to get a unified picture has been a research topic for
decades [10]. An often used sensor fusion technique is
(Extended) Kalman Filtering and variants, which has been
shown useful for use cases such as object position estimation [23], robot pose estimation [3], localization [15] and
navigation [14]. In these cases the desired state representation
based on the sensor input (e.g. the robot pose) is known and
fixed upfront. However, in our approach we want to learn a
policy end-to-end and the optimal feature representation to
fuse the sensor input to is unknown upfront.
Work has been done on fusing multiple inputs using a
deep learning approach, by learning a shared representation
between all inputs [17]. Eitel et al. [5] present a twostream Convolutional Neural Network (CNN) for RGB-D
object recognition. Both RGB and D data are first separately
processed by a pretrained CNN, whose outputs are then
concatenated and converged in a fully connected layer with
softmax classifier. In [12], RGB-D data is used for detecting robotic grasps. A structured regularization penalty is
added when concatenating multi-modal features. Srivastava
et al. [22] fuse bi-modal image-text and audio-video data
using deep Boltzman machines. In this paper, we use neural
networks to fuse lidar data from multiple sensors, which
are trained end-to-end using a deep reinforcement learning
approach.
B. Deep Reinforcement Learning
Deep reinforcement learning uses deep neural network
models as function approximators. The neural network can
either represent the policy itself by mapping observations to
actions directly, or a value function from which a policy can
be derived. In this paper we mainly consider the action-value
function Q (s, a), defined as
"
#
T
1 Authors
are with Ghent University - imec, IDLab, Department of
Information Technology. [email protected]
Q (s, a) = Eπ
∑ γ k−t rk st = s, at = a
k=t
.
(1)
(a)
(b)
Fig. 1. Our real setup (a) with a Kuka Youbot in a rectangular cage, equipped with a Hokuyo lidar. Two more Hokuyo sensors are deployed in the corners
of the cage forming the diagonal. We replicated the same setup in the V-REP simulator (b) for training.
Q (s, a) represents the expected discounted return, or Qvalue, of executing an action a in a given state s and
executing policy π afterwards. The reward r in each time
step is discounted with a factor γ ∈ [0, 1], which gives more
weight to short-term rewards. The policy π (a|s) determines
which action to take in each state. The optimal policy
is found when following the optimal Q function Q∗ =
maxπ Eπ [Q(s, a)]. In Q-learning, Q∗ is typically found by
iteratively taking the action with the highest Q-value, i.e.
π (a|s) = arg maxa Q (s, a), until convergence.
Deep Q-Networks (DQNs) extend Q-learning by approximating the Q-function by a deep neural network and
have achieved state-of-the-art results in playing ATARI
games [16]. This method uses the game pixel data as
input and outputs discrete actions. The performance of this
method was further improved using double Q-learning [8]
and prioritized experience replay [19].
Recently, deep Q-learning was extended for continuous
actions using Normalized Advantage Functions (NAFs) [6].
Other methods that work on continuous action spaces have
an explicit policy function and directly optimize the expected
reward using the policy gradient [20] or use an actor-critic
approach in which a separate actor and critic neural network
are used [13]. The actor network converts input to actions,
while the critic network evaluates the actor network and
provides gradient information. This approach can also be
extended to stochastic policies [9].
In the remainder of this paper we will use the DQN
algorithm for training our policies. However, our neural
network architecture for fusing multiple inputs might equally
well be applied to the other deep reinforcement learning
algorithms.
III. ENVIRONMENT SETUP
As a test case for our sensor fusion approach, we have
created a controlled environment for a search and pick task.
Our setup is depicted in Figure 1(a). We created a 1.8 by 2.7
meter rectangular arena in which a Kuka YouBot is placed.
The robot consists of an omni-directional base and a 5-DOF
manipulator. A two-finger soft gripper is mounted on the
manipulator to facilitate object picking. As a pick target we
use soda cans. Sensor information is provided by Hokuyo
URG-04LX-UG01 1D lidar sensors, of which one is mounted
on the robot base, and two more are mounted on nonadjacent
the arena corners. All sensors are configured to have a 180
degree field of view. An Nvidia Jetson TX1 embedded GPU
is mounted inside the YouBot base and provides the compute
power required for evaluating the policy. Each environment
lidar is attached to a Raspberry Pi 2 that is connected
over WiFi to the Jetson board. Both robot and sensors are
controlled using the Robot Operating System (ROS) [18].
We have also replicated this environment in simulation as
shown on Figure 1(b), using the V-REP simulator [4]. Within
the simulator, the robot and sensors expose exactly the same
interface via ROS. This allows us to control the simulated
environment in the same way as the real world.
IV. NEURAL NETWORK ARCHITECTURES FOR
DEEP RL SENSOR FUSION
We will use a DQN [16] to approximate and maximize
the action-value function Q and derive the policy.
In our case the state s is determined by the observations of
the three lidar sensors. Hence, our DQN will consist of three
inputs, one for each sensor, and of one output: the Q values
for each of the actions. For simplicity we have discretized the
action space to seven actions for the robot: move left/right,
move forward/backward, rotate clockwise/counterclockwise
and execute a (hard-coded) grip. Not only do we want to fuse
the input coming from the three sensors, we also want to be
robust against sensor failures, since two of the sensors are not
wired to the robot. However, we want to avoid unnecessarily
increasing the number of weights when new sensors are
added.
There are multiple ways to combine multiple inputs in
a neural network. As also pointed out by [12], inputs can
be concatenated to a single input for a neural network.
This approach will be denoted as “early fusion”. A second
approach is to first process the inputs in separate neural
16
32
conv-1
conv-2
width 5
stride 2
padding 2
width 5
stride 2
padding 2
(i) concat
16
32
16
32
(ii) conv
width 1
stride 1
padding 0
7
256
fc-1
fc-2
7
(iii) acc
(a) early fusion
16
32
+
+
256
(b) late fusion
Fig. 2. The early fusion architectures (a) concat all inputs as feature planes. The late fusion architectures (b) start with two convolution layers for each
sensor input, which is then merged for the fully connected layers. We present three types of late fusion: (i) simply concatenating the convolution outputs,
(ii) reducing the concatenated feature planes with a 1x1 convolution, and (iii) accumulating the convolution outputs.
network stacks, that are later combined into a number of
common layers, which we will call “late fusion”.
A. Early fusion
We start by defining a baseline model, that uses only the
sensor mounted on the robot as input. Similar to [16], this
model starts with two convolution layers, the first one with
16 filters, the second one with 32 filters, followed by a
fully connected layer with 256 hidden units and a final fully
connected layer that produces 7 outputs. In each layer we
use ReLU activations.
Next, we add the two additional sensor inputs by early
fusing. A first, small model is similar to single-sensor model,
but now the additional sensor inputs are added as additional
feature planes to the first convolution. A second, large model
triples the number of filters and hidden units in the fully
connected layer to compensate for the additional inputs.
B. Late fusion
In our late fusion models we dedicate the first two convolution layers to each sensor input, and fuse the outcoming feature representations before forwarding to the fully connected
layers, as shown in Figure 2. In this case, the ReLU activation
of the second convolution layer is applied after the merge,
before the fully connected layer. A naive approach of fusing
is to concatenate the convolution outputs into a single feature vector. This has the main disadvantage of dramatically
increasing the number of weights of the next fully-connected
layer as the number of sensors increases. This is because
we fuse at the layer with the highest number of activations.
In order to overcome this effect somewhat, we introduce
as second model that adds an additional convolutional layer
with width 1 to first collapse the concatenated feature planes.
Finally, we propose a model that accumulates the convolution
outputs directly, which requires no additional parameters but
also limits the connectivity between the feature maps from
different sensors. The intuition is that by training this model
end-to-end, the output of the second convolution layer of
each sensor will be a suitable state representation where each
sensor can add weight to each of the features. Features that
get enough weight from one or more sensor inputs will be
activated for the fully connected layer.
An overview of all models and the number of parameters
is listed in Table I.
C. DQN training and DropPath regularization
To train each model, we let agents interact with the
environment and store each tuple (s,a,r,s0 ) in an experience
pool P. The Q-network is then updated by sampling from
the experience pool and minimizing the loss
2
Li (θi ) = E(s,a,r,s0 )∼P r + γ max Q s0 , a0 ; θ̄i − Q (s, a; θi ) ,
a0
(2)
where θi are the Q-network parameters and θ̄i are the
parameters of the slow moving target as in [16].
In our environment, chances are that one or more sensor
inputs are unavailable at a given moment in time. In order
to have a policy that is robust against unavailable sensor
information, we introduce DropPath regularization, similar
to what is used in training FractalNets [11]. We apply
DropPath in our Late (acc) model, by dropping each of the
environment sensors input with a certain probability before
the accumulation.
However, applying DropPath during the policy training
with Equation 2 did not yield good results. Instead, we
first train a Q-network until convergence using Equation 2,
resulting in trained parameters θ ∗ . Next, we add DropPath,
and further train the Q-network by minimizing the loss
TABLE I
OVERVIEW OF THE DIFFERENT POLICY ARCHITECTURES
Model
Architecture
#Params
Single
conv(16)>conv(32)>fc(256)>fc(7)
266887
Early (small)
conv(16)>conv(32)>fc(256)>fc(7)
267047
Early (large)
conv(48)>conv(96)>fc(256)>fc(7)
812391
Late (concat)
concat after 3x conv(32)
796551
Late (conv)
conv (width 1) (32) after concat
275367
Late (acc)
accumulate after 3x conv(32)
272263
Acc + DP
add DropPath before accumulate
272263
0
−1
(3)
where Q (s, a; θ ∗ ) is the output of the original converged
Q-network with all sensor information available. So intuitively, we first approximate the optimal Q function using
DQN, and next we try to reconstruct this Q function with
missing sensor input.
−2
−3
Q
L j (θ j ) = E(s,a,r,s0 )∼P [Q (s, a; θ ∗ ) − Q (s, a; θ j )]2
−5
Single
Early (small)
Early (large)
Late (concat)
Late (conv)
Late (acc)
−6
V. EXPERIMENTAL RESULTS
−7
−8
0.0
0.3
0.6
0.9
1.2
Iterations
1.5
×106
Fig. 3. Average maxa Q (s, a) of mini-batches during training. After 1.5M
iterations the Q-networks converge. Also note that the Single and Early
(small) converge to a slightly lower Q-value.
and the Single model that is trained on the robot’s frontfacing sensor only. To evaluate the models’ effectiveness, we
randomly generated 100 start poses for the robot and the can,
evaluated each policy for each of these start configurations
and capture the total return collected by the agent in each
roll-out. Since the absolute reward depends on the start
situation, we compare box plots showing the distribution of
the rewards per roll-out.
Our results are shown in Figure 4. It is clear that the Single
model performs much better than the pure random baseline.
Adding information from the two additional lidar sensors
introduces an additional leap in performance, showing that
the models are indeed able to fuse the information. More
specific, we see that in starting situations where the can is
positioned behind the YouBot (out of view of the front-facing
sensor) the Single model is unable to correctly navigate
towards the can, whereas this is more often the case in
the fused models. We observed no significant difference in
0
−20
−40
Return
To speed up the learning process, we simulate multiple
agents that are uploading concurrently to a single, shared
experience pool, and limit the resolution of the simulated
lidar sensors to 128 scan points. The simulator steps at
100ms, which is the update rate of the lidar sensors, and each
simulation step is recorded as experience pool sample. We
run simulations on a compute cluster of 40 virtual machines
with 2 CPUs and 4GB RAM each. We don’t require GPU
infrastructure for training since the limiting factor is the
experience generation using the V-REP simulator.
For each simulation, a random start position and orientation for both the robot and soda can are generated. An
optimal target location for the soda can is right in front of
the robot, where a hard-coded grip action is able to pick
up the can. The goal then becomes for the robot to position
itself that the can is on the optimal target spot.
As the reward function, we use the normalized negative
distance of the can position to the optimal target location of
the robot. This results in a reward of -1 when the robot is
located at one end of the arena and the can is at the opposite
end, and a reward of 0 when the robot can actually fetch
the can with the hard coded grip action. A fixed reward of
-1 is given when the robot collides with either the border
or the can. Every roll-out is capped at 100 simulation steps,
which is enough for the robot to fetch the can from any start
configuration.
All models are trained with DIANNE [2], a distributed
deep learning framework built on Torch [1], that has builtin support for deep reinforcement learning. We update the
neural network weights according to Equation 2, but we use
double Q learning as described in [8], and pseudo-Huber loss
instead of Mean Squared Error (MSE) as it is more resilient
to outliers. For the gradient updates we use RMSProp with a
learning rate of 0.0001. All agents use ε-greedy exploration
with ε = 1 at the start and exponentially annealed to ε =
0.1. Our experience pool has a maximum size of 1 million
samples, and is updated in a first-in-first-out manner. We train
our models for 1.5 million batches of size 32. The average
maxa Q (s, a) per mini-batch is plotted on Figure 3. Note that
the Single and Early (small) converge to a slightly lower
Q-value, hinting at a lower success rate.
−4
−60
−80
−100
−120
om
nd
Ra
gle
Sin
rly
Ea
all)
(s m
rly
Ea
)
rge
(la
(
te
La
at)
nc
co
(
te
La
)
nv
co
(
te
La
c
ac
)
A. Early vs late fusion
To compare whether our models are able to fuse information coming from multiple sensors, we compare them
against two baselines: one taking purely random actions,
Fig. 4. Box plot of the returns achieved on 100 roll-outs for each of the
models. The Q-network clearly benefits from additional sensor information,
although there is little difference between late or early fusion. The models
with more parameters perform slightly better.
B. DropPath regularization
To evaluate DropPath regularization, we add DropPath to
our Late (acc) model, with a drop rate of 0.5 for the remote
sensors. The front facing sensor input is never dropped, as we
assume this one is always available since it is directly wired
into the robot. This balances equally between only having
the front facing sensor, front facing + sensor 1, front facing
+ sensor 2 and all sensors available. We refine the Late (acc)
model with additional training using Equation 3 for another
1M mini-batches on 5M experience pool samples generated
with Late (acc), resulting in the Acc + DP model.
We compare the regularized Acc + DP model to Late
(acc) when using all three or only the front facing sensor in
Figure 5. Since Late (acc) was trained on input data from all
three sensors, we see that dropping the environment sensors
has serious impact on the performance. When dropping the
environment sensors, Late (acc) performs even worse than
random. After regularization with DropPath however, we see
that Acc + DP performs on par with Single when only
the front facing sensor is available. When all sensors are
available, there is only a negligible performance difference
with the original Late (acc) model. This shows that our Acc +
DP model is indeed able to fuse information from multiple
sensor sources, while still being robust to missing sensor
information at runtime.
To evaluate the final position of the robot, we have plotted
on Figure 6 the cumulative distribution of the final distance
of the can to the “optimal” grip location where a reward
of 0 is received. All methods end in a 10 cm radius of this
location in more than half of the cases. The models that only
0
−20
Return
−40
−60
0.8
0.6
0.4
Single
Early (large)
Late (acc)
DP (3 sensors)
DP (1 sensor)
0.2
0.0
10−2
10−1
Final distance (m)
100
Fig. 6. Cumulative distribution of the final distance of the can to the
optimal grip location.
have a single sensor input perform worse than the methods
fusing multiple sensors, often missing out in the cases that
the can is positioned behind the robot.
C. Real-world results
When transferring our trained models to our real-world
setup, we experienced a number of problems. First, with a
fixed grip action it is difficult for the robot to actually pick
up a can with the two-finger gripper, as the robot has to
be positioned very accurately. Second, the lidars’ rays don’t
reflect on all material, and reflection also depends on the
angle. Especially the YouBot’s base seemed not to reflect the
lidar beams well, resulting in missing values in the sensor
inputs (which are set to 0).
We mitigated the first issue by adapting the grip action
to take feedback from the last lidar scan to estimate the
exact position of the can to the robot base. To make the
YouBot more “visible” by the environment sensors, we
covered the base with masking tape. However, imperfect
scans with missing measurements still occurred. Therefore,
we also added DropOut [21] on the sensor inputs during our
DropPath refinement, dropping lidar rays from the input with
a probability of 0.025. This made our Acc + DP model more
robust against the real-world lidar data. Results of our realworld experiments can be seen in the supplementary video
material.
VI. CONCLUSION
−80
−100
−120
1.0
Cumulative distribution
performance between early or late fusion models. The two
models with a larger amount of parameters in the first linear
layer (Early (large) and Late (concat)) perform slightly better
than the others. Also note the large negative tails of all box
plots, indicating that there are certain start configurations that
none of the models are able to solve correctly.
om
nd
Ra
gle
Sin
c
Ac
l)
(al
c
Ac
+D
P(
all)
c
Ac
le)
ng
(s i
c
Ac
P
+D
le)
ng
(s i
Fig. 5. When removing the remote sensor information from Late (acc),
this performs worse than random. However, when refining with DropPath
regularization, we get similar results as training on single sensor data, while
still performing better when all sensors are present.
In this paper, we presented methods for fusing sensor input
from multiple sources into a robot control policy using deep
reinforcement learning. We evaluated different architectures
on a search and pick task with a Kuka YouBot and multiple
lidar sensors. We have shown that neural networks are indeed
able to fuse information from multiple sources. We have
found no significant performance difference between fusing
early or late in the neural network. However, combining
our late fusion approach by means of accumulation with
DropPath regularization, we are able to train a policy network
that leverages all sensor inputs, while still being robust when
sensor input is missing.
As future work we will extend our evaluation to a wide
variety of deep reinforcement learning algorithms, more
specifically exploring recurrent policies that can better capture the environment dynamics, as well as using methods
that work on continuous action spaces as this would be
more suited for controlling the robot. Next, we would also
like to expand our approach to multi-modal sensor input,
for example combining the lidar data with RGB(-D) camera
input.
ACKNOWLEDGMENT
Steven Bohez is funded by Ph.D. grant of the Agency of
Innovation by Science and Technology in Flanders (IWT).
We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Jetson TX1 used in this work.
R EFERENCES
[1] R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A matlablike environment for machine learning. In BigLearn, NIPS Workshop,
2011.
[2] Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez,
Sam Leroux, and Pieter Simoens. Dianne: Distributed artificial neural
networks for the internet of things. In Proceedings of the 2Nd
Workshop on Middleware for Context-Aware Applications in the IoT,
M4IoT 2015, pages 19–24, 2015.
[3] Y. Dobrev, S. Flores, and M. Vossiek. Multi-modal sensor fusion for
indoor mobile robot pose estimation. In IEEE/ION Position, Location
and Navigation Symposium (PLANS), pages 553–556, April 2016.
[4] M. Freese E. Rohmer, S. P. N. Singh. V-REP: a Versatile and Scalable
Robot Simulation Framework. In IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), 2013.
[5] A. Eitel, J. T. Springenberg, L. Spinello, M. Riedmiller, and W. Burgard. Multimodal deep learning for robust RGB-D object recognition.
In IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), pages 681–687, Sept 2015.
[6] Shixiang Gu, Ethan Holly, Timothy P. Lillicrap, and Sergey Levine.
Deep Reinforcement Learning for Robotic Manipulation. CoRR,
abs/1610.00633, 2016.
[7] R. Hadsell, A. Erkan, P. Sermanet, M. Scoffier, U. Muller, and Yann
LeCun. Deep belief net learning in a long-range vision system for
autonomous off-road driving. In IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), pages 628–633, 2008.
[8] Hado V. Hasselt. Double Q-learning. In Advances in Neural
Information Processing Systems 23, pages 2613–2621, 2010.
[9] Nicolas Heess, Gregory Wayne, David Silver, Tim Lillicrap, Tom Erez,
and Yuval Tassa. Learning Continuous Control Policies by Stochastic
Value Gradients. In Advances in Neural Information Processing
Systems 28, pages 2944–2952, 2015.
[10] Bahador Khaleghi, Alaa Khamis, Fakhreddine O. Karray, and
Saiedeh N. Razavi. Multisensor data fusion: A review of the state-ofthe-art. Information Fusion, 14(1):28 – 44, 2013.
[11] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. FractalNet: Ultra-Deep Neural Networks without Residuals. arXiv preprint
arXiv:1605.07648, 2016.
[12] Ian Lenz, Honglak Lee, and Ashutosh Saxena. Deep Learning for
Detecting Robotic Grasps. International Journal of Robotics Research,
34(4-5):705–724, April 2015.
[13] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas
Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra.
Continuous control with deep reinforcement learning. In International
Conference on Learning Representations (ICLR), 2016.
[14] S. Lynen, M. W. Achtelik, S. Weiss, M. Chli, and R. Siegwart. A
robust and modular multi-sensor fusion approach applied to MAV
navigation. In IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), pages 3923–3929, Nov 2013.
[15] V. Malyavej, W. Kumkeaw, and M. Aorpimai. Indoor robot localization by RSSI/IMU sensor fusion. In 10th International Conference
on Electrical Engineering/Electronics, Computer, Telecommunications
and Information Technology, pages 1–6, May 2013.
[16] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu,
Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller,
Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control
through deep reinforcement learning. Nature, 518(7540):529–533,
2015.
[17] Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak
Lee, and Andrew Y. Ng. Multimodal Deep Learning. In Proceedings
of the 28th International Conference on Machine Learning (ICML),
pages 689–696, 2011.
[18] Morgan Quigley, Ken Conley, Brian P. Gerkey, Josh Faust, Tully Foote,
Jeremy Leibs, Rob Wheeler, and Andrew Y. Ng. ROS: an opensource Robot Operating System. In ICRA Workshop on Open Source
Software, 2009.
[19] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver.
Prioritized Experience Replay. CoRR, abs/1511.05952, 2015.
[20] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and
Philipp Moritz. Trust Region Policy Optimization. In Proceedings
of the 32nd International Conference on Machine Learning (ICML),
pages 1889–1897, 2015.
[21] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever,
and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural
networks from overfitting. Journal of Machine Learning Research,
15:1929–1958, 2014.
[22] Nitish Srivastava and Ruslan Salakhutdinov. Multimodal Learning
with Deep Boltzmann Machines. Journal of Machine Learning
Research, 15:2949–2980, 2014.
[23] A. W. Stroupe, M. C. Martin, and T. Balch. Distributed sensor
fusion for object position estimation by multi-robot systems. In IEEE
International Conference on Robotics and Automation (ICRA), pages
1092–1098, 2001.
| 9cs.NE
|
Deep Sparse Coding for Invariant Multimodal Halle Berry Neurons
Edward Kim1,2 , Darryl Hannan1,2 , Garrett Kenyon2
1
Department of Computing Sciences, Villanova University, PA
2
Los Alamos National Laboratory, Los Alamos, NM
arXiv:1711.07998v1 [cs.CV] 21 Nov 2017
[email protected],[email protected],[email protected]
1. Introduction
Abstract
In the past several decades, neuroscientists have been
studying the response of the brain to sensory input and theorized that humans have neurons that are colloquially known
as grandmother cells. A grandmother cell is a single neuron that responds to a specific concept or object and activates upon seeing, hearing, or sensibly discriminating that
entity, such as his or her grandmother. In 2005, neuroscientists Quiroga et al. [28] conducted a study on eight patients that exhibited pharmacologically intractable epilepsy.
These patients had electrodes implanted in their brains, enabling precise recording from multiple isolated neurons in
the medial temporal lobe (MTL). The MTL has many functions including long-term memory, language recognition,
and processing sensory input including auditory and visual
signals. Quiroga et al. sought to answer the question of
whether MTL neurons represent concept-level information
invariant to metric characteristics of images. In other words,
do MTL neurons fire selectively on individuals, landmarks,
or objects? Their results demonstrated that invariant neurons do exist, suggesting a sparse and explicit neural code.
For example, a woman had a single neuron fire when shown
a picture of Jennifer Aniston, but not on other pictures of
people, places, or things. Another patient had a different
neuron fire when shown a picture of Halle Berry, as well as
the text string “Halle Berry”, demonstrating sparse neural
codes and invariance of a neuron to concepts and specific
modalities.
Despite this research, on the computational side, neural networks have gradually moved away from biological
thematics. This has largely been due to engineering breakthroughs in the past several years that have transformed the
field of computer vision. Deep feed-forward convolutional
neural networks (CNNs) have become ubiquitous in virtually all vision challenges, including classification, detection, and segmentation. But further engineering of these
networks is reaching a saturation point where incremental
novelty in the number of layers, activation function, parameter tuning, gradient function, etc., is only producing
incremental accuracy improvements. Although there is ev-
Deep feed-forward convolutional neural networks
(CNNs) have become ubiquitous in virtually all machine
learning and computer vision challenges; however, advancements in CNNs have arguably reached an engineering
saturation point where incremental novelty results in minor
performance gains. Although there is evidence that object
classification has reached human levels on narrowly defined
tasks, for general applications, the biological visual system
is far superior to that of any computer. Research reveals
there are numerous missing components in feed-forward
deep neural networks that are critical in mammalian vision.
The brain does not work solely in a feed-forward fashion,
but rather all of the neurons are in competition with each
other; neurons are integrating information in a bottom up
and top down fashion and incorporating expectation and
feedback in the modeling process. Furthermore, our visual
cortex is working in tandem with our parietal lobe, integrating sensory information from various modalities.
In our work, we sought to improve upon the standard
feed-forward deep learning model by augmenting them with
biologically inspired concepts of sparsity, top-down feedback, and lateral inhibition. We define our model as a
sparse coding problem using hierarchical layers. We solve
the sparse coding problem with an additional top-down
feedback error driving the dynamics of the neural network.
While building and observing the behavior of our model, we
were fascinated that multimodal, invariant neurons naturally emerged that mimicked, “Halle Berry neurons” found
in the human brain. These neurons trained in our sparse
model learned to respond to high level concepts from multiple modalities, which is not the case with a standard feedforward autoencoder. Furthermore, our sparse representation of multimodal signals demonstrates qualitative and
quantitative superiority to the standard feed-forward joint
embedding in common vision and machine learning tasks.
1
idence that object classification has reached human levels
on certain narrowly defined tasks [31], for general applications, the biological visual system is far superior to that of
any computer. Research reveals there are numerous missing components in feed-forward deep neural networks that
are critical in mammalian vision. The brain does not work
solely in a feed-forward fashion, but rather all of the neurons are in competition with each other; neurons are integrating information in a bottom up and top down fashion
and incorporating expectation and feedback into the inference process. Furthermore, our visual cortex is working in
tandem with our parietal lobe, integrating sensory information from various modalities.
In the following sections, we describe our motivation
and research towards improving the standard feed-forward
deep learning model by augmenting them with biologically
inspired concepts of sparsity, top-down feedback, and lateral inhibition. Section 3.1 describes the formulation of the
problem as a hierarchical sparse coding problem, Section
3.2 explains how our model incorporates lateral inhibition
among neurons, and Section 3.3 defines the dynamics of
top-down feedback and demonstrates the effects on a toy
problem. While building and observing the behavior of our
model, we were fascinated that multimodal, invariant neurons naturally emerged that mimicked the Halle Berry neurons found in the human brain. We describe the emergence
of these invariant neurons and present experiments and results that demonstrate that our sparse representation of multimodal signals is qualitatively and quantitatively superior
to the standard feed-forward joint embedding in common
vision and machine learning tasks.
also describes the invariance of neurons to modality, i.e. a
neuron would fire not only from the picture, but also from
the text. Further evidence of invariant neurons is provided
by Giudice et al. [6] with what are called “mirror” neurons.
Mirror neurons are a result of Hebbian learning optimized
to support the association of the perception of actions and
the corresponding motor program. The biological motivation for sparsity is clear, and the computational benefits for
sparsity are many. Glorot et al. [7] describes the advantages of sparsity in a neural network. These include information disentangling, efficient variable-size representation,
and evidence that sparse representations are more linearly
separable. Glorot et al. also notes that sparse models are
more neurologically plausible and computationally efficient
than their dense counterparts. Goodfellow et al. [8] found
that sparsity and weight decay had a large effect on the invariance of a single-layer network. As a prime example,
Le et al. [19] was able to build distinctive high-level features using sparsity in unsupervised large scale networks.
Lundquist et al. [21] was able to use sparse coding to successfully infer depth selective features and detect objects
in video. Our computational network is related to these
works; we formulate our model as an unsupervised hierarchical sparse coding problem.
Aside from sparsity, other biological concepts are extremely important to consider. Both lateral and feedback
connections are utilized to perform neural computation
[29]. Lateral inhibition was discovered decades ago between neighboring columns in the visual cortex [3, 4]. Early
visual neurons (for example in V1 - the primary visual cortex) do not act as simple linear feature detectors as they
do in artificial neural networks [27]. They transform retinal signals and integrate top-down and lateral inputs, which
convey prediction, memory, attention, expectation, learning, and behavioral context. Such higher processing is fed
back to V1 from cortical and subcortical sources [25]. Topdown feedback is also thought to transmit Bayesian inferences of forthcoming inputs into V1 to facilitate perception
and reinforce the representation of rewarding stimuli in V1
[16]. In fact, there are many more feedback connections in
the brain than there are feed-forward. There is a clear motivation to include some sort of feedback connection in deep
neural networks. Recent work by Hinton’s group [32] using capsule networks describe a complex architecture where
capsules are essentially providing top-down feedback to the
lower layers. For our model, we build in both lateral inhibition and top-down feedback in a simple and elegant way.
2. Background
The current feed-forward deep neural network model
has been extremely popular and successful in recent years
spanning from the seminal works of LeCun et al. [20] to
Krizhevsky et al. [18] to He et al. [11] deep residual neural
networks. Inspiration for this model had biological underpinnings in the work from David Marr’s functional model
of the visual system [23] where levels of computation e.g.
primal sketch to 2.5D to 3D representation, mimicked the
cortical areas of the primate visual system. Indeed, CNNs
possess remarkable similarities with the mammalian visual
cortex, as shown by Hubel and Wiesel’s receptive field experiments [15].
However it’s clear that the dense feed-forward model is
not the architecture in the visual system. Jerry Lettvin [9]
first postulated the existence of grandmother cells, hyper
specific neurons that responds to complex and meaningful
stimuli. R. Quiroga et al. [28] extended this research further
by demonstrating the existence of the Jennifer Aniston neuron and Halle Berry neuron, providing evidence for extreme
sparsity when processing sensory input. Quiroga’s study
Finally, our model is multimodal as we process two distinct multimedia streams in the same network. Others have
built and trained multimodal models in neural networks
mostly incorporating vision and language [2, 17, 26, 34].
Yang et al. [35] trained a multimodal model for super resolution using high and low resolution images. Some have
2
(a) Deep Sparse Coding model
(b) Deep Feed-Forward Autoencoder
Figure 1. Illustration of the (a) deep sparse coding model and competing (b) feed forward autoencoder model. In (a), our multimodal
representation alternates between the optimization of a sparse signal representation and optimization of the dictionary elements. Within
each layer, the representation is influenced by lateral inhibition from the competing neurons as well as top down feedback from the
hierarchical layers. The reconstruction of the inputs are straightforward convolutions after an internal threshold of the membrane potentials
(activations). Then in (b) we create an equivalent architecture using a feed-forward convolutional neural network with an addition layer to
combine the two modalities, ReLU activations, and L1 regularization.
experienced problems in their network since there is no
explicit objective for the models to discover correlations
across the modalities. Neurons learn to separate internal
layer representations for the different modalities. Generation of missing modalities, when trained on two modalities,
has the additional issue that one would need to integrate out
the unobserved visible variables to perform inference [34].
We demonstrate that the neurons in our network learn invariant, joint representations between modalities. We also
show missing modality generation is trivial with our sparse
coding model.
for the text input. The last layer e.g. the “P1 Linked Dictionary”, is a joint layer of both vision and text. We build a
standard feed-forward autoencoder as a comparison model
with the identical architecture seen in Figure 1(b).
Our network can be formulated as a multimodal reconstruction minimization problem which can be defined as follows. In the sparse coding model, we have some input variable x(n) from which we are attempting to find a latent representation a(n) (we refer to as “activations”) such that a(n)
is sparse e.g. contains many zeros, and we can reconstruct
the original input, x(n) as well as possible. Mathematically,
this can be defined as,
3. Methodology
min
3.1. Deep Sparse Coding
Φ
We formulate the problem as a sparse coding problem
where we are attempting to reconstruct two modalities. The
first modality, the vision input, is a 64x64 RGB image of a
person’s face. The second modality is the text of the person’s name. To better simulate the conditions of the real
world, we decided not to represent the text as a one-hot vector typically used in language modeling. Instead, we represent the text as raw input i.e. the text modality is a 128x16
grayscale image of the printed name of the person. The
full model we developed is a highly recurrent multimodal
network seen in Figure 1(a). Our sparse coding hierarchy
consists of three layers for the visual input and two layers
N
X
1
min kx(n) − Φa(n) k22 + λka(n) k1
a(n) 2
n=1
(1)
Where Φ is the dictionary, and Φa(n) = x̂(n) , or the reconstruction of x(n) . The λ term controls the sparsity penalty,
balancing the reconstruction versus sparsity term. N is the
total training corpus, where n is one element of training.
We reformulate the reconstruction of a signal by substituting the Φ · a(n) term with Φ ~ a(n) , where ~ denotes
the transposed convolution (deconvolution) operation, and
Φ now represents a dictionary composed of small kernels
that share features across the input signal.
We use the Locally Competitive Algorithm (LCA) [30]
to minimize the mean-squared error (MSE) with sparsity
3
fine the internal state of the neuron as u(t) and the active
coefficients as a(t) = Tλ (u(t)). The internal state and active coefficients are related by a monotonically increasing
function, allowing differential equations of either variable
to descend the energy of the network. The energy of the
system can be represented as the mean squared error of reconstruction and a sparsity cost penalty C(·),
X
1
C(am (t))
(2)
E(t) = kx(t) − Φa(t)k22 + λ
2
m
cost function as described in Equation 1. The LCA algorithm is a biologically informed sparse solver that uses principles of thresholding and local competition between neurons. The LCA model is governed by dynamics that evolve
the neuron’s internal state when presented with some input image. The internal state, i.e. “membrane potential”,
charges up like a leaky integrator and when it exceeds a certain threshold, will activate that neuron. This activation will
then send out inhibitory responses to units within the layer
to prevent them from firing. The input potential to the state
is proportional to how well the image matches the neuron’s
receptive field, while the inhibitory strength is proportional
to the activation and the similarity of the two neuron’s convolutional patches, forcing the neurons to be decorrelated.
We will derive the lateral inhibition term in the following
section.
The idea of thresholding the internal state of a neuron
(membrane) is important when building deep, hierarchical
sparse coding models. Sparse coding a signal that is already
a sparse code is difficult and challenging from both a mathematical and logical point of view. Thus, stacked sparse implementations attempt to densify a sparse code before generating another sparse code. For example, some do a pooling after a sparse layer [19, 10, 24, 5] or other operations to
densify the sparse layer [12].
Our model does not require such operations. Our reconstruction activation maps are thresholded and sparse; however, the input signal passed to the next hierarchical layer is
the dense membrane potential of the current layer. Figure 2
illustrates this concept.
Thus the dynamics of each node is determined by the set of
coupled ordinary differential equations,
dum
= −um (t) + (ΦT x(t))m − (ΦT Φa(t) − a(t))m (3)
dt
where the equation is related to leaky integrators. The
−um (t) term is leaking the internal state, the (ΦT x(t)) term
is “charging up” the the state by the inner product (match)
between the dictionary element and input patch, and the
(ΦT Φa(t)−a(t)) term represents the inhibition signal from
the set of active neurons proportional to the inner product between dictionary elements. The −a(t) in this case
is eliminating self interactions. In summary, neurons that
match the input image charge up faster, then pass a threshold of activation. Once they pass the threshold, other neurons in that layer are suppressed proportional to how similar the dictionary elements are between competing neurons.
This prevents the same image component from being redundantly represented by multiple nodes.
3.3. Top-Down Feedback
When we view the world, our brains are continuously
trying to understand what it perceives. Rarely are we confused about our environment, and that is a function of our
higher cognitive functions providing feedback to the lower
sensory receptors. If there is some sort of discordance i.e.
error between the levels of cognition in our brain, our neural pathways will rectify feed-forward perception, and feedback error, gradually settling into confident inference. Similarly, we incorporate top down feedback into the model as
the error caused by the upper layers of the hierarchy. The
error propagates down to the lower layers forcing the lower
layers to adjust their representation so that it matches both
the reconstruction of the signal while mitigating the error of
the upper levels.
In the neural network, the top layer is attempting to reconstruct the bottom layer’s internal state, i.e. membrane
potential at layer N , uN (t), using the same MSE reconstruction energy. When adding this term to the lower layer,
we can take the gradient of the MSE with respect to the
membrane potential uN (t) such that,
Figure 2. Sparse codes can be extracted from any Layer N using a
thresholding of the membrane potential. The membrane potential
is passed to the next layer as a dense input for hierarchical sparse
coding.
3.2. Lateral Inhibition
The LCA model is an energy based model similar to a
Hopfield network [13] where the neural dynamics can be
represented by a nonlinear ordinary differential equation.
Let us consider a single input image at time t, x(t). We de-
dE
= eN +1 (t) = uN (t) − ΦN +1 aN +1 (t)
duN (t)
4
(4)
Figure 3. An illustrative example of top-down feedback on a toy problem. We train a small sparse coding network on 50 handwritten
characters and 50 printed characters. The handwritten characters B and 13 can be easily confused as shown in the test input. Given only
the handwritten input, the network speculates whether the character is a B or 13. Given that the network thinks this is a B, we then provide
the network with a contradictory 13 in print. The network at T1 and P1 naturally change their minds as to the label of the test input, but the
most interesting part is that H1 also changes its representation based upon the expectation from top-down feedback.
This error can thus be used as an inhibitory signal to the
driving mechanics of the lower layer, such that equation 4
is modified to,
ding (t-SNE) [22] technique to reduce the dimensionality of
the representations to a 2D plot, see Figure 4.
dum
= −um (t)+(ΦT x(t))m −(ΦT Φa(t)−a(t))m −eN +1 (t)
dt
(5)
To further elucidate the concept of top-down feedback in
our network, we present a toy example as shown in Figure
3. Our example is a small sparse coding network that takes
in a handwritten modality and its associated printed text.
The network is trained on two types of characters, “B” and
“13”. We train a small sparse coding network on 50 handwritten characters and 50 printed characters. The handwritten characters B and 13 can be easily confused as shown in
the test input. Based upon nearest neighbor clustering, the
test image would be classified as a “B” in both P1 and H1
feature space. The test input image’s euclidean distance in
P1’s 128-dim feature space to class cluster centers is (B =
0.3509 < 13 = 0.4211). At P1’s 4096-dim feature space, the
test image is very much at the border with (B = 0.6123 < 13
= 0.6131). Given that the network believes that the input image is a “B”, we introduce a contrary “13” input to the text
modality, T1. The text branch informs P1 of its extremely
strong belief that the input image is a “13” which drastically
sways P1’s opinion. This is no surprise, nor is it a feat of our
sparse coding top-down feedback. Any feed-forward neural
network has the ability to change its prediction at a higher
level based upon additional information. What makes our
model extraordinary, is that it not only changes its prediction at the high levels, but then forces H1 to conform to
its belief. After top-down feedback, P1 (B = 0.5013 > 13
= 0.3756) and H1 (B = 0.4271 > 13 = 0.4020). This has
biological basis [27] where feedback at higher areas in the
brain have been shown to suppress predictable inputs in the
visual cortex [1]. We can visualize the effect of top-down
feedback by plotting the feature representations at P1 and
H1. We use the t-Distributed Stochastic Neighbor Embed5
(a) t-SNE of P1
(b) t-SNE of H1
Figure 4. Low dimensional plots of the activation features created
by B’s and 13’s in (a) P1 and (b) H1. Originally, the test image lies
in the middle of the B cluster in P1 and closer to the B cluster in
H1. With the introduction of the printed 13, the test image changes
class drastically in P1 to the 13 cluster, and also shifts classes in
H1 with top-down feedback.
4. Experiments and Results
For our experimentation, we explored the use of faces
and names as the two modalities in the Deep Sparse Coding
(DSC) model. We evaluate our model against a standard
Feed-Forward Autoencoder (FFA) and perform an analysis
in regard to multimodal representation, feature extraction,
and missing data generation.
4.1. Dataset
We use a subset of the Labeled Faces in the Wild (LFW)
dataset [14]. This dataset consists of face photographs collected from the web and labeled with the name of the person pictured. The only constraint was that these faces were
detected by the Viola-Jones face detector. We further augment the faces in the wild dataset with more images of Halle
Berry simulating that fact that the more frequent the image,
the more likely that it will produce a grandmother cell and
Figure 5. Sample images from the Labeled Faces in the Wild
dataset. We augment the dataset with more images of Halle Berry.
(a) DSC Activation Halle Berry. (b) DSC Activation on Random.
Note the strong spike on Neuron Random input distirbutes activation
326.
across neurons.
an invariant neuron in the model. The number of images of
Halle Berry has been augmented to 370 using Google images and images scraped from IMDB, see Figure 5. However, even with this data augmentation Halle Berry is only
the second most frequently represented person; George W
Bush remains the most frequent face in the dataset with a
total of 530 images. The total number of faces used for
training and testing was 5,763 consisting of 2,189 unique
persons.
Each face image is a 64x64x3 color image. We use the
name of the person to create the corresponding image containing the text of the name in 20 point Arial font. Each text
image is a 128x16 grayscale image.
(c) FFA Activation Halle Berry.
Strong spikes on numerous neurons.
4.2. Model Implementation
(d) FFA Activation Random.
Strong spikes on the same
neurons as Halle Berry faces.
Figure 6. Average activation of faces of 75 Halle Berry and 75
Random faces on the Deep Sparse Coding (DSC) model and the
Feed-Forward Autoencoder (FFA) model. The distinct spike in
the DSC model (neuron 326) fires strongly on both the picture and
text of Halle Berry; whereas, in the FFA model, the activations are
shared by different inputs.
The Deep Sparse Coding model was implemented using OpenPV1 . OpenPV is an open source, object oriented
neural simulation toolbox optimized for high-performance
multi-core computer architectures. The model was run on
an Intel Core i7 with 12 CPUs and an Nvidia GeForce GTX
1080 GPU. We run the LCA solver for 400 iterations per
input. If we consider one epoch of the training data to be
4000 images, the training time per epoch is approximately
7 hours, or about 6.3 seconds per input image/text pair. The
Feed-forward Autoencoder was implemented using Keras2
on top of TensorFlow. We use convolutional layers with
ReLU activation functions and an addition layer to merge
the two modalities into one embedding. Activity regularization using the L1 norm is added to the model. The time
required to run through an epoch of the training data is 47
seconds, or about 0.012 seconds per input image/text pair.
We train the Deep Sparse Coding model for 3 epochs and
the Feed Forward Autoencoder for 25 epochs.
(a) N-326 activation on Vision only
(b) N-326 activation on Text only
Figure 7. The activation of neuron 326 is shown on the test set
when using a single modality as input, (a) Vision (face) only, (b)
Text (name) only. In nearly all test cases, the Halle Berry input
activates neuron 326 stronger than a non-Halle Berry input.
4.3. Multimodal Representation
ted the multimodal representation at the joint layer of the
Deep Sparse Coding model (DSC) and Feed Forward Autoencoder (FFA). The average activation of these test images can be seen in Figure 6.
In the DSC model, we found that neuron 326 actively
fires for Halle Berry input but not on other people or other
names. Upon further inspection, using only inputs from one
modality (just the face or just the text), N-326 activates with
When analyzing the joint representations created by our
model, we were fascinated that invariant neurons naturally
emerged. To explore this further, we selected 75 image/text
pairs of Halle Berry and 75 random inputs from our test
set and fed them into the network. We extracted and plot1 https://github.com/PetaVision/OpenPV
2 https://keras.io/
6
either modality, see Figure 7. Our results demonstrate the
invariance of this neuron to modality, an exciting property
that naturally emerged from the introduction of biological
concepts in the network. We call N-326 the Halle Berry
neuron of our Deep Sparse Coding model.
In contrast, the top three spikes in the FFA are not multimodal neurons. Neurons 146, 230, and 275 are the top
three activations in the FFA, none of which activate on the
text modality. It appears that the FFA is doing a nonlinear
variant of principle component analysis where the model is
encoding variance components shared across many image
and text inputs. The FFA is also learning to separate the
joint embedding layer as only 27 of the 512 (5.27%) embedding layer have multimodal representations and respond
to both face and name inputs. This is in stark contrast to our
DSC model where 306 of the 512 (59.76%) neurons activate
for both visual and text input.
The multimodal nature of the DSC neurons has interesting implications. For example, one could simply threshold
the activation of N-326 and get nearly 90% accuracy on the
task of face classification in our test set. The features extracted are also more informative as we discuss in the next
section.
Finally, to visualize the activation of the joint embedding neurons i.e. see what the neuron responds to, we use
a technique called activity triggered averaging. We create a
compilation of the activation of a neuron at the joint layer
multiplied by the input. We continually add to the compilation and visualize the average input after an epoch of
training. The activity triggered averages of vision and text
of selected neurons can be seen in Figure 8.
(a) N-326
(b) N-493
(c) N-121
machine learning tasks. As noted by Glorot et al. [7], evidence suggests that sparse representations are more linearly
separable. We measure the sparsity of our joint representation in the DSC with the sparsity shown in the FFA. Our
model is on average 20.2% sparse at the joint layer compared to 47.8% sparse in the FFA. Tweaking the FFA’s L1
activity regularization to encourage more sparsity resulted
in drastically reduced reconstruction performance. When
visualizing the extracted features from the test input, we are
able to confirm that the DSC model produces more easily
separable features than the FFA, see Figure 9.
(a) DSC t-SNE
(b) FFA t-SNE
Figure 9. t-SNE plot of the joint embedding layer extracted from
our deep sparse coding network (DSC) vs a standard feed-forward
autoencoder (FFA). Our DSC is perfectly separable whereas the
FFA has intermingling of classes. The same parameters (PCA=30,
perplexity=30) were used to generate the plots.
4.5. Generate Missing Modalities
In a multimodal network, a common task is the omission
of one input to a trained network and forcing it to reconstruct that input as well as generate the missing modality.
Sparse coding as a generative model is naturally adept in
“hallucinating” a missing modality and we can force the decoder of the feed-forward model to render some output as
well. Figure 10 shows the removal of the text portion of the
input while maintaining the vision and text output. Since
the joint layer in the FFA is an addition of the two streams,
we inject the zero vector to simulate a missing input. Qualitative results can be seen in Figure 11. Both the DSC and
FFA are able to generate the same number of legible text
outputs with relatively high accuracy, approximately 70%.
Both models fail on the same input images; one example of
a failure can be seen in last column of Figure 11. However,
one notable difference is the in the reconstruction quality of
the visual output. The DSC model reconstructs with greater
detail and resemblance to the input. This a powerful testimony to sparse coding which only trained over the data for
3 epochs versus the 25 epochs for the feed-forward autoencoder.
(d) N-220
Figure 8. Examples of the activity triggered average corresponding
to various neurons in the Deep Sparse Coding model. Neuron 326
is the invariant Halle Berry neuron, which activates strongly to
both the face of Halle Berry and the text, “Halle Berry”. Other
invariant neurons emerged from our dataset including the George
Bush neuron (N-493), Colin Powell neuron (N-121), and Gerhard
Schroeder neuron (N-220).
4.4. Feature Extraction
Given that our joint embedding produces multimodal
neurons, we wanted to explore the uses of the embedding as
a feature representation of the data. The features extracted
can be used in classification, clustering, and various other
4.6. Limitations
Our method is a novel approach to incorporating biologically inspired concepts into a deep neural network and has
7
sparse coding model was shown to be more easily separable and robust for classification tasks. The neurons in our
model also had the property of being invariant to modality, with neurons showing activations for both modalities,
whereas, the standard feed-forward model simply segregates the modality streams.
Finally, for completeness, Quiroga et al. [28] noted that
the Halle Berry neuron, “...was also activated by several pictures of Halle Berry dressed as Catwoman, her character in
a recent film, but not by other images of Catwoman that
were not her”. As shown in Figure 12, our model similarly
distinguishes between catwomen.
Figure 10. Neural network used to generate the text missing
modality. For our DSC model, we can simply remove the input
chain, whereas in the FFA model, we fill in the missing input with
the zero vector.
Figure 12. Activation on neuron 326 with various pictures of catwoman. N-326 activates more when tested with Halle Berry as
Catwoman versus Anne Hathaway and Michele Pfeiffer.
References
Figure 11. Output from the generating network. (a) The original
input image, (b) the face reconstruction and text hallucination from
the FFA, (c) the face reconstruction and text hallucination from our
DSC.
[1] A. Alink, C. M. Schwiedrzik, A. Kohler, W. Singer, and
L. Muckli. Stimulus predictability reduces responses in primary visual cortex. Journal of Neuroscience, 30(8):2960–
2966, 2010.
[2] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra,
C. Lawrence Zitnick, and D. Parikh. Vqa: Visual question
answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433, 2015.
[3] C. Blakemore, R. H. Carpenter, and M. A. Georgeson. Lateral inhibition between orientation detectors in the human
visual system. Nature, 228(5266):37–39, 1970.
[4] C. Blakemore and E. A. Tobin. Lateral inhibition between
orientation detectors in the cat’s visual cortex. Experimental
Brain Research, 15(4):439–440, 1972.
[5] M. Cha, Y. Gwon, and H. Kung. Multimodal sparse
representation learning and applications. arXiv preprint
arXiv:1511.06238, 2015.
[6] M. D. Giudice, V. Manera, and C. Keysers. Programmed
to learn? the ontogeny of mirror neurons. Developmental
science, 12(2):350–363, 2009.
[7] X. Glorot, A. Bordes, and Y. Bengio. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics,
pages 315–323, 2011.
not been tested on many datasets; however, we believe these
concepts make important contributions to the field. Time
and computation in the sparse coding model is most obvious limitation. The FFA model is fast and highly optimized,
whereas, sparse coding has an expensive inference step. Although this is a current limitation, specialized hardware is
being developed that can do sparse coding with memristor
networks suitable for neuromorphic algorithms [33].
5. Conclusion
In conclusion, we augmented a standard feed-forward
autoencoder with biologically inspired concepts of sparsity, lateral inhibition, and top-down feedback. Inference
is achieved by optimization via the LCA sparse solver,
rather than an autoencoder model. Our deep sparse coding
model was trained on multimodal data and the joint representation was extracted for comparison against the standard feed-forward encoding. The joint embedding using our
8
[25] L. Muckli and L. S. Petro. Network interactions: Nongeniculate input to v1. Current Opinion in Neurobiology,
23(2):195–201, 2013.
[26] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng.
Multimodal deep learning. In Proceedings of the 28th international conference on machine learning (ICML-11), pages
689–696, 2011.
[27] L. S. Petro, L. Vizioli, and L. Muckli. Contributions of cortical feedback to sensory processing in primary visual cortex.
Frontiers in psychology, 5, 2014.
[28] R. Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, and I. Fried.
Invariant visual representation by single neurons in the human brain. Nature, 435(7045):1102–1107, 2005.
[29] H. Roebuck, P. Bourke, and K. Guo. Role of lateral and
feedback connections in primary visual cortex in the processing of spatiotemporal regularity- a tms study. Neuroscience,
263:231–239, 2014.
[30] C. Rozell, D. Johnson, R. Baraniuk, and B. Olshausen. Locally competitive algorithms for sparse approximation. In
Image Processing, 2007. ICIP 2007. IEEE International
Conference on, volume 4, pages IV–169. IEEE, 2007.
[31] 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.
International Journal of Computer Vision, 115(3):211–252,
2015.
[32] S. Sabour, N. Frosst, and G. E. Hinton. Dynamic routing
between capsules. arXiv preprint arXiv:1710.09829, 2017.
[33] P. M. Sheridan, F. Cai, C. Du, W. Ma, Z. Zhang, and W. D.
Lu. Sparse coding with memristor networks. Nature Nanotechnology, 2017.
[34] N. Srivastava and R. R. Salakhutdinov. Multimodal learning with deep boltzmann machines. In Advances in neural
information processing systems, pages 2222–2230, 2012.
[35] J. Yang, J. Wright, T. S. Huang, and Y. Ma. Image superresolution via sparse representation. IEEE transactions on
image processing, 19(11):2861–2873, 2010.
[8] I. Goodfellow, H. Lee, Q. V. Le, A. Saxe, and A. Y. Ng. Measuring invariances in deep networks. In Advances in neural
information processing systems, pages 646–654, 2009.
[9] C. G. Gross. Genealogy of the ?grandmother cell? The
Neuroscientist, 8(5):512–518, 2002.
[10] Y. Gwon, M. Cha, and H. Kung. Deep sparse-coded network
(dsn). In Pattern Recognition (ICPR), 2016 23rd International Conference on, pages 2610–2615. IEEE, 2016.
[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
770–778, 2016.
[12] Y. He, K. Kavukcuoglu, Y. Wang, A. Szlam, and Y. Qi.
Unsupervised feature learning by deep sparse coding. In
Proceedings of the 2014 SIAM International Conference on
Data Mining, pages 902–910. SIAM, 2014.
[13] J. J. Hopfield. Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of the national academy of sciences,
81(10):3088–3092, 1984.
[14] G. B. Huang, M. Mattar, H. Lee, and E. Learned-Miller.
Learning to align from scratch. In NIPS, 2012.
[15] D. H. Hubel and T. N. Wiesel. Receptive fields, binocular
interaction and functional architecture in the cat’s visual cortex. The Journal of physiology, 160(1):106–154, 1962.
[16] H. Kafaligonul, B. G. Breitmeyer, and H. Öğmen. Feedforward and feedback processes in vision. Frontiers in psychology, 6, 2015.
[17] R. Kiros, R. Salakhutdinov, and R. Zemel. Multimodal neural language models. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pages
595–603, 2014.
[18] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in neural information processing systems, pages
1097–1105, 2012.
[19] Q. V. Le, R. Monga, M. Devin, K. Chen, G. S. Corrado,
J. Dean, and A. Y. Ng. Building high-level features using
large scale unsupervised learning. In In ICML. Citeseer,
2012.
[20] Y. LeCun, Y. Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory
and neural networks, 3361(10):1995, 1995.
[21] S. Y. Lundquist, M. Mitchell, and G. T. Kenyon. Sparse
coding on stereo video for object detection. arXiv preprint
arXiv:1705.07144, 2017.
[22] L. v. d. Maaten and G. Hinton. Visualizing data using t-sne.
Journal of Machine Learning Research, 9(Nov):2579–2605,
2008.
[23] D. Marr and T. Poggio. A computational theory of human
stereo vision. Proceedings of the Royal Society of London B:
Biological Sciences, 204(1156):301–328, 1979.
[24] J. Masci, U. Meier, D. Cireşan, and J. Schmidhuber. Stacked
convolutional auto-encoders for hierarchical feature extraction. Artificial Neural Networks and Machine Learning–
ICANN 2011, pages 52–59, 2011.
9
| 1cs.CV
|
arXiv:1712.03610v2 [math.PR] 27 Dec 2017
LOGARITHMIC DIVERGENCES FROM OPTIMAL TRANSPORT
AND RÉNYI GEOMETRY
TING-KAM LEONARD WONG
Abstract. Divergences, also known as contrast functions, are distance-like
quantities defined on manifolds of non-negative or probability measures. Using a duality motivated by optimal transport, we introduce and study a parameterized family of L(±α) -divergences which includes the Bregman divergence corresponding to the Euclidean quadratic cost, and the L-divergence
introduced by Pal and Wong [27, 28] in connection with portfolio theory and
a logarithmic cost. They admit natural generalizations of exponential family
that are closely related to the α-family and q-exponential family. In particular,
the L(±α) -divergences of the corresponding potential functions are Rényi divergences. Using this unified framework we prove that the induced geometries
are dually projectively flat with constant curvatures, and satisfy a generalized
Pythagorean theorem. Conversely, we show that if a statistical manifold is
dually projectively flat with constant curvature ±α with α > 0, then it is
locally induced by an L(∓α) -divergence. We define in this context a canonical
divergence which extends the one for dually flat manifolds.
1. Introduction
In this paper we study geometric properties of statistical manifolds defined by
solutions to some optimal transport problems. We show that they characterize
dually projectively flat manifolds with constant curvatures and provide natural
geometries to generalizations of exponential family. Before stating our main results
let us describe the background and motivations of our study.
1.1. Background. Information geometry studies manifolds of probability distributions and measures using geometric ideas. For modern introductions to its theory
and applications we refer the reader to the excellent recent monographs [3] and [9].
Typically the geometry is induced by a divergence which is a distance-like quantity
D [q : p] ≥ 0 defined on the underlying manifold M . Of particular importance is
the Bregman divergence. Given a differentiable concave function ϕ(ξ) defined on a
convex domain Ω in Rd , the Bregman divergence of ϕ is defined for ξ, ξ ′ ∈ Ω by
(1.1)
D(0+) [ξ : ξ ′ ] = Dϕ(ξ ′ ) · (ξ − ξ ′ ) − (ϕ(ξ) − ϕ(ξ ′ )) ,
where D is the Euclidean gradient and a · b is the dot product. (The meaning
of the superscript 0+ will become clear in Section 1.2.) For example, if M is an
exponential family of probability densities where ξ is the natural parameter, the
relative entropy, also known as the Kullback-Leibler divergence, can be expressed
as the Bregman divergence of the cumulant generating function ϕ(ξ) (which is
Date: December 29, 2017.
Key words and phrases. Information geometry, statistical manifold, optimal transport, space of
constant curvature, projective flatness, Rényi divergence, α-divergence.
1
2
TING-KAM LEONARD WONG
convex, so in (1.1) we may consider −ϕ or use D(0−) to be defined in Section 2).
The differential geometry of Bregman divergence was first studied by Nagaoka and
Amari in [23].
Following the general framework established by Eguchi [15, 16], the geometry
induced on M by a given divergence D [· : ·] consists of a Riemannian metric g
and a dual pair (∇, ∇∗ ) of torsion-free affine connections (see Definition 4.1). We
call the quadruplet (M, g, ∇, ∇∗ ) a dualistic structure or a statistical manifold.
The Riemannian metric captures the quadratic approximation of D [q : p] when
q ≈ p, whereas the two affine connections capture higher order local properties
and allow us to define primal and dual geodesics that are in some sense compatible
with the divergence. In contrast with usual Riemannian geometry where the LeviCivita connection is the canonical connection defined by the metric, in information
geometry we usually use a pair of connections (∇, ∇∗ ) to quantify the asymmetry
of the divergence: generally we have D [p : q] 6= D [q : p]. When the divergence is
symmetric (for example when D [q : p] = 12 d(p, q)2 where d is a metric) both ∇ and
∇∗ coincide with the Levi-Civita connection. From this geometric perspective, the
Bregman divergence is fundamental in the sense that it is the canonical divergence
which generates a dually flat geometry, i.e., both the primal and dual connections
∇ and ∇∗ have zero curvature (see [3, Section 6.6] and [9, Section 4.2]; this is
also a limiting case of Theorem 1.3). Moreover, the Bregman divergence satisfies a
generalized Pythagorean theorem which enables explicit computation and geometric
interpretation of projections based on the divergence. By now these results are
classical and numerous applications can be found in the two monographs cited
above.
Motivated by portfolio theory, in a series of papers [26, 27, 35, 28, 36] we introduced and studied the L-divergence (L stands for logarithmic) defined by
(1.2)
D(1) [ξ : ξ ′ ] = log(1 + ∇ϕ(ξ ′ ) · (ξ − ξ ′ )) − (ϕ(ξ) − ϕ(ξ ′ )) ,
where now the potential function ϕ is exponentially concave, i.e, Φ = eϕ is concave.
Being the logarithm of the concave function Φ, ϕ is itself concave and the Ldivergence has a logarithmic correction term that captures the extra concavity.
An important example of L-divergence is the free energy defined for probability
measures ν, ν ′ ≪ µ by
D [ν : ν ′ ] = log
Z
Z
dν ′
dν ′
dµ − log
dµ,
dν
dν
R
dν
where ϕ(ν) = log dµ
dµ. Exponentially concave functions appear naturally in
many recent applications in analysis, probability and statistics. In particular, let
us mention the paper [17] on Bochner’s inequality on metric measure spaces. Other
examples can be found in the references of [28].
In [28] we established the following results where the underlying manifold is the
open unit simplex:
(i) The dualistic structure induced by the L-divergence (1.2) is dually projectively flat with constant curvature −1, and the generalized Pythagorean
theorem (which has a financial interpretation in terms of optimal trading
frequency) holds true.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
3
(ii) The L-divergence can be expressed in terms of the optimal transport map
with respect to a logarithmic cost function, whereas the Bregman divergence corresponds to the classical quadratic cost. Furthermore, displacement interpolations for these transport problems are related to the dual
geodesics of the induced dualistic structure.
It is natural to ask if there is a unified framework that covers both Bregman and Ldivergences. Moreover, can we characterize statistical manifolds that satisfy these
(very strong) geometric properties? On the other hand, Amari and Nagaoka [2] considered α-families of probability distributions and the corresponding α-divergences
that generalize the exponential family and the Kullback-Leibler divergence. We
will see that these questions are intimately related.
The connection between optimal transport and information geometry is a recent
topic that has started to receive serious attention. For example, the recent paper
[4] studies the manifold of entropy-relaxed optimal transport plans and defines
a divergence which interpolates between the Kullback-Leibler divergence and the
Wasserstein metric (optimal transport cost). A dynamic approach is given in [13].
Also see [25] which builds on the ideas in [28] and embeds a large family of optimal
transport problems in a generalized simplex.
1.2. Summary of main results. The key objects of study in this paper are the
L(±α) -divergences where α > 0. The L(α) -divergence is defined by
1
log(1 + α∇ϕ(ξ ′ ) · (ξ − ξ ′ )) − (ϕ(ξ) − ϕ(ξ ′ )) ,
α
where the function ϕ is α-exponentially concave, i.e., eαϕ is concave. To cover
the case of positive curvature, we will also consider a local L(−α) -divergence where
ϕ is α-exponentially convex, i.e., eαϕ is convex. When α = 1 we have the L(1) divergence (1.2), and when α → 0+ it reduces to the Bregman divergence (1.1) of
a concave function (D(0−) is the Bregman divergence of a convex function). Based
on financial arguments, the L(α) -divergence was introduced recently in [36] as the
canonical interpolation between the Bregman and L-divergences. In this paper we
consider also the L(−α) -divergence show that they are of fundamental importance.
While we worked on the open unit simplex in our previous paper [28], here we
consider a general open convex domain Ω in Rd . This answers a question asked in
[28, Section 1] about extending the results to general domains. By a translation
if necessary, we may assume that 0 ∈ Ω (this is used for the normalization of the
function c(α) in (1.4)). Nevertheless, we note that exponential concavity/convexity
impose certain restrictions on the domains of ϕ and the L(−α) -divergence.
In Section 2 we prove the following result which connects the L(±α) -divergence
with optimal transport and generalizes the classical Legendre duality for Bregman
divergence.
(1.3)
D(α) [ξ : ξ ′ ] =
Theorem 1.1. For α > 0, consider the function
(1.4)
c(α) (x, y) =
1
log (1 + αx · y) ,
α
where x · y is the Euclidean dot product. Then the L(α) -divergence (2.5) admits the
self-dual representation
(1.5)
D(α) [ξ : ξ ′ ] = c(α) (ξ, η ′ ) − ϕ(ξ) − ψ(η ′ ),
4
TING-KAM LEONARD WONG
where ψ is the α-conjugate of ϕ and is (locally) α-exponentially concave, and η =
D(α) ϕ(ξ) is the α-gradient which generalizes the Legendre transformation.
Similarly, the L(−α) -divergence has the self-dual representation
(1.6)
D(−α) [ξ : ξ ′ ] = ϕ(ξ) + ψ(η ′ ) − c(α) (ξ, η ′ ),
where now ϕ and ψ are α-exponentially convex.
Precise statements corresponding to Theorem 1.1 are given in Theorems 2.12,
2.13, 2.15 and Corollary 2.23. The self-dual expressions (1.5) and (1.6) are motivated by optimal transport which identifies the dual coordinate system η as the
image of ξ under the optimal transport map. The connection with optimal transport is described in detail in Section 2.4.
Motivated by the close relationships between Bregman divergence and exponential family (see [10]), in Section 3 we consider generalizations of exponential family
that are closely related to the α-family and q-exponential family [2, 24, 5]. We
show that the analogue of the cumulant generating function is α-exponentially concave/convex (Propositions 3.3 and 3.4), and the corresponding L(±α) -divergence
is the Rényi divergence (Theorem 3.6). Moreover, the dual function is the Rényi
entropy. Our results thus provide a new approach to the geometry of Rényi and
α-divergences. When α → 0 we recover the dually flat geometry of exponential
family. The case of finite simplex is discussed in Section 5.4.
In Sections 4 and 5 we study the statistical manifold (M = Ω, g, ∇, ∇∗ ) induced
by the L(±α) -diverence. Our approach, based on the duality in Theorem 1.1, gives
a unified treatment covering both the Bregman and L-divergences and simplifies
the proofs in [28]. The following result summarizes Theorems 5.2, 5.9 and 5.10.
Theorem 1.2. For α > 0, the dualistic structure induced by the L(±α) -divergence
defined on an open convex set Ω ⊂ Rd with d ≥ 2 is dually projectively flat, has
constant curvature ∓α, and the generalized Pythagorean theorem holds. (When
α → 0+ we reduce to the dually flat case.)
By dual projective flatness, we mean that there exist ‘affine’ coordinate systems
(ξ and η) under which the primal and dual geodesics are straight lines up to time
reparameterizations (for the precise technical statement see Definition 5.1). Since
the properties in Theorem 1.2 are very strong and spaces of constant curvatures
play fundamental roles in differential geometry, it is natural to ask if Theorem 1.2
has a converse. Indeed there is one and it will be proved in Section 6 (see Theorems
6.1 and 6.2).
Theorem 1.3. Consider a dualistic structure (M, g, ∇, ∇∗ ) which is dually projectively flat and has constant curvature −α < 0. Then locally there exist affine
coordinate systems ξ and η for ∇ and ∇∗ respectively, and α-exponentially concave
functions ϕ(ξ) and ψ(η) which satisfy the generalized Fenchel identity
ϕ(ξ) + ψ(η) ≡ c(α) (ξ, η).
Moreover, the self-dual representation (1.5) defines locally a canonical divergence
which induces the given dualistic structure. Analogous statements hold when the
curvature is α > 0.
Our results establish the L(±α) -divergence as the canonical divergence for a dually projectively flat statistical manifold with constant curvature, thus generalizing
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
5
the Bregman divergence for dually flat manifolds (see [19] for another characterization based on affine differential geometry). Recently, Ay and Amari [8] defined
a canonical divergence for an arbitrary statistical manifold and showed that on a
dually flat manifold (i.e., α = 0) it reduces to the Bregman divergence. It is interesting to know if their canonical divergence is consistent with our L(±α) -divergences.
Since the L(±α) -divergences have properties analogous to those of the Bregman
divergence, explicit computations are tractable and a natural question is to study
how they perform in applied problems such as clustering [10] and statistical estimation. Projectively flat connections, statistical manifolds with constant curvatures
and relations with affine differential geometry have been studied in the literature;
see in particular [19, 14, 20, 22, 12]. Using ideas of optimal transport and the
L(±α) -divergence, we consider these properties as a whole and are able to give a
new and elegant characterization of the geometry.
In optimal transport, it is well known that the quadratic cost, which corresponds
in our setting to the Bregman divergence, is intimately related to Brownian motion
and the heat equation (via the Wasserstein gradient flow of entropy, see [18, 7]) as
well as large deviations [1] and the Schrödinger problem [21]. A program suggested
in [28] is to study the analogues for our logarithmic cost functions. We plan to
address some of these questions in future research. Another intriguing problem is
to connect other information-geometric structures with optimal transport.
2. Logarithmic divergences and optimal transport
We begin by defining exponential concavity and convexity as well as the associated L(±α) -divergences D(±α) . They admit generalized self-dual representations
that are motivated by the theory of optimal transport.
2.1. Exponential concavity and convexity.
Definition 2.1. Let Ω ⊂ Rd be an open convex set, ϕ : Ω → R, and fix α > 0.
Write Φ = eαϕ .
(i) We say that ϕ is α-exponentially concave if Φ is concave on Ω. For α = 0+
(that is, the limit as α → 0+ ), we say that ϕ is 0+ -exponentially concave
if ϕ is concave.
(ii) We say that ϕ is α-exponentially convex if Φ is convex on Ω. For α = 0+ ,
we say that ϕ is 0+ -exponentially convex if ϕ is convex.
Assuming ϕ is twice continuously differentiable on Ω and denoting respectively
by Dϕ and D2 ϕ the Euclidean gradient and Hessian, we have
(2.1)
D2 Φ = D2 eαϕ = αeαϕ D2 ϕ + α(Dϕ)(Dϕ)⊤ ,
where (Dϕ)⊤ is the transpose of Dϕ (regarded as a column vector). From this we
have the following elementary lemma which justifies the limiting cases in Definition
2.1.
Lemma 2.2. Suppose ϕ is twice continuously differentiable on Ω and α > 0.
(i) ϕ is α-exponentially concave if and only if
(2.2)
− D2 ϕ − α(Dϕ)(Dϕ)⊤ ≥ 0.
(ii) ϕ is α-exponentially convex if and only if
(2.3)
D2 ϕ + α(Dϕ)(Dϕ)⊤ ≥ 0.
6
TING-KAM LEONARD WONG
Here the inequalities are in the sense of positive semidefinite matrix.
Note that because of the product term in (2.1), we cannot pass between exponential concavity and convexity simply by considering −ϕ. In particular, from
Lemma 2.2 we see that an α-exponentially concave function is always concave, but
an α-exponentially convex function is not necessarily convex. In Section 4 we will
regard (2.2) and (2.3) as the Riemannian matrices of D(α) and D(−α) respectively.
Remark 2.3. Given an α-exponentially concave function ϕ on Ω, we cannot extend
ϕ to Rd simply by letting ϕ = −∞ on Rd \ Ω. This is because e−∞ = 0 but Φ = eαϕ
is required to be non-negative and concave.
2.2. L(±α) -divergences. By convention we always let α > 0. Suppose ϕ is differentiable and α-exponentially concave. By the concavity of Φ = eαϕ , for any
ξ, ξ ′ ∈ Ω we have
′
(2.4) Φ(ξ ′ ) + DΦ(ξ ′ ) · (ξ − ξ ′ ) ≥ Φ(ξ) ⇒ 1 + αDϕ(ξ ′ ) · (ξ − ξ ′ ) ≥ eα(ϕ(ξ)−ϕ(ξ )) .
This motivates the following definition of the L(α) -divergence. It is different from
the Bregman divergence in that our divergence is given by the ratio rather than
the difference in (2.4).
Definition 2.4 (L(α) -divergence). If ϕ is differentiable and α-exponentially concave, we define the L(α) -divergence of ϕ by
1
(2.5) D(α) [ξ : ξ ′ ] = log (1 + αDϕ(ξ ′ ) · (ξ − ξ ′ )) − (ϕ(ξ) − ϕ(ξ ′ )) , ξ, ξ ′ ∈ Ω.
α
For α = 0+ , we define the L(0+) -divergence by the Bregman divergence (1.1).
From (2.4) we see that D(α) [ξ : ξ ′ ] ≥ 0, and if Φ is strictly concave, then
D [ξ : ξ ′ ] = 0 only if ξ = ξ ′ .
For an α-exponentially convex function, the analog of (2.4) is
(α)
(2.6)
′
1 + αDϕ(ξ ′ ) · (ξ − ξ ′ ) ≤ eα(ϕ(ξ)−ϕ(ξ )) .
Unfortunately, now the left hand side of (2.6) may become negative and so its logarithm may not exist. Nevertheless, we can define the L(−α) -divergence locally when
ξ and ξ ′ are close. This is sufficient for defining the dualistic structure (g, ∇, ∇∗ ).
Definition 2.5 (L(−α) -divergence). If ϕ is differentiable and α-exponentially convex, we define the L(−α) -divergence of ϕ locally by
1
(2.7)
D(−α) [ξ : ξ ′ ] = (ϕ(ξ) − ϕ(ξ ′ )) − log (1 + αDϕ(ξ ′ ) · (ξ − ξ ′ )) ,
α
for all pairs (ξ, ξ ′ ) in Ω such that the logarithm in (2.7) exists. For α = 0+ ,
ϕ is convex and we define (globally) the L(0−) -divergence as the usual Bregman
divergence
(2.8)
D(0−) [ξ : ξ ′ ] = (ϕ(ξ) − ϕ(ξ ′ )) − Dϕ(ξ ′ ) · (ξ − ξ ′ ).
2.3. Duality of L(±α) divergence. In this subsection we develop a duality theory
for α-exponentially concave/convex functions. Some of these ideas are implicit in
[28]. In particular, we generalize the Legendre transform and conjugate which will
be used to formulate the self-dual representations of the L(±α) -divergences. In
Section 2.4 we consider an abstract version of this construction terms of of optimal
transport.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
7
To ensure that the potential function ϕ is well-behaved and the differential geometric objects associated with the L(±α) -divergences are well-defined, we will impose throughout this paper some regularity conditions.
2.3.1. Duality for α-exponentially concave functions. We first consider the case
where ϕ is α-exponentially concave. By convention we always assume α > 0.
Condition 2.6 (conditions for L(α) -divergence). We assume ϕ is smooth on Ω and
the matrix −D2 ϕ − α(Dϕ)(Dϕ)⊤ is strictly positive definite on Ω. Moreover, by
considering a translation of the variable ξ if necessary, we assume without loss of
generalizty that for all ξ ∈ Ω we have
(2.9)
1 − αDϕ(ξ) · ξ = 1 − α
d
X
ξℓ
ℓ=1
Note that by strict concavity of Φ = e
αφ
∂ϕ
(ξ) > 0.
∂ξ ℓ
and (2.4), (2.9) holds whenever 0 ∈ Ω.
Observe that there does not exist any positive and strictly concave function Φ =
eαϕ on the line (infinite in both directions). Thus, in order that Condition 2.6 holds,
the domain Ω must not contain any line (rays are fine). This is different from the
classical Bregman case where the concave function may be globally defined on Rd .
Examples of feasible domains in our setting include the open simplex (considered
in [28]) and the positive quadrant Rd++ = (0, ∞)d .
Condition (2.9) will be used in the following construction. For each ξ ′ ∈ Ω (which
is open), the supporting tangent hyperplane of Φ = eαϕ at ξ ′ is given by
′
ξ ∈ Ω 7→ Φ(ξ ′ ) + DΦ(ξ ′ ) · (ξ − ξ ′ ) = eαϕ(ξ ) (1 − αDϕ(ξ ′ ) · ξ ′ + αDϕ(ξ ′ ) · ξ) .
Thus (2.9) is equivalent to positivity of the constant coefficient. By concavity of Φ,
for ξ ∈ Ω we have
(2.10)
′
eαϕ(ξ) = min
eαϕ(ξ ) ((1 − αDϕ(ξ ′ ) · ξ ′ ) + αDϕ(ξ ′ ) · ξ) .
′
ξ ∈Ω
Note that the minimum is attained uniquely by ξ ′ = ξ. This and the logarithm
is the basis of Theorem 2.12 which is our duality result. To prepare for it let us
introduce some notations.
Definition 2.7 (α-gradient). We define the α-gradient D(α) ϕ(ξ) at ξ ∈ Ω by
(2.11)
D(α) ϕ(ξ) =
1
Dϕ(ξ),
1 − αDϕ(ξ) · ξ
where Dϕ = D(0) ϕ is the Euclidean gradient.
Definition 2.8 (dual coordinate). We define the dual coordinate η by
(2.12)
η = D(α) ϕ(ξ).
We let Ω′ = (D(α) ϕ)(Ω) be the range of η.
This terminology is justified by the following result.
Proposition 2.9. The α-gradient D(α) ϕ is a diffeomorphism from Ω onto Ω′ which
is an open set in Rd .
8
TING-KAM LEONARD WONG
Proof. It is clear from (2.11) that the map D(α) ϕ is smooth.
To show that D(α) ϕ is injective, suppose towards a contradiction that D(α) ϕ(ξ) =
(α)
D ϕ(ξ ′ ) for distinct ξ, ξ ′ ∈ Ω. By definition, we have
Dϕ(ξ ′ )
Dϕ(ξ)
.
=
1 − αDϕ(ξ) · ξ
1 − αDϕ(ξ ′ ) · ξ ′
It follows that for any ξ ′′ ∈ Ω, we have
1+
αDϕ(ξ) · ξ ′′
αDϕ(ξ ′ ) · ξ ′′
.
= 1+
1 − αDϕ(ξ) · ξ
1 − αDϕ(ξ ′ ) · ξ ′
Rearranging, we have the identity
1 + αDϕ(ξ ′ ) · (ξ ′′ − ξ ′ )
1 + αDϕ(ξ) · (ξ ′′ − ξ)
.
=
1 − αDϕ(ξ) · ξ
1 − αDϕ(ξ ′ ) · ξ ′
Letting ξ ′′ be ξ and ξ ′ respectively, we see after some computation that
(1 + αDϕ(ξ) · (ξ ′ − ξ)) · (1 + αDϕ(ξ ′ ) · (ξ − ξ ′ )) = 1.
Observe that this is equivalent to
D(α) [ξ ′ : ξ] + D(α) [ξ : ξ ′ ] = 0,
which is a contradiction since D(α) [ξ ′ : ξ] , D(α) [ξ : ξ ′ ] > 0 by Condition 2.6. This
proves that the α-gradient is injective.
Finally, from Condition 2.6 again we have −D2 ϕ − α(Dϕ)(Dϕ)⊤ > 0. This can
be used to show that the Jacobian ∂η
∂ξ is invertible. By the inverse function theorem,
(α)
′
D ϕ : Ω → Ω is a diffeomorphism and Ω′ is open. (Essentially the same result is
proved in [28, Theorem 3.2] so we do not provide the details. Also see (4.6) below
which gives an explicit expression of the Jacobian.)
Remark 2.10. Note that the range Ω′ of the α-gradient may not be convex even
when α = 0+ . An explicit example is the concave function ϕ(ξ) = log(1 − kξk2 )
defined on say the open square Ω = (−0.5, 0.5)2. To ensure convexity of Ω′ so that
we have a complete correspondence, we may consider analogues of convex functions
of Legendre type (see [30, p.258]). Nevertheless, as we have shown the dual variable
η is always a global coordinate system of the manifold M = Ω.
Definition 2.11 (α-conjugate). Consider the function c(α) defined for x, y ∈ Rd
with 1 + αx · y > 0 by
1
(2.13)
c(α) (x, y) = log(1 + αx · y),
α
We define the α-conjugate ψ of ϕ on Ω′ = D(α) ϕ(Ω) by
−1
(2.14)
ψ(η) = c(α) (ξ, η) − ϕ(ξ), ξ = D(α) ϕ
(η).
By (2.19) below 1 + αξ · η > 0, hence c(α) (ξ, η) and ψ(η) are well-defined.
Theorem 2.12 (α-duality). For ξ ∈ Ω, η ∈ Ω′ we have
(α)
′
′
c
(ξ,
η
)
−
ψ(η
)
,
(2.15)
ϕ(ξ) = min
′
′
η ∈Ω
(2.16)
(α) ′
′
c
(ξ
,
η)
−
ϕ(ξ
)
.
ψ(η) = min
′
ξ ∈Ω
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
9
Furthermore, ψ is locally α-exponentially concave on Ω′ in the sense that −D2 ψ(η)−
α(Dψ(η))(Dψ(η))⊤ > 0 for η ∈ Ω′ , and we have 1 − αDψ(η) · η > 0 on Ω′ so that
1
Dψ(η) is well-defined. We have
the α-gradient D(α) ψ(η) = 1−αDψ(η)·η
(2.17)
−1
= D(α) ψ.
D(α) ϕ
Proof. Taking logarithm in (2.10) and rearranging, we have
(2.18)
1
ϕ(ξ) = min
ϕ(ξ ′ ) + log (1 − αDϕ(ξ ′ ) · ξ ′ + αDϕ(ξ ′ ) · ξ)
ξ ′ ∈Ω
α
1
Dϕ(ξ ′ )
1
·
ξ
+ ϕ(ξ ′ ) + log (1 − αDϕ(ξ ′ ) · ξ ′ )
log
1
+
α
= min
′
′
′
ξ ∈Ω α
1 − αDϕ(ξ ) · ξ
α
1
= min
c(α) (ξ, η ′ ) + ϕ(ξ ′ ) + log (1 − αDϕ(ξ ′ ) · ξ ′ ) .
ξ ′ ∈Ω
α
Note that since η ′ = D(α) ϕ(ξ ′ ), from (2.11) we have the useful identity
1 + αξ ′ · η ′ =
(2.19)
1
.
1 − αDϕ(ξ ′ ) · ξ ′
With this observation, we see that
ϕ(ξ ′ ) +
1
log (1 − αϕ(ξ ′ ) · ξ ′ ) = − c(α) (ξ ′ , η ′ ) − ϕ(ξ ′ ) = −ψ(η ′ ).
α
Putting this into (2.18), we obtain (2.15). Note that this derivation explains our
definitions of c(α) and the α-gradient.
From (2.15), for any η, ξ ′ we have
ϕ(ξ ′ ) ≤ c(α) (ξ ′ , η) − ψ(η) ⇒ ψ(η) ≤ c(α) (ξ ′ , η) − ϕ(ξ ′ ).
By definition we have ψ(η) = c(α) (ξ, η)− ϕ(ξ), so (2.16) holds as well. In particular,
for any ξ, ξ ′ ∈ Ω we have
ψ(η) = c(α) (ξ, η) − ϕ(ξ) ≤ c(α) (ξ ′ , η) − ϕ(ξ ′ ).
(2.20)
Using (2.19), we compute
c
(α)
′
(ξ , η) − c
(α)
1
1 + αξ ′ · η
(ξ, η) = log
α
1 + αξ · η
η
1
1
′
+ αξ ·
= log
α
1 + αξ · η
1 + αξ · η
1
′
= log (1 + αDϕ(ξ) · (ξ − ξ)) .
α
It follows that
(2.21)
c(α) (ξ ′ , η) − c(α) (ξ, η) − (ϕ(ξ ′ ) − ϕ(ξ)) = D(α) [ξ ′ : ξ]
is the L(α) -divergence of ϕ.
By (2.15), for any ξ ∈ Ω, η is the unique minimizer (over Ω′ ) of
η ′ 7→
1
log (1 + αξ · η ′ ) − ψ(η ′ ).
α
10
TING-KAM LEONARD WONG
ξ
= Dψ(η). Rearranging,
Since Ω′ is open, we have the first order condition 1+αξ·η
we have
1
1 − αDψ(η) · η =
.
1 + αξ · η
Thus D(α) ψ(η) is well-defined, D(α) ψ(η) = ξ, and we have (2.17).
It remains to show that
− D2 ψ(η) − α(Dψ(η))(Dψ(η))⊤
(2.22)
is strictly positive definite on Ω′ . Expressing (2.21) in terms of ψ using the definition
(2.14) of α-conjugate, we can show that
1
D(α) [ξ : ξ ′ ] = log (1 + αDψ(η) · (η ′ − η)) − (ψ(η ′ ) − ψ(η)) ,
α
so the L(α) divergence of ϕ is the L(α) -divergence of ψ with the arguments interchanged (see Corollary 2.22 for a more general result). It follows that (2.22) is the
Riemannian matrix of D(α) in the dual coordinates η. Since ϕ satisfies Condition
2.6 by assumption, the matrix (2.22) is strictly positive definite as well.
From (2.14) and (2.21), we immediately obtain
Theorem 2.13 (self-dual representation). The L(α) -divergence of ϕ admits the
self-dual representation
D(α) [ξ : ξ ′ ] = c(α) (ξ, η ′ ) − ϕ(ξ) − ψ(η ′ ),
(2.23)
ξ, ξ ′ ∈ Ω.
The self-dual representation (2.25) generalizes [3, Theorem 1.1] which addresses
the case of Bregman divergence (α = 0+ ). This completes the circle of ideas
and the duality theory for α-exponentially concave functions. A general self-dual
representation for c-divergence is proved in Theorem 2.21.
2.3.2. Duality for α-exponentially convex functions. Now we state without proof
the analogous results for the L(−α) -divergence defined by an α-exponentially convex function. The only difference is that because eαϕ is convex, we have to take
maximums instead of minimums and switch the inequalities accordingly. Also, as
the L(−α) -divergence may not be globally defined, here we only consider the local
geometry.
Condition 2.14 (conditions for L(−α) -divergence). We assume ϕ is smooth and
the matrix D2 ϕ + α(Dϕ)(Dϕ)⊤ is strictly positive definite on Ω. By a translation
if necessary, we assume that 0 ∈ Ω. By restricting the domain if necessary, we
assume that
1 + αDϕ(ξ ′ ) · (ξ − ξ ′ ) > 0
(2.24)
for all ξ, ξ ′ ∈ Ω. This implies that the L(−α) -divergence is defined for all ξ, ξ ′ ∈ Ω.
Theorem 2.15 (duality for L(−α) -divergence). Consider the α-gradient and c(α)
defined as in (2.11) and (2.13). Then the map ξ 7→ η = D(α) ϕ(ξ) is a diffeomorphism, and the (−α)-conjugate ψ defined by (2.14) is locally α-exponentially
convex. The functions ϕ and ψ are related by
(α) ′
′
(α)
′
′
c
(ξ
,
η)
−
ϕ(ξ
)
.
c
(ξ,
η
)
−
ψ(η
)
,
ψ(η)
=
max
ϕ(ξ) = max
′
′
′
η ∈Ω
The L
(2.25)
(−α)
ξ ∈Ω
-divergence (2.7) admits the self-dual representation
D(−α) [ξ : ξ ′ ] = ϕ(ξ) + ψ(η ′ ) − c(α) (ξ, η ′ ),
ξ, ξ ′ ∈ Ω.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
11
2.4. c-divergence induced by an optimal transport map. The ideas in Section 2.3 can be developed in a rather general framework. In this subsection we
review some basic results of optimal transport and introduce a c-divergence defined by an optimal transport map. This c-divergence was first introduced in [28]
and here we give a more systematic exposition. For introductions to the theory of
optimal transport we refer the reader to standard references such as [6, 31, 33, 34].
2.4.1. Preliminaries in optimal transport. Let X and Y be Polish spaces (i.e., complete and separable metric spaces) interpreted respectively as the source and target
spaces of the optimal transport problem. Let c : X × Y → R be a continuous
cost function. Given Borel probability measures µ ∈ P(X ) and ν ∈ P(Y), the
Monge-Kantorovich optimal transport problem is
Z
c(x, y)dγ(x, y),
(2.26)
inf
γ
X ×Y
where the infimum is taken over joint distributions γ ∈ P(X × Y) whose marginals
are µ and ν respectively. We call γ a coupling of the pair (µ, ν) and write γ ∈
Π(µ, ν). If γ attains the infimum in (2.26), we say that it is an optimal coupling.
For certain cost functions and under suitable conditions on µ and ν, the optimal
coupling has the form
γ = (Id × T )# µ
for some measurable map T : X → Y. Here we use F# µ = µ ◦ F −1 to denote the
pushfoward of a measure. In this case the optimal coupling is deterministic, and
we call T an optimal transport map.
Remark 2.16. Note that if we replace the cost function c by
(2.27)
c̃(x, y) = c(x, y) + h(x) + k(y),
where h and k and real-valued functions, then for any coupling γ ∈ Π(µ, ν) we have
Z
Z
Z
Z
c̃dγ = cdγ + hdµ + kdν.
The last two integrals, if they exist, are determined once µ and ν are fixed and
is independent of the coupling γ. This means that we are free to modify the cost
function in the manner of (2.27) without changing the optimal couplings.
Before discussing some general results about the Monge-Kantorovich optimal
transport problem (2.26), it is helpful to review an important and motivating special
case.
Example 2.17 (quadratic cost). Let X = Y = Rd and let
d
(2.28)
c(x, y) =
1X i
1
(x − y i )2
kx − yk2 =
2
2 i=1
be the quadratic cost (here we write x = (x1 , . . . , xd ) ∈ Rd ). Suppose µ, ν ∈ P(Rd )
have finite second moments, i.e.,
Z
Z
kxk2 dµ(x) < ∞,
kyk2 dν(y) < ∞,
and suppose µ is absolutely continuous with respect to the Lebesgue measure on Rd .
By Brenier’s theorem [11], there exists a convex function ϕ : Rd → R ∪ {+∞} such
12
TING-KAM LEONARD WONG
that its gradient ∇ϕ (which is well-defined µ-a.e.) pushforwards µ to ν. Moreover,
the deterministic coupling
γ = (Id × ∇ϕ)# µ
solves the Monge-Kantorovich problem. In fact, it can be shown that the optimal
transport map T (x) = ∇ϕ(x), which is a Legendre transformation, is unique µ-a.e.
Note that since
1
1
1
kx − yk2 = −x · y + kxk2 + kyk2 ,
2
2
2
by Remark 2.16 we may consider instead the cost function c̃(x, y) = −x · y, or even
c̃(x, y) = x · y by a change of variable y 7→ −y. This gives the limiting case for the
parameterized cost function c(α) (x, y) given in (2.13).
In general, solutions to the optimal transport problem (2.26) can be characterized
using generalized concepts of convex analysis. In the following definition we fix a
cost function c : X × Y → R.
Definition 2.18. Let f : X → R ∪ {−∞} and g : Y → R ∪ {−∞}.
(i) The c-transforms of f and g are defined respectively by
(2.29)
f c (y) = inf (c(x, y) − f (x)) ,
y ∈ Y,
g c (x) = inf (c(x, y) − g(x)) ,
x ∈ X.
x∈X
y∈Y
(ii) We say that f (respectively g) is c-concave if f cc = f (respectively g cc = g).
(iii) If f and g are c-concave, their c-superdifferentials are defined by
(2.30)
∂f = {(x, y) ∈ X × Y : f (x) + f c (y) = c(x, y)},
∂g = {(x, y) ∈ X × Y : g c (x) + g(y) = c(x, y)}.
If f is c-concave on X , then
(2.31)
f (x) + f c (y) ≤ c(x, y)
for all (x, y) ∈ X × Y, and equality holds if and only if (x, y) ∈ ∂f . We call (2.31)
the (generalized) Fenchel inequality (or identity when equality holds).
Theorem 2.19. Suppose the cost function c is continuous and bounded below, and
there exist a ∈ L1 (µ) and b ∈ L1 (ν) such that c ≤ a + b. Then, a coupling γ of
the pair (µ, ν) solves the Monge-Kantorovich problem (2.26) if and only if there
exists a c-concave function f on X such that supp(γ) ⊂ ∂ c f . We call the c-concave
function f a Kantorovich potential.
Proof. See [6, Theorem 1.13] where this result is called the Fundamental Theorem
of Optimal Transport.
By Theorem 2.19, a c-concave function f on X (or, rather, its c-superdifferential)
can be regarded as encoding the solution to an optimal transport problem. If the
cost function c is not bounded below but satisfies suitable convexity properties, we
may modify c in the manner of (2.27) such that c(x, y) + h(x) + k(y) is bounded
below and Theorem 2.19 is still applicable.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
13
2.4.2. c-divergence. Let f be a c-concave function on X . Assume that f is cdifferentiable in the following sense: for all x ∈ X there is a unique element y ∈ Y
such that (x, y) ∈ ∂ c f . We call y = Dc f (x) the c-gradient of f . This is a strong
assumption, but is satisfied for the cost function c(α) as seen in the proof of Theorem
2.12. In this case, we may regard Dc f : X → Y as an optimal transport map. For
a detailed discussion of the smoothness properties of optimal transport maps see
[34, Chapter 11].
It is easy to show using the definition of c-concavity that for any x, x′ ∈ X we
have
(2.32)
f (x′ ) + c(x, y ′ ) − c(x′ , y ′ ) ≥ f (x),
y ′ = Dc f (x′ ).
Analogous definitions and results hold for a c-concave function g on Y.
The inequality (2.32) motivates the definition of c-divergence (a more suggestive
name may be transport divergence) which was first introduced in [28]. From (2.21)
we see that the L(±α) -divergences are special cases.
Definition 2.20 (c-divergence). Let f be c-concave and c-differentiable on X . The
c-divergence of f is the functional Df : X × X → [0, ∞) defined by
(2.33) Df [x : x′ ] = c(x, y ′ )−c(x′ , y ′ )−(f (x)−f (x′ )),
x, x′ ∈ X ,
y ′ = Dc f (x′ ).
If g is a c-differentiable c-concave function on Y, we define the dual c-divergence
Dg : Y × Y → [0, ∞) by
(2.34) D∗g [y : y ′ ] = c(x′ , y) − c(x′ , y ′ ) − (g(y) − f (y ′)),
y, y ′ ∈ Y,
x′ = Dc g(y ′ ).
Intuitively, the c-divergence Df [x : x′ ] quantifies the deviation of (x, y ′ ) from
the optimal coupling (x′ , y ′ = Dc f (x′ )) ∈ ∂ c f .
Theorem 2.21 (self-dual representations). Let y ′ = Dc f (x′ ). Then
(2.35)
Df [x : x′ ] = c(x, y ′ ) − f (x) − f c (y ′ ).
Similarly, if x = Dc g(y). Then
(2.36)
D∗g [y : y ′ ] = c(x′ , y) − g(y) − g c (x′ ).
Proof. By the generalized Fenchel identity (2.31), we have
f (x′ ) + f c (y ′ ) = c(x′ , y ′ ) ⇒ f c (y ′ ) = c(x′ , y ′ ) − f (x′ ).
It follows that
Df [x : x′ ] = c(x, y ′ ) − c(x′ , y ′ ) − (f (x) − f (x′ ))
= c(x, y ′ ) − f (x) − f c (y ′ ).
This gives (2.36). The proof of the second identity (2.36) is similar.
Corollary 2.22 (Duality of c-divergence). Suppose f c is c-differentiable on the
range of Dc f . Let y = Dc f (x) and y ′ = Dc f (x′ ). Then
(2.37)
Df [x′ : x] = D∗f c [y : y ′ ] .
Thus, if we identify x with y and x′ with y ′ , then the dual divergence of f c is simply
the divergence of f with the arguments interchanged.
We conclude this section by specializing to the case of L(α) -divergence. Because
of the logarithm, to state the result in the framework of c-divergence we need to
pick a specific domain so that c(α) (x, y) is well-defined. A natural domain for this
purpose is the positive quadrant Rd++ .
14
TING-KAM LEONARD WONG
Corollary 2.23. Consider the space X = Y = Rd++ = (0, ∞)d . (To make it
Pd
Polish, we may use for example the metric d(x, y) = i=1 (log xi − log y i )2 .) On
Rd++ consider the continuous cost function c(α) (x, y) = α1 log(1 + αx · y). If ϕ is
α-exponentially concave on Rd++ satisfying Condition 2.6, then it is c(α) -concave.
The c(α) -gradient of ϕ is the α-gradient, and the c(α) -divergence of ϕ is the L(α) divergence D(α) [· : ·].
Proof. Once the cost function c(α) is identified, the development follows the same
lines as in [28, Section 3] and has been covered using a slightly different language
in Section 2.2.
In the limit α → 0+ , we see that the Bregman divergence D(0+) corresponds to
the c-divergence of c(x, y) = x · y, and D(0−) corresponds to c(x, y) = −x · y or
equivalently the quadratic cost 21 kx − yk2 .
3. Generalizations of exponential family and Rényi divergence
It is well known that Bregman divergences (D(0±) ) arise naturally in applications
involving exponential families. In fact, it was shown by Banerjee et al in [10]
that there is a one-to-one correspondence between what they call regular Bregman
divergences and regular exponential families. Are there parameterized families of
probability distributions that correspond to the L(±α) divergences?
In this section we consider generalizations of exponential family that are closely
related to the α-family [2, Section 2.6] and the q-exponential family [3, Section
4.3]. We show that a suitably defined potential function is α-exponentially concave/convex, and the corresponding L(±α) -divergences give the Rényi divergence.
This gives a natural framework to study the geometry induced by the Rényi divergence.
3.1. Motivations. First let us recall the main idea in [10]. Consider an exponential
family of probability densities of the form
(3.1)
log p(x, ξ) = ξ · h(x) − ϕ(ξ),
where h(x) = (h1 (x), . . . , hd (x)) is a vector of sufficient statistics, and ϕ is the
convex cumulant generating function. Let D = D(0−) be the Bregman divergence
of ϕ. By the self-dual representation (induced by the function c(ξ, h(x)) = −ξ ·h(x);
see Section 2.4), we have
(3.2)
log p(x, ξ) = −D [ξ : ξ ′ ] + ψ(h(x)),
where ψ is the conjugate function and h(x) plays the role of the dual variable η ′ .
Remarkably, the conjugate ψ is the negative Shannon entropy (see [3, Section 2.1]).
Now let us consider an L(α) -divergence D(α) with α > 0 and domain Ω = Rd++ ,
where ϕ is α-exponentially concave. Its self-dual representation (Theorem 2.13) is
D(α) [ξ : ξ ′ ] =
1
log (1 + αξ · η ′ ) − ϕ(ξ) − ψ(η ′ ).
α
Imitating (3.2), let us consider a parameterized density of the form
(3.3)
log p(x, ξ) = −D(α) [ξ : ξ ′ ] − ψ(h(x)) = −
1
log(1 + αξ · h(x)) + ϕ(ξ),
α
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
15
where h(x) = η ′ ∈ Rd++ as well. Exponentiating both sides, we have
(3.4)
p(x, ξ) = (1 + αξ · h(x))
1
−α
eϕ(ξ) .
If µ is the dominating measure of the family, the normalizing function ϕ(ξ) in (3.4)
is given by
Z
1
(3.5)
ϕ(ξ) = − log (1 + αξ · h(x))− α dµ(x).
It is the analogue of the cumulant generating function. For lack of a better name,
let us call (3.4) an F (α) -family.
Similarly, the L(−α) -divergence motivates the F (−α) -family
(3.6)
1
p(x, ξ) = (1 + αξ · h(x)) α e−ϕ(ξ) ,
where
(3.7)
ϕ(ξ) = log
Z
1
(1 + αξ · h(x)) α dµ(x).
Remark 3.1. For α′ 6= 1, Amari and Nagaoka [2, Section 2.6] defined an α′ -family
of the form
X
1−α′
2
ξ i F i (x), L(α) (u) =
L(α) (p(x, ξ)) = C(x) +
u 2 ,
′
1−α
i
!# 2 ′
"
(3.8)
1−α
X
1 − α′
ξ i F i (x)
C(x) +
or p(x, ξ) =
,
2
i
where C and F i are functions on X . Comparing this with (3.5) and (3.7), we see
that our F (±α) -families are essentially the their α′ -family with a different parameterization and a difference choice of the potential function. We decide to keep our
formulation so as to be consistent with (3.3) and the framework of the paper.
3.2. Exponential concavity/convexity of ϕ. In order that the densities (3.4)
and (3.6) are well-defined, we need to check that the potential functions (3.5) and
(3.7) define α-exponentially concave/convex functions.
We impose the following conditions which formalize the above constructions.
Condition 3.2. Let µ be a probability measure on a measurable space X . Let
h(x) = (h1 (x), . . . , hd (x)) be a vector function defined on X and with values in
Rd++ , such that if ξ · h(x) = ξ ′ · h(x) µ-almost everywhere on X then ξ = ξ ′ .
For the F (α) -case, we assume that the integral in (3.5) exists for all ξ in an open
convex set Ω ⊂ Rd++ and we can differentiate with respect to ξ under the integral
sign. Analogous conditions are imposed for the F (−α) -case.
3.2.1. F (α) -family.
Proposition 3.3. For α > 0 fixed, the potential function ϕ(ξ) of the F (α) -family
defined by (3.5) is α-exponentially concave on Ω.
Proof. We need to check that Φ = eαϕ is concave on Ω. From (3.5), we have
−α
Z
1
,
(3.9)
Φ(ξ) =
(1 + αξ · h(x))− α dµ(x)
which can be regarded as a generalized harmonic average.
16
TING-KAM LEONARD WONG
Let X be the random element X(x) = x, where x ∈ X and X is the sample
space. Given ξ ∈ Ω, let Eξ denote the expectation with respect to p(x, ξ), and let
Zξ be the random vector defined by
(3.10)
Zξ =
h(X)
.
1 + αξ · h(X)
Differentiating (3.9) and simplifying, we have
Z
−α−1 Z
1
1
∂Φ
−α
−1 i
−α
(ξ) = α
(1 + αξ · h(x)) dµ
h (x)dµ
(1 + αξ · h(x))
∂ξ i
−α Z
Z
1
p(x, ξ)zξi dµ(x)
=α
(1 + αξ · h(x))− α dµ(x)
= αΦ(ξ)Eξ Zξi .
In particular, we have
(3.11)
Dϕ(ξ) = Eξ [Zξ ]
which can be interpreted as an expectation parameter. Differentiating one more
time and calculating carefully, we get
∂2Φ
(ξ) = α(1 + α) Eξ [Zξi ]Eξ [Zξj ] − Eξ [Zξi Zξj ]
i
j
∂ξ ∂ξ
= −α(1 + α)Cov(Zξi , Zξj ).
Since D2 Φ is a negative multiple of the covariance matrix of Zξ , Φ is concave and
ϕ is α-exponentially concave.
3.2.2. F (−α) -family. Next we consider the F (−α) -family. Interestingly, the result
is a little different and depends on the value of α.
Proposition 3.4. Consider the potential function ϕ(ξ) of the F (−α) -family defined
by (3.7).
(i) If 0 < α < 1, then ϕ is α-exponentially convex.
(ii) If α > 1, then ϕ is α-exponentially concave.
Proof. Now we heave
Φ(ξ) = eαϕ(ξ) =
Z
1
(1 + αξ · h) α dµ
α
.
A computation similar to that in the proof of Proposition 3.3 shows that
h(X)
Dϕ(ξ) = Eξ Zξ = Eξ
1 + αξ · h(X)
and
∂2Φ
(ξ) = α(1 − α)Covξ (Zξi , Zξj ).
∂ξ i ∂ξ j
Now Φ is convex if 0 < α < 1 and concave if α > 1.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
17
3.3. Rényi entropy and divergence. Based on information-theoretic reasonings,
Rényi entropy and divergence were first introduced by Rényi in [29] and they have
numerous applications in information theory and statistical physics among other
areas. See [32] for a survey of their remarkable properties. We show that these
quantities are naturally related to our F (±α) -families.
First let us recall the definitions of Rényi entropy and divergence. Same as above
we fix a measure space (X , µ).
Definition 3.5. Let α̃ ∈ (0, 1) ∪ (1, ∞). Let P , Q be probability measures on X
dQ
that are absolutely continuous with respect to µ. We let p = dP
dµ and q = dµ be
their densities.
(i) The Rényi entropy of P of order α̃ (and with reference measure µ) is
defined by
Z
1
(3.12)
Hα̃ (P ) =
log pα̃ dµ.
1 − α̃
We also denote the above by Hα̃ (p).
(ii) The Rényi divergence or order α̃ between P and Q is defined by
Z
1
log pα̃ q 1−α̃ dµ.
(3.13)
Dα̃ (P ||Q) =
α̃ − 1
Note that Hα̃ (P ) = −Dα̃ (P ||µ). We use a different notation to distinguish the
Rényi divergence from the L(±α) -divergences D(±α) [ξ : ξ ′ ]. It is well known that
as α̃ → 1, the Rényi entropy and divergence converge respectively to the Shannon
entropy and the Kullback-Leibler divergence.
The following is the main result of this section.
Theorem 3.6.
(i) For an F (α) -family with α > 0, the L(α) -divergence of ϕ is the Rényi
entropy of order α̃ := 1 + α:
(3.14)
D(α) [ξ : ξ ′ ] = Dα̃ (p(·, ξ ′ )||p(·, ξ)) .
Moreover, the α-conjugate ψ of ϕ is the Rényi entropy of order α̃:
(3.15)
ψ(η) = Hα̃ (p(·, ξ)),
η = D(α) ϕ(ξ).
(ii) Consider an F (−α) -family. If 0 < α < 1, the L(−α) -divergence of ϕ is the
Rényi entropy of order α̃ := 1 − α:
(3.16)
D(−α) [ξ : ξ ′ ] = Dα̃ (p(·, ξ ′ )||p(·, ξ)) ,
and the (−α)-conjugate is ψ(η) = −Hα̃ (p(·, ξ)) (note the minus sign).
If α > 1, the L(α) -divergence of ϕ is given by
(3.17)
α−1
Dα (p(·, ξ)||p(·, ξ ′ )) ,
α
Note that here α̃ = α and the order of the arguments is switched. The
conjugate is
Z
−1
log p(x, ξ)1−α dµ.
ψ(η) =
α
Note that it is not a Rényi entropy since 1 − α < 0.
D(α) [ξ : ξ ′ ] =
18
TING-KAM LEONARD WONG
Proof. (i) We want to compute
1
D(α) [ξ : ξ ′ ] = log (1 + αDϕ(ξ ′ ) · (ξ − ξ ′ )) − (ϕ(ξ) − ϕ(ξ ′ )).
α
First, using (3.10) and (3.11), we have
(3.18)
1 + αDϕ(ξ ′ ) · (ξ − ξ ′ ) = 1 + αEξ′ Zξ′ · (ξ − ξ ′ )
h(X)
′
= 1 + αEξ′
· (ξ − ξ )
1 + αξ ′ · h(X)
1 + αξ · h(X)
= Eξ′
.
1 + αξ ′ · h(X)
From (3.4), we have
′
1 + αξ ′ · h(X) =
eαϕ(ξ )
.
p(X, ξ ′ )α
It follows from (3.18) that
1
D(α) [ξ : ξ ′ ] = log (1 + αDϕ(ξ ′ ) · (ξ − ξ ′ )) − (ϕ(ξ) − ϕ(ξ ′ ))
α
eαϕ(ξ) /p(X, ξ)α
1
− (ϕ(ξ) − ϕ(ξ ′ ))
= log Eξ′ αϕ(ξ′ )
α
e
/p(X, ξ ′ )α
1
p(X, ξ ′ )α
= log Eξ′
.
α
p(X, ξ)α
Expressing the divergence in terms of the integral, we have
Z
1
D(α) [ξ : ξ ′ ] = log p(x, ξ ′ )1+α p(x, ξ)−α dµ
α
Z
1
log p(x, ξ ′ )α̃ p(x, ξ)1−α̃ dµ,
=
α̃ − 1
where α̃ = 1 + α > 1. This shows that D(α) is the Rényi divergence of order α̃.
Next we consider the α-conjugate
1
ψ(η) = log (1 + αξ · η) − ϕ(ξ).
α
Using the identity (2.19) and following the argument in (3.18), we have
−1
1 + αξ · η = (1 − αDϕ(ξ) · ξ)
−1
= Eξ (1 + αξ · h(X))−1
i−1
h
.
= Eξ p(X, ξ)α e−αϕ(ξ)
From this we get
Z
Z
1
1
1+α −αϕ(ξ)
ψ(η) = − log p(x, ξ)
e
dµ − ϕ(ξ) =
p(x, ξ)α̃ dµ = Hα̃ (p(·, ξ)).
α
1 − α̃
This proves (i), and the proof of (ii) is similar.
Remark 3.7. It appears from the above results that the F (−α) -family with α > 1
is different from the other cases. Nevertheless, the F (α) -family for α > 0 and the
F (−α) -family for 0 < α < 1 together give all Rényi entropy/divergence of order
α̃ ∈ (0, 1) ∪ (1, ∞) (see Figure 1).
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
F (−α)
α̃ = 1 − α
19
F (α)
α̃ = 1 + α
α̃
0
1
Figure 1. The order α̃ of the Rényi entropy/divergence for an
F (±α) family.
Example 3.8 (finite simplex). Let X = {0, 1, . . . , d} be a finite set and let µ be the
counting measure on X . Fix α > 0. Let hi (x) = δ i (x) be the indicator function
for point i, i = 1, . . . , d. Then (3.4) gives, for α > 0, the F (α) -family of probability
mass functions on X given by
(3.19)
p(i, ξ) = (1 + αξ i )−1/α eϕ(ξ) ,
i = 1, . . . , d,
p(0, ξ) = eϕ(ξ) ,
where the α-exponentially concave function ϕ is
d
X
(1 + αξ j )−1/α ,
ϕ(ξ) = log p(0, ξ) = − log 1 +
j=1
and the domain is Ω = {ξ ∈ Rd : 1 + αξ j > 0}. The coordinate ξ is aPglobal
coordinate system of the open simplex Sd = {p = (p0 , . . . , pd ) ∈ (0, 1)1+d : i pi =
1}. From (3.19), we have
α
1
p(0, ξ)
i
ξ =
− 1 , i = 1, . . . , d.
α
p(i, ξ)
By Theorem 3.6, the L(α) -divergence of ϕ is the discrete Rényi divergence of order
α̃ = 1 + α > 0.
Analogously, we may also express the simplex Sd as an F (−α) -family where
1
p(i, ξ) = (1 + αξ i ) α e−ϕ(ξ) ,
i = 1, . . . , d, p(0, ξ) = e−ϕ(ξ) ,
d
X
(1 + αξ j )1/α ,
ϕ(ξ) = − log p(0, ξ) = log 1 +
j=1
and
ξi =
1
α
p(i, ξ)
p(0, ξ)
α
−1 ,
i = 1, . . . , d.
When 0 < α < 1, the L(−α) -divergence of ϕ is the discrete Rényi divergence of
order α̃ = 1 − α > 0.
The corresponding dualistic structures will be considered in Section 5.4.
4. Dualistic structure of L(±α) -divergence
In this and the next sections we study the dualistic structure induced by a given
L(±α) -divergence. First we state the general definition of dualistic structure. For
preliminaries in differential geometry see [3, 9].
Definition 4.1. Let M be a smooth manifold. A dualistic structure on M is a
quadruplet (M, g, ∇, ∇∗ ) where g (denoted by h·, ·i) is a Riemannian metric, and
20
TING-KAM LEONARD WONG
(∇, ∇∗ ) is a pair of torsion-free affine connections that are dual with respect to g:
for any vector fields X, Y and Z, the covariant derivatives satisfy the identity
(4.1)
ZhX, Y i = h∇Z X, Y i + hX, ∇∗Z Y i.
We also call M equipped with a dualistic structure a statistical manifold.
For a dualistic structure, it is easy to see that the average ∇ = 12 (∇ + ∇∗ ) is the
Levi-Civita connection of the metric g. By the general results of Eguchi [15, 16],
any divergence (see [3, Definition 1.1]) on a manifold M defines a dualistic structure
on M . It is this structure induced by D(±α) that we study in this section.
We will derive explicit coordinate representations of this dualistic structure. The
geometric consequences will be studied in Section 5. Let us stress that while any
divergence induces a geometric structure, the self-dual representations of the L(±α) divergences motivated by optimal transport suggest the appropriate coordinate
systems.
By the results of Section 3, for an F (±α) -family this gives the geometry induced by the Rényi divergence. The Rényi divergence is closely related to the
α-divergence.1 Our framework is more general as it does not depend on the probabilistic representation. We discuss the connection with the α-divergences in Section
5.4.
4.1. Notations. We focus on the case of L(α) -divergence for an α-exponentially
concave function ϕ. The arguments for the L(−α) -divergence are essentially the
same (with some changes of signs) and will be left for the reader. The main results
for the L(−α) -divergence will be stated in Section 4.3.
We will use the self-dual representation derived in Theorem 2.13:
1
(4.2)
D [ξ : ξ ′ ] = D(α) [ξ : ξ ′ ] = log (1 + αξ · η ′ ) − ϕ(ξ) − ψ(η ′ ).
α
Here η = D(α) ϕ(ξ) is the dual coordinate system given by the α-gradient. To
simplify the notations we also use the notation
(4.3)
Π(ξ, η ′ ) = Π(α) (ξ, η ′ ) = 1 + αξ · η ′
for ξ ∈ Ω, η ′ ∈ Ω′ .
We consider the smooth manifold M = Ω where the primal coordinate ξ (identity
map) and the dual coordinate η = D(α) ϕ(ξ) are global coordinate systems. Note
that by symmetry of the cost function c(α) and the associated α-duality, we only
need to consider the primal coordinate system ξ. Consider the coordinate vector
fields ∂ξ∂ 1 , . . . , ∂ξ∂ d . By definition, the coordinate representations of (g, ∇, ∇∗ ) under
ξ are given by
∂ ∂
∂
∂
= − i ′ j D [ξ : ξ ′ ]
,
,
gij (ξ) =
∂ξ i ∂ξ j
∂ξ ∂ξ
ξ=ξ ′
∂2
∂
∂
∂
Γijk (ξ) = ∇ ∂ i j , k = − i j ′ k D [ξ : ξ ′ ]
,
(4.4)
∂ξ ∂ξ
∂ξ
∂ξ ∂ξ ∂ξ
ξ=ξ ′
∂
∂2
∂
∂
∗
∗
Γijk (ξ) = ∇ ∂
=
−
,
D [ξ : ξ ′ ]
.
′
j ∂ξ k
∂ξ i ∂ξ ′ j ∂ξ k
∂ξ i ∂ξ
ξ=ξ ′
1The author thanks Shun-ichi Amari for pointing this out.
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
21
We also define
(4.5)
Γij k (ξ) = Γijm (ξ)g mk (ξ),
Γ∗ij k (ξ) = Γ∗ijm (ξ)g mk (ξ),
where g ij (ξ) is the inverse of (gij (ξ)) and the Einstein summation convention (see
for example [3, p.21]) is used. This implies that
∂
∂
∇ ∂ i j = Γij k (ξ) k
∂ξ ∂ξ
∂ξ
and similarly for the dual
∇∗ .
connection
∂η
∂ξ
∂ξ i
∂η i
be the Jacobian matrices of the
We also let ∂η = ∂ηj and ∂ξ = ∂ξ
j
transition maps. They are inverses of each other.
4.2. Dualistic structure of L(α) -divergence.
4.2.1. The Riemannian metric.
Proposition 4.2. The Riemannian matrix G(ξ) = (gij (ξ)) is given by
−1
α
∂η
(4.6)
G(ξ) = (α)
.
I − (α)
ηξ ⊤
∂ξ
Π (ξ, η)
Π (ξ, η)
Its inverse is given by
∂ξ
I + αηξ ⊤ .
(4.7)
G−1 (ξ) = g ij (ξ) = −Π(α) (ξ, η)
∂η
Proof. Using the self-dual representation (4.2), we have
η ′i
∂ϕ
∂
′
D
[ξ
:
ξ
]
=
− i (ξ),
i
′
∂ξ
Π(ξ, η ) ∂ξ
(4.8)
1
∂2
D [ξ : ξ ′ ] =
i
′j
∂ξ ∂ξ
Π(ξ, η ′ )2
d
X ∂η ′ℓ
∂η ′i
Π(ξ, η ) ′j − αη ′i
ξ ℓ ′j
∂ξ
∂ξ
′
ℓ=1
!
.
Setting ξ = ξ ′ and expressing in matrix form, we obtain (4.6).
We obtain (4.7) by taking the inverse of (4.6). For the middle term, we may use
the Sherman-Morrison formula (see (4.10) below) to get
−1
α
I−
ηξ ⊤
= I + αηξ ⊤ .
Π(ξ, η)
Remark 4.3. In the above proposition the Riemannian matrix is given in terms
of the Jacobian matrix ∂η
∂ξ . This is to emphasize the role of duality (compare
(4.6) with [3, (1.66)]) and to enable explicit expressions of the Christoffel symbols
Γij k = Γijm g mk which appear in the primal geodesic equations.
It is not difficult to show directly using the formula (2.5) of the L(α) -divergence
that
(4.9)
G(ξ) = −D2 ϕ(ξ) − α(Dϕ(ξ))(Dϕ(ξ))⊤ ,
which is exactly the positive definite matrix in (2.2). Using the general ShermanMorrison formula
−1
A−1 uv ⊤ A−1
(4.10)
A + uv ⊤
= A−1 −
,
1 + v ⊤ A−1 u
22
TING-KAM LEONARD WONG
we can invert (4.9) to get
(D2 ϕ)−1 (Dϕ)(Dϕ)⊤ (D2 ϕ)−1
.
1 + α(Dϕ)⊤ (D2 ϕ)−1 (Dϕ)
While this formula is explicit, it is difficult to use in differential-geometric computations.
G−1 (ξ) = −(D2 ϕ)−1 + α
4.2.2. Affine connections.
Proposition 4.4. The Christoffel symbols of the primal connection are given by
(4.11)
d
i
j
X
α
∂η ℓ
α
2α2
j ∂η
i ∂η
i j
Γijk (ξ) = (α)
η
η
η
η
ξℓ k ,
+
−
k
k
2
(α)
2
(α)
3
∂ξ
∂ξ
∂ξ
Π (ξ, η)
Π (ξ, η)
Π (ξ, η)
ℓ=1
∂ϕ
∂ϕ k
−α
(4.12)
δ + j δik .
η i δjk + η j δik = −α
Γij k (ξ) = (α)
∂ξ i j
∂ξ
Π (ξ, η)
where δ·· is Dirac’s delta function.
Proof. From our previous computation (4.8), we have
d
X ∂η ′ℓ
∂ ∂
∂η ′i
1
α
D [ξ : ξ ′ ] =
−
η ′i
ξ ℓ ′k .
i
′k
′
′k
′
2
∂ξ ∂ξ
Π(ξ, η ) ∂ξ
Π(ξ, η )
∂ξ
ℓ=1
′
Differentiating one more time and writing Π = Π(ξ, η ), we get
d
X ∂η ′ℓ
∂
−α
∂η ′i
α
∂η ′j
2α
∂2
D [ξ : ξ ′ ] = 2 η ′j ′k − 2 η ′i ′k + 3 η ′i η ′j
ξ ℓ ′k .
i
j
′k
∂ξ ∂ξ ∂ξ
Π
∂η
Π
∂ξ
Π
∂ξ
ℓ=1
′
Setting ξ = ξ , we obtain (4.11).
To prove (4.12), we first use (4.7) to write
d
g
mk
(ξ) = −Π(ξ, η)
X ∂ξ m
∂ξ m
+α
ηℓ ξ k
k
∂η
∂η ℓ
ℓ=1
It follows that
Γij k (ξ) = Γijm (ξ)g mk (ξ)
=−
α i ∂η j
2α2 i j X ℓ ∂η ℓ
α j ∂η i
+
−
ηη
η
η
ξ
Π ∂ξ m
Π ∂ξ m
Π2
∂ξ m
ℓ
2
=
2
!
!
.
d
X ∂ξ m
∂ξ m
+
α
ηℓ ξ k
∂η k
∂η ℓ
ℓ=1
3
−α j k α i k 2α i j k 2α i j k 2α i j k
η δi − η δj + 2 η η ξ −
ηη ξ + 2ηη ξ
Π
Π
Π
Π
Π
d
X
!
ξ ℓ ηℓ .
ℓ=1
We observe that the last three terms cancel out and the resulting expression gives
the first equality in (4.11). The second equality follows from the definition of η as
the α-gradient (2.11) as well as the identity (2.19).
Remark 4.5. Same as the Riemannian matrix, we may represent Γijk (ξ) using the
function ϕ. For completeness, it is given by
2
∂ 2 ϕ ∂ϕ
∂ ϕ ∂ϕ
2 ∂ϕ ∂ϕ ∂ϕ
+α
+ j k i .
Γijk (ξ) = 2α
∂ξ i ∂ξ j ∂ξ k
∂ξ i ∂ξ k ∂ξ j
∂ξ ∂ξ ∂ξ
Without using the self-dual representation, it is not easy to see how this formula
leads to (4.12).
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
23
4.3. Dualistic structure of L(−α) -divergence. Here we state the corresponding
results for the L(−α) -divergence.
Proposition 4.6. For the dualistic structure generated by D(−α) where ϕ is αexponentially convex, we have
(4.13)
(4.14)
gij (ξ) =
k
∂ 2ϕ
∂ϕ
∂ϕ
(ξ) + α i (ξ) j (ξ),
i
j
∂ξ ∂ξ
∂ξ
∂ξ
Γij (ξ) = −α
∂ϕ
∂ϕ
k
k
(ξ)δj + j (ξ)δi .
∂ξ i
∂ξ
Note that (4.12) and (4.14) have the same form (even though here ϕ is αexponentially convex). This is because Γij k = Γijm g mk and the changes of signs
cancel out in the product.
5. Geometric consequences
With all the coefficients available, we derive in this section the geometric properties of the dualistic structure (g, ∇, ∇∗ ). We state the results for any D(±α) with
α > 0, but the proofs will again be given for the case of D(α) . In Section 5.4 we
specialize to the unit simplex and explain the connections with the α-divergence.
5.1. Dual projective flatness. The following definition is taken from [22].
e be torsion-free affine connecDefinition 5.1 (Projective flatness). Let ∇ and ∇
tions on a smooth manifold. They are projectively equivalent if there exists a differential 1-form τ such that
(5.1)
e X Y + τ (X)Y + τ (Y )X
∇X Y = ∇
for any smooth vector fields X and Y . We say that ∇ is projectively flat if ∇ is
projectively equivalent to a flat connection (a connection whose Riemann-Christoffel
curvature tensor vanishes).
If we write τ = ai (ξ)dξ i , say, using the primal coordinate system, then (5.1) is
equivalent to
(5.2)
e ij k (ξ) + ai (ξ)δjk + aj (ξ)δik .
Γij k (ξ) = Γ
Theorem 5.2. For any D(±α) , the primal connection ∇ and the dual connection
∇∗ are projectively flat. Thus we say that the induced dualistic structures are dually
projectively flat.
Proof. We only consider the primal connection. Consider the 1-form defined by
τ (X) = −αXϕ. In primal coordinates, we have
(5.3)
τ =−
d
X
i=1
α
∂ϕ i
dξ .
∂ξ i
e be the flat Euclidean connection with respect to the coordinate system ξ
Let ∇
e ij k (ξ) ≡ 0. From (5.2), we see that ∇ is
such that its Christoffel symbols satisfy Γ
e
projectively equivalent to ∇ and thus ∇ is projectively flat.
24
TING-KAM LEONARD WONG
Projective flatness can be related to the behaviors of the geodesics. Recall that
a (smooth) curve γ : [0, 1] → M is said to be a primal geodesic if ∇γ̇(t) γ̇(t) ≡ 0.
Equivalently, its primal coordinate representation ξ(t) = (ξ 1 (t), . . . , ξ d (t)) satisfies
the primal geodesic equations
(5.4)
ξ̈ k (t) + ξ˙i (t)ξ˙j (t)Γij k (ξ(t)) = 0, k = 1, . . . , d,
where the dots denote derivatives with respect to time. Similarly, the curve γ is a
dual geodesic if ∇∗γ̇(t) γ̇(t) ≡ 0.
The following result is well-known in differential geometry.
e are projectively equivalent, then a ∇-geodesic is a
Proposition 5.3. If ∇ and ∇
e
∇-geodesic up to a time reparameterization, and vice versa.
Corollary 5.4. For the dualistic structure induced by D(±α) , if γ is a primal (dual)
geodesic, then its trace in the primal (dual) coordinate system is a straight line.
Thus the primal geodesics are straight lines under the primal coordinate system
but run at non-constant speeds. Let us consider the time reparameterization. Using
(4.11), we note that the primal geodesic equation (5.4) is equivalent to the single
vector equation
˙ dϕ(ξ(t)) .
(5.5)
ξ̈(t) = 2αξ(t)
dt
From this we immediately see that the trace of the primal geodesic is a straight
line under the coordinate ξ.
Proposition 5.5. Let γ : [0, 1] → M be a primal geodesic. Then, in primal
coordinates, we have
(5.6)
ξ(t) = (1 − h(t))ξ(0) + h(t)ξ(1),
where h : [0, 1] → [0, 1] is given by
Rt
h(t) = R 01
(5.7)
0
e2αϕ(ξ(s)) ds
e2αϕ(ξ(s)) ds
.
Proof. We already know that a primal geodesic can be written in the form (5.6).
Plugging this into the geodesic equation (5.5), we have
d
dϕ(ξ(t))
dϕ(ξ(t))
⇒
log h′ (t) = 2α
.
dt
dt
dt
Integrating and using the fact that h(0) = 0 and h(1) = 1, we obtain the desired
formula (5.7).
h′′ (t) = 2αh′ (t)
A natural question is the relationship between the primal and dual coordinate
vector fields. This is studied in the following proposition.
Proposition 5.6.
(i) For D(α) , the inner product between ∂ξ∂ i and ∂η∂ j is
∂
−1
∂
α
(5.8)
=
,
ξ j ηi .
δij +
∂ξ i ∂η j
Π(ξ, η)
Π(ξ, η)2
(ii) For D(−α) , the inner product between ∂ξ∂ i and ∂η∂ j is
1
∂
α
∂
=
,
ξ j ηi .
δij −
(5.9)
∂ξ i ∂η j
Π(ξ, η)
Π(ξ, η)2
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
Proof. Consider (i). Using Proposition 4.2, we compute
+
*
X ∂ξ m ∂
∂ X ∂ξ m ∂
∂
∂
∂
=
=
,
,
,
∂ξ i ∂η j
∂ξ i m ∂η j ∂ξ m
∂η j ∂ξ i ∂ξ m
m
=
X ∂ξ m
m
∂η j
X ∂η ℓ
−1 ∂η i
α
i
+
η
ξℓ m
Π(ξ, η) ∂ξ m
Π(ξ, η)2
∂ξ
ℓ
We obtain (5.8) under some simplification.
25
!
.
The formulas (5.8) and (5.9) are quite interesting. In the dually flat case (i.e.,
D(0±) , alsoDsee [3, Theorem
6.6]), the two coordinate fields are orthonormal in the
E
sense that ∂ξ∂ i , ∂η∂ j = ±δij . When α > 0, the first term of the inner product is
conformal to ±δij , but there is an extra term related to duality. This formula is
the key ingredient in the proof of the generalized Pythagorean theorem.
5.2. Curvatures. Next we consider the primal Riemann-Christoffel curvature tensor R defined by
(5.10)
R(X, Y )Z = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z,
where X, Y, Z are vector fields and [X, Y ] = XY − Y X is the Lie bracket. In the
primal coordinate system, we use the notation
∂
∂
∂
∂
,
= Rijk ℓ ℓ .
R
∂ξ i ∂ξ j ∂ξ k
∂ξ
It can be shown (see for example [3, Section 5.8]) that the coefficients are given by
(5.11)
Rijk ℓ = ∂i Γjk ℓ − ∂j Γik ℓ + Γjk m Γim ℓ − Γik m Γjm ℓ .
The dual curvature tensor R∗ is defined analogously.
Proposition 5.7.
(i) For D(α) , the primal curvature tensor is given by
(5.12)
Rijk ℓ (ξ) = α gik (ξ)δjℓ − gjk (ξ)δiℓ .
(ii) For D(−α) , the primal curvature tensor is given by
(5.13)
Rijk ℓ (ξ) = −α gik (ξ)δjℓ − gjk (ξ)δiℓ .
Proof. This is a straightforward but lengthy computation, so we omit the details.
Definition 5.8. When dim M ≥ 2, a torsion-free connection has constant sectional
curvature k ∈ R with respect to the metric g if its curvature tensor R satisfies the
identity
(5.14)
R(X, Y )Z = k (hY, ZiX − hX, ZiY )
for all vector fields X, Y and Z.
Using coordinate vector fields, we see that R has constant curvature k if and
only if in any coordinate system we have
(5.15)
Rijk ℓ = k gjk δiℓ − gik δjℓ .
Comparing this with (5.12) and (5.13), we obtain
26
TING-KAM LEONARD WONG
Theorem 5.9. When dim Ω = d ≥ 2, for D(±α) the primal and dual connections
have constant sectional curvature ∓α with respect to the induced metric.
It is known that for a general dualistic structure (M, g, ∇, ∇∗ ), ∇ has constant
sectional curvature k with respect to g if and only if the same statement holds for
∇∗ . For a proof see [12, Proposition 8.1.4]. Some general properties of dualistic
structures with constant curvatures are given in [12, Theorem 9.7.2]. For example, it
is known that such a manifold is conjugate symmetric, conjugate Ricci-symmetric,
and the connections ∇ and ∇∗ are equiaffine. In Section 6, we show that if we
assume in addition that the manifold is dually projectively flat in the sense of
Theorem 5.2, then the geometry can be characterized elegantly in terms of the
L(±α) -divergence.
5.3. Generalized Pythagorean theorem. The generalized Pythagorean theorem is a fundamental result in information geometry and has numerous applications
in information theory, statistics and machine learning. It was first proved for the
Bregman divergence (D(±α) in our context) which induces a dually flat geometry.
For an exposition of this classical result see [3, Chapter 1]. In [27, 28] we introduced
the L-divergence (L(α) -divergence when α = 1) on the unit simplex and proved the
generalized Pythagorean theorem. In [28] the proof is quite different from that of
the Bregman case and is rather involved because there we expressed the dualistic
structure in terms of the exponential coordinate system. Here we give a unified and
simplified treatment covering all L(±α) -divergences.
Theorem 5.10. Consider any L(±α) -divergence D(±α) [· : ·] and the induced dualistic structure. Let p, q, r ∈ Ω and suppose that the dual geodesic from q to p exists
(this amounts to say that the line segment between ηq and ηp lies in Ω′ . We need
this assumption because Ω′ may not be convex). Then the generalized Pythagorean
relation
(5.16)
D(±α) [q : p] + D(±α) [r : p] = D(±α) [r : p]
holds if and only if the primal geodesic from q to r and the dual geodesic from q to
p meet g-orthogonally at q.
Proof. Again we consider the case of D(α) ; the proof for D(−α) is similar. By the
self-dual representation (Theorem 2.13), we have
1
D(α) [q : p] = log (1 + αξq · ηp ) − ϕ(ξq ) − ψ(ηp ),
α
1
D(α) [r : q] = log (1 + αξr · ηq ) − ϕ(ξr ) − ψ(ηq ),
α
1
D(α) [r : p] = log (1 + αξr · ηp ) − ϕ(ξr ) − ψ(ηp ).
α
Using these expressions and the Fenchel identity, we see that the Pythagorean
relation (5.16) holds if and only if
(1 + αξq · ηp ) (1 + αξr · ηq ) = (1 + αξr · ηp ) (1 + αξq · ηq ) .
Expanding and simplifying, we have
(5.17)
(ξr − ξq ) · (ηp − ηq ) = α(ξq · ηp )(ξr · ηq ) − α(ξr · ηp )(ξq · ηq ).
On the other hand, consider the primal geodesic from q to r. By projective
flatness, in primal coordinates, the initial velocity vector is proportional to ξr − ξq .
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
27
Similarly, in dual coordinates, the initial velocity of the dual geodesic from q to p
is proportional to ηp − ηq . By Proposition 5.6, we see that the two geodesics are
orthogonal at q if and only if
∂
j
j
i
i ∂
0 = (ξr − ξq ) i , (ηp − ηq ) j
∂ξ
∂η
X
α i j
−1
i
i
j
j
δij + 2 ηq ξq
(ξr − ξq )(ηp − ηq )
=
Π
Π
i,j
!
X
X
−1
α
(ξri − ξqi )ηqi (ηpj − ηqj )ξqj ,
=
(ξr − ξq ) · (ηp − ηq ) + 2
Π
Π
j
i
where Π = Π(ξq , ηq ). Rearranging, we have
(1 + αξq · ηq )(ξr · ηp − ξr · ηq − ξq · ηp + ξq · ηq )
= α(ξr · ηq − ξq · ηq )(ξq · ηp − ξq · ηq ).
The proof is completed by verifying that this and (5.17) are equivalent.
5.4. The α-divergence. In this subsection we specialize to the unit simplex
Sd = {p = (p0 , p1 , . . . , pd ) ∈ (0, 1)1+d : p0 + · · · + pd = 1}.
The following definition is taken from [3, (3.39)]. Also see [9, Definition 2.9].
Definition 5.11 (α-divergence). Let α 6= ±1. The α-divergence is defined by
!
d
X
1+α
4
i 1−α
i
, p, q ∈ Sd .
(p ) 2 (q ) 2
1−
(5.18)
Dα [p : q] =
1 − α2
i=0
We also consider the discrete Rényi divergence given by
!
d
X
1
α̃ 1−α̃
.
(5.19)
Dα̃ (p||q) =
pi qi
log
α̃ − 1
i=0
Note the difference in the notations to avoid confusion. The following elementary
lemma (whose proof is omitted) shows that the two divergences are monotone
transformations of each other.
Lemma 5.12. For α̃ ∈ (0, 1) ∪ (1, ∞), let α = 1 − 2α̃ ∈ (−∞, −1) ∪ (−1, 1). Then
for p, q ∈ Sd we have
1
(5.20)
Dα̃ (p||q) =
log (1 + α̃(α̃ − 1)Dα [p : q]) .
α̃ − 1
As in Example 3.8, we may express Dα̃ (p||q) as the L(α̃−1) -divergence of the
F
-potential function when α̃ > 1, and as the L(−(1−α̃)) -divergence of the
−(1−α̃)
F
-potential function when 0 < α̃ < 1.
Let α and α̃ be as in Lemma 5.12. Let (g, ∇, ∇∗ ) be the dualistic structure
˜ ∇
˜ ∗ ) be that induced
induced on Sd by the α-divergence Dα [p : q], and let (g̃, ∇,
by the Rényi divergence Dα̃ (p||q). Since the two divergences are related by a
monotone transformation, from (4.4), (5.20) and the chain rule we have gij = α̃g̃ij ,
Γijk = α̃Γ̃ijk and Γ∗ijk = α̃Γ̃∗ijk . (Here the multiplier α̃ is the derivative of the
transformation (5.20) at 0.) This implies that Γij k = Γ̃ij k and Γ∗ij k = Γ̃∗ij k , so
(α̃−1)
28
TING-KAM LEONARD WONG
the two dualistic structures have the same primal and dual geodesics and the same
primal and dual curvature tensors (see (5.11)).
Theorem 5.13. For α = 1 − 2α̃ ∈ (−∞, −1) ∪ (−1, 1), the statistical manifold
(Sd , g, ∇, ∇∗ ) induced by the α-divergence (5.18) is dually projectively flat with con2
stant curvature 1−α
4 , and a generalized Pythagorean theorem holds.
Proof. We only indicate how the curvature is determined, as the rest follows immediately from other results of this section. We have α̃ = 1−α
2 . Suppose α ∈
′
(−∞, −1) > 0. The Rényi divergence Dα̃ (p||q) is then an L(α ) -divergence where
α̃ = 1 + α′ ∈ (1, ∞). By Theorem 5.9, we have
R̃(X, Y )Z = −α′ (g̃(Y, Z) − g̃(X, Z)Y ) .
But R = R̃ and g̃ = α̃g. So we have
R(X, Y )Z = −α′ α̃ (g(Y, Z) − g(X, Z)) =
This shows that the curvature is
1−α2
4 .
1 − α2
(g(Y, Z) − g(X, Z)) .
4
The case α ∈ (−1, 1) is similar.
In this context, our geometry reduces to the geometry of the α-divergence. See in
2
is determined using affine differential
particular [2, p.179] where the curvature 1−α
4
geometry. Their (8.67) is the product form of our (5.16) for the Rényi divergence
(5.20). Thus, our results provide a new approach to study the geometry (compare
with [2, Section 3.6]). We plan to carry out a deeper study in future papers.
6. Dually projectively flat manifolds with constant curvatures
Using the expressions given in (4.4), any divergence D [· : ·] on a manifold M
defines a dualistic structure (M, g, ∇, ∇∗ ). This operation is not one-to-one; it is
easy to construct examples where two different divergences induce the same dualistic
structure on M . An interesting question is whether we can single out a divergence
which is in some sense most natural for a given dualistic structure. If so, we call it
a canonical divergence. For dually flat manifolds, it is possible to show that there
is a canonical divergence of Bregman type (see [3, Section 6.6]). Thus, a dualistic
structure is dually flat if and only if it is (locally) induced by a Bregman divergence.
In this section we ask and solve the same question where the manifold is dually
projectively flat with constant curvature. We assume that the manifold M has dimension d ≥ 2, and both ∇ and ∇∗ are projectively flat and have constant sectional
curvature ±α where α > 0. (Note that if α = 0 the manifold is dually flat and we
reduce to the classical setting.) Our aim is to characterize the geometry based on
these properties. The key idea is to use these properties to extend in a novel way
the proofs of [3, Theorem 6.2] and [9, Theorem 4.3] which address the dually flat
case. Using tools of affine differential geometry, an alternative characterization is
given in [19, Corollary 1].
6.1. Dual coordinates and potential functions.
Theorem 6.1. Consider a satistical manifold (M, g, ∇, ∇∗ ) such that both ∇ and
∇∗ are projectively flat and have constant non-positive sectional curvature −α where
α > 0. Then, near each point of M , there exist local coordinate systems ξ and η
such that the following statements hold:
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
29
(i) The primal geodesics are straight lines in the ξ coordinates up to time
reparameterizations, and the dual geodesics are straight lines in the η coordinates up to time reparameterizations.
Pd
(ii) The coordinates satisfy the constraint 1 + α ℓ=1 ξ ℓ η ℓ > 0.
(iii) There exist local α-exponentially concave functions ϕ(ξ) and ψ(η) such
that the generalized Fenchel identity
(6.1)
ϕ(ξ) + ψ(η) ≡ c(α) (ξ, η) =
1
log (1 + αξ · η)
α
holds true.
(iv) Under the coordinate systems ξ and η respectively, the Riemannian metric
is given by
∂ϕ
∂ϕ
∂2ϕ
(ξ) − α i (ξ) j (ξ),
i
j
∂ξ ∂ξ
∂ξ
∂ξ
∂2ψ
∂ψ
∂ψ
gij (η) = − i j (η) − α i (η) j (η).
∂η ∂η
∂η
∂η
gij (ξ) = −
(6.2)
If the constant curvature is +α > 0, analogous statements hold where the functions
ϕ and ψ in (iii) are α-exponentially convex, and in (6.2) the negative signs are
replaced by positive signs.
Proof. Let the curvature be −α < 0 (the proof for positive curvature is similar).
Consider the primal connection ∇ which is projectively flat. By Definition 5.1,
e and a 1-form τ such that (5.1) holds. As ∇
e is
there exists a flat connection ∇
1
d
flat, near each point there exists an affine coordinate system ξ = (ξ , . . . , ξ ) under
e
which the ∇-geodesics
are constant velocity straight lines. By projective equivalence, the primal geodesics are straight lines under the ξ coordinates up to time
reparameterizations. Furthermore, if we write τ = ai (ξ)dξ i where the ai (ξ)’s are
smooth functions of ξ, by (5.2) we have
(6.3)
Γij k (ξ) = ai (ξ)δjk + aj (ξ)δik .
Motivated by our previous result (5.3), we want to show that
(6.4)
ai (ξ) = −α
∂ϕ
(ξ)
∂ξ i
for some α-exponentially concave function ϕ. To this end we will use the assumption
that ∇ has constant sectional curvature −α with respect to g.
Using the representation (6.3) of the primal Christoffel symbols, we compute
(6.5)
Rijk ℓ (ξ) = ∂i Γjk ℓ − ∂j Γik ℓ + Γjk m Γim ℓ − Γik m Γjm ℓ
= (∂i aj − ∂j ai )δkℓ + (∂i ak − ak ai )δjℓ − (∂j ak − ak aj )δiℓ .
On the other hand, since ∇ has constant curvature −α, by (5.15) we have
(6.6)
Rijk ℓ (ξ) = α gik (ξ)δjℓ − gjk (ξ)δiℓ .
Equating (6.5) and (6.6) and using the assumption that dim M = d ≥ 2 (so that
we can pick different indices), we see that
(6.7)
∂ai
∂aj
(ξ) = j (ξ)
∂ξ i
∂ξ
30
TING-KAM LEONARD WONG
and
(6.8)
αgij (ξ) = ∂i aj (ξ) − ai (ξ)aj (ξ).
By (6.7), the 1-form τ = ai (ξ)dξ i is closed. So, locally, it is exact. Thus, there
exists a locally defined function ϕ(ξ) such that our claim (6.4) holds. In particular,
the 1-form τ is locally given by τ (X) = −αXϕ.
Now we may write the metric (6.8) in the form
(6.9)
gij (ξ) = −
∂ϕ
∂ϕ
∂2ϕ
(ξ) − α i (ξ) j (ξ).
i
j
∂ξ ∂ξ
∂ξ
∂ξ
Since the Riemannian matrix (gij (ξ)) is strictly positive definite, from Remark 4.3
we know that D2 eαϕ < 0. In particular, ϕ is locally α-exponentially concave.
Without loss of generality we may assume 0 belongs to the domain of ϕ. This
implies
1 − αDϕ(ξ) · ξ > 0,
and the α-gradient
(6.10)
η = ∇(α) ϕ(ξ) =
1
∇ϕ(ξ)
1 − αDϕ(ξ) · ξ
is well-defined. This implies as in (2.19) that 1 + αξ · η > 0.
With the dual coordinate η defined by (6.10), we define the dual function ψ(η)
by the generalized Fenchel identity:
(6.11)
ψ(η) = c(α) (ξ, η) − ϕ(ξ).
By α-duality ψ is locally α-exponentially concave, and the Fenchel identity (6.1)
holds by construction.
Since the dual connection ∇∗ is uniquely determined given g and ∇ (see [3,
(6.6)]), with the results in Section 4 it is a now routine exercise to verify that η is
an affine coordinate system for the projectively flat dual connection ∇∗ , and the
metric is also given by the second formula in (6.2).
6.2. Canonical divergence. Using Theorem 6.1, we can define locally a canonical
divergence which generalizes the one for a dually flat manifold. It is interesting to
know whether this canonical divergence coincides with the one defined by Ay and
Amari in [8] for a general dualistic structure.
Theorem 6.2 (canonical divergence). Consider the setting of Theorem 6.1. Let ξ,
η, ϕ and ψ be given as in the theorem.
(i) If the curvature is −α < 0, we define the local L(α) -divergence
(6.12)
D(α) [q : p] =
1
log (1 + αξq · ηp ) − ϕ(ξq ) − ψ(ηp ).
α
(ii) If the curvature is α > 0, we define the local L(−α) -divergence
1
log (1 + αξq · ηp ) .
α
Then these divergences are independent of the choices of ξ, η, ϕ and ψ. Furthermore, the canonical divergence induces the given Riemannian metric and the primal
and dual connections.
(6.13)
D(−α) [q : p] = ϕ(ξq ) + ψ(ηp ) −
LOGARITHMIC DIVERGENCES AND RÉNYI GEOMETRY
31
Proof. Note that the primal affine coordinate system ξ is unique up to an affine
transformation. The rest are then determined by (6.10) and (6.11). By a direct
computation, we can verify that (6.12) and (6.13) remain invariant.
Acknowledgment
Some of the results were obtained when the author was teaching a topics course
on optimal transport and information geometry at the University of Southern California. He thanks his students for their interest and feedbacks. He also thanks
John Man-shun Ma, Jeremy Toulisse and Soumik Pal for helpful discussions. He
is grateful to Shun-ichi Amari for his insightful comments and for pointing out the
connection with the α-divergence.
References
[1] S. Adams, N. Dirr, M. A. Peletier, and J. Zimmer. From a large-deviations principle to the
Wasserstein gradient flow: a new micro-macro passage. Communications in Mathematical
Physics, 307(3):791–815, 2011.
[2] S. Amari, H. Nagaoka, and D. Harada. Methods of Information Geometry. Translations of
mathematical monographs. American Mathematical Society, 2007.
[3] S.-I. Amari. Information Geometry and Its Applications. Springer, 2016.
[4] S.-I. Amari, R. Karakida, and M. Oizumi. Information geometry connecting Wasserstein
distance and Kullback-Leibler divergence via the entropy-relaxed transportation problem.
arXiv preprint arXiv:1709.10219, 2017.
[5] S.-I. Amari and A. Ohara. Geometry of q-exponential family of probability distributions.
Entropy, 13(6):1170–1185, 2011.
[6] L. Ambrosio and N. Gigli. A users guide to optimal transport. In Modelling and optimisation
of flows on networks, pages 1–155. Springer, 2013.
[7] L. Ambrosio, N. Gigli, and G. Savaré. Gradient Flows: In Metric Spaces and in the Space of
Probability Measures. Springer, 2008.
[8] N. Ay and S.-i. Amari. A novel approach to canonical divergences within information geometry. Entropy, 17(12):8111–8129, 2015.
[9] N. Ay, J. Jost, H. Lê, and L. Schwachhöfer. Information geometry. Springer, 2017.
[10] A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh. Clustering with bregman divergences.
Journal of machine learning research, 6(Oct):1705–1749, 2005.
[11] Y. Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Communications on pure and applied mathematics, 44(4):375–417, 1991.
[12] O. Calin and C. Udrişte. Geometric modeling in probability and statistics. Springer, 2014.
[13] L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. Unbalanced optimal transport: geometry and kantorovich formulation. arXiv preprint arXiv:1508.05216, 2015.
[14] F. Dillen, K. Nomizu, and L. Vranken. Conjugate connections and radon’s theorem in affine
differential geometry. Monatshefte für Mathematik, 109(3):221–235, 1990.
[15] S. Eguchi. Second order efficiency of minimum contrast estimators in a curved exponential
family. The Annals of Statistics, 11(3):793–803, 1983.
[16] S. Eguchi et al. Geometry of minimum contrast. Hiroshima Mathematical Journal, 22(3):631–
647, 1992.
[17] M. Erbar, K. Kuwada, and K.-T. Sturm. On the equivalence of the entropic curvaturedimension condition and bochners inequality on metric measure spaces. Inventiones mathematicae, 201(3):993–1071, 2015.
[18] R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker–Planck
equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998.
[19] T. Kurose. Dual connections and affine geometry. Mathematische Zeitschrift, 203(1):115–121,
1990.
[20] T. Kurose. On the divergences of 1-conformally flat statistical manifolds. Tohoku Mathematical Journal, Second Series, 46(3):427–433, 1994.
[21] C. Léonard. From the Schrödinger problem to the Monge–Kantorovich problem. Journal of
Functional Analysis, 262(4):1879–1920, 2012.
32
TING-KAM LEONARD WONG
[22] H. Matsuzoe. Geometry of contrast functions and conformal geometry. Hiroshima Mathematical Journal, 29(1):175–191, 1999.
[23] H. Nagaoka and S.-i. Amari. Differential geometry of smooth families of probability distributions. Technical Report METR 82–7, University of Tokyo, 1982.
[24] J. Naudts. The q-exponential family in statistical physics. In Journal of Physics: Conference
Series, volume 201, page 012003. IOP Publishing, 2010.
[25] S. Pal. Embedding optimal transports in statistical manifolds. arXiv preprint
arXiv:1708.08148, 2017.
[26] S. Pal and T.-K. L. Wong. Energy, entropy, and arbitrage. arXiv preprint arXiv:1308.5376,
2013.
[27] S. Pal and T.-K. L. Wong. The geometry of relative arbitrage. Mathematics and Financial
Economics, 10(3):263–293, 2016.
[28] S. Pal and T.-K. L. Wong. Exponentially concave functions and a new information geometry.
Annals of Probability, forthcoming.
[29] A. Rényi. On measures of entropy and information. In Proceedings of the Fourth Berkeley
Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the
Theory of Statistics. The Regents of the University of California, 1961.
[30] R. T. Rockafellar. Convex Analysis. Princeton Landmarks in Mathematics. Princeton University Press, 1997.
[31] F. Santambrogio. Optimal transport for applied mathematicians. Birkäuser, NY, 2015.
[32] T. Van Erven and P. Harremos. Rényi divergence and Kullback-Leibler divergence. IEEE
Transactions on Information Theory, 60(7):3797–3820, 2014.
[33] C. Villani. Topics in Optimal Transportation. American Mathematical Society, 2003.
[34] C. Villani. Optimal Transport: Old and New. Springer, 2008.
[35] T.-K. L. Wong. Optimization of relative arbitrage. Annals of Finance, 11(3-4):345–382, 2015.
[36] T.-K. L. Wong. On portfolios generated by optimal transport. arXiv preprint
arXiv:1709.03169, 2017.
University of Southern California, Los Angeles, CA 90089, United States
E-mail address: [email protected]
| 10math.ST
|
Manuscript
Types of Cognition and its Implications for future High-Level
Cognitive Machines
Camilo Miguel Signorelli
Department of Experimental and Health Sciences, Universidad Pompeu Fabra, Barcelona, Spain
Centro Interdisciplinario de Neurociencias de Valparaíso, Universidad de Valparaíso, Chile
[email protected]
Abstract
This work summarizes part of current knowledge on Highlevel Cognitive process and its relation with biological
hardware. Thus, it is possible to identify some paradoxes
which could impact the development of future technologies
and artificial intelligence: we may make a High-level Cognitive Machine, sacrificing the principal attribute of a machine, its accuracy.
Introduction
Intelligence is one of the main characteristics of the Human Brain. For example, cognitive abilities such as problem-solving, language skills, social skills, among others,
are considered as High-level cognitive processes with a
strong impact on different types of intelligences. Understanding how these abilities work can help us to implement
desirable social behaviors in the new generation of robots
with Human-Machine interactions.
How do cognitive, abstract abilities and subjective experience emerge from a physical system like the brain?
(Chalmers, 1995; Dehaene, et al. 2014; Tononi, et al. 2016)
How can we implement these abilities in a natural form in
machines? Is it possible? What are the requirements to
achieve this? These are some of the most complex questions in science and to resolve them, we need to understand
the biological basis, the cellular dynamics as well as the
physical and mathematics basis to know the fundamental
properties of matter which could eventually be involved in
these emergent properties.
One requirement for High-level cognitive abilities seems
to be what it is called subjective experience (Barron &
Klein, 2016) or in a more complex order: consciousness.
For example, we first need to be conscious to take some
Manuscript to be presented to Association for the Advancement of
Artificial Intelligence (www.aaai.org).
complex rational decisions, to plan, and to do something
with intent (Baars, 2005; Tononi & Koch, 2008). How
biology implements rational or high intelligences, is completely different from how we started to implement it in
computer science (Moravec, 1988).
In order to improve machines socials skills, one should
try to understand how these abilities emerge in the brain
and also, in order to understand how these abilities work
and emerge from the brain, we should study how to implement them into machines. This is why the last test or
full understanding of subjective experience and consciousness will be to make a conscious machine.
This work will show a brief of what it is known about
conscious states and some paradoxes that emerge from the
idea of creating a conscious machine: it may be possible to
make a conscious robot, but it does not come without a
cost.
Consciousness and Intelligence
Consciousness, a brief notion
If consciousness could be defined, the questions above
would be easily answered. We may also have a recipe with
different “ingredients” to explain consciousness in humans,
animals, and accept or reject the option to implement it in
machines.
However, this is not the case. All of us understand what
consciousness is, what it is to be aware, what it is to have a
subjective experience and on which "elements" our experience is based. Sometimes these concepts are considered
as synonyms, other times as different ideas with different
meanings. Thus, it is needed to make some distinctions.
The modern approach to consciousness has made a differentiation between conscious states as different levels of
awareness (vegetative, sleep, anesthesia, altered states,
aware); contents of consciousness as elements or information in the external or internal world which at times are
part of our conscious state; and conscious processing as
the operations applied to these contents (Dehaene, et al.
2014). Thanks to these definitions, theories as Global
Workspace (Baars, 2005), Integrated Information (Tononi,
et al. 2016) and Dynamic theories (Varela, et al. 2001)
have developed many tools leading to a greater understanding of consciousness. In these approaches, the consciousness and eventually intentionality are understood as an
emergent property from brain networks dynamics.
Nevertheless, these definitions have some difficulties
explaining how this emergency works, explaining subjectivity experiences (Chalmers, 1995; Lycan & Dennett,
1993) or sometimes avoiding the importance of previous
learning to develop consciousness at early stage of childhood (Cleeremans, 2011).
Even though this article will refer to the above definitions, as we are referring to previous works that use these
definitions; our own understanding of consciousness, unlike these theories, might be defined as a process of
processes which mainly interferes with the neural integration. In this case, consciousness would be an operation
more than a container where one might put something. It
implies that consciousness could emerge from neural network dynamics as an autonomous system (e.g. brain oscillations), at a different space of action than the original system, but with sub-emergent properties interfering with the
space from which it originates. This work will interpret
certain results in this direction with the intention of expanding this notion in future projects. Also, while the author agrees with many of the conclusions, results and
progresses of the theories mentioned, it is reasonable to
consider that a new focus that integrates all of them is
needed, since they share many similarities and are at the
same time complementary.
The concept of consciousness involves awareness, emotions, subjectivity, intentionality, and attention, among
others. Consciousness is formed from all of these processes
like a differentiated and unified whole, but it is not any of
them. For example, it could be necessary to be aware to
have emotions and subjective experiences, or maybe vice
versa, and we will need them to show intentionality and
attention. We also distinguish that these are different
processes as, for example, awareness and attention; while
understanding all of them as constituent parts of what we
define as consciousness.
Consciousness and its relation with Intelligence
Intelligence is something a little simpler to define; at least,
many scientists think so. Intelligence can be understood as
the ability to solve problems in an efficient way. This
means, the maximization of the positive results in our solution while minimizing the negative impacts, for instance,
waste of time. To do that, other processes as Learning and
Memory are also needed and associated with the definition
of intelligence. Learning may be understood as the process
to gain new knowledge or improve some behavior, while
Memory is the storage of this knowledge. To solve problems efficiently, it is necessary to access a certain memory
that was acquired thanks to a specific learning that will be
modifying again the memory of the system. The more intelligent is the system, the more it learns.
This framework also allows finding mathematical implementations of intelligence and its use in machines. Why
could consciousness be necessary in this context? First,
many different types of intelligences have already recognized, between them, the emotional intelligence that helps
us with social interaction. Secondly, emotions play a crucial role in learning, consolidation of memories, retrieved
memory and cognition in general (Cleeremans, 2011). In
other words, emotions, as subjective experiences, and cognition are deeply related into human intelligence
(Haladjian & Montemayor, 2016).
Finally, emotions and subjective experiences are also
constitutive parts of consciousness, as stated above. This
means consciousness and High-level cognitive abilities are
also related. In consequence, to improve complex intelligence and social behavior in machines, we should understand constitutive parts of consciousness, and how its
processes affect the computability and accuracy of the
most powerful system that we know to date, our brains.
What we already know
First: The Hardware
One simplification about the brain code is to think in the
neuron as the principal informative unit for the brain codification and the spikes as the language of its code. It assumes one kind of codification between 0, without spike,
and 1 with a spike.
However, now it is known that the electrical transmission through gap junctions (Bennett & Zukin, 2004) are not
only between other neurons but also between cells as astrocytes (Alvarez-maubecin, et al. 2000). It is also known
that neuromodulatory substances can reconfigure neuronal
circuits within minutes and hours (Nusbaum, et al. 2001).
In some cases, these substances applied in axon (Bucher,
Thirumalai, & Marder, 2003; Goaillard, et al. 2004) can
spontaneously initiate action potentials in a non-classical
mode of integration. Another research has shown that action potential in some neurons can travel backward from
the axon and soma regions into dendrites (Stuart, Dodt, &
Sakmann, 1993), in contrary to Polarization law. Other
examples are: The functional complexity of dendrites and
the roles they play, the influence of non-neuronal cells in
the action potential propagation (Dupree, et al. 2005) and
axon-glia communication, where the information is transduced through cells that are not neurons (Fields & Stevensgraham, 2005). Astrocytes synapses regulation (Coggan, et
al. 2005; Dani, Chernjavsky, & Smith, 1992; Fields &
Stevens-graham, 2005; Matsui & Jahr, 2004) and its influ-
ence on neurons and so on. All this new knowledge is not
included in the idea of spike codification or is definitely
contrary to it.
These findings have shown us that the neuron is not the
unique codification unit in the brain code and in consequence the spike is not the unique key to understand its
codification (Llinas, et al. 1998). It implies that the binary
code is not enough to understand the brain.
Second: Computability
Turing and Gödel were the first in demonstrating the theorem of incompleteness (Gödel, 1931; Turing, 1937). This
theorem showed that it is not possible to resolve all the
problems with a deductive process or algorithm. From their
demonstrations, one of the consequences suggested is the
potential impossibility to build conscious machines with
algorithms only. Physiological and Psychological data also
seem to show that this is not possible (Haladjian &
Montemayor, 2016). This, together with point one, means
that if we can reach the gap to build a conscious machine,
it will not be with the algorithmic computation. A new
computational framework will be needed.
Third: The two Systems
When we try to explain, on one side, the biological mechanisms in the brain, and on the other, the human psychological behavioral, some paradoxes appear. Some research
and theories have shown that the dynamic of neural systems can be interpreted in a classic probabilities framework
(Pouget, et al. 2000; Quiroga & Panzeri, 2009), like good
estimator and predictor of external stimuli. While other
results, principally from economic psychology, show cognitive fallacies (Gilovich, et al. 2002, Ellsberg, 1961;
Machina, 2009; Moore, 2002). These results are incompatible with the classical probability theories (Pothos &
Busemeyer, 2013).
Hence, these disconnections between the behavior of the
human brain and the cognitive fallacies may help to understand better how the human brain works. How can some
cognitive capabilities, with apparently non-classical dynamic, emerge from apparently classical systems as neural
networks? It is necessary to answer this question as a requirement to start to think in conscious machines.
Forth: Four types of cognition
A recent paper by Shea and Frith (Shea & Frith, 2016)
presents a valuable differentiation between contents and
cognitive processes (computations from some contents to
others). These categories can be conscious or not. From the
experimental evidence and with this framework, they define four cognitive categories (Table 1): i) Type 0 Cognition: non-conscious contents and non-conscious (noncontrolled) processing. Optimal choice is computationlight but learning-heavy (e.g. Motor Control). ii) Type 1
Cognition: conscious contents but automatic and noncontrolled processing (e.g. Fallacy questions). The system
accesses to a wider range of information (Holistic information), however some optimal calculations may become
computationally intractable. iii) Type 2 Cognition: contents
are conscious and the cognitive process is deliberate and
controlled (e.g. Reasoning). It is computation-heavy, learning-light and interferes with Type 0. iv) Type ∞: nonconscious content but conscious process. Yet no further
definition is given as they are not sure if this category
could exist (Supra Reasoning information).
Non conscious Processing
Conscious Processing
Contents non conscious
Type 0 Cognition
Type ∞ Cognition
Contents conscious
Type 1 Cognition
Type 2 Cognition
Table 1. Types of cognition according different contents and
processing exposed above.
Fifth: The perception is a discrete process
The perception needs time. We are conscious of a perceptual content in more o less 200 ms up 400 ms. Indeed, evidence suggests that the conscious perception is a discrete
process (Dehaene & Changeux, 2011; Dehaene, et al.
2014; VanRullen & Koch, 2003) in contrast to a continuous process. The differences between fast time
processing for cognition type 0 (order of 40 ms) and a slow
time processing for type 1 (order of 200 ms) have stimulated the idea of the Two-Stage Model (Herzog, Kammer,
& Scharnowski, 2016). The first stage would be a fast,
continuous and classic probabilistic process (type 0 cognition) while the second stage would correspond to a slow
and discrete process (type 1 and 2 cognition) due to loops
in information processing (Figure 1). In relation with the
discussion above, one may also say that the second stage
would obey non-classical probabilities in some context.
Implications and Paradoxes
The four types of cognitive categories, together with the
other points that were mentioned previously, open the door
for some new interpretations and classifications on the
types of machines (Figure 1):
• The Machine-Machine: It could correspond to the type 0
cognition. The examples are the robots that we are making today with a high learning curve.
• The Conscious Machine: A machine with the ability to be
conscious of its contents but not of the process. In this
case, we will sacrifice accuracy in favor of consciousness. The type 1 cognition requires integration of information and it is a time consuming process. This is why
the accuracy is lost instead of consciousness. However,
is this machine useful for something? Here is the first
Paradox. If we can make a Conscious Machine Type 1
Cognition, this machine will lose the most important
attribute of being a machine: the accuracy in the calculation. In this case, the machine does not work how we
expect: it is a machine without being a machine.
• The Super Machine: A machine with the ability to be
conscious of its contents and own processes. This machine will have subjective experience at the same time
that it will have the option to control the accuracy of its
own rational process. However, the type 2 cognition
needs more time to find accuracy solutions than type 1.
This process also interferes with type 0 cognition, and in
consequence robots will have less precision in motor
tasks. In this case, one observation appears: we cannot
expect an algorithm to control the process of emergency
of consciousness (point 2) and we will not be able to
control it. In other words, the machine can do whatever
it wants; it has the power to do it and the intention to do
it. It could be considered biological new species, more
than a machine. The paradox is that the Super Machine
is not useful anymore unless the machine wants to collaborate with us.
• The Subjective Machine: Machine type ∞ cognition.
Some modern progress in artificial intelligence might fit
in this category (Bringsjord, et al. 2015). The machine
will be conscious of each computation process but not
conscious of the contents, at least, they will not be aware
in the sense that we are. What does this mean? Probably
the Machine will be able to control the computational
processing; however it will not be able to take meanings
of contents from these computations. In other words, the
machine will not be able to “feel something”, to have
“emotions”, the machine will have a type of subjective
experience that we do not know yet. This is probably the
evolution of machine type 0 that we are making these
days but without the intermediate steps type 1 and 2
cognitions. However, if this machine does not feel like a
human, will it be able to understand us? It could be the
kind of machine in science fiction. Again, the machine
could be useless for our purpose and even worst, it could
be contrary to our purposes.
Based on these observations, it is possible to summarize
all the paradoxes in one type: The Conscious Machine is
not a useful machine anymore. In other words, the machine
losses the meaningful characteristics of being a machine: to
resolve problems with accuracy, speed and obedience.
Some Alternatives
It could be interesting for Artificial Intelligence to make
Machine type 1 or type 2 cognition. Based on the principles exposed above from biological systems, the requirements are: i) new non-binary codification principles,
ii) no deterministic computation, iii) two different systems
of integration of information, which should be related as
the non-classical probabilistic system emerging from the
classical probabilistic system, iv) enough computation capability to support the first time-consuming system and the
second non-classical system. Is it possible to respond to
these requirements with the actual framework in Artificial
Intelligence?
Figure 1. Types of cognition related with potential Systems of
processing in the brain.
Quantum computation is one of the alternatives due to
the increase of computational capability thanks to qubits,
non-classical probabilities and non-classical computability.
Some alternative tools aiming to implement Quantum
computation in analogy with brain process, might be new
approaches of contextuality from Quantum Physics to Psychology (Abramsky, 2015), like contextual semantics, natural language, and semantics of computation (Kartsaklis, et
al. 2013). Starting from the explanation of the cognitive
fallacies above; concept combinations, human judgments
and decision making under uncertainty can be better modeled and described by using the mathematical formalism of
quantum theory (Aerts, et al. 2013; Busemeyer, et al. 2011;
Sozzo, 2015; Wang, et al. 2014). This emerging domain
has been called "quantum cognition" (Bruza, Wang, &
Busemeyer, 2015) and these models may help us understand the connection between new computational frameworks, psychophysics, recent advances in neuroscience on
conscious perception in the brain, and also how to make
conscious machines.
Conclusions
These ideas are part of a work in progress. This work has
developed a hypothesis based on previous works: we may
build conscious machines but with computational costs.
These costs reflect some paradoxes when we examine the
requirements to build a conscious machine: it will be possible to make a conscious machine without good rational
abilities or a good rational conscious machine with more
time to respond accuracy. In both cases, the price of nonobedience will be paid. The next step in this research will
be to quantify the decrease in computational efficiency,
time-consuming of these processes and the mathematical
formulation of these propositions.
Acknowledgments
This work was partially supported by CONICYT. The author appreciates valuable comments of Marlène Morvan
and Miriam Potter in the final version.
References
Abramsky, S. 2015. Contextual Semantics: From Quantum Mechanics to Logic, Databases, Constraints, and Complexity. In
Contextuality from Quantum Physics to Psychology (Vol. 6, pp.
23–50).
Aerts, D., Gabora, L., & Sozzo, S. 2013. Concepts and their dynamics: a quantum-theoretic modeling of human thought. Topics
in Cognitive Science, 5(4), 737–72.
Aerts, D., & Sozzo, S. 2011. Quantum structure in cognition:
Why and how concepts are entangled. In Quantum Interaction.
Lecture Notes in Computer Science (Vol. 7052, pp. 116–127).
Alvarez-maubecin, V., Garcı, F., Williams, J. T., & Bockstaele,
E. J. Van. 2000. Functional Coupling between Neurons and Glia.
The Journal of Neuroscience, 20(11), 4091–4098.
Baars, B. J. 2005. Global workspace theory of consciousness:
toward a cognitive neuroscience of human experience. Progress
in Brain Research, 150, 45–53.
Barron, A. B., & Klein, C. 2016. What insects can tell us about
the origins of consciousness. Proceedings of the National Academy of Sciences of the United States of America, 113(18), 4900–8.
Bennett, M. V. L., & Zukin, R. S. 2004. Neuronal Synchronization in the Mammalian Brain. Neuron, 41, 495–511.
Bringsjord, S., Licato, J., Sundar, N., Rikhiya, G., & Atriya, G.
2015. Real Robots that Pass Human Tests of Self-Consciousness.
In Proceeding of the 24th IEEE International Symposium on Robot and Human Interactive Communication (pp. 498–504).
Bruza, P. D., Wang, Z., & Busemeyer, J. R. 2015. Quantum cognition: a new theoretical approach to psychology. Trends in Cognitive Sciences, 19(7), 383–93.
Bucher, D., Thirumalai, V., & Marder, E. 2003. Axonal Dopamine Receptors Activate Peripheral Spike Initiation in a Stomatogastric Motor Neuron. The Journal of Neuroscience, 23(17),
6866–6875.
Busemeyer, J. R., Pothos, E. M., Franco, R., & Trueblood, J. S.
2011. A quantum theoretical explanation for probability judgment
errors. Psychological Review, 118(2), 193–218.
Chalmers, D. 1995. The puzzle of conscious experience. Scientific American, 273(6), 80–86.
Cleeremans, A. 2011. The Radical Plasticity Thesis: How the
Brain Learns to be Conscious. Frontiers in Psychology, 2(May),
86.
Coggan, J. S., Bartol, T. M., Esquenazi, E., Stiles, J. R., Lamont,
S., Martone, M. E., … Sejnowski, T. J. 2005. Evidence for Ectopic Neurotransmission at a Neuronal Synapse. Science, 309(5733),
446–451.
Dani, J. W., Chernjavsky, A., & Smith, S. J. 1992. Neuronal activity triggers calcium waves in hippocampal astrocyte networks.
Neuron, 8(3), 429–440.
Dehaene, S., & Changeux, J.-P. 2011. Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200–27.
Dehaene, S., Charles, L., King, J.-R., & Marti, S. 2014. Toward a
computational theory of conscious processing. Current Opinion
in Neurobiology, 25(1947), 76–84.
Dupree, J. L., Mason, J. L., Marcus, J. R., Stull, M., Levinson, R.,
Matsushima, G. K., & Popko, B. 2005. Oligodendrocytes assist in
the maintenance of sodium channel clusters independent of the
myelin sheath. Neuron Glia Biology, 1(03), 179–192.
Fields, R. D., & Stevens-graham, B. 2005. New Insights into
Neuron-Glia Communication. Science, 298(5593), 556–562.
Gilovich, T., Griffin, D., & Kahneman, D. 2002. Heuristics and
biases: The psychology of intuitive judgment. Cambridge University Press.
Goaillard, J.-M., Schulz, D. J., Kilman, V. L., & Marder, E. 2004.
Octopamine modulates the axons of modulatory projection neurons. The Journal of Neuroscience, 24(32), 7063–73.
Gödel, K. 1931. Über formal unentscheidbare Sätze der Principia
Mathematica und verwandter Systeme I. Monatshefte Für Mathematik Und Physik, 38-38(1), 173–198.
Haladjian, H. H., & Montemayor, C. 2016. Artificial consciousness and the consciousness-attention dissociation. Consciousness
and Cognition, 45, 210–225.
Herzog, M. H., Kammer, T., & Scharnowski, F. 2016. Time Slices: What Is the Duration of a Percept? PLoS Biology, 14(4),
e1002433.
Kartsaklis, D., Sadrzadeh, M., Pulman, S., & Coecke, B. 2014.
Reasoning about meaning in natural language with compact
closed categories and frobenius algebras. Logic and Algebraic
Structures in Quantum Computing and Information, (pp 199–
222). Cambridge University Press.
Llinas, R., Ribary, U., Contreras, D., & Pedroarena, C. 1998. The
neuronal basis for consciousness. Philosophical Transactions of
the Royal Society of London B, 353, 1841–1849.
Lycan, W. G., & Dennett, D. C. 1993. Consciousness Explained.
The Philosophical Review, 102(3), 424.
Matsui, K., & Jahr, C. E. 2004. Differential control of synaptic
and ectopic vesicular release of glutamate. The Journal of Neuroscience, 24(41), 8932–9.
Moravec, H. P. 1988. Mind children : the future of robot and
human intelligence (p. 214). Harvard University Press.
Nusbaum, M. P., Blitz, D. M., Swensen, A. M., Wood, D., &
Marder, E. 2001. The roles of co-transmission in neural network
modulation. Trends in Neurosciences, 24(3), 146–154.
Pothos, E. M., & Busemeyer, J. R. 2013. Can quantum probability provide a new direction for cognitive modeling? The Behavioral and Brain Sciences, 36(3), 255–74.
Pouget, A., Dayan, P., & Zemel, R. 2000. Information processing
with population codes. Nature Reviews Neuroscience, 1(11), 125–
132.
Quian Quiroga, R., & Panzeri, S. 2009. Extracting information
from neuronal populations: information theory and decoding
approaches. Nature Reviews. Neuroscience, 10(3), 173–85.
Shea, N., & Frith, C. D. 2016. Dual-process theories and consciousness: the case for “Type Zero” cognition. Neuroscience of
Consciousness, 2016(1), 1–10.
Sozzo, S. 2015. Conjunction and negation of natural concepts: A
quantum-theoretic modeling. Journal of Mathematical Psychology, 66, 83–102.
Stuart, G. J., Dodt, H. U., & Sakmann, B. 1993. Patch-clamp
recordings from the soma and dendrites of neurons in brain slices
using infrared video microscopy. Pflugers Archv European Journal of Physiology, 423, 511–518.
Tononi, G., Boly, M., Massimini, M., & Koch, C. 2016. Integrated information theory: from consciousness to its physical
substrate. Nature Reviews. Neuroscience, 17(7), 450–61.
Tononi, G., & Koch, C. 2008. The neural correlates of consciousness: an update. Annals of the New York Academy of Sciences,
1124, 239–61.
Turing, A. M. 1937. On Computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London
Mathematical Society, 42(1), 230–265.
VanRullen, R., & Koch, C. 2003. Is perception discrete or continuous? Trends in Cognitive Sciences, 7(5), 207–213.
Varela, F., Lachaux, J., Rodriguez, E., & Martinerie, J. 2001. The
brainweb: phase synchronization and large-scale integration.
Nature Reviews Neuroscience, 2(April), 229–239.
Wang, Z., Busemeyer, J. R., Atmanspacher, H., & Pothos, E. M.
2013. The potential of using quantum theory to build models of
cognition. Topics in Cognitive Science, 5(4), 672–88.
Wang, Z., Solloway, T., Shiffrin, R. M., & Busemeyer, J. R.
2014. Context effects produced by question orders reveal quantum nature of human judgments. Proceedings of the National
Academy of Sciences of the United States of America, 111(26),
9431–6.
| 2cs.AI
|
1
Capacity of a Full-Duplex Wirelessly Powered
Communication System with Self-Interference
and Processing Cost
arXiv:1710.02273v1 [cs.IT] 6 Oct 2017
Ivana Nikoloska, Student Member, IEEE, Nikola Zlatanov, Member, IEEE, and
Zoran Hadzi-Velkov, Senior Member, IEEE
Abstract
In this paper, we investigate the capacity of a point-to-point, full-duplex (FD), wirelessly powered
communication system impaired by self-interference. This system is comprised of an energy transmitter
(ET) and an energy harvesting user (EHU), both operating in a FD mode. The ET transmits energy
towards the EHU. The EHU harvests this energy and uses it to transmit information back to the ET. As a
result of the FD mode, both nodes are affected by self-interference. The self-interference has a different
effect at the two nodes: it impairs the decoding of the received signal at the ET, however, it provides an
additional source of energy for the EHU. This paper derives the capacity of this communication system
assuming a processing cost at the EHU and additive white Gaussian noise channel with block fading.
Thereby, we show that the capacity achieving scheme is relatively simple and therefore applicable to
devices with limited resources. Moreover, our numerical results show significant improvements in terms
of data rate when the capacity achieving strategy is employed compared to half-duplex transmission.
Moreover, we show the positive and negative effects of the self-interference at the EHU and the ET,
respectively. Furthermore, we show the crippling effect of the processing cost and demonstrate that
failing to take it into consideration gives a false impression in terms of achievable rate.
I. I NTRODUCTION
In recent years, wireless communication, among many others, has been highly affected by
emerging green technologies, such as green radio communications and energy harvesting (EH).
Ivana Nikoloska and Nikola Zlatanov are with the Department of Electrical and Computer Systems Engineering, Monash
University, Melbourne, Australia. Emails: [email protected], [email protected]
Zoran Hadzi-Velkov is with the Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius
University, Skopje, Macedonia. Email: [email protected]
2
While the former aims at minimizing the use of precious radio resources, the latter relies on
harvesting energy from renewable and environmentally friendly sources such as, solar, thermal,
vibration or wind, [1], [2], and using this energy for the transmission of information. Thereby, EH
promises a perpetual operation of communication networks. Moreover, EH completely eliminates
the need for battery replacement and/or using power cords, making it a highly convenient option
for communication networks with nodes which are hard or dangerous to reach. This is not to say
that EH communication networks lack their own set of challenges. For example, ambient energy
sources are often intermittent and scarce, which can put in danger the continuous reliability of the
communication session. Possible solution to this problem is harvesting of radio frequency (RF)
energy from a dedicated energy transmitter, which gives rise to the so called wireless powered
communication networks (WPCNs) [3], [4].
EH technology and WPCNs are also quite versatile, and have been studied in many different
contexts. Specifically, the capacity of the EH additive white Gaussian noise (AWGN) channel
where the transmitter has no battery for energy storage, was studied in [5]. Transmitters with
finite-size batteries have been considered in [6], [7]. In [6], authors use stationarity and ergodicity
to derive a series of computable capacity upper and lower bounds for the general discrete EH
channel. Authors in [7] investigate the capacity of EH binary symmetric channels with deterministic energy arrival processes. Both small and large battery regimes have been investigated
in [8], [9]. The capacity of a sensor node with an infinite-size battery has been investigated in
[10]. Infinite-size batteries have also been adopted in [11], [12]. In [11], the authors derive the
capacity of the EH AWGN channel without processing cost or storage inefficiencies, while the
authors in [12] take into account the processing cost as well as the energy storage inefficiencies.
The authors in [13] provide two capacity achieving schemes for the AWGN channel with random
energy arrivals at the transmitter. The outage capacity of a practical EH circuit model with a
primary and secondary energy storage devices has been studied in [14]. The authors in [15]
derive the minimum transmission outage probability for delay-limited information transfer and
the maximum ergodic capacity for delay unlimited information transfer versus the maximum
average energy harvested at the receiver. The capacity of the Gaussian multiple access channel
(MAC) has been derived in [16], [17]. Relaying (i.e., cooperative) networks with wireless energy
transfer have also been extensively analysed due to their ability to guarantee longer distance
communication than classical point-to-point EH links [18], [19]. Information theory has also
been paired with queuing theory in order to compute the capacity of EH links [20].
3
Most of the considered EH system models in the literature are assumed to operate in the halfduplex (HD) mode. This is logical since up until recently it was assumed that simultaneous
reception and transmission on the same frequency, i.e., full-duplex (FD) communication is
impossible. However, recent developments have shown this assumption to be false, i.e., have
shown that FD communication is possible. To accomplish FD communication, a radio has to
completely cancel or significantly reduce the inevitable self-interference. Otherwise, the selfinterference increases the amount of noise and thereby reduces the achievable rate. Research
efforts have made significant progress when it comes to tackling this problem and both active
and passive cancelation methods have been proposed. The former, refers to techniques which
introduce signal attenuation when the signal propagates from the transmit antenna to the receiver
one [21], [22]. The latter exploits the knowledge of the transmit symbols by the FD node in
order to cancel the self-interference [23]-[25]. Combinations of both methods have also been
considered [26].
Motivated by the idea that the role of the self-interference can be redefined in WPCNs, such
as in [27]-[30], as well as by the advances in FD communication, in this work, we investigate the
capacity of a FD wirelessly powered communication system comprised of an energy transmitter
(ET) and an energy harvesting user (EHU) that operate in an AWGN block-fading environment.
In this system, the ET sends RF energy to the EHU, whereas, the EHU harvests this energy and
uses it to transmit information back to the ET. Both the ET and the EHU work in the FD mode,
hence, both nodes transmit and receive RF signals in the same frequency band and at the same
time. As a result, both are affected by self-interference. The self-interference has opposite effects
at the ET and the EHU. Specifically, the self-interference signal has a negative effect at the ET
since it hinders the decoding of the information signal received from the EHU. However, at the
EHU, the self-interference signal has a positive effect since it increases the amount of energy
that can be harvested by the EHU. In this paper, we derive the capacity of this system model.
Thereby, we prove that the capacity achieving distribution at the EHU is Gaussian, whereas,
the input distribution at the ET is degenerate and takes only one value with probability one.
This allows simple capacity achieving schemes to be developed, which are applicable to nodes
with limited resources such as EH devices. Our numerical results show significant gains in term
of data rate when the capacity achieving scheme is employed, as opposed to a HD benchmark
scheme.
The results and schemes presented in this paper provide clear guidelines for increasing the
4
Fig. 1. System model.
data rates of future wirelessly powered communication systems. Moreover, due to their relative
simplicity, the proposed schemes are applicable to future Internet of Things systems.
The rest of the paper is organized as follows. Section II provides the system and channel
models as well as the energy harvesting policy. Section III presents the capacity as well as the
optimal input probability distributions at the EHU and the ET. Section IV provides the converse
as well as the achievability of the capacity. In Section V, we provide numerical results and a
short conclusion concludes the paper in Section VI. Proofs of theorems are provided in the
Appendix.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
We consider a system model comprised of an EHU and an ET, c.f. Fig. 1. The ET transmits
RF energy to the EHU and simultaneously receives information from the EHU. On the other
hand, the EHU harvests the energy transmitted from the ET and uses it to transmit information
back to the ET.
In order to improve the spectral efficiency of the considered system, both the EHU and the
ET are assumed to operate in the FD mode, i.e., both nodes transmit and receive RF signals
simultaneously and in the same frequency band. Thereby, the EHU receives energy signals from
5
the ET and simultaneously transmits information signals to the ET in the same frequency band.
Similarly, the ET transmits energy signals to the EHU and simultaneously receives information
signals from the EHU in the same frequency band. Due to the FD mode of operation, both the
EHU and the ET are impaired by self-interference. The self-interference has opposite effects at
the ET and the EHU. More precisely, the self-interference signal has a negative effect at the ET
since it hinders the decoding of the information signal received from the EHU. As a result, the
ET should be designed with a self-interference suppression apparatus, which can suppress the
self-interference at the ET and thereby improve the decoding of the desired signal received from
the EHU. On the other hand, at the EHU, the self-interference signal has a positive effect since
it increases the amount of energy that can be harvested by the EHU. Hence, the EHU should be
designed without a self-interference suppression apparatus in order for the energy contained in
the self-interference signal of the EHU to be fully harvested by the EHU, i.e., the EHU should
perform energy recycling as proposed in [27].
A. Channel Model
We assume that both the EHU and the ET are impaired by AWGNs, with variances σ12 and
σ22 , respectively. Let H12i and H21i model the fading channel gains of the EHU-ET and ET-EHU
channels in channel use i, respectively. We assume that the channel gains, H12i and H21i , follow
a block-fading model, i.e., they remain constant during all channel uses in one block, but change
from one block to the next, where each block consists of (infinitely) many channel uses. Due
to the FD mode of operation, the channel gains H12i and H21i are assumed to be identical, i.e.,
H12i = H21i = Hi . We assume that the EHU and the ET are able to estimate the channel gain
Hi perfectly.
In the i-th channel use, let the transmit symbols at the EHU and the ET be modeled as
random variables (RVs), denoted by X1i and X2i , respectively. Moreover, in channel use i, let
the received symbols at the EHU and the ET be modeled as RVs, denoted by Y1i and Y2i ,
respectively. Furthermore, in channel use i, let the RVs modeling the AWGNs at the EHU and
the ET be denoted by N1i and N2i , respectively, and let the RVs modeling the additive selfinterferences at the EHU and the ET by denoted by I1i and I2i , respectively. As a result, Y1i
6
and Y2i can be written as
Y1i = Hi X2i + I1i + N1i ,
(1)
Y2i = Hi X1i + I2i + N2i .
(2)
We adopt the first-order approximation for the self-interference, justified in [31] and [32], where
the authors show that the higher-order terms carry significantly less energy than the first-order
terms. Thereby, we only consider the linear components of the self-interference and, similar to
[32], approximate I1i and I2i as
I1i = G̃1i X1i ,
(3)
I2i = G̃2i X2i ,
(4)
where G̃1i and G̃2i denote the self-interference channel gains in channel use i. The self-interference
channel gains, G̃1i and G̃2i , are time-varying and the statistical properties of these variations are
dependent of the hardware configuration and the adopted self-interference suppression scheme. In
order for us to obtain the worst-case performance in terms of self-interference, we assume that G̃1i
and G̃2i are independent and identically distributed (i.i.d.) Gaussian RVs across different channels
uses, which is a direct consequence of the fact that the Gaussian distribution has the largest
entropy under a second moment constraint, see [33]. Hence, we assume that G̃1i ∼ N {ḡ1 , α1 }
and G̃2i ∼ N {ḡ2 , α2 }, where N {ḡ, α} denotes a Gaussian distribution with mean ḡ and variance
α.
Inserting (3) and (4) into (1) and (2), respectively, we obtain
Y1i = Hi X2i + G̃1i X1i + N1i ,
(5)
Y2i = Hi X1i + G̃2i X2i + N2i .
(6)
Now, since G̃1i and G̃2i can be written equivalently as G̃1i = G1i + ḡ1 and G̃2i = G2i + ḡ2 ,
respectively, where G1i ∼ N {0, α1 } and G2i ∼ N {0, α2 }, without loss of generality, (5) and
(6) can also be written equivalently as
Y1i = Hi X2i + ḡ1 X1i + G1i X1i + N1i
(7)
Y2i = Hi X1i + ḡ2 X2i + G2i X2i + N2i ,
(8)
and
7
Fig. 2. Block diagram of the system model.
respectively.
Since the ET knows X2i in channel use i, and since given sufficient time it can always estimate
the mean of its self-interference channel, ḡ2 , the ET can remove ḡ2 X2i from its received symbol
Y2i , given by (8), and thereby reduce its self-interference. In this way, the ET obtains a new
received symbol, denoted again by Y2i , as
Y2i = Hi X1i + G2i X2i + N2i .
(9)
Note that since G2i in (9) changes i.i.d. randomly from one channel use to the next, the ET
cannot estimate and remove G2i X2i from its received symbol. Thus, G2i X2i in (9) is the residual
self-interference at the ET. On the other hand, since the EHU benefits from the self-interference,
it does not remove ḡ1 X1i from its received symbol Y1i , given by (7), in order to have a selfinterference signal with a much higher energy, which it can then harvest. Hence, the received
symbol at the EHU is given by (7).
In this paper, we investigate the capacity of a channel given by the input-output relations
in (7) and (9), where we are only interested in the mutual-information between X1n and Y2n
subject to an average power constraint on X2n , where the notation an is used to denote the
vector an = (a1 , a2 , ..., an ). An equivalent block diagram of the considered system model is
presented in Fig 2.
B. Energy Harvesting Model
We assume that the energy harvested by the EHU in channel use i is given by [27]
Ein,i = η(Hi X2i + ḡ1 X1i + G1i X1i )2 ,
(10)
8
where 0 < η < 1 is the energy harvesting inefficiency coefficient. The EHU stores Ein,i in its
battery, which is assumed to be infinitely large. Let Bi denote the amount of harvested energy
in the battery of the EHU in the i-th channel use. Moreover, let Eout,i be the extracted energy
from the battery in the i-th channel use. Then, Bi , can be written as
Bi = Bi−1 + Ein,i − Eout,i .
(11)
Since in channel use i the EHU cannot extract more energy than the amount of energy stored
in its battery during channel use i − 1, the extracted energy from the battery in channel use i,
Eout,i , can be obtained as
2
Eout,i = min{Bi−1 , X1i
+ Pp },
(12)
2
is the transmit energy of the desired transmit symbol in channel use i, X1i , and Pp is
where X1i
the processing cost of the EHU. The processing cost, Pp , is the energy spent for processing and
the energy spent due to the inefficiency and the power consumption of the electrical components
in the electrical circuit such as AC/DC convertors and RF amplifiers.
Remark 1: Note that the ET also requires energy for processing. However, the ET is assumed
to be equipped with a conventional power source which is always capable of providing the
processing energy without interfering with the energy required for transmission.
We now use the results in [11] and [33], where it was proven that if the total number of
channel uses satisfies n → ∞, the battery of the EHU has an unlimited storage capacity, and
2
E{Ein,i } ≥ E{X1i
} + Pp
(13)
holds, where E{·} denotes statistical expectation, then the number of channel uses in which the
extracted energy from the battery is insufficient and thereby Eout,i = Bi−1 holds is negligible
compared to the number of channel uses when the extracted energy from the battery is sufficient
2
and thereby Eout,i = X1i
+ Pp holds. In other words, when the above three conditions hold, in
almost all channel uses there will be enough energy to be extracted from the EHU’s battery for
2
both processing, Pp , and for the transmission of the desired transmit symbol X1i , X1i
.
III. C APACITY
The channel in Fig. 2, modeled by (7) and (9), is a discrete-time channel with inputs X1 and
X2 , and outputs Y2 and Y1 . Furthermore, the probability of observing Y2 or Y1 is only dependent
on the input X1 and/or X2 at channel use i and is conditionally independent of previous channel
9
inputs and outputs, making the considered channel a discrete-time memoryless channel [34]. For
this channel, we propose the following theorem.
Theorem 1: Assuming that the average power constraint at the ET is PET , the capacity of the
considered EHU-ET communication channel is given by
X X
I(X1 ; Y2 |X2 = x2 , H = h)p(x2 |h)p(h)
C=
max
p(x1 |x2 ,h),p(x2 |h)
x2 ∈X2 h∈H
Subject to
C1 :
X X
x22 p(x2 |h)p(h) ≤ PET
x2 ∈X2 h∈H
Z
X X
C2 :
(x21 + Pp )p(x1 |x2 , h)p(x2 |h)p(h)dx1 ≤
x1 x ∈X h∈H
2
2
Z
X X
Ein p(x1 |x2 , h)p(x2 |h)p(h)dx1
x1 x ∈X h∈H
2
2
C3 :
X
p(x2 |h) = 1
x2 ∈X2
Z
p(x1 |x2 , h)dx1 = 1,
C4 :
(14)
x1
where I(; |) denotes the conditional mutual information between X1 and Y2 when the EHU
knows the transmit symbols of the ET and both the EHU and the ET have full channel state
information (CSI). In (14), lower case letters x2 and h represent realizations of the random
variables X2 and H and their support sets are denoted by X2 and H, respectively. Constraint C1
in (14) constrains the average transmit power of the ET to PET , and C2 is due to (13), i.e., due
to the fact that EHU has to have harvested enough energy for both processing and transmission
of symbol X1 . The maximum in the objective function is taken over all possible conditional
probability distributions of x1 and x2 , given by p(x1 |x2 , h) and p(x2 |h), respectively.
Proof: The proof is in two parts. In Subsection IV-A we prove the converse and in Subsection IV-B we provide the achievability of the capacity given by (14).
A. Optimal Input Distribution and Simplified Capacity Expression
The optimal input distributions at the EHU and the ET which achieve the capacity in (14)
and the resulting capacity expressions are provided by the following theorem.
10
Theorem 2: We have two cases for the channel capacity and the optimal input distributions.
Case 1: If
!
√
h2 PEHU
PET , h
p(h) = λ1 PET + µ1
log 1 +
2
+
P
2
σ
ET α2
2
h∈H
X1
+ λ2
1 − η(ḡ12 + α1 )
X
(15)
!
p
X
PET , h p(h) − ηPET
h2 p(h)
PEHU
h∈H
h∈H
holds, where λ1 , λ2 , and µ1 are the Lagrangian multipliers associated with constraints C1, C2,
and C3 in (14), respectively, the optimal input distribution at the EHU is zero-mean Gaussian
√
√
PET , h , i.e., p(x1 |x2 , h) ∼ N 0, PEHU
PET , h , where
with variance PEHU
+
p
1
σ22 + PET α2
PEHU
PET , h =
−
,
(16)
λ2 (1 − η(ḡ12 + α1 ))
h2
and λ2 and is chosen such that
p
X
X
PEHU
1 − η(ḡ12 + α1 )
PET , h p(h) + Pp = ηPET
h2 p(h)
h∈H
(17)
h∈H
holds.
On the other hand, the optimal input distribution at the ET is given by
p
p(x2 |h) = δ x2 − PET .
(18)
Finally, the capacity for this case is given by
!
√
PET , h
h2 PEHU
C=
log 1 +
p(h).
2
σ22 + PET α2
h∈H
X1
(19)
Case 2: If (15) does not hold, the optimal input distribution at the EHU is zero-mean Gaussian
with variance PEHU (x0 (h), h), i.e, p(x1 |x2 , h) ∼ N (0, PEHU (x0 (h), h)), where
+
1
σ22 + x20 (h)α2
PEHU (x0 (h), h) =
−
,
λ2 (1 − η(ḡ12 + α1 ))
h2
(20)
where x0 (h) is given by
v
!
+
u
h2
u
2 ln 2 (λ1 −λ2 ηh2 )
−1
2
u
λ2 (1−η(ḡ1 +α1 ))α2
u h2 W
2 −λ ) σ 2 +µ
2
ln
2
1+(λ
ηh
(
2
2
1 α2
1)
u
e
σ
x0 (h) = u
− 2 ,
u
2
2
2
u 2 ln 2((λ1 − λ2 ηh )h − λ2 (1 − η(ḡ1 + α1 ))α2 ) α2
t
(21)
11
and λ2 is chosen such that
1 − η(ḡ12 + α1 )
X
PEHU (x0 (h), h)p(h) + Pp = η
h∈H
X
x20 (h)h2 p(h)
(22)
h∈H
holds. In (21), W (.) denotes the Lambert W function.
On the other hand, the optimal input distribution at the ET is given by
p(x2 |h) = δ(x2 − x0 (h)).
Finally, the channel capacity for this case is given by
X1
h2 PEHU (x0 (h), h)
p(h).
C=
log 1 +
2
σ22 + x20 (h)α2
h∈H
(23)
(24)
Proof: Please refer to Appendix A.
Essentially, depending on the average fading power, we have two cases. Case 1 is when the
√
ET transmits the symbol PET in each channel use and in all fading blocks. Whereas, the
EHU transmits a Gaussian codeword and adapts its power to the fading realisation in each
fading block, i.e., it performs classical waterfiling. Thereby, the stronger the fading channel, h,
the stronger the average transmit power of the EHU during that fading realisation. Conversely,
the weaker the fading channel, h, the lower the average transmit power of the EHU during
that fading realisation. In cases when the fading channel is too weak, the EHU remains silent
during that fading realisation and only harvests the energy transmitted by the ET. In Case 2, the
ET transmits the same symbol x0 (h) during all channel uses of the fading block with fading
realisation h. Since now x0 (h) depends on the fading, as shown in Fig. 3, the ET transmits
different symbols in fading blocks with different fading realisations. In particular, if there is a
strong fading gain h, x0 (h) is higher. If h is low so is x0 (h). When h is very low, then x0 (h) = 0,
i.e., the ET becomes silent during that fading block. On the other hand, the EHU in this case
transmits a Gaussian codeword with an average power PEHU (x0 (h), h), where PEHU (x0 (h), h)
is adapted to the fading gain in that fading block. If the fading gain is strong, PEHU (x0 (h), h)
is higher and if the fading gain is weak, PEHU (x0 (h), h) is low. If the fading gain is too low,
PEHU (x0 (h), h) = 0, i.e., the EHU becomes silent.
Also, since λ2 is chosen such that constraint C2 in (14) holds, the expressions in (16), (20), and
(21) are dependent on the processing cost Pp . As a result, the optimal solution takes into account
Pp , which hinders the system’s performance, i.e., a larger Pp results in a lower capacity. This is
particularly important in networks constrained by their power supply, since their performance in
12
14
12
x 0 (h)
10
8
6
4
2
0
0.5
1
1.5
2
2.5
3
x10-4
h2
Fig. 3. ET’s transmit symbol x0 (h), given by (21), as a function of the fading realisation h.
terms of achievable rates and/or lifetime can be easily overestimated by ignoring the processing
cost, as shown in the numerical examples, cf. Fig. 8.
B. Special Case of Rayleigh Fading
When the channel fading gain follows a Rayleigh probability distribution, i.e., their probability
density function is given by
pH (h) =
h − Ωh2
e h,
2Ωh
(25)
the channel capacity in (19) has a closed-form solution as
2 +α ))(σ 2 +P
λ (1−η(ḡ1
1
ET α2 )
2
1
λ2 (1 − η(ḡ12 + α1 ))(σ22 + PET α2 )
1
− 2
Ωh
C=
E1
+e
ln
,
ln 2
Ωh
Ωh
(26)
where λ2 can be found from the following equation
2 +α ))(σ 2 +P
λ (1−η(ḡ1
1
ET α2 )
2
1
− 2
Ωh
e
− (σ22 + PET α2 )E1
λ2 (1 − η(ḡ12 + α1 ))
ηPET Ωh − Pp
=
,
(1 − η(ḡ12 + α1 ))
λ2 (1 − η(ḡ12 + α1 ))(σ22 + PET α2 )
Ωh
(27)
13
where E1 denotes the exponential integral function given by E1 (x) =
R∞
x
e−t
dt.
t
When condition
(15) does not hold, the integral which gives the channel capacity can be evaluated numerically
using software such as Mathematica.
Corollary 1: When the channel is not impaired by fading, the capacity can be derived as
follows
√
1
PEHU ( PET )
C = log 1 + 2
,
2
σ2 + PET α2
(28)
ηPET
.
1 − η(ḡ12 + α1 )
(29)
√
where PEHU ( PET ) is given by
PEHU (
p
PET ) =
The capacity in (28) is achieved when X1 follows a Gaussian probability distribution as
√
p(x1 |x2 ) ∼ N 0, PEHU
PET and the probability distribution of X2 is degenerate and given
√
by p(x2 ) = δ x2 − PET .
IV. C ONVERSE AND ACHIEVABILITY OF THE C HANNEL C APACITY
In this section, we prove the converse and the achievability of the channel capacity.
A. Converse of the Channel Capacity
Let W be the message that the EHU wants to transmit to the ET. Let this message be uniformly
selected at random from the message set {1, 2, ..., 2nR }, where n → ∞ is the number of channel
uses that will be used for transmitting W from the EHU to the ET, and R denotes the data rate
of message W . We assume a priori knowledge of the CSI, i.e., Hi is known for i = 1....n before
the start of the communication at both nodes. Then, we have
nR ≤ H(W |H n ) = H(W |H n ) − H(W |H n , Y2n , X2n ) + H(W |H n , Y2n , X2n )
= I(W ; Y2n , X2n |H n ) + H(W |H n , Y2n , X2n ),
(30)
which follows from the definition of the mutual information. By Fano’s inequality we have
(a)
H(W |H n , Y2n , X2n ) ≤ H(W |Y2n , H n ) ≤ Pe nR + 1,
(31)
14
where (a) follows from the fact that conditioning reduces entropy and Pe is the average probability of error of the message W . Inserting (31) into (30), and dividing both sides by n, we
have
R≤
1
I(W ; Y2n , X2n |H n ) + Pe R + 1/n.
n
(32)
Assuming that n → ∞, and assuming that Pe → 0 as n → ∞, (32) can be written as
R≤
1
I(W ; Y2n , X2n |H n ).
n
(33)
We represent the right hand side of (33) as
I(W ; Y2n , X2n |H n ) = I(W ; Y2n |X2n , H n ) + I(W ; X2n |H n ).
(34)
Now, since the transmit message W is uniformly drawn from the message set {1, 2, ..., 2nR } at
the EHU, and the ET does not know which message the EHU transmits, the following holds
I(W ; X2n |H n ) = 0.
(35)
Inserting (35) into (34), we have
I(W ; Y2n , X2n |H n ) = I(W ; Y2n |X2n , H n )
(a)
≤
(b)
=
(c)
≤
(d)
≤
n
X
i=1
n
X
i=1
n
X
i=1
n
X
i=1
I(W ; Y2i |Y2i−1 , X2n , H n )
H(Y2i |Y2i−1 , X2n , H n )
−
n
X
H(Y2i |Y2i−1 , X2n , H n , W )
i=1
H(Y2i |X2n , H n ) −
H(Y2i |X2n , H n ) −
n
X
i=1
n
X
H(Y2i |Y2i−1 , X2n , H n , W )
H(Y2i |Y2i−1 , X1i , X2n , H n , W ),
(36)
i=1
where (a) follows from the fact that the entropy of a collection of random variables is less
than the sum of their individual entropies [34], (b) is a consequence of the chain rule, and (c)
and (d) follow from the fact that conditioning reduces entropy. Now, due to the memoryless
assumption, it easy to see that Y2i is independent of all elements in the vectors X2n and H n
except the elements X2i and Hi . Thereby, the following holds
H(Y2i |X2n , H n ) = H(Y2i |X2i , Hi ).
(37)
15
Moreover, it is easy to see that given X1i and X2i , Y2i is independent of Y2i−1 , of all elements
in the vector X2n except the element X2i , of all the elements of the vector H n except Hi , and
of W . Thereby, the following holds
H(Y2i |Y2i−1 , X1i , X2n , H n , W ) = H(Y2i |X1i , X2i , Hi ).
(38)
Inserting (37) and (38) into (36), we obtain
I(W ; Y2n , X2n |H n )
≤
=
n
X
i=1
n
X
H(Y2i |X2i , Hi ) −
n
X
H(Y2i |X1i , X2i , Hi )
i=1
I(X1i ; Y2i |X2i , Hi ).
(39)
i=1
Now, inserting (39) into (33), we have
n
R≤
1X
I(X1i ; Y2i |X2i , Hi ) = I(X1 ; Y2 |X2 , H).
n i=1
(40)
Hence, an upper bound on the capacity is given by (40) when no additional constraints on X1
and X2 exist. However, in our case, we have two constraints on X1 and X2 . One constraint is that
E{X22 } ≤ PET , expressed by C1 in (14). The other constraint is given by (13), expressed by C2 in
(14), which limits the average power of the EHU to be less than the maximum average harvested
power. Constraints C3 and C4 in (14) come from the definitions of probability distributions.
Hence, by inserting C1, C2, C3, and C4 from (14) into (40) and maximizing with respect to
p(x1 , x2 |h) = p(x1 |x2 , h)p(x2 |h), we obtain that the capacity is upper bounded by (14). This
proves the converse. In Subsection IV-B, we prove that this upper bound can be achieved. Thus,
the capacity of the considered channel is given by (14).
B. Achievability of the Channel Capacity
The capacity achieving coding scheme for this channel is similar to the coding scheme for the
AWGN fading channel with EH given in [12]. The proposed scheme is outlined in the following.
The EHU wants to transmit message W to the ET using the harvested energy from the ET.
Message W is assumed to be drawn uniformly at random from a message set {1, 2...2nR }.
Thereby, message W carries nR bits of information, where n → ∞.
In the following, we describe a method for transferring nR bits of information from the EHU
to the ET in n + b channel uses, where R = C − , and → 0 and n/(n + b) → 1 as n → ∞.
As a result, the information from the EHU to the ET is transferred at rate R = C, as n → ∞.
16
For the proposed achievability scheme, we assume that the transmission is carried out in N +B
time slots, where N/(N + B) → 1 as N → ∞. In each time slot, we use the channel k times,
where k → ∞. The numbers N , B, and k are chosen such that n = N k and b = Bk hold.
Moreover, we assume that message W is represented in a binary form as a sequence of bits that
is stored at the EHU.
Transmissions at the ET: In each time slot, the ET transmits the same symbol x2 during the k
channel uses of the considered time slot. The value of the symbol x2 depends only on the fading
gain of the channel h during the corresponding time slot, and it can be found from Theorem 2.
Receptions and transmissions at the EHU: During the first few time slots, the EHU is silent
and only harvests energy from the ET. The EHU will transmit for the first time only when it
has harvested enough energy both for processing and transmission, i.e., only when its harvested
energy accumulates to a level which is higher than Pp + PEHU (x2 , h), where h is the fading gain
in the time slot considered for transmission. In that case, the EHU extracts kR(h) bits from its
storage, maps them to a Gaussian codeword with rate R(h) and transmits that codeword to the
ET. The rate of the codeword R(h) is given by
!
√
h2 PEHU
PET , h
1
−
R(h) = log 1 +
2
σ22 + PET α2
if (15) holds. Otherwise, if (15) does not hold, rate of the codeword R(h) is given by
1
h2 PEHU (x0 (h), h)
R(h) = log 1 +
− .
2
σ22 + x20 (h)α2
(41)
(42)
Receptions at the ET: The ET is able to decode the transmitted codeword from the EHU since
it is received via an AWGN channel with total AWGN variance of σ22 + PET α2 and σ22 + x20 (h)α2
for the rates in (41) and (42), respectively.
The EHU and the ET repeat the above procedure for N + B time slots.
Let N denote a set comprised of the time slots during which the EHU has enough energy
harvested and thereby transmits and let B denote a set comprised of the time slots during which
the EHU does not have enough energy harvested and thereby it is silent. Let N = |N | and
B = B, where || denotes the cardinality of a set. Moreover, let h(i) denote the outcome of H
in the i-th time slot. Using the above notations, the rate achieved during the N + B time slots
is given by
R=
N X
1 X
R(h(i)) = lim
R(h)p(h).
(N +B)→∞ N + B
(N +B)→∞ N + B
i∈N
h∈H
lim
(43)
17
Now, it is proven in [33] that when the EHU is equipped with an unlimited battery capacity and
when (13) holds, N/(N + B) → 1 as (N + B) → ∞. Using this, (43) simplifies to
X
R=
R(h)p(h),
(44)
h∈H
which is the channel capacity.
V. N UMERICAL R ESULTS
In this section, we illustrate examples of the capacity of the system model outlined above,
and compare it with the achievable rates of a chosen benchmark scheme. In the following, we
outline the system parameters, than we introduce the benchmark scheme, and finally we provide
the numerical results.
A. System Parameters
We use the standard path loss model given by
2
E{H } = ΩH =
c
fc 4π
2
d−γ
(45)
in order to compute the average power of the channel fading gains of the ET-EHU/EHU-ET
link, where c denotes the speed of light, fc is the carrier frequency, d is the length of the link,
and γ is the path loss exponent. We assume that γ = 3. The carrier frequency is equal to 2.4
GHz, a value used in practice for sensor networks, and d = 10 m or d = 20 m. We assume
a bandwidth of B = 100 kHz and noise power of −160 dBm per hertz, which for 100 kHz
adds-up to a total noise power of 10−14 Watts. The energy harvesting efficiency coefficient η is
assumed to be equal to 0.8. The system parameters are summarized in Table I. Throughout this
section, we assume Rayleigh fading with average power ΩH given by (45).
B. Benchmark Schemes
For the benchmark scheme, we divide the transmission time into slots of length T . We assume
that the EHU is silent and only harvests energy during a portion of the time slot, denoted by
t. In the remainder of the time slot, (T − t), the EHU only transmits information to the ET
and does not harvest energy. Similarly, the ET transmits energy during t, but remains silent and
receives information during T − t. In other words, both the EHU and the ET work in a HD
18
TABLE I
S IMULATION PARAMETERS
Parameter
Value
Speed of light c
299 792 458 m / s
Carrier frequency fc
2.4 GHz
Bandwidth B
100 kHz
Noise power σ
2
-160 dBm per Hz
EH efficiency η
0.8
Path loss exponent γ
3
Distance d
10 m ∨ 20 m
Processing cost Pp
-10 dBm ∼ 10 dBm
ET transmit power PET
0 dBm ∼ 35 dBm
mode. In this mode, the nodes are not impaired by self-interference, thus the harvested energy
in channel use i is given by
Ei = η(Hi X2i )2 .
(46)
Again, we assume CSI knowledge at the ET and at the EHU, and in addition the EHU is also
equipped with a battery with an unlimited storage capacity. Therefore, as per [33], the EHU can
also choose any amount of power for information transmission as long as its average extracted
energy from the battery is smaller than E{Ei }, where Ei is given by (46). Considering the HD
nature of the channel, the EHU can achieve its maximum rate given by
h2 PEHU (h)
= max t log 1 +
t
σ22
RHD
,
(47)
where
n 1 σ2 o
PEHU (h) = max 0, − 2 ,
λ h
where λ is found such that
t Pp +
holds.
e
− λσ
Ω
2
H
λ
σ 2 E1
−
λσ 2
ΩH
ΩH
(48)
= (T − t)ηPET ΩH
(49)
19
250
C
C
Achievabe rates (kbps)
200
, , = -100 dB
FD
2
, , = -80 dB
FD
2
HHD
150
100
50
0
0
5
10
15
20
25
30
35
ET average transmit power (dBm)
Fig. 4. Comparison of the capacity and the achievable rates of the benchmark scheme as a function of the ET average transmit
power for a link distance of d = 10m.
C. Numerical Examples
Figs. 4 and 5 illustrate the data rates achieved with the proposed capacity achieving scheme
and the benchmark scheme for link distances of 10 m and 20 m, respectively, and average
power of the ET that ranges from 0 dBm to 35 dBm. The processing cost at the EHU is set to
Pp = −10 dBm. It can be clearly seen from Figs. 3 and 4 that the achievable rates of the HD
benchmark scheme are much lower than the derived channel capacity. The poor performance of
the benchmark scheme is a consequence of the following facts. Firstly, self-interference energy
recycling in the HD mode is impossible at the EHU since there is no self-interference. Secondly,
the FD mode of operation is much more spectrally efficient than the HD mode, i.e., using part of
the time purely for energy harvesting without transmitting information has a big impact on the
system’s performance. As it can be seen from comparing Figs. 4 and 5, doubling the distance
of the node, has a severe effect on performance.
In Fig. 6, we present the ratio between the capacity and the rate of the benchmark scheme,
CF D /RHD , as a function of the self-interference suppression factor at the ET. The distance
20
30
CFD, ,2 = -100 dB
C
25
, , = -80 dB
FD
2
Achievabe rates (kbps)
RHD
20
15
10
5
0
0
5
10
15
20
25
30
35
ET average transmit power (dBm)
Fig. 5. Comparison of the capacity and the achievable rates of the benchmark scheme as a function of the ET average transmit
power for a link distance of d = 20m.
between the ET and the EHU is set to d = 10 m and the average transmit power of the ET is
set to PET = 30 dBm. The self-interference suppression factor at the ET can be found as the
reciprocial of the self-interference amplification factor α2 , i.e., as 1/α2 . The ratio CF D /RHD
can be interpreted as the gain in terms of data rate obtained by using the proposed capacity
achieving scheme compared to the data rate obtained by using the benchmark scheme. When
the self-interference suppression factor is very small, i.e., around 40 dB, it cripples the FD
capacity and the FD rate converges to the HD rate. Naturally, as the self-interference is more
efficiently suppressed, i.e., ≥ 50 dB the FD capacity rate becomes significantly larger than the
HD rate. An interesting observation can be made for suppression factor around 70 dB, which
are available in practice. In this case, the capacity achieving scheme results in a rate which is
approximately 50% larger than the HD rate. Another interesting result is that for self-interference
suppression of more than 85 dB, which is also available in practice, the proposed FD system
model achieves a rate which is more than double the rate of the HD system. This effect is a
21
4
3.5
CFD / R HD
3
2.5
2
1.5
1
0.5
40
50
60
70
80
90
100
110
120
130
Self-interefence suppression factor 1 / ,2 (dB)
Fig. 6. CF D /RHD ratio as a function of the self-interference suppression factor.
result of the energy recycling at the EHU.
Fig. 7 presents the capacity as a function of the mean and the variance of the self-interference
channel at the EHU, (α1 + g12 ). For this figure, the distance between the ET and the EHU is
d = 10 m and the average transmit power of the ET is PET = 30 dBm. The self-interference
suppression factor at the ET is 100 dB. As the average self-interference channel gain at the EHU
increases, i.e., (α1 + g12 ) increases, the EHU can recycle a larger amount of its transmit energy.
As a consequence, this results in a large capacity increase. Therefore, Figs. 6 and 7 illustrate
that the self-interference at the EHU can be transformed from a deleterious factor, to an aide,
or even an enabler of communication.
To illustrate the effect the processing energy cost has on the achievable rate, in Fig. 8, we
present the capacity in the case when the processing cost is zero and non-zero, for a distance
of d = 10 m. The Y axes in Fig. 8 is given in the logarithmic scale in order to better observe
the discrepancy between the two scenarios. It can clearly be seen that when the processing cost
is high, it has a detrimental effect on the capacity. This confirms that the energy processing
22
210
Achievable rates (kbps)
200
190
180
170
160
150
140
130
-15
-14
-13
-12
-11
-10
-9
g 21
-8
-7
-6
-5
-4
-3
+ ,1
Fig. 7. Capacity as a function of (α1 + g12 ).
cost must be considered in EH networks. Failing to do so might result in overestimating the
achievable rates, which in reality would only represent very loose upper bounds that can never
be achieved.
VI. C ONCLUSION
We studied the capacity of the point-to-point FD wirelessly powered communication system,
comprised of an ET and an EHU. Because of the FD mode of operation, both the EHU and the
ET experience self-interference, which impairs the decoding of the information-carrying signal
at the ET, but serves as an additional energy source at the EHU. We showed that the capacity
is achieved with a relatively simple scheme, where the input probability distribution at the EHU
is zero-mean Gaussian and where the ET transmits only one symbol. Numerical results showed
huge gains in terms of data rate when the proposed capacity achieving scheme is employed
compared to a HD benchmark scheme as well as the indisputable effect that the processing cost
and the self-interference have on the performance of WPCNs.
23
104
P = -1
p
P = -10 dBm
p
Achievable rates (kbps)
P p = 0 dBm
103
P = 10 dBm
p
102
101
100
0
5
10
15
20
25
30
35
ET transmit power
Fig. 8. Comparison of the capacity for different processing costs as a function of the ET’s average transmit power for a link
distance of d = 10m.
ACKNOWLEDGEMENT
The authors would like to thank Prof. P. Popovski, Aalborg University, Denmark, for comments
that greatly improved the manuscript.
A PPENDIX A
P ROOF OF T HEOREM 2
Let us assume that the optimal p(x2 |h) is a discrete and that the optimal p(x1 |x2 , h) is a
continuous probability distribution, which will turn out to be valid assumptions. In order to find
both input distributions, in the following, we solve the optimization problem given by (14).
Since I(X1 ; Y2 |X2 = x2 , H = h) is the mutual information of an AWGN channel with channel
gain h and AWGN with variance σ22 +x22 α2 , the optimal input distribution at the EHU, p(x1 |x2 , h),
is Gaussian with mean zero and variance PEHU (x2 , h), which has to satisfy constraint C2 in
2
(x2 ,h)
(14). Thereby, I(X1 ; Y2 |X2 = x2 , H = h) = 21 log 1 + h PσEHU
. Now, since G1 and X1
2 +x2 α
2
2
2
24
are zero-mean Gaussian RVs, the left-hand side of constraint C2 can be transformed into
Z X X
(50)
(x21 + Pp )p(x1 |x2 , h)p(x2 |h)p(h)dx1
x1 x ∈X h∈H
2
2
X X
=
PEHU (x2 , h)p(x2 |h)p(h) + Pp .
x2 ∈X2 h∈H
Whereas, the right-hand side of C2 can be rewritten as
Z X X
Ein p(x1 |x2 , h)p(x2 |h)p(h)dx1
(51)
x1 x ∈X h∈H
2
2
Z Z
=
g1
=
X X
η(hx2 + ḡ1 x1 + g1 x1 )2 p(x1 |x2 , h)p(x2 |h)p(h)p(g1 )dx1 dg1
x1 x ∈X h∈H
2
2
X X
ηh2 x22 p(x2 |h)p(h)
Z
+
+
g1
+ ηḡ12
ηḡ12 x21 p(x1 |x2 , h)p(x2 |h)p(h)dx1
x1 x ∈X h∈H
2
2
x2 ∈X2 h∈H
Z Z
X X
X X
ηg12 x21 p(x1 |x2 , h)p(x2 |h)p(h)p(g1 )dx1 dg1 =
x1 x ∈X h∈H
2
2
X X
X X
ηh2 x22 p(x2 |h)p(h)
x2 ∈X2 h∈H
PEHU (x2 , h)p(x2 |h)p(h) + ηα1
x2 ∈X2 h∈H
X X
PEHU (x2 , h)p(x2 |h)p(h),
x2 ∈X2 h∈H
where g1 represents the realizations of the random variable G1 . Combining (50) and (51)
transforms (14) into
X X1
h2 PEHU (x2 , h)
log 1 +
p(x2 |h)p(h)
max
2
2
PEHU (x2 ,h),p(x2 |h)
2
σ
+
x
α
2
2
2
x ∈X h∈H
2
∈
Subject to
C1 :
X X
x22 p(x2 |h)p(h) ≤ PET
x2 ∈X2 h∈H
C2 :
X X
PEHU (x2 , h)p(x2 |h)p(h) + Pp ≤
x2 ∈X2 h∈H
X X
ηh2 x22 p(x2 |h)p(h) + η(g¯1 2 + α1 )
x2 ∈X2 h∈H
C3 :
X
X X
PEHU (x2 , h)p(x2 |h)p(h)
x2 ∈X2 h∈H
p(x2 |h) = 1
x2 ∈X2
.
C4 :PEHU (x2 , h) ≥ 0.
(52)
25
Now, (52) can be solved in a straightforward manner using the Lagrange duality method. Thereby,
we write the Lagrangian of (52) as
X X1
h2 PEHU (x2 , h)
p(x2 |h)p(h)
L=
log 1 +
2
σ22 + x22 α2
x2 ∈X∈ h∈H
!
!
X X
X
2
− λ1
x2 p(x2 |h)p(h) − PET − µ1
p(x2 |h) − 1 − µ2 PEHU (x2 , h)
x2 ∈X∈ h∈H
(53)
x2 ∈X∈
!
− λ2 (1 − η(g¯1 2 + α1 ))
X X
PEHU (x2 , h)p(x2 |h)p(h) + Pp −
x2 ∈X2 h∈H
X X
ηh2 x22 p(x2 |h)p(h) .
x2 ∈X2 h∈H
In (52), we assume that 0 < η(g¯1 2 + α1 ) < 1, since η(g¯1 2 + α1 ) ≥ 1 would practically imply that
the EHU recycles the same or even a larger amount of energy than what has been transmitted
by the EHU, which is not possible in reality. In (53), λ1 , λ2 , µ1 , and µ2 are the Lagrangian
multipliers associated with C1, C2, C3, and C4 in (14), respectively. Differentiating (53) with
respect to the optimization variables, we obtain
∂L
=
∂PEHU (x2 , h)
1+
h2
σ22 +x22 α2
h2 PEHU (x2 ,h)
σ22 +x22 α2
− λ2 (1 − η(g¯1 2 + α1 )) − µ2 = 0
(54)
X
∂L
1X
h2 PEHU (x2 , h)
=
log 1 +
p(h)
−
λ
x22 p(h) − µ1
1
∂p(x2 |h)
2 h∈H
σ22 + x22 α2
h∈H
!
− λ2
(1 − η(g¯1 2 + α1 ))
X
h∈H
PEHU (x2 , h)p(h) − η
X
h2 x22 p(h)
= 0.
(55)
h∈H
When PEHU (x2 , h) > 0, then µ2 = 0. In consequence, we can use (54) to find PEHU (x2 , h) as
given by Theorem 2.
When µ1 > 0, then (55) has only two possible solutions ±x∗2 . In order for p∗ (x2 |h) to be a valid
P
probability distribution p(x∗2 |h)+p(−x∗2 |h) = 1 has to hold. Furthermore, x2 ∈X2 x22 p(x2 |h)p(h) ≤
PET also has to hold. So, one possible solution for the optimal input distribution is p∗ (x2 |h) =
1
δ(x2
2
− x∗2 ) + 21 δ(x2 + x∗2 ). However, x2 does not need to carry any information to the EHU,
thus uniformly choosing x2 between x∗2 and −x∗2 brings no benefit to the EHU. Moreover, the
complexity of the ET will be reduced if it only transmits one symbol, the symbol x∗2 . Thus,
P
p∗ (x2 |h) = δ(x2 − x∗2 ) is chosen. Now for x2 ∈X2 x22 p(x2 |h)p(h) ≤ PET to hold, we can choose
26
√
PET . Thus possible solution for p∗ (x2 |h) is given by (18) in Theorem 2. Using (55), we
√
find the condition for p(x2 |h) = δ(x2 − PET ) as
√
X1
h2 PEHU ( PET , h)
p(h) = λ1 PET + µ1
(56)
log 1 +
2
+
P
2
σ
ET α2
2
h∈H
!
p
X
X
+ λ2 1 − η(g¯1 2 + α1 )
PET , h p(h) − ηPET
PEHU
h2 p(h) .
x∗2 =
h∈H
h∈H
In this case, the capacity is given by (19). On the other hand, when (56) does not hold, another
P
possible way to satisfy x2 ∈X2 x22 p(x2 |h)p(h) ≤ PET is to have x∗2 = x0 (h), where in addition
to C1 in (52), the following has to be satisfied
X1
X
h2 PEHU (x0 (h), h)
p(h)
=
λ
x20 (h)p(h) + µ1 +
log 1 +
1
2
2
(h)α
+
x
2
σ
2
0
2
h∈H
h∈H
λ2
(57)
!
X
X
2
2
1 − η(g¯1 + α1 )
PEHU (x0 (h), h)p(h) − η
x0 (h)h p(h) .
2
h∈H
h∈H
Using (57) and C3 in (52), we can find the optimal x0 (h) as given by (21) in Theorem 2.
R EFERENCES
[1] D. Gunduz, K. Stamatiou, N. Michelusi, and M. Zorzi, ”Designing intelligent energy harvesting communication systems,”
in IEEE Commun. Magazine, vol. 52, no. 1, pp. 210 - 216, Jan. 2014
[2] C. K. Ho, and R. Zhang, ”Optimal energy allocation for wireless communications with energy harvesting constraints,” in
IEEE Trans. Signal Proccessing, vol. 60, no. 9, pp. 4808 - 4818, Sept. 2012
[3] L. R. Varshney, ”Transporting information and energy simultaneousely,” in Proc. IEEE ISIT, Toronto, ON, Canada, July
2008
[4] H. Yu, and R. Zhang, ”Througnput maximization in wireless powered communication networks,” in Proc. IEEE
GLOBECOM, Atlanta, GA, USA, Dec. 2014
[5] O. Ozel and S. Ulukus, ”AWGN channel under time-varying amplitude constraints with causal information at the
transmitter,” in Proc. IEEE ASILOMAR, Pacific Grove, CA, 2011
[6] W. Mao and B. Hassibi, ”Capacity Analysis of Discrete Energy Harvesting Channels,” in IEEE Trans. Inf. Theory, vol.
63, no. 9, pp. 5850 - 5885, Sept. 2017
[7] Z. Chen, G. C. Fen-ante, H. H. Yang and T. Q. S. Quek, ”Capacity bounds on energy harvesting binary symmetric channels
with finite battery,” in Proc. IEEE ICC, Paris, 2017
[8] D. Shaviv, P. M. Nguyen and A. Ozgur, ”Capacity of the Energy-Harvesting Channel With a Finite Battery,” IEEE Trans.
Inf. Theory, vol. 62, no. 11, pp. 6436-6458, Nov. 2016
[9] M. Ashraphijuo, V. Aggarwal and X. Wang, ”On the Capacity of Energy Harvesting Communication Link,” in IEEE
Journal on Selected Areas in Communications, vol. 33, no. 12, pp. 2671 - 2686, Dec. 2015
27
[10] R. Rajesh, V. Sharma and P. Viswanath, ”Information capacity of energy harvesting sensor nodes,” in Proc. IEEE ISIT,
St. Petersburg, 2011
[11] O. Ozel and S. Ulukus, ”Achieving AWGN capacity under stochastic energy harvesting,” in IEEE Trans. Inf. Theory, vol.
58, no. 10, pp. 6471 - 6483, Oct. 2012
[12] R. Rajesh, V. Sharma, and P. Viswanath, ”Capacity of Gaussian channels with energy harvesting and processing cost,” in
IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2563 -2575, May 2014
[13] O. Ozel and S. Ulukus, ”Information-theoretic analysis of an energy harvesting communication system,” in Proc. IEEE
International Symposium on Personal, Indoor and Mobile Radio Communications Workshops, Instanbul, Turkey, 2010
[14] S. Luo, R. Zhang and T. J. Lim, ”Optimal Save-Then-Transmit Protocol for Energy Harvesting Wireless Transmitters,” in
IEEE Trans. on Wireless Commun., vol. 12, no. 3, pp. 1196 - 1207, March 2013
[15] L. Liu, R. Zhang, and K-C. Chua, ”Wireless information transfer with opportunistic energy harvesting,” in IEEE Trans.
on Wireless Commun., vol. 12, no. 1, pp. 288-300, Jan. 2013
[16] Y. Dong, Z. Chen and P. Fan, ”Capacity Region of Gaussian Multiple-Access Channels With Energy Harvesting and Energy
Cooperation,” in IEEE Access, vol. 5, pp. 1570-1578, 2017
[17] H. A. Inan, D. Shaviv and A. Ozgur, ”Capacity of the energy harvesting Gaussian MAC,” in Proc. IEEE ISIT, Barcelona,
2016
[18] N. Zlatanov, D. Wing Kwan Ng and R. Schober, ”Capacity of the Two-Hop Relay Channel With Wireless Energy Transfer
From Relay to Source and Energy Transmission Cost,” in IEEE Trans. Wireless Commun., vol. 16, no. 1, pp. 647 - 662,
Jan. 2017
[19] Y. Gu and S. Assa, ”RF-Based Energy Harvesting in Decode-and-Forward Relaying Systems: Ergodic and Outage
Capacities,” in IEEE Trans. on Wireless Commun., vol. 14, no. 11, pp. 6425 - 6434, Nov. 2015
[20] V. Sharma and R. Rajesh, ”Queuing theoretic and information theoretic capacity of energy harvesting sensor nodes,” in
Proc. IEEE ASILOMAR, Pacific Grove, CA, 2011
[21] E. Everett, A. Sahai, and A. Sabharwal, ”Passive Self-Interference Suppression for Full-Duplex Infrastructure Nodes,” in
IEEE Trans. on Wireless Commun., vol. 13, no. 2, pp. 680-694, Feb. 2014
[22] C. Anderson, S. Krishnamoorthy, C. Ranson, T. Lemon, W. Newhall, T. Kummetz, and J. Reed, ”Antenna isolation,
wideband multipath propagation measurements, and interference mitigation for on-frequency repeaters,” in Proc. IEEE
SoutheastCon, pp. 110-114, March 2004
[23] E. Everett, A. Sahai, and A. Sabharwal, ”Experiment-Driven Characterization of Full-Duplex Wireless Systems,” in IEEE
Trans. on Wireless Commun., vol. 11, no. 12, pp. 4296 - 4307, Dec. 2012
[24] A. Sahai, G. Patel, C. Dick, and A. Sabharwal, ”On the impact of phase noise on active cancelation in wireless full-duplex,”
in IEEE Trans. on Vehicular Technology,vol. 62, no. 9, pp. 4494-4510, Nov. 2013
[25] B. Day, A. Margetts, D. Bliss, and P. Schniter, ”Full-duplex MIMO relaying: Achievable rates under limited dynamic
range,” in IEEE ASILOMAR, Pacific Grove, CA, 2012
[26] M. Duarte, A. Sabharwal, V. Aggarwal, R. Jana, K. K. Ramakrishnan, C. Rice and N. K. Shankaranarayanan, ”Design
and Characterization of a Full-Duplex Multiantenna System for WiFi Networks,” in IEEE Trans. on Vehicular Technology,
28
vol. 63, no. 3, pp. 1160-1177, March, 2014
[27] Y. Zeng and R. Zhang, ”Full-Duplex Wireless-Powered Relay With Self-Energy Recycling,” inIEEE Wireless Communications Letters, vol. 4, no. 2, pp. 201-204, April 2015
[28] M. Maso, C. F. Liu, C. H. Lee, T. Q. S. Quek and L. S. Cardoso, ”Energy-Recycling Full-Duplex Radios for NextGeneration Networks,” in IEEE Journal on Selected Areas in Communications, vol. 33, no. 12, pp. 2948 - 2962, Dec.
2015
[29] Z. Hu, C. Yuan, F. Zhu and F. Gao, ”Weighted Sum Transmit Power Minimization for Full-Duplex System With SWIPT
and Self-Energy Recycling,” in IEEE Access, vol. 4, pp. 4874-4881, 2016
[30] J. Chae, H. Lee, J. Kim and I. Lee, ”Self energy recycling techniques for MIMO wireless communication systems,” in
Proc. IEEE ICC, Paris, 2017
[31] D. Bharadia, E. McMilin, and S. Katti, ”Full duplex radios,” in Proc. ACM SIGCOMM, Hong Kong, China, 2013,
[32] N. Zlatanov, E. Sippel, V. Jamali, and R. Schober, ”Capacity of the Gaussian Two-Hop Full-Duplex Relay Channel with
Residual Self-Interference,” in IEEE Trans. on Commun., vol.65 , no.3 , pp. 1005 - 1021 , March 2017
[33] N. Zlatanov, R. Schober, and Z. Hadzi-Velkov, ”Asymptotically Optimal Power Allocation for Energy Harvesting
Communication Networks”, in IEEE Trans. on Vehicular Technology, vol. 66, no. 8 , pp. 7286 - 7301, Aug. 2017
[34] T. M. Cover and J. A. Thomas, ”Elements of Information Theory”, John Wiley and Sons, 2012
[35] D. Tse and P. Viswanath, ”Fundamentals of Wireless Communication” Cambridge University Press, 2005
| 7cs.IT
|
arXiv:1604.07049v1 [math.OC] 24 Apr 2016
Fast Approximation Algorithms for the Generalized Survivable
Network Design Problem
Andreas Emil Feldmann1 , Jochen Könemann2 , Kanstantsin Pashkovich2 , and Laura Sanità2
1
SZTAKI, Hungarian Academy of Sciences, Hungary & KAM, Charles University in
Prague, Czechia. [email protected]
2
Department of Combinatorics and Optimization, University of Waterloo, Canada.
{jochen, kpashkovich, laura.sanita}@uwaterloo.ca
April 26, 2016
Abstract
In a standard f -connectivity network design problem, we are given an undirected graph
G = (V, E), a cut-requirement function f : 2V → N, and non-negative costs c(e) for all e ∈ E.
We are then asked to find a minimum-cost vector x ∈ NE such that x(δ(S)) ≥ f (S) for all S ⊆ V .
We focus on the class of such problems where f is a proper function. This encodes many wellstudied NP-hard problems such as the generalized survivable network design problem.
In this paper we present the first strongly polynomial time FPTAS for solving the LP relaxation of the standard IP formulation of the f -connectivity problem with general proper functions f . Implementing Jain’s algorithm, this yields a strongly polynomial time (2 + ǫ)-approximation for the generalized survivable network design problem (where we consider rounding up
of rationals an arithmetic operation).
1
Introduction
The input in a typical network design problem consists of a directed or undirected graph G = (V, E),
non-negative unit-capacity installation costs c(e) for all e ∈ E, and a collection of connectivity requirements among the vertices in V . The goal is then to find a minimum-cost capacity installation
in G that satisfies the connectivity requirements. The above abstract problem class captures many
practically relevant optimization problems, many of which are NP-hard. Therefore, maybe not surprisingly, there has been a tremendous amount of research in the area of approximation algorithms
for network design problems throughout the last four decades (e.g., see [17, 18]).
X
Connectivity requirements can be modelled in
c(e)x(e)
(IP)
min
many ways, but we will adopt the f -connectivity
e∈E
viewpoint in this paper. Here, one is given a cuts.t. x(δ(S)) ≥ f (S) ∀S ⊂ V
requirement function f : 2V → N, and one wants to
x ≥ 0, x integer.
find a minimum-cost non-negative integer vector x
such that for each S ⊆ V , the sum of variables x(e)
for edges e crossing the cut S is at least f (S). In other words, we are interested in problems that
can be encoded by integer program (IP); here
P δ(S) denotes the set of edges incident to a vertex in
S and a vertex outside S, and x(δ(S)) := e∈δ(S) x(e).
1
Restricting even further, we will henceforth only be concerned with instances of (IP) where f is
proper: f is proper if it satisfies the three properties of maximality (i.e., f (A∪B) ≤ max{f (A), f (B)}
for all disjoint sets A, B ⊆ V ), symmetry (i.e., f (S) = f (V \ S) for all S ⊆ V ), and f (V ) = 0.
Program (IP) with proper cut-requirement functions f captures (among others) the special case
where the goal is to find a minimum-cost network that has r(u, v) edge-disjoint paths connecting
any pair u, v of vertices (for given non-negative integer parameters r). The implicit cut-requirement
function in this case is then given by f (S) := maxu∈S,v∈V \S r(u, v) for all S ⊆ V .
Based on the primal-dual method, Goemans and Williamson [12] first gave a 2H(fmax )-approximation algorithm for (IP) with proper cut-requirement functions where one is allowed to pick edges
multiple times in the solution. Goemans et al. [11] later obtained the same performance ratio
for the setting where multiple copies of edges are not allowed. More recently, in a breakthrough
result, Jain [19] obtained a 2-approximation for the more general class of skew-supermodular cutrequirement functions based on iterative rounding. Jain’s algorithm iteratively fixes the value of a
subset of variables in (IP). To aid this, he first defines a slightly more general version of the IP,
where the value of certain variables is fixed. Specifically, for a set I ⊆ E of edges, assume that the
value of variable x(e) is fixed to z(e) ∈ N. The LP relaxation of the IP for the corresponding residual problem is given in (LP1 ). Jain’s key observation was that the extreme points of the feasible
region of (LP1 ) are sparse, and have at least one variable with value at least 1/2.
Capitalizing on this insight, his algorithm then itX
eratively solves O(|V |) instantiations of (LP1 ) while
min
c(e)x(e)
(LP1 )
intermittently rounding up the values of large varie∈E
ables in the computed solutions. In order to solve
s.t. x(δ(S)) ≥ f (S) − z(δ(S)) ∀S ⊆ V
(LP1 ) one needs to employ the Ellipsoid method [16]
x(e) = 0 ∀e ∈ I
together with a polynomial-time seperation oracle
x≥0
for the LP’s constraints (see [7]).
Our work is motivated by Open Problem 4 in Williamson and Shmoys’ recent book [23] where
the authors point out that solving (LP1 ) for general (proper) functions f may be computationally
quite demanding despite the fact that it can be done efficiently in a theoretical sense. The authors
leave as an open problem the design of a primal-dual 2-approximation for the survivable network
design problem. Our main result is a replacement of the Ellipsoid-based exact LP-solver calls in
Jain’s algorithm by approximate ones that are based on the (in a sense) primal-dual multiplicativeweights method of [10]. We realize that the likely intended meaning of primal-dual in Williamson
and Shmoys’ open problem statement is the primal-dual method for approximation algorithms (as
in [12]). However we believe that the contribution made in this paper is in line with the motivation
given for Open Problem 4 in [23]: we substantially speed up LP computations in Jain’s algorithm
at the expense of an inconsequential loss in performance guarantee of the algorithm.
Theorem 1. For any ǫ > 0, there is a (1 + ǫ)-approximation algorithm for (LP1 ) that runs in
strongly polynomial time1 independently of the values of c and f .
In contrast to our result, Jain [19] observes that the relaxations of (IP) and (LP1 ) are of
a combinatorial nature, and hence can be solved in strongly polynomial-time via Tardos’ algorithm [22] whenever their number of variables and constraints are polynomially bounded (in the
problem dimension). For example, (IP) and (LP1 ) have an equivalent compact representation when
f (S) = maxu∈S,v6∈S r(u, v) for all S ⊆ V . We also note that one can argue that the Ellipsoid method
1
An algorithm is strongly polynomial if its number of arithmetic operations, i.e. the number of additions, subtractions, multiplications, divisions and comparisons, is bounded by a polynomial in the dimension of the problem (i.e.,
the number of data items in the input), and the length of the numbers occurring during the algorithm is bounded by
a polynomial in the length of the input.
2
applied to (LP1 ) and the linear relaxation of (IP) terminates in a strongly polynomial number of
steps whenever function f (S) is polynomially bounded (in the problem dimension), for all S ⊆ V
as this implies small encoding-length of vertices of the feasible region of (IP) and (LP1 ).
To achieve the result in Theorem 1, we rely on the multiplicative weights method of Garg and
Könemann [10] (henceforth referred to as GK). This is a natural idea as (LP1 ) belongs to the class
of positive covering LPs. As such, [10] applies to the LP dual of (LP1 ). The algorithm can therefore
be used to compute an approximate pair of primal and dual solutions in strongly-polynomial time
as long as we are able to provide it with a strongly-polynomial time (approximation) algorithm for
the so called shortest row problem. For (LP1 ) this boils down to computing
min
f (S)−z(δ(S))≥1
S⊆V
x(δ(S))
,
f (S) − z(δ(S))
E
for given x ∈ RE
+ and z ∈ Z+ , i.e., finding a corresponding set S. The above shortest row problem
is solved directly by Gabow et al.’s strongly-polynomial time separation oracle for the constraints
of (LP1 ) (see [7]) in the case where I = ∅, and hence z = 0. Once I 6= ∅, Gabow et al.’s algorithm
can not be used directly to give a strongly polynomial-time algorithm, and a more subtle approach
is needed. In fact, in this case, we provide only a (1 + ζ)-approximate solution to the shortest
row problem (for appropriate ζ > 0). As is well-known (e.g., see [6, 10]), the exact shortest-row
subroutine used in GK may be replaced by an α-approximate one, sacrificing a factor of α in
the overall performance ratio of the algorithm in [10]. We obtain the following direct corollary of
Theorem 1.
Corollary 1. Combining Theorem 1 with Jain’s algorithm, we obtain a strongly polynomial-time2
(2 + ε)-approximation algorithm for (IP), that does not use linear programming solvers.
We once again stress that the above results hold for any not necessarily bounded proper cutrequirement function f .
Further related work The past 30 years have seen significant research on solving linear programs
efficiently; e.g., see the work by Shahrokhi & Matula [21], Luby & Nisan [20], Grigoriadis &
Khachian [14, 15], Young [24, 25], Garg & Könemann [10], Fleischer [6], and Iyengar & Bienstock [3].
We refer the reader to two recent surveys by Bienstock [2] and Arora, Hazan & Kale [1].
Particularly relevant to this paper is the work by Fleischer [5] who previously proposed a
Lagrangian-based approximation scheme for positive covering LPs with added variable upper bounds.
Their algorithm builds on [10] and [6], and achieves a performance ratio of (1 + ǫ) for any positive ǫ
using O(ǫ−2 m log(Cm)) calls to a separation oracle for the given covering problem; here m denotes
the number of variables, and C is bounded by the maximum objective function coefficient. Garg
and Khandekar [9] later addressed the same problem, and presented an improved algorithm with
O(mǫ−2 log m + min{n, log log C}) calls to an oracle for the most violated constraint.
The algorithms in [5, 9] naturally apply to solving LP relaxations of various network design IPs
where the multiplicity x(e) of each edge e is limited to some given upper bound. As the approaches
in [5, 9] need to approximate the same type of the shortest row problem, as an immediate corollary
of our result, we obtain a strongly polynomial-time (2 + ε)-approximation algorithm for (IP) with
constant upper bounds on the variables. This captures in particular the interesting case in which
we have binary constraints for x.
2
if rounding up numbers is considered an arithmetic operation.
3
Finally, we also mention the work of Garg & Khandekar [8] who present a fully polyonimal-time
approximation algorithm for the fractional Steiner forest problem. The algorithm also applies to
the more general problem of finding a minimum-cost fractional hitting set of a given collection of
clutters.
Organization We first provide some more details on how to implement the iterative rounding
algorithm of Jain. We continue and provide a detailed description of GK with approximate oracles in
Section 3 for completeness, and describe the shortest-row oracles in Section 4. Finally, in Section 5
we put together all ingredients to prove our main result.
2
Iterative rounding
Recall that Jain’s key structural insight was to observe that extreme points x ∈ RE
+ of (LP1 ) have
x(e) ≥ 1/2 for at least one e ∈ E. Jain also noted that, in an implementation of his algorithm, the
computation of extreme points may be circumvented. In fact, he suggests obtaining LP (LPg ) from
(LP1 ) by adding the constraint x(g) ≥ 1/2 for some edge g ∈ E. Let optg be the objective function
value of an optimum solution to (LPg ). Jain’s structural lemma now implies that ming∈E optg is
at most the optimum value of (LP1 ). Jain’s algorithm can now be implemented by replacing the
computation of an optimum basic solution to the residual problem in each iteration, by computing
optimal solutions to linear programs of type (LPg ) for all edges g ∈ E.
Of course, we can also replace computing an optimal solution to (LPg ) with computing an
approximate one, at the expense of a slight increase of the final approximation factor. Jain’s
method in this case is summarized for completeness in Algorithm 1.
Algorithm 1 A 2(1 + ζ)|E|- approximation algorithm for (IP)
1: I0 ← ∅, z0 (e) ← 0 for all e ∈ E, k ← 0
2: while f (S) − zk (δ(S)) > 0 for some S ⊆ V do
3:
k ←k+1
4:
for all g ∈ E \ Ik−1 do
5:
find a (1 + ζ)-approximation xk,g for (LPg ) with z := zk−1 and I := Ik−1
6:
end for
P
7:
let xk be a vector xk,g corresponding to ming∈E\Ik−1 e∈E c(e)xk,g (e)
8:
Ik ← Ik−1 ∪ {e ∈ E \ Ik−1 |xk (e) = 0 or xk (e) ≥ 1/2}
9:
for all e ∈ E, let zk (e) ← ⌈xk (E)⌉ if xk (e) ≥ 1/2, and zk (e) ← zk−1 (e) otherwise
10: end while
11: return zk
Claim 1. Given ζ > 0, Algorithm 1 is a 2(1 + ζ)|E|-approximation algorithm for (IP). Moreover,
Algorithm 1 terminates after at most |E| iterations of step 2.
Proof. Jain’s structural lemma immediately implies that Ik−1 ( Ik ⊆ E for every k. Hence the total
number of iterations is at most |E|. In iteration k we fix the values of variable x(e), e ∈ Ik \ Ik−1 ,
and due to the definition of Ik \ Ik−1 we have zk (e) ≤ 2xk (e), e ∈ Ik \ Ik−1 . The remaining values
xk (e), e ∈ E \ Ik form a valid solution for (LP1 ) with z(e) := zk (e), e ∈ E, which is solved with
approximation guarantee (1 + ζ) in the (k + 1)-st iteration. Since there are at most |E| iterations
and in step 5 the found solution is a (1 + ζ)-approximation of (LPg ), we know that the objective
4
value of the output is at most 2(1 + ζ)|E| times the objective value of the linear relaxation of (IP),
finishing the proof.
Note that by Claim 1, if ζ ≤ ln(1 + ε)/|E| then Algorithm 1 gives a 2(1 + ε)-approximation for
(IP).
3
Multiplicative weights method
In this section, we briefly review the multiplicative weights method [10] of GK, when applied to a
positive covering LP of the form
X
min
c(j)x(j)
(LP2 )
j∈[n]
s.t.
X
A(i, j)x(j) ≥ b(i)
∀i ∈ [m],
j∈[n]
x ≥ 0,
where A(i, j) ≥ 0 for all i ∈ [m], j ∈ [n], b(i) > 0 for all i ∈ [m] and c(j) > 0 for all j ∈ [n].
Note, that the linear program (LPg ) is a positive LP of the above form when simply eliminating
variables x(e) for e ∈ I. In the same way, we can exclude the inequalities corresponding to S ⊆ V
with f (S) − z(δ(S)) ≤ 0.
Given i ∈ [m] and a vector x ∈ Rn+ we define the length len(i, x) of row i with respect to x as
len(i, x) :=
X
A(i, j)x(j)/b(i) ,
(1)
j∈[n]
and we denote by len(x) the shortest length of a row in A with respect to x, i.e. len(x) :=
mini∈[m] len(i, x). Now it is straightforward to reformulate (LP2 ) as
min
x≥0,x6=0
X
c(j)x(j)/ len(x) .
(2)
j∈[n]
The multiplicative weights method of GK applied to the dual of (LP2 ) computes an approximate
pair of primal and dual solutions in strongly-polynomial time, as long as it is provided with a
strongly-polynomial time oracle for determining the row q of shortest length (the shortest row )
with respect to given lengths x ∈ Rn+ as in (1).
It is implicit in the work of [10, 6] that exact oracles can be replaced by approximate ones
(incurring a corresponding degradation in performance ratio, of course). Such a modification is
described from a packing point of view in [4], for example. Algorithm 2 shows the pseudo code of
the algorithm for completeness. In step 4 of the algorithm a (1 + ζ)-approximation q of the shortest
row with respect to some vector x ∈ Rn+ is computed. That is, q is a row for which len(q, x) ≤
(1 + ζ) len(x). Section 4 describes how to obtain this approximation in strongly-polynomial time.
We give a proof of the next lemma for completeness.
Lemma 1 (implicit in [10, 6]). Algorithm 2 is a (1 + 4ζ)-approximation for (LP2 ). Moreover,
Algorithm 2 terminates after at most ζ1 log1+ζ (1 + ζ)n iterations.
5
Algorithm 2 The multiplicative weights algorithm to solve (LP2 ).
− 1
1: δ ← (1 + ζ) (1 + ζ)n ζ , x0 (j) ← δ/c(j) for all j ∈ [n], y0 (i) ← 0 for all i ∈ [m], k ← 0
P
2: while
j∈[n] c(j)xk (j) < 1 do
3:
k ←k+1
4:
determine a (1 + ζ)-approximation for the shortest row with respect to xk−1 , let it be row
qk
5:
determine j ∈ [n] with the minimum value c(j)/A(qk , j), let it be column pk
6:
for all i ∈ [m] do
(
yk−1 (i) + c(pk )/A(qk , pk ) if i = qk
yk (i) ←
yk−1 (i)
otherwise .
end for
c(pk )A(qk ,j)
xk−1 (j) for all j ∈ [n]
8:
xk (j) ← 1 + ζ c(j)A(q
k ,pk )
9: end while
P
10: return xk / len(xk ) corresponding to mink
j∈[n] c(j)xk (j)/ len(xk )
7:
P
c(j)xk (j)
Proof. Let us define β := mink j∈[n]
len(xk )
for the problem given by (2).
For every k ≥ 1 we have
X
X
c(j)xk (j) −
c(j)xk−1 (j) =
j∈[n]
, below we show that β provides a good approximation
j∈[n]
ζ len(qk , xk−1 )c(pk )b(qk )/A(qk , pk ) ≤ ζ(1 + ζ) len(xk−1 ) ×
X
b(i)(yk (i) − yk−1 (i)) .
i∈[m]
Hence,
X
c(j)xk (j) ≤
j∈[n]
X
c(j)x0 (j) + ζ(1 + ζ) ×
j∈[n]
c(j)xk (j) ≤
j∈[n]
X
b(i)(yh (i) − yh−1 (i)) len(xh−1 ) .
h∈[k] i∈[m]
Due to the definition of β we have len(xh−1 ) ≤
X
X X
c(j)x0 (j) +
P
j∈[n] c(j)xh−1 (j)/β
and thus
X X
X
ζ(1 + ζ)
×
b(i)(yh (i) − yh−1 (i))
c(j)xh−1 (j) .
β
h∈[k] i∈[m]
j∈[n]
j∈[n]
P
To show that the right-hand side of (3) is at most nδ eζ(1+ζ) i∈[m] b(i)yk (i)/β , we use induction.
Indeed, the case k = 0 is clear, and to show the statement consider
X
X X
X
ζ(1 + ζ)
c(j)x0 (j) +
×
b(i)(yh (i) − yh−1 (i))
c(j)xh−1 (j) ,
β
j∈[n]
h∈[k] i∈[m]
j∈[n]
which equals
X
j∈[n]
c(j)x0 (j) +
ζ(1 + ζ) X
×
β
X
b(i)(yh (i) − yh−1 (i))
h∈[k−1] i∈[m]
X
c(j)xh−1 (j)+
j∈[n]
X
i∈[m]
6
X
c(j)xk−1 (j)
b(i) yk (i) − yk−1 (i)
j∈[n]
Due to (3) we conclude that the last expression is at most
X
ζ(1 + ζ)
ζ(1 + ζ) X
c(j)x0 (j) +
b(i) yk (i) − yk−1 (i) ×
×
1+
β
β
j∈[n]
i∈[m]
X X
X
c(j)xh−1 (j) .
b(i) yh (i) − yh−1 (i)
h∈[k−1] i∈[m]
j∈[n]
Using the inequality (1 + α) ≤ eα , α ∈ R and the induction hypothesis we upper-bound the
expression above by
P
P
P
ζ(1+ζ) i∈[m] b(i) yk (i)−yk−1 (i) /β
e
×nδ eζ(1+ζ) i∈[m] b(i)yk−1 (i)/β = nδ eζ(1+ζ) i∈[m] b(i)yk (i)/β .
Now let us consider the last iteration t, where we have
P
X
1≤
c(j)xt (j) ≤ nδ eζ(1+ζ) i∈[m] b(i)yt (i)/β ,
j∈[n]
and thus
P
whenever nδ < 1.
Now let us show that yt / log1+ζ
to show that
max
j
β
ζ(1 + ζ)
≤
ln((nδ)−1 )
i∈[m] b(i)yt (i)
1+ζ
δ
(3)
is a feasible solution for the dual of (LP2 ). It is enough
X A(i, j)yt (i)
1+ζ
≤ log1+ζ
.
c(j)
δ
i∈[m]
To see this note that for every j ∈ [n] and every k
X A(i, j)yk (i)
X A(i, j)yk−1 (i)
A(qk , j) c(pk )
−
=
≤ 1,
c(j)
c(j)
c(j) A(qk , pk )
i∈[m]
and
P
i∈[m]
A(i,j)y0 (i)
c(j)
i∈[m]
= 0. On the other hand for every j ∈ [n] and every k
xk (j)
c(pk )A(qk , j)
=1+ζ
≤1+ζ,
xk−1 (j)
c(j)A(qk , pk )
x0 (j) = δ/c(j) and due to the termination condition xt−1 (j) < 1/c(j) and hence xt (j) < (1 +
ζ)/c(j). This implies that the algorithm terminates after at most log1+ζ 1+ζ
δ iterations. Thus,
1+ζ
yt / log1+ζ δ is a feasible solution for the dual of (LP2 ).
Hence, the algorithm provides a feasible solution for (2) with value β, which is an approximation
with guarantee
1+ζ
1+ζ
ζ(1 + ζ) ln δ
ζ(1 + ζ)
log
=
,
1+ζ
ln((nδ)−1 )
δ
ln(1 + ζ) ln((nδ)−1 )
is a feasible solution for the dual of (LP2 ). Thus, we
due to (3) and the fact that yt / log1+ζ 1+ζ
δ
obtain
1+ζ
ζ(1 + ζ) ln δ
ζ(1 + ζ)
ζ(1 + ζ)
1+ζ
=
≤
≤
,
ln(1 + ζ) ln((nδ)−1 )
(1 − ζ) ln(1 + ζ)
(1 − ζ)(ζ − ζ 2 /2)
(1 − ζ)2
which is at most (1 + 4ζ) for ζ ≤ 0.15.
7
4
The shortest row problem
In this section we describe how to (approximately) solve the shortest row problem needed in
Algorithm 2 when applied to (LP1 ). We start by stating the following simple remark, that we
will need at the end of our analysis.
Remark 1. For every k ≥ 1 and every S ⊆ V , we have f (S) − zk (δ(S)) ≤ |E|/2 in Algorithm 1.
Proof. Define x ∈ RE
+ by letting x(e) := xk (e) whenever xk (e) < 1/2, and let x(e) := 0 otherwise.
Then note that |E|/2 ≥ x(δ(S)) ≥ f (S) − zk (δ(S)), due to the feasibility of x in (LP1 ) with z := zk
and I := Ik .
E
Let us recall that, for given x ∈ RE
+ , z ∈ Z+ , and proper function f , the shortest row problem
we need to solve is the following:
min
f (S)−z(δ(S))≥1
S⊆V
x(δ(S))
.
f (S) − z(δ(S))
The above shortest row problem is quite easy to solve when z = 0. In this case, Gabow et al. [7] give
a strongly-polynomial time separation oracle based on the construction of Gomory-Hu trees [13],
as we are now going to explain.
Given a graph G = (V, E) and values x(e) ∈ R+ for each e ∈ E, a Gomory-Hu tree [13] is a
capacitated tree T = (V, J) such that for any two vertices v, u ∈ V the minimum x-value of a cut
in G separating v and u equals the minimum x-value among the u-v cuts induced by the edges of
T . More concretely, let Se and V \ Se induce connected components in T after removing e from T .
We then have
min x(δ(S)) = min x(δ(Se )) .
u∈S,v6∈S
S⊆V
u∈Se ,v6∈Se
e∈J
The next lemma shows that in order to find the shortest row in the first iteration of step 2 in
Algorithm 1 (i.e. when z = 0), it is enough to compute a Gomory-Hu tree with respect to values
x(e) ∈ R+ , e ∈ E.
Lemma 2 ([7]). Given a graph G = (V, E), a proper function f : 2V → Z+ and a Gomory-Hu tree
T = (V, J) with respect to values x(e) ∈ R+ , e ∈ E, we have
min x(δ(S))/f (S) = min x(δ(Se ))/f (Se ) .
f (S)6=0
S⊆V
f (Se )6=0
e∈J
Proof. Consider S ⊆ V and the edges δT (S) in the Gomory-Hu tree T defined by S. By definition
of a Gomory-Hu tree x(δ(S)) ≥ x(δ(Se )) for every e ∈ δT (S), due to the cut in T incurred by the
vertices incident to e. Thus, to prove the claim it is enough to show that
f (S) ≤ max f (Se ) .
e∈δT (S)
(4)
To show the last inequality let V1 ,. . . ,Vk be vertex sets of the connected components after removing
S in T . Thus, V1 ,. . . , Vk form a partition of V \S, and so maxi∈[k] f (Vi ) ≥ f (V \S) = f (S). Choose
i ∈ [k]. Replacing Se by V \ Se , we can assume that Se and Vi are disjoint for every e ∈ δT (Vi ).
Thus the sets Se with e ∈ δT (Vi ) partition V \ Vi , showing that f (Vi ) ≤ maxe∈δT (Vi ) f (Se ). Since,
δT (Vi ), i ∈ [k] partition δ(S) we get (4), finishing the proof.
8
In the later iterations of steps 2 in Algorithm 1, the inequality corresponding to S ⊆ V has
the form x(δ(S)) ≥ g(S), where g : 2V → Z+ is such that g(S) = f (S) − z(δ(S)) for some
z(e) ∈ Z+ , e ∈ E, and a proper function f . Once z 6= 0, g(S) is not a proper function any more
and unfortunately, Gabow et al.’s algorithm can not be used directly. We do not know how to solve
this problem exactly in strongly-polynomial time, but we can approximate it using the following
observation.
Fix a value γ > 0, and let us check whether the optimal solution of the shortest row problem
has a value less than γ. The crucial fact is that given x(e) ∈ R+ , e ∈ E and γ > 0 checking whether
min
f (S)−z(δ(S))≥1
S⊆V
x(δ(S))
<γ.
f (S) − z(δ(S))
is equivalent to checking whether
x(δ(S))/γ + z(δ(S)) < f (S)
for some S ⊆ V , i.e. it can be reduced to finding
x(δ(S))/γ + z(δ(S))
.
f (S)
f (S)6=0
min
S⊆V
Therefore, we can apply Lemma 2 after replacing x(e) with x(e)/γ + z(e). This enables us to
use binary search to find a (1 + ζ)-approximation for the shortest row indexed by vertex subsets
whenever we have a lower bound γmin and an upper bound γmax on the length of the shortest row.
Giving trivial bounds on such a value (e.g. 1 and (|E|·maxS f (S))) is of course easy. However, given
an interval [γmin , γmax ] for binary search we have to construct a Gomory-Hu tree ⌈log1+ζ γmax /γmin ⌉
times, and therefore we need that γmax /γmin is independent of the size of f in order to achieve strong
polynomiality. To this aim, we propose Algorithm 3.
Algorithm 3 Determining upper and lower bounds γmin , γmax .
1: G0 = (V0 , E0 ) ← G = (V, E), f0 (S) ← f (S) for all S ⊆ V , k ← 0
2: while fk (S) − z(δGk (S)) > 0 for some S ⊆ Vk do
3:
find a set S ⊆ Vk such that fk (S) − z(δGk (S)) ≥ 1, let it be Sk
4:
determine e ∈ δGk (Sk ) with maximum x(e), let it be ek ← {uk , vk }
5:
contract ek in Gk (keeping multiple copies of edges) to obtain Gk+1 , and set
(
fk (S ∪ {uk , vk } \ wk ) if wk ∈ S
fk+1 (S) ←
fk (S)
otherwise ,
6:
7:
8:
9:
10:
for all S ⊆ Vk+1 , where wk is the vertex in Gk+1 corresponding to the contracted edge ek .
k ←k+1
end while
γmin ← mink 2x(ek )/|E|, and let p be the index for which this minimum is achieved
γmax ← x(δ(Up )), where Up is the vertex subset of V corresponding to the vertex subset Sp of
Vp
return γmin , γmax
9
Lemma 3. Algorithm 3 computes an interval [γmin , γmax ], which contains the shortest row length
with respect to x(e) ∈ R+ , e ∈ E. Moreover, γmax /γmin ≤ |E|2 /2, and the algorithm runs in
strongly-polynomial time.
Proof. Algorithm 3 works as follows. It does a sequence of at most |V | iterations. In iteration k, it
takes an arbitrary subset S corresponding to a violated cut, i.e. such that fk (S) − z(δGk (S)) > 0,
and contracts the edge ek in this cut of maximum x-value. Contracting this edge naturally yields
a graph Gk+1 and a function fk+1 to use in the next iteration. Note that a violated subset S can
be computed efficiently given that f is a proper function [12].
Our first claim is that γmin is a valid lower bound on the shortest row length. In other words,
we claim that for every S : f (S) − z(δ(S)) ≥ 1, we have
x(ep )
x(δ(S))
≥
= γmin .
f (S) − z(δ(S))
|E|/2
Due to the termination condition, for every S ⊆ V with f (S) − z(δ(S)) ≥ 1 the edge set
δ(S) contains at least one of the edges e1 ,. . . ,et selected by the algorithms during its t iterations.
Therefore, x(δ(S)) ≥ x(ep ), by the choice of p in step 8. Moreover, by Remark 1 f (S) − z(δ(S)) ≤
|E|/2. The claim then follows.
Our second claim is that γmax is a valid upper bound on the shortest row length. To see this,
note that f (Up ) − z(δ(Up )) ≥ 1 because f (Up ) = fp (Sp ) and z(δ(Up )) = z(δGp (Sp )), proving that
γmax is a valid upper bound for the shortest length of a row indexed by S ⊆ V .
Finally, recalling that ep satisfies x(ep ) = maxe∈δ(Up ) x(e) (step 4), we have
γmax /γmin =
5
x(δ(Up ))
≤ |E|2 /2 .
2x(ep )/|E|
Concluding remarks
We are now ready to put all pieces together and give a proof of Theorem 1 and Corollary 1 stated
in the introduction.
Proof of Theorem 1. Given an ε > 0, we apply Algorithm 2 to (LP1 ) with ζ = ln(1 + ε)/|E|.
Algorithm 2 in its turn approximates the shortest row at most O((ln |V |)/ζ 2 ) times3 . It makes a
call to Algorithm 3, computing at most |V | Gomory-Hu trees and afterwards the binary search needs
O((ln |E|)/ζ) computations of a Gomory-Hu tree in G = (V, E). Recall, that ζ = ln(1 + ε)/|E| =
Θ(ε/|E|) and hence each linear program appearing in Algorithm 1 is solved in time dominated by
finding O(|E|3 (ln |E|)2 /ε3 ) Gomory-Hu trees. Note that a Gomory-Hu tree for G = (V, E) with
respect to values x(e) ∈ R+ , e ∈ E can be found by |V | computations of the minimum cut in G [13],
so a Gomory-Hu tree can be found in strongly-polynomial time.
The number of times our algorithm solves the Gomory-Hu tree problem is substantially smaller
than the corresponding number for the Ellipsoid method given the classical estimation for the
encoding length of vertices or given that maxS f (S) is sufficiently large, because this number for
the Ellipsoid method grows proportionally with the logarithm of maxS f (S).
Proof of Corollary 1. To obtain a (2 + ε)-approximation for (IP) we apply Algorithm 1. Algorithm 1
solves O(|E|2 ) linear programs, i.e. there are O(|E|2 ) calls of Algorithm 1 to Algorithm 2 to solve
linear programs, and it makes at most |E| roundings. Considering rounding as a basic operation,
the result follows.
3
Here, we use the inequality that 1 + ζ < eζ < 1 + 74 ζ for 0 < ζ < 1.
10
We conclude the paper with some open questions. It remains open whether one is able to
provide a 2-approximation algorithm for (IP), which does not need to solve linear programs. This
question is among the top 10 open questions in the theory of approximation algorithms according
to Shmoys and Williamson [23]. In our opinion, a good intermediate question is whether it is
possible to give an algorithm with a constant approximation guarantee such that the number of
linear programs solved in its course is bounded by a constant. One way to prove this could be to
exploit that after each rounding in the algorithm of Jain [19] we have a sufficiently “good” feasible
point for the new linear program.
References
[1] S. Arora, E. Hazan, and S. Kale. The multiplicative weights update method: a meta-algorithm
and applications. Theory of Computing, 8(1):121–164, 2012.
[2] D. Bienstock. Potential function methods for approximately solving linear programming problems: theory and practice, volume 53. Springer Science & Business Media, 2006.
[3] D. Bienstock and G. Iyengar. Solving fractional packing problems in o ast (1/ε) iterations.
In Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, pages
146–155. ACM, 2004.
[4] K. Elbassioni, K. Mehlhorn, and F. Ramezani. Towards more practical linear programmingbased techniques for algorithmic mechanism design. SAGT, pages 98–109, 2015.
[5] L. Fleischer. A fast approximation scheme for fractional covering problems with variable upper
bounds. In Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete algorithms,
pages 1001–1010. Society for Industrial and Applied Mathematics, 2004.
[6] L. K. Fleischer. Approximating fractional multicommodity flow independent of the number of
commodities. SIAM Journal on Discrete Mathematics, 13(4):505–520, 2000.
[7] H. N. Gabow, M. X. Goemans, and D. P. Williamson. An efficient approximation algorithm
for the survivable network design problem. Mathematical Programming, 82(1-2, Ser. B):13–40,
1998. Networks and matroids; Sequencing and scheduling.
[8] N. Garg and R. Khandekar. Fast approximation algorithms for fractional steiner forest and
related problems. In Foundations of Computer Science, 2002. Proceedings. The 43rd Annual
IEEE Symposium on, pages 500–509. IEEE, 2002.
[9] N. Garg and R. Khandekar. Fractional covering with upper bounds on the variables: Solving
lps with negative entries. In Algorithms–ESA 2004, pages 371–382. Springer, 2004.
[10] N. Garg and J. Könemann. Faster and simpler algorithms for multicommodity flow and other
fractional packing problems. SIAM J. Comput., 37(2):630–652, 2007.
[11] M. X. Goemans, D. B. Shmoys, A. V. Goldberg, É. Tardos, S. Plotkin, and D. P. Williamson.
Improved approximation algorithms for network design problems. In Proceedings of the Fifth
Annual ACM-SIAM Symposium on Discrete Algorithms (Arlington, VA, 1994), pages 223–232.
ACM, New York, 1994.
[12] M. X. Goemans and D. P. Williamson. A general approximation technique for constrained
forest problems. SIAM Journal on Computing, 24(2):296–317, 1995.
11
[13] R. E. Gomory and T. C. Hu. Multi-terminal network flows. J. Soc. Indust. Appl. Math.,
9:551–570, 1961.
[14] M. D. Grigoriadis and L. G. Khachiyan. Fast approximation schemes for convex programs with
many blocks and coupling constraints. SIAM Journal on Optimization, 4(1):86–107, 1994.
[15] M. D. Grigoriadis and L. G. Khachiyan. Approximate minimum-cost multicommodity flows
in õ(ǫ − 2knm) time. Mathematical Programming, 75(3):477–482, 1996.
[16] M. Grötschel, L. Lovász, and A. Schrijver. Geometric algorithms and combinatorial optimization, volume 2 of Algorithms and Combinatorics: Study and Research Texts. Springer-Verlag,
Berlin, 1988.
[17] A. Gupta and J. Könemann. Approximation algorithms for network design: A survey. Surveys
in Operations Research and Management Science, 16(1):3–20, 2011.
[18] D. S. Hochbaum. Approximation algorithms for NP-hard problems. PWS Publishing Co., 1996.
[19] K. Jain. A factor 2 approximation algorithm for the generalized Steiner network problem.
Combinatorica, 21(1):39–60, 2001.
[20] M. Luby and N. Nisan. A parallel approximation algorithm for positive linear programming.
In Proceedings of the twenty-fifth annual ACM symposium on Theory of computing, pages
448–457. ACM, 1993.
[21] F. Shahrokhi and D. W. Matula. The maximum concurrent flow problem. Journal of the ACM
(JACM), 37(2):318–334, 1990.
[22] E. Tardos. A strongly polynomial algorithm to solve combinatorial linear programs. Operations
Research, 34(2):250–256, 1986.
[23] D. P. Williamson and D. B. Shmoys. The design of approximation algorithms. Cambridge
University Press, Cambridge, 2011.
[24] N. E. Young. Randomized rounding without solving the linear program. In SODA, volume 95,
pages 170–178, 1995.
[25] N. E. Young. Sequential and parallel algorithms for mixed packing and covering. In Foundations of Computer Science, 2001. Proceedings. 42nd IEEE Symposium on, pages 538–546.
IEEE, 2001.
12
| 8cs.DS
|
arXiv:1604.02342v1 [math.AG] 8 Apr 2016
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
EDOARDO BALLICO AND ALESSANDRA BERNARDI
Abstract. Let X(R) be a geometrically connected variety defined over R and
such that the set of all its (also complex) points X(C) is non-degenerate. We
introduce the notion of admissible rank of a point P with respect to X to be the
minimal cardinality of a set of points of X(C) such that P ∈ hSi that is stable
under conjugation. Any set evincing the admissible rank can be equipped
with a label keeping track of the number of its complex and real points. We
show that in the case of generic identifiability there is an open dense euclidean
subset of points with certain admissible rank for any possible label. Moreover
we show that if X is a rational normal curve than there always exists a label
for the generic element. We present two examples in which either the label
doesn’t exists or the admissible rank is strictly bigger than the usual complex
rank.
Introduction
A very important problem in the framework of Tensor Decomposition is to understand when a given real tensor T can be written as a linear combination of real
rank 1 tensors with a minimal possible number of terms, and that number is called
the rank of T . For the applications it is also often very useful to know which are the
typical ranks, i.e. the ranks for which it is possible to find an euclidean open set of
tensors with that given rank. For the specific case of tensors, computing the rank
of T corresponds to find the smallest projective space s-secant to the real Segre
variety containing [T ]. One can clearly define all these concepts for any variety X
and saying that the real rank of P ∈ hXi is the minimal cardinality of a set of points
of X whose span contains P . In order to be as much as general as possible we will
always work with a real field K instead of over R and its real closure R. We will
indicate with C := R(i) the algebraic closure of R, but the reader not interested in
abstract fields may take R instead of R and C instead of C.
We introduce the notion of admissible rank rX,R (P ) of a point P (Definition 12)
to be the minimal cardinality of a set S ⊂ X(C) that is stable under the conjugation
action and such that P ∈ hSi. Any such a set can be labelled by keeping track of
the number of the real points. Clearly if S evinces the rank of P than the label of
S will be also a label for P .
One can ask various questions regarding those notions of admissible rank and
label of a decomposition of a point.
First of all when a point has a unique decomposition, then clearly the label is
uniquely associated to it. In this case it is even possible to prove that if the generic
2010 Mathematics Subject Classification. 15A69;14N05; 14P99.
Key words and phrases. tensor rank; symmetric tensor rank; real symmetric tensor rank.
The author was partially supported by MIUR and GNSAGA of INdAM (Italy).
1
2
EDOARDO BALLICO AND ALESSANDRA BERNARDI
element is identifiable, then the set of all real points with that rank is dense for the
euclidean topology (see Theorem 3).
A “ partial particular case ” of this situation is the rational normal curve. When
the curve has odd degree, then the generic homogeneous bivariate polynomial of
that degree is identifiable, so in this case we can apply the result just described.
But if the rational normal curve is of even degree, then Theorem 3 does not apply.
Anyway we can completely describe the situation for rational normal curves with
the following theorem.
Theorem 1. Let X ⊂ Pd , d ≥ 2, be a rational normal curve defined over R. Then
there is a non-empty open subset U ⊂ Pd (R) such that Pd (R) \ U has euclidean
dimension < d and each q ∈ U has admissible rank ⌈(d + 1)/2⌉.
The behaviur of rational normal curves is peculiar. In fact it is not always true
that when we don’t have generic identifiability one can find a label for the generic
element (remark that the identifiability of the generic element is quite rare). In
Secion 3.3 we show an example by using an elliptic curve: we can explicitly build
an euclidean neighborhood of a generic point in P3 (R) where no points in that
neighborhood have a label with respect to the Variety of Sum of Powers V SP (P )
(see Definition 16). However in this case V SP (P ) is finite and we do not have
positive dimensional examples that could be very interesting if they exist.
The last question that we like to address is what happens when the complex
rank is smaller than the admissible rank. For a given P ∈ Pr (R) consider the
set S(P, X, C) of all S ⊂ X(C) evincing rX(C) (P ). First of all it is a constructible
subset of Pr (C) invariant by the conjugation action, hence it is defined over R and
S ∈ S(P, X, C)(R) if and only if S is fixed by the conjugation
k action, i.e. if and
j
rX(C) (P )
. Therefore P has
only if S has a label (s, a) for some a with 0 ≤ a ≤
2
admissible rank rX(C) (P ) if and only S(P, X, C)(R) 6= ∅, not even for P outside a
subset of Pr (R) with euclidean dimension < r. Therefore if rX(C) (P ) < rX,R (P )
then S(P, X, C)(R) = ∅. In Example 2 we show this behavior by constructing a
very special homogeneous polynomial in n ≥ 2 variables and even degree d ≥ 6
such that rX(C) (P ) = 3d/2 < rX,R (P ), ♯(S(P, X, C)) = 2 and S(P, X, C)(R) = ∅.
Acknowledgments: We want to thank Giorgio Ottaviani for very helpful and
constructive remarks.
1. K-dense and K-typical ranks
Let K be any field with characteristic zero and let K denote its algebraic closure.
Let X be a geometrically integral projective variety defined over K such that X(K)
is Zariski dense in X := X(K). We fix an inclusion X ⊂ Pr defined over K and
such that X spans Pr , i.e. no hyperplane defined over K contains X(K). Since
X(K) is assumed to be Zariski dense in X, the non-degeneracy of X is equivalent
to assuming that X(K) spans Pr (K) over K.
Now, Pr (K) and Pr (K) have their own Zariski topology and the Zariski topology
of Pr (K) is the restriction of the one on Pr (K).
Definition 1 (X-rank and X(K)-rank). For each point P ∈ Pr (K) (resp. P ∈
Pr (K)) the X-rank rX (P ) (resp. the X(K)-rank rX(K) (P )) of P is the minimal
cardinality of a subset S ⊂ X (resp. S ⊂ X(K)) such that P ∈ hSi, where h i
denote the linear span. We say that S evinces rX (P ).
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
3
Definition 2. For each integer a > 0 let
(1)
UK,a := P ∈ Pr (K) | rX(K) (P ) = a
be the subset in Pr (K) of the points of fixed X-rank equal to a as in (1). We say
that a is a K-dense rank if the set UK,a is Zariski dense in Pr (K).
Since Pr (K) is Zariski dense in Pr (K), then UK,a is a dense subset of Pr (K).
Definition 3 (Generic X(K)-rank). The generic X(K)-rank of Pr (K) is the Xrank of the generic element of Pr (K).
The minimal K-dense rank is just the generic X(K)-rank of Pr (K) (Lemma 1).
P
Definition 4. A field K is real if x1 , . . . , xn ∈ K are such that ni=1 x2i = 0 then
xi = 0 for all i = 1, . . . , n.
We recall that a field K admits an ordering if and only if −1 is not a sum of
squares ([8, Theorem 1.1.8]). It is possible to prove (cfr [8, Chapter 4]) that when
K admits an ordering, then the filed K is real.
In the real field R the notion of typical rank has been introduced by various
authors (see e.g. [2], [3], [4], [5], [11]): An integer r is said to be a X-typical rank
if the set UR,r contains a non-empty subset for the euclidean topology of Pr (R).
We can also introduce this notion of typical rank into our setting, but we have
to assume that K is real closed.
Definition 5. A real field is closed if it does not have trivial real algebraic extensions.
We recall that for any ordering ≤ of K, there is a unique inclusion K → R of
ordered field with R real closed ([8, Theorem 1.3.2]).
If K is real closed, then the sets X(K) and Pr (K) have the euclidean topology
in the sense of [8, page 26] where the euclidean topology of K n comes from the
ordering structure of K, i.e. the euclidian topology on K n is the topology for which
open balls form a basis of open subsets (cft. [8, Definition 2.19]).
Definition 6 (Typical X(K)-rank). We say that a is a typical X(K)-rank if UK,a
contains a non-empty subset for the euclidean topology of Pr (K).
As in the case K = R the minimal typical rank is the generic rank ([5, Theorem
2]). The set of all typical ranks has no gaps, i.e. if a and b ≥ a + 2 are typical
ranks, then c is typical if a < c < b ([2, Theorem 2.2]).
Notation 1. If K is contained in a field F and S ⊆ Pr (K), let hSiF denote the
linear span of S in Pr (F ).
Lemma 1. Each set UK,a is constructible. If K is real closed, then each set UK,a
is semialgebraic.
Proof. Since UK,1 = X(K) is Zariski closed, we may assume a > 1 and that
G := ∪c<a UK,c is constructible.
There is an obvious morphism from the set E ⊂ X(K)a of all a-uple of linearly
independent points to the Grassmannian G(a − 1, r)(K) of all (a − 1)-dimensional
K-linear subspaces of Pr (K):
φ : E → G(a − 1, r)(K).
4
EDOARDO BALLICO AND ALESSANDRA BERNARDI
As usual let I := {(x, N ) ∈ Pr (K) × G(a − 1, r)(K) | x ∈ N } be the incidence
correspondence and let π1 : I → Pr (K) and π2 : I → G(a − 1, r)(K) denote the
restrictions to I of the two projections.
Now U (K, a) is the intersection with Pr (K) \ G of π1 (π2−1 (φ(E)). Obviously the
counterimage by a continuous map for the Zariski topology of a constructible set
is constructible. Use that the image of a constructible set is constructible by a
theorem of Chevalley [12, Exercise II.3.19]. If K is real closed, it is sufficient to
quote [8, Proposition 2.2.3], instead of Chevalley’s theorem.
Proposition 1. If K is a real closed field, then an integer is K-typical if and only
if it is K-dense.
Proof. Since UK,a is semialgebraic (Lemma 1), it is Zariski dense in Pr (K) if and
only if it has dimension r (or, equivalently, if and only if it contains a non-empty
open subset in the euclidean topology). Since K is infinite, Pr (K) is Zariski dense
in Pr (K). Hence UK,a is Zariski dense in Pr (K) if and only if it is Zariski dense in
Pr (K). Thus a is K-dense if and only if it is K-typical.
Remark 1. Assume K real closed and let L ⊃ K be any real closed field containing
K. By Proposition 1 the K-typical ranks of X(K) ⊂ Pr (K) and of X(L) ⊂ Pr (L)
are the same (this also may be proved directly from the Tarski-Seidenberg principle).
In particular this means that the typical ranks of real tensors and the typical rank of
real homogeneous polynomials are realized over the real closure of Q. Moreover, if
a is typical, then UK,a is dense in the interior of U (L, a) for the euclidean topology.
A very important result (see [8, Chapter 4]) is that for any real field K, there
exists a unique real closed field R such that K ⊂ R.
Theorem 2. Assume that K admits an ordering, ≤, and let R be the real closure
of the pair (K, ≤). Assume that X(K) is dense in X(R) in the euclidean topology.
Then every R-typical rank of X(R) is a K-dense rank for X(K).
Proof. Let U ⊂ Pr (R) be an open subset for the euclidean topology formed by
points with fixed X(R)-rank equal to a. Let E ⊂ X(R)a be the set of all linearly
independent a-ples (Q1 , . . . , Qa ) of distinct points. For each (Q1 , . . . , Qa ) ∈ E, we
have an (a − 1)-dimensional R-linear space h{Q1 , . . . , Qa }iR . By assumption there
is an open subset F ⊂ E in the euclidean topology such that the union Γ of all
h{Q1 , . . . , Qa }iR contains all points of U . Since X(K) is dense in X(R) in the
euclidean topology, the set E(K) is dense in E in the euclidean topology. Hence
the subset Γ′ of Γ formed by the R-linear spans of elements of E(K) is dense in Γ
and hence its closure in the euclidean topology contains U . Since the closure in the
euclidean topology of the set Γ′′ ⊂ Γ′ formed by the K-linear spans of elements of
E(K) contains Γ′ , every R-typical rank is a K-dense rank.
Remark 2. The condition that X(K) is Zariski dense in X(K) is very restrictive
if K is a number field and X has general type. For example if X is a curve of
geometric genus ≥ 2, then it is never satisfied. But it is not restrictive in the two
more important cases: tensors (where X is a product of projective spaces with the
Segre embedding) and degree d-homogeneous polynomials (where X = Pn , and the
n+d
inclusion X ⊂ P( n )−1 is the order d Veronese embedding). It applies also to
Segre-Veronese embeddings of multiprojective spaces (the so-called Segre-Veronese
varieties).
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
5
If K has an ordering ≤ and R is the real closure of (K, ≤), then X(K) is dense
in X(R) for the euclidean topology, because K n is dense in Rn for the euclidean
topology. The set of typical X-ranks may be very large ([3], [4, Theorem 1.7] and
[2, Theorem 2.2]).
2. Join, set-theoretic K-join and K-join
We describe now the situation of more than only one variety. First of all we
need to distinguish if the join of two ore more varieties is defined over Pr (K) or
over Pr (K). This will allow to introduce the notion of rank with respect to join
varieties and the “ label ” associated to a decomposition of an element. We will
label a decomposition of a point with the number of its Pr (K) elements.
Definition 7 (Join). Let X, Y ⊂ Pr (K) be integral varieties over K. We define
the join [X; Y ] of X and Y to be the closure in Pr (K) of the union of all lines
spanned by a point of X(K) and a different point of Y (K) (if X and Y are the
same point Q we set [X; Y ] := {Q}).
The set [X; Y ] is an integral variety of dimension at most min{r, dim(X) +
dim(Y ) + 1}.
Definition 8. If we have s ≥ 3 integral varieties Xi ⊂ Pr (K), 1 ≤ i ≤ s, we define
inductively their join [X1 ; · · · ; Xs ] by the formula
[X1 ; . . . ; Xs ] := [[X1 ; . . . ; Xs−1 ]; Xs ].
The join is symmetric in the Xi ’s. If X and Y are defined over K, then [X; Y ]
is defined over K and [X; Y ](K) contains the closure (in the Zariski topology) of
[X; Y ] ∩ Pr (K), but it is usually larger (even if K is real closed and X(K) and
Y (K) are dense in X(K) and Y (K)).
Definition 9 (Set theoretic K-join and K-join). Assume that Xi (K) is Zariski
dense in Xi (K) for all i. The set-theoretic K-join ((X1 ; . . . ; Xs ))K ⊆ Pr (K) of
X1 , . . . , Xs is the union of K-linear subspaces spanned by points Q1 , . . . , Qs with
Qi ∈ Xi (K) for all i.
The K-join [X1 ; . . . ; Xs ]K is the Zariski closure in Pr (K) of ((X1 ; . . . ; Xs ))K .
Take geometrically integral projective varieties Xi , i ≥ 1, defined over K and
equipped with an embedding Xi ⊂ Pr defined over K. We allow the case Xi = Xj
for some i 6= j (in the case Xi = X for all i’s we would just get the set-up of
Proposition 1 and Theorem 2). We assume that each Xi (K) is Zariski dense in
Xi (K) and that ((X1 ; . . . ; Xh ))K = Pr (K) for some h. The latter condition implies
that [X1 ; . . . ; Xh ] = Pr (K).
Definition 10 (Rank and label with respect to join). Fix a point Q ∈ Pr (K). The
rank rXi ,i≥1 (Q) is the minimal cardinality of a finite set I ⊂ N \ {0} such that there
is Qi ∈ Xi (K) ⊂ Pr (K), i ∈ I, with Q ∈ h{Qi }i∈I i. Any I as above will be called
a label of Q.
In this new setup we can re-define UK,a of (1) more generally.
Notation 2. Let UK,a denote the set of all points of Pr (K) with fixed rank a with
respect to the sequence of varieties Xi , i ≥ 1 as in Definition 10.
With this extended notion of UK,a the Definitions 2 and 6 of K-dense and typical
rank respectively can be re-stated here verbatim.
6
EDOARDO BALLICO AND ALESSANDRA BERNARDI
Remark 3. The statements and proof of Proposition 1 and Theorem 2 work verbatim in this more general setting (for the generalization of Theorem 2 we require
that each Xi (K) is dense in X(R) in the euclidean topology).
The more general set-up clearly covers the tensor cases where X is the Segre or
the Segre-Veronese variety. For example it applies to any Xi being closed subvarieties of Pn1 × · · · × Pnk and then considering the Segre embedding of the multiprojective space Pn1 × · · · × Pnk into Pr with r + 1 = Πki=1 (ni + 1). For instance, each
Xi may be a smaller multiprojective space (depending on less multihomogeneous
variables). If Xi = Pn1 ×· · ·×Pnk for sufficiently many indices, any rank is achieved
by a set I such that Xi = Pn1 × · · · × Pnk for all i, but even in this case there may
be cheaper sets J (i.e. with ♯(J) = ♯(I), but Xi ( Pn1 × · · · × Pnk for some i ∈ J).
The case of Segre-Veronese embedding of multidegree (d1 , . . . , dk ) of Pn1 × · · · Pnk
Qk
i
), is completely analogous.
(here r + 1 = i=1 nin+d
i
3. On the “ generically identifiable ” case
Let R be a real closed field as in Definitions 4 and 5 and take C := R(i) to be the
algebraic closure of R (for these fundamental facts we always refer to [8]). Now X
is a geometrically connected variety defined over R with a fiexd embedding X ⊂ Pr
and we assume that X(C) is non-degenerate, i.e. X(C) spans Pr (C).
Definition 11 (X-rank). For each P ∈ Pr (C) the X-rank of P is a minimal
cardinality of a set S ⊂ X(C) such that P ∈ hSi.
We want to consider now the “ conjugation action ” on the elements appearing
in a decomposition of P .
Notation 3. Let σ : C → C be the field automorphism with σ(x) = x for all x ∈ R
and σ(i) = −i.
Remark that the map σ just introduced acts (not algebraically) on X(C) and
Pr (C) with σ 2 as the identity and with X(R) and Pr (R) as its fixed point set.
Note that if dim(X) = n and R = R, then dim(X(C) \ X(R)) = 2n and hence,
topologically, a pair of complex conjugate points of X(C) \ X(R) “ costs ” as two
points of X(R). It seems to us that the same should be true for the algorithms
used in the case of the tensor decomposition or the decomposition of degree d
homogeneous polynomial as a sum (with signs if d is even) of d powers of linear
forms.
Definition 12 (Admissible rank). The admissible rank rX,R (P ) of a real point
P ∈ Pr (R) is the minimal cardinality of a possibly complex set S ⊂ X(C) such that
P ∈ hSi and σ(S) = S.
As we did in Definition 10, we can again keep track of the elements appearing
in a decomposition of a point by labelling those that are real.
Definition 13 (Label of S). In any finite set S ⊂ X(C) such that
k = S there
j σ(S)
♯(S)
(the other
are a pairs of σ-conjugated points of X(C) \ X(R), with 0 ≤ a ≤ 2
(♯(S) − 2a) points of S are in X(R)).
We say that (♯(S), a) is the label of S. If S evinces the admissible rank of P ,
then we say that (rX,R (P ), a) is a label of P .
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
7
Remark 4. Fix a finite set S ⊂ X(C). The set σ(S) is finite and ♯(σ(S)) = ♯(S).
It is very natural to say that S is defined over R if and only if σ(S) = S. Clearly
if σ(S) = S, σ sends bijectively S onto itself and the label a of Definition 4 is the
number of pairs {p, σ(p)}, p ∈ X(C) \ X(R), contained in S, while ♯(S) − 2a =
♯(X(R) ∩ S).
If P ∈ Pr (R) has a unique decomposition, then we can directly associate a unique
label to P itself.
Notation 4 (Label of P ). Fix P ∈ Pr (R) such that there is a unique set S ⊂ X(C)
evincing the X-rank, s, of P . Since σ(P ) = P , the uniqueness of S implies σ(S) =
S. Hence S has a label (s, a) and we say that P has label (s, a). If X(R) = ∅, then
each label is of the form (2a, a).
Definition 14. For every integer t ≥ 1 the t-secant variety Sect (X(C)) of X(C) is
the closure in Pr (C) of the set of all points with X-rank t.
The set Sect (X(C)) is an integral variety defined over R and we are interested
in its real locus σt (X(C)) ∩ Pr (R).
Definition 15. We say that Sect (X(C)) is generically identifiable if for a general
P ∈ Sect (X(C)) there is a unique S ⊂ X(C) such that ♯(S) = t and P ∈ hSi (this
notion has been already widely introduced in literature, see e.g. [7], [9], [10]).
We have C = R + Ri and hence we may see Pr (C) as an R-algebraic variety of
dimension 2r. Hence Pr (C) has an euclidean topology and this topology is inherited
by all subsets of Pr (C). This topology on Sect (X(C)) is just the euclidean topology
obtained seeing it as a real algebraic variety of dimension twice the dimension of
Sect (X(C)).
Theorem 3. Fix an integer t > 0 and assume that Secs (X(C)) is generically
identifiable. Then the set of all real points P ∈ Secs (X(C)) ∩ Pr (R) with one of the
labels (s, a), 0 ≤ a ≤ ⌊s/2⌋ is dense in the smooth part Secs (X(C)) ∩ Pr (R) for the
euclidean topology and its complementary is contained in a proper closed subset of
Secs (X(C)) ∩ Pr (R) for the Zariski topology.
Proof. We may assume s ≥ 2, because Sec1 (X(C)) = X(C). Since Secs (X(C)) is
generically identifiable, it has the expected dimension (s + 1) dim(X) − 1.
Let E be the set of all subsets of X(C) formed by s linearly independent points. For
each S ∈ E, the map S 7→ hSi defines a morphism φ from E to the Grassmannian
G(s − 1, r)(C) of all (s − 1)-dimensional C-linear subspaces of Pr (C). Let I :=
{(x, N ) ∈ Pr × G(s − 1, r) | x ∈ N } be the incidence correspondence and let
π1 : I → Pr (C) and π2 : I → G(s − 1, r)(C) denote the restriction to I of the two
projections. The set UC,a of all points with rank exactly equal to s is the intersection
with {P ∈ Pr (C) | rk(P ) ≥ s} with π1 (π2−1 (φ(E)) and hence it is constructible.
Since X is real and the embedding X ⊂ Pr is defined over R, we have σ(UC,a ) =
UC,a . By assumption there is a non-empty open subset VC,s of UC,s for the Zariski
topology such that each P ∈ VC,s comes from a unique point of π1 (π2−1 (E)) and in
particular it is associated to a unique SP ∈ E. Since the embedding is real, each
point of σ(VC,s ) has the same property. Since VC,s is Zariski dense in Secs (X(C)),
VC,s ∩ Pr (R) is Zariski dense in Secs (C) ∩ Pr (R). Fix any P ∈ VC,s ∩ Pr (R). Since
SP is uniquely determined by P and σ(P ) = P , we have σ(SP ) = SP . Hence SP
′
′
has a label. Let UC,s
be the subset of UC,s
corresponding only to sets S ⊂ Xreg (C)
8
EDOARDO BALLICO AND ALESSANDRA BERNARDI
′
and at which the map π1 has rank s(dim(X) + 1) − 1. Note that UC,s
is contained
′
′
′
in the smooth part of Secs (C). Set VC,s := UC,s ∩ VC,s . Since VC,s
is smooth,
′
′
VC,s
∩ Pr (R) is a smooth real algebraic variety. We saw that each point of VC,s
′
has a label (s, a). VC,s
is dense in the smooth part of the real algebraic variety
Secs (X(C))(R) = Secs (X(C)) ∩ Pr (R) for the euclidean topology and hence it is
Zariski dense in Secs (X(C)) and Secs (X(C)) ∩ Pr (R).
Remark 5. When the set S(P, X, C) of all S ⊂ X(C) evincing rX(C) (P ), is finite,
in order to have S(P, X, C)(R) 6= ∅ it is sufficient that ♯(S(P, X, C)(R)) is odd (we
use this observation to prove Theorem 1 for odd d).
There are many uniqueness results for submaximal tensors (see for example [7],
[9], [10]) and so Theorem 3 applies in many cases.
A first interesting case where to study the admissible rank is the one of rational
normal curves. If the degree if the curve is odd, then a general point is identifiable so
Theorem 3 assures the existence of a dense set of points with any label (1 + d/2, a),
while if d is even, the euclidean dimension of points with admissible rank ⌈(d+1)/2⌉
has to be studied. This is the purpose of Theorem 1. We need to recall the following
definition.
Definition 16. Fix any P ∈ V and let V SP (P ) denote the set of all S ⊂ X(C)
such that P ∈ hSiC and ♯(S) = ρ.
Proof of Theorem 1: If d is odd, then a general P ∈ Pd (C) has rank ⌊(d + 1)/2⌋ and
Sec⌊(d+1)/2⌋ (X(C)) is generically identifiable ([13, Theorem 1.40]). Thus we may
apply Theorem 3 in this case.
Now assume that d is even. In this case a general q ∈ Pd (C) has rank 1 + d/2
and the set V SP (q) has dimension 1. Fix p ∈ X(R).
Claim : There is a non-empty open subset U of Pd (R) such that Pd (R) \ U
has euclidean dimension < d and each q ∈ U has admissible rank 1 + d/2 with label
(1 + d/2, a) and 2a ≤ d/2, computed by a set S with σ(S) = S and p ∈ S.
Proof of the Claim: If d = 2 we can simply take as U the subset of P2 (R) \
X(R) formed by the points that are not on the tangent line to X(R) at p.
Therefore assume d ≥ 4. Let ℓ : Pd (C) \ {p} → Pd−1 (C) denote the linear
projection from p. Let Y (C) denote the closure of ℓ(X(C) \ {p}). Since p ∈ Pd (R),
ℓ is defined over R and Y (C) is defined over R. By construction the curve Y (C) is
a rational normal curve of degree d − 1 and Y (C) \ ℓ(X(C) \ {p}) is a unique point,
p′ , corresponding to the tangent line of X(C) at p. Since p ∈ X(R) and ℓ is defined
over R, then p′ ∈ Y (R).
Now d − 1 is odd, so Sec⌊d/2⌋ (Y (C)) is generically identifiable and we can use
Theorem 3 to find a non-empty open subset V ⊂ Pd−1 (R) such that Pd−1 (R) has
euclidean dimension ≤ d−2, each q ∈ V has admissible rank d/2 and ♯(V P S(q)) = 1
for all q ∈ V. Since d/2 > 1, there is an open subset V ′ of V such that p′ ∈
/ Sq
for all q ∈ V ′ , where Sq is the unique element of V SP (q). Note that V \ V ′ has
euclidean dimension ≤ d − 2 and so Pd−1 (R) \ V ′ has euclidean dimension ≤ d − 2.
Now we have simply to lift up V ′ : consider U ′ := ℓ−1 (V ′ ), the set Pr (C) \ U ′ has
clearly euclidean dimension ≤ d − 1. Fix any a ∈ U ′ , call b := ℓ(a) and take
{Sb } := V P S(b). We have σ(Sb ) = Sb and p′ ∈
/ Sb . Since p′ ∈
/ Sb , there is a unique
set Sa ⊂ X(C) \ {p} such that ℓ(Sa ) = Sb . Now the set S we are looking for is
nothing else than S := {p} ∪ Sa . In fact since σ(Sb ) = Sb , ℓ is defined over R and
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
9
p ∈ X(R), we have σ(S) = S. Hence each q ∈ U ′ has admissible rank ≤ 1 + d/a.
To get U it is sufficient to intersect U ′ with the set of all q ∈ Pr (R) with C-rank
1 + d/2.
Landsberg and Teitler gave an upper bound concerning the X-rank over C ([14,
Proposition 5.1]). Several examples in [4] show that this upper bound is not always
true, not even for typical ranks, over R. The case for labels is easier and we adapt
the proof of [14, Proposition 5.1] in the following way.
Proposition 2. Let X be a geometrically integral variety defined over R and
equipped with an embedding X ⊂ Pr defined over R and of dimension m.
• If either r − m + 1 is even or X(R) is Zariski dense in X(C), then each
P ∈ Pr (R) has a label (s, a) with s ≤ r − m + 1.
• If r − m + 1 is odd and X(R) is not Zariski dense in X(C), then P has a
label (s, a) with either s ≤ r − m + 1 or s = r − m + 1 and a = 0.
Proof. Fix P ∈ Pr (R). If P ∈ X(R) it has (1, 0) as its unique label.
Now assume P ∈
/ X(R). Let U be the set of all linear spaces H ⊂ Pr (C) defined over
C, containing P and transversal to X(C). By Bertini’s theorem, U is a non-empty
open subset of the Grassmannian GC := G(r−m−1, r−1) of all (r−m)-dimensional
linear subspaces of Pr (C) containing P . Since X and P are defined over R, GC is
also defined over R.
By definition of U , for each H ∈ U we have that P ∈ H and H intersects X(C) in
deg(X) distinct points. There is a non-empty open subset V of U such that every
H ∈ V has the following property: any S1 ⊆ H ∩ X(C) with ♯(S1 ) = r − m + 1
spans H. Since GC is a Grassmannian, this implies that V (R) is Zariski dense in
V and in particular V (R) 6= ∅. Take H ∈ V (R). The set S := H ∩ X(C) is formed
by deg(X) points of X(C) and σ(S) = S.
Assume for the moment that either r − m + 1 is even or S ∩ X(R) 6= ∅. We may
find S1 ⊆ S such that ♯(S1 ) = r − m + 1 and σ(S1 ) = S1 . Hence S1 is a set with a
label (r − m + 1, a) for some integer a.
If r−m+1 is odd and H∩X(R) = ∅, then we have a set S2 ⊆ S with ♯(S2 ) = r−m+2
with σ(S2 ) = S2 and hence P has a label (r − m + 2, (r − m + 2)/2).
Finally if r − m + 1 is odd and X(R) is Zariski dense in X(R), the set of all
H ∈ V (R) with H ∩ X(R) 6= ∅ is non-empty (and Zariski dense in V ).
3.1. Typical a-ranks. As above, let X ⊂ Pr be a geometrically integral nondegenerate variety defined over R and with a smooth real point, i.e. with X(R)
Zariski dense in X(C).
Notation 5. We use h i or h iC to denote linear span over C inside Pr (C). For any
S ⊆ Pr (R) let hSiR be its linear span in Pr (R). Note that hSiC ∩ Pr (R) = hSiR .
Here we consider the case in which Secs (X) = Pr and hence almost never we
have generic uniqueness and we won’t be able to apply Theorem 3. The notion of
typical rank may be generalized in the following way.
Fix an integer a ≥ 0. For any set of complex points P1 , . . . , Pa ∈ X(C) \ X(R)
the linear space L := h{P1 , σ(P1 ), . . . , Pa , σ(Pa )}iC is defined over R and hence
L ∩ Pr (R) is a linear space with dimR L ∩ Pr (R) = dimC L and L = (L ∩ Pr (R))C .
Definition 17 (a-X(R)-rank). For any P ∈ Pr (R), the a-X(R)-rank of P is the
minimal integer c such that there are P1 , . . . , Pa ∈ X(C) \ X(R) and Q1 , . . . , Qc ∈
X(R) such that P ∈ h{P1 , σ(P1 ), . . . , Pa , σ(Pa ), Q1 , . . . , Qc }iC .
10
EDOARDO BALLICO AND ALESSANDRA BERNARDI
The a-typical X(R)-ranks are the integers occurring as a-ranks a non-empty
euclidean open subset of Pr (R).
Note that 0 is typical if and only if Seca (X(C)) = Pr (C). The proof of [2,
Theorem 1.1] easily prove the following result.
Proposition 3. All the integers between two different a-typical ranks are a-typical
ranks.
One may wonder if a label for a general element in Pr (R) of certain typical rank
always exists. The answer is “ no ” and we present an example in the following
section by using varieties of sum of powers.
3.2. Variety of Sum of Powers. We indicate with ρ the generic X-rank of Pr (C)
and with U ⊂ Pr (C) a non-empty open subset of Pr (C) of points of generic rank:
rX(C) (P ) = ρ for all P ∈ U. Since X is defined over R, also all point P ∈ σ(U)
have the same generic rank. Now if
U := U ∩ σ(U) and V := U ∩ Pr (R),
then the real part V of U is a non-empty open subset of Pr (R) whose complement
has dimension smaller than ρ. For all the points P of generic rank, the variety of
sum of powers V SP (P ), recalled in Definition 16, is non-empty. Since P ∈ Pr , we
have σ(V SP (P )) = V SP (P ) and hence the constructible set V SP (P ) is defined
over R. Now the real part of V SP (P ), V SP (P )(R) is non empty if and only if
there is an integer a with 0 ≤ a ≤ ρ/2 and P has a label (ρ, a). In a few cases
V SP (P ) is known (see e.g. [6], [7], [10]).
The following example says that V SP (P ) may not have a label for a general P ∈
Pr (R) with generic rank over C. However in this case V SP (P ) is finite and we do not
have positive dimensional examples. If V SP (P ) is finite, we have V SP (P )(R) 6= ∅
(and so P has a label) if ♯(V SP (P )) is odd.
Example 1. Let C be a smooth elliptic curve defined over R. Take two points
P1′ , P2′ ∈ C(C) \ C(R) and set Pi′′ := σ(Pi′ ). By assumption Pi′′ 6= Pi′ . For general
P1′ , P2′ , we have P1′′ 6= P2′ and so P1′ 6= P2′′ .
Let E be a geometrically integral curve of arithmetic genus 3 and with exactly 2
singular points O′ and O′′ , both of them ordinary nodes, where C has its normalization and with O′ obtained gluing together the set {P1′ , P2′ } and O′′ obtained gluing
together the set {P1′′ , P2′′ }. The involution σ acts on E with σ(O′ ) = O′′ . For general P1′ , P2′ the divisors P1′ + P2′ and P1′′ + P2′′ are not linearly equivalent and hence
E is not hyperelliptic. Hence ωE is very ample and embeds E in P2 (C) as a degree
4 curve with 2 ordinary nodes, Q1 , Q2 , and no other singularity. Since σ(O′ ) = O′′
and O′ 6= O′′ , we have that for i = 1, 2, σ(Qi ) = Q3−i and Qi ∈ P2 (C) \ P2 (R).
Let u : C → E denote the normalization map. Set OC (1) := u∗ (ωE ). Since
ωC ∼
= OC , we have P1′ + P1′′ + P2′ + P2′′ ∈ |OC (1)|. Thus OC (1) is a degree 4 line
bundle on C defined over R and the complete linear system |OC (1)| induces an
embedding j : C → P3 (C) defined over R and with X := j(C) a smooth elliptic
curve of degree 4. The pull-back by u of the linear system |ωE | is a codimension 1
linear subspace of the 3-dimensional C-linear space |OC (1)|, the one associated to
the C-vector space of all rational 1-forms on C with poles only at the points P1′ , P2′ ,
P1′′ , P2′′ , each of them at most of order one and such that the sum of their residues at
the 4 points P1′ , P2′ , P1′′ , P2′′ is zero. Since σ({P1′ , P2′ , P1′′ , P2′′ }) = {P1′ , P2′ , P1′′ , P2′′ },
this 2-dimensional linear subspace of |OC (1)| is defined by one equation ℓ = 0 with
TYPICAL AND ADMISSIBLE RANKS OVER FIELDS
11
σ(ℓ) = ℓ, i.e. it is defined over R. Since this is the linear system |ωE | whose image
is the curve that we indicate with D, then D is defined over R. Since D is defined
over R, the curve D is the image of X(C) by the linear projection from a point
P ∈ P3 (R)\ X(R). By construction X has exactly 2 secant lines passing through Q
and they are the lines L′ , L′′ with L′ spanned by j(P1′ ) and j(P2′ ) and L′′ spanned
by j(P1′′ ) and j(P2′′ ). Now, since L′ 6= L′′ , we get that Q has no label. Now take
P ∈ P3 (R) near Q in the euclidean topology. The linear projection from Q is a
geometrically integral curve Dp defined over R, with C as its normalization and
near D and so with two ordinary nodes Q1 (P ) and Q2 (P ) near Q1 and Q2 . Thus
none of these nodes is defined over R. Therefore there is an euclidean neighborhood
V of Q in P3 (R) such that no P ∈ V has a label.
Remark 6. If V SP (P ) is finite, we have V SP (P )(R) 6= ∅ (and so P has a label)
if ♯(V SP (P )) is odd. If r = 3 and X is a smooth curve of genus g and degree
d the genus formula for plane curves gives ♯(V SP (P )) = (d − 1)(d − 2)/2 − g.
When S(P, X, C) is infinite, it is not clear that at least one irreducible component
of S(P, X, C) is σ-invariant and hence defined over R.
3.3. Polynomials with admissible rank bigger than complex rank. We use
elliptic curves and [6, Example 3.4] to construct an example of a pair (X, P ) with
♯(S(P, X, C)) = 2 and S(P, X, C)(R) = ∅ with P a symmetric tensor and with
X a d-Veronese embedding of Pn , n ≥ 2 and d even (here r = n+d
− 1) and
n
rX(C) (P ) = 3d/2.
Example 2. We consider the Veronese variety X of dimension n ≥ 2 embedded
n+d
into P( n )−1 with the complete linear system |O(d)| with d ≥ 6 and d even. Set
k := d/2. The projective space Pn and the embedding are defined over R.
Let E be a smooth curve of genus 1 defined over R and with E(R) 6= ∅ (so if
R = R, then E(R) is homeomorphic either to a circle or to the disjoint union of 2
circles). Fix O ∈ E(R). By Riemann-Roch the complete linear system |OE (3O)|
defines an embedding j : E → P2 defined over R and with j(E) a smooth plane
cubic. With an R-linear change of homogeneous coordinates x, y, z we may assume
that j(O) = (0 : 1 : 0) and that j(E) = {y 2 z = x3 +Axz 2 +Bz 3} for some A, B ∈ R.
The restriction to j(E) \ {O} of the linear projection P2 \ {O} → P1 induces the
morphism φ : E → P1 which is induced by the complete linear system |OE (2O)|.
Since O ∈ E(R), φ is defined over R. If t ∈ R and −t ≫ 0, then t3 + At + B < 0
and so y 2 = t3 + At + B has two solutions q ′ , q ′′ ∈ C with q ′′ = σ(q ′ ). Hence we may
find S ′ = {q1 , . . . , q3k } ⊂ E(C) \ E(R) such that ♯(S ′ ∪ σ(S ′ )) = 6k and qi + σ(qi ) ∈
P3k
P3k
|OE (2O)| for all i. Note that i=1 qi + i=1 σ(qi ) ∈ |OE (6kO)| = |OE (3dO)|.
Now fix a plane M ⊆ Pn defined over R and identify the P2 where j(E) is
embedded with M so that j(E) ⊂ Pn . If νd is the Veronese embedding that maps Pn
n+d
to X, we can now embedd E into X ⊂ Pr:=( n )−1 and define C := νd (j(E)), Q :=
νd (j(O)), S := νd (S ′ ) and consider the space spanned by C: ΛC := hCiC . Since the
smooth plane curve j(E) is projectively normal, the embedding E → ΛC is induced
by the complete linear system |OC (3dQ)| and so dimC ΛC = 3d − 1 = 6k − 1. Since
νd , the inclusion M ⊆ Pn and j are defined over R, ΛC ∩ Pr (R) is an R-projective
space of dimension 6k − 1. Let Z ⊂ C(C) be any set with a := ♯(Z) ≤ 6k. Since C
is embedded in ΛC by the complete linear system |OC (6kQ)|, Riemann-Roch gives
that dimC (hZiC ) = a − 1, unless a = 6k and Z ∈ |OC (6kQ)|. If Z ∈ |OC (6kQ)|,
then hZiC is a hyperplane of ΛC . Hence dimC (hSi) = dim(hσ(S)i) = 3k − 1 and
12
EDOARDO BALLICO AND ALESSANDRA BERNARDI
dimC (hS ∪ σ(S)i) = 3k − 2. By the Grassmann’s formula the set hSi ∩ hσ(S)i is a
unique point, P . Since σ(P ) ∈ hSi ∩ hσ(S)i, we have σ(P ) = P , i.e. P ∈ Pr (R).
Since P ∈ hSiC and C ⊂ X, we have rX(C) (P ) ≤ 3k. Over C this is the construction
of [6, Example 3.2]. In [6, Example 3.2] it is proved that rX(C) (P ) = 3k and
that ♯(S(P, X, C)) = 2. Hence S(P, X, C) = {S, σ(S)}. Since S 6= σ(S), we have
S(P, X, C)(R) = ∅ and hence P has admissible rank > 3k.
References
[1] E. Ballico and L. Chiantini, Sets computing the symmetric tensor rank, Mediterr. J. Math.
10 (2013), 643–654.
[2] A. Bernardi, G. Blekherman and G. Ottaviani, On real typical ranks, arXiv:1512.01853.
[3] G. Blekherman, Typical real ranks of binary forms, Found. Comput. Math. 15 (2015) , 793–
798.
[4] G. Blekherman and R. Sinn, Real rank with respect to varieties, arXiv:1511.07044.
[5] G. Blekherman and Z. Teitler, On maximum, typical and generic ranks, Math. Ann., 362
(2015), 1021–1031.
[6] C. Bocci and L. Chiantini, On the identifiability of binary Segre products, J. Algebraic
Geometry 22 (2011), no. 1, 1–11.
[7] C. Bocci, L. Chiantini and G. Ottaviani, Refined methods for the identifiability of tensors,
Ann. Mat. Pura Appl. (4) 193 (2014), no. 6, 1691–1702.
[8] J. Bochnak, M. Coste and M.-F. Roy, Real Algebraic Geometry, Springer, Berlin, 1998.
[9] L. Chiantini and G. Ottaviani, On generic identifiability of 3-tensors of small rank, SIAM J.
Matrix Anal. Appl. 33 (2012), no. 3, 1018–1037.
[10] L. Chiantini, G. Ottaviani and N. Vannieuwenhoven, An algorithm for generic and low-rank
specific identifiability of complex tensors, SIAM J. Matrix Anal. Appl. 35 (2014), no. 4,
1265–1287.
[11] P. Comon and G. Ottaviani, On the typical rank of real binary forms, Linear Multilinear
Algebra 60 (2011), no. 6, 657–667.
[12] R. Hartshorne, Algebraic Geometry, Springer, Berlin, 1977.
[13] A. Iarrobino and V. Kanev, Power sums, Gorenstein algebras, and determinantal loci, Lecture
Notes in Mathematics, vol. 1721, Springer-Verlag, Berlin, 1999, Appendix C by Iarrobino and
Steven L. Kleiman.
[14] J. M. Landsberg and Z. Teitler, On the ranks and border ranks of symmetric tensors, Found.
Comput. Math. 10 (2010), no. 3, 339–366.
Dept. of Mathematics, University of Trento, 38123 Povo (TN), Italy
E-mail address: [email protected]
Dept. of Mathematics, University of Trento, 38123 Povo (TN), Italy
E-mail address: [email protected]
| 0math.AC
|
arXiv:1603.04068v4 [cs.DB] 22 Jun 2017
A Signaling Game Approach to Databases Querying and
Interaction
Ben McCamish
Arash Termehchy
Behrouz Touri
Oregon State University
Oregon State University
University of Colorado Boulder
[email protected]
[email protected]
[email protected]
ABSTRACT
As most users cannot precisely express their information needs in
form of queries, it is challenging for database management systems
to understand the true information needs behind users’ queries.
Query interfaces leverage user’s feedback on the returned answers
for a query to improve their understanding of the true information
need behind the query. Current query interfaces generally assume
that a user follows a fixed strategy of expressing her information
needs, that is, the likelihood by which a user submits a query to express a certain information need remains unchanged over a potentially long period of time. Nevertheless, users may learn from their
interactions with the database system and gradually choose more
precise queries to express their intents. In this paper, we introduce
a novel formal framework that models database querying as a collaboration between two active and potentially rational agents: the
user and the database management system. These agents follow the
identical interest of establishing a mutual language for representing
information needs. We formalize this collaboration as a signaling
game, where each mutual language is an equilibrium for the game.
A query interface is more effective if it establishes a less ambiguous
mutual language faster. We explore some important characteristics
of the equilibria of the game. Using an extensive empirical analysis over a real-world query workload, we show that users follow a
reinforcement learning method to improve the articulation of their
information needs. We also propose and analyze a reinforcement
learning mechanism for the database query interface. We prove
that this adaptation mechanism for the query interface improves
the effectiveness of answering queries stochastically speaking, and
converges almost surely, for both the cases where users follows a
fixed and reinforcement learning strategy for expressing their information needs.
1.
INTRODUCTION
Most users do not know the structure and/or content of
databases and cannot precisely express their queries [35, 13,
21, 34]. Hence, it is challenging for a database management system (DBMS), to understand and satisfy users’ information needs. Database researchers have proposed models, methods and systems to help users specify their queries
more precisely and DBMS understand users’ intents more
accurately [44, 34, 13, 46, 35, 25, 8, 22, 37, 21, 2, 9, 52].
Current models mainly focus on improving user satisfaction
for a single information need. Given a user’s information
need e, the DBMS estimates e by various methods including
showing potential results to the user and collecting her feedback [40, 17, 9, 52, 60], asking questions from her [2], or
suggesting potential queries to her [36].
Nevertheless, many users may explore a database to find
answers for various queries and information needs over a
rather long period of time. For these users, database querying is an inherently interactive and continuing process. This
setting extends the problem of answering a single information need in two aspects. First, the DBMS can improve its
understanding of how the user expresses her intents progressively over the course of many queries. The DBMS may
learn the desired answers for a submitted query q from the
user’s explicit or implicit feedback on the returned results.
Types of the user’s feedback may include clicking on the relevant answers [30, 65], the amount of time the user spends
on reading the results [29], or user’s eye movements [33].
Second, the user may leverage her experience from previous
interactions with the DBMS to formulate her future queries.
As the user submits queries and inspect or consume their results, she may gain a better understanding of the database
content, which may impact how she formulates queries in
the future. For example, to find the answers for a particular information need, the user may submit some initial and
underspecified query, observes its results, and reformulate it
according to her observations.
Ideally, we would like the user and DBMS to gradually develop some degree of mutual understanding over the course
of several queries and interactions: the DBMS should better
understand how the user expresses her intents and the user
may get more familiar with the content of the database. Of
course, the user and DBMS would like to establish a perfect
or near-perfect mutual understanding, where the DBMS returns all or a majority of the desired answers to all or most
user queries. An important question is if there are inherent
limits to establishing these mutual understandings. In other
words, one would like to know the degrees to which these
mutual understandings are possible. One would also like explore the characteristics of information needs, queries, and
behavior of the DBMS in such limits. Subsequently, one
may leverage these results to design methods and strategies
for DBMS to help establishing more effective mutual understandings in rather small amounts of time.
Unfortunately, current models assume that a user follows
a fixed strategy of expressing her information needs in form
of queries. In other words, they maintain that a user submits a query with a fixed probability to express an intent
throughout her interaction with the database system. Hence,
one cannot use current models and techniques to answer the
aforementioned questions and help the database system to
understand the users’ information need in rather long periods of interaction. Furthermore, it has been shown that
certain learning strategies that may be useful in a static setting do not converge to a desired outcome in a setting where
both agents modify their strategies [24]. Thus, current query
interfaces may not be able to effectively satisfy users’ information needs in a long-term interaction. In this paper, we
propose a novel framework that formalizes the collaboration
between user and the DBMS over the course of many interactions. Our framework models this collaboration as a game
between two active and potentially rational agents: the user
and DBMS. The common goal of the user and DBMS is
to reach a mutual understanding on expressing information
needs in the form of queries. The user informs the DBMS
of her intents by submitting queries. The DBMS returns
some results for the query. Both players receive some reward
based on the degree by which the returned answers satisfy
the information need behind the query. The equilibria and
stable states of the game model the inherent limits in establishing a mutual understanding between the user and DBMS.
We further explore the strategies that a DBMS may adapt to
improve user satisfaction. We believe that this framework
naturally models the long term interactions between the user
and DBMS. More specifically, we make the following contributions:
• We model the long term interaction between the user
and DBMS as a particular type of game called a signaling game in Section 2 [15, 49]. The user’s strategy
in the game is a stochastic mapping from her intents to
queries, which reflects the user’s decision in choosing
a query to express an information need. The DBMS
strategy is a stochastic mapping from queries to results.
After each interaction, the DBMS may update its strategy based user feedback and/or the user may update its
strategy based on the returned results.
• We analyze the equilibria of the game and show that
the game has some Nash and strict Nash equilibria in
Section 3. We find the number of Nash and strict Nash
equilibria in some interesting settings. We also show
that the Nash and strict Nash equilibria of the game
may not provide a desired level of mutual understanding between the user and DBMS. Hence, the game may
stabilize in an undesirable state.
• Using extensive empirical studies over a real-world query
workload, we show that users’ learning can be accurately modeled by Roth and Erv’s reinforcement learning algorithm [51] in Section 4.
• We propose a reinforcement learning algorithm that
updates the DBMS strategy in Section 5 and prove that
it improves the effectiveness of answering queries stochastically speaking, and converges almost surely. Our proof
shows that the proposed learning rule is robust to the
choice of effectiveness measure which is a highly desirable property. We show that our results hold for both
the cases where the user adapts her strategy using Roth
and Erv’s reinforcement learning algorithm a the case
where she follows a fixed strategy.
2. FORMAL FRAMEWORK
Users and DBMSs typically achieve a common understanding gradually and using a querying/feedback paradigm. After submitting each query, the user may revise her strategy
of expressing intents based on the returned result. If the returned answers satisfy her intent to a large extent, she may
keep using the same query to articulate her intent. Otherwise, she may revise her strategy and choose another query
to express her intent in the hope that the new query will provide her with more relevant answers. The user may also
inform the database system about the degree by which the
returned answers satisfy the intent behind the query using
explicit or implicit feedback, e.g., click-through information
[29]. The database system may update its interpretation of
the query according to the user’s feedback. Intuitively, one
may model this interaction as a game between two agents
with identical interests in which the agents communicate via
sharing queries, results and feedback on the results. In each
interaction, both agents will receive some reward according to the degree by which the returned result for a query
matches the intent behind the query. The user will receive
her rewards in form of answers relevant to her intent and
database system will receive its reward through getting positive feedback on the returned results. The final goal of both
agents is to maximize the amount of reward they receive during the course of their interaction. They will get the maximum reward if they establish a common mapping between
intents and queries.
2.1 The Data Interaction Game
Next, we describe the components and structure of the
data interaction game for relational databases. A schema
S is a set of relation symbols. Each relation symbol R has
a corresponding arity. Let dom be an infinite set constants.
A database instance I over S associates with every n-ary
relation symbol R ∈ S a finite subset IR of domn .
2.1.1 Intent
An intent represents an information need sought after by
the user. Current usable query interfaces over databases usually assume that each intent is a query in a sufficiently expressive query language in the domain of interest, e.g., SQL,
over the underlying database instance [14, 35, 21]. Our
framework and results are orthogonal to the language that
precisely describes the users’ intents. In our examples, we
consider the intent language to be conjunctive queries [1].
Table 1 illustrates a database instance with schema Univ(Name,
Abbreviation, State, Type, Ranking) that contains information about university rankings. A user may want to find
the ranking for university MSU in Michigan, which is precisely represented by the intent e2 in Table 2(a): ans(z) ←
U niv(x, ‘M SU ’, ‘M I’, y, z).
2.1.2 Query
Users’ articulations of their intents are queries. Users do
not often know relatively complex languages, such as SQL
or Datalog, that precisely describe their intents. Thus, they
prefer to articulate their intents in languages that are easy-touse and relatively less complex, such as keyword query language [35, 14, 21]. In this instantiation of signaling games
for database querying, we assume that the user expresses her
intents as keyword queries. More formally, we fix a countably infinite set of terms T . A keyword query (query for
short) is a nonempty (finite) set of terms in T . Consider the
database instance in Table 1. Table 2 depicts a set of queries
and intents over this database. Suppose the user wants to
find the ranking for Michigan State University in Michigan,
i.e. the intent e2 . Because the user may not know how
to write conjunctive queries and not be sufficiently familiar
with the content of the database, she may express intent e2
using query q2 : ‘MSU’. Some users may know and submit
their queries in the language that is sufficiently expressive to
represent their intents. Nevertheless, because they may not
know precisely the content and structure of the data, their
submitted queries may not always be the same as their intents [12, 36]. For example, a user may know how to write a
union of conjunctive query. But, since she may not know the
state abbreviation MI, she may articulate intent e2 as ans(t)
← U niv(x, ‘M SU ’, y, z, t), which is different from e2 . Our
framework can be extended for these scenarios.
Table 3: Two strategy profiles
3(a) A strategy profile
e1
e2
e3
q1
q2
q1
0
0
0
e1
0
0
q2
1
1
1
e2
1
1
3(b) Another strategy profile
e1
e2
e3
e3
0
0
q1
q2
q1
0
1
0
e1
0
0.5
q2
1
0
1
e2
1
0
e3
0
0.5
of user’s intents and the language in which those intents are
precisely expressed. The DBMS usually selects a query language for the interpreted intents according to the information in the domain of interest [21, 31, 14]. For example,
some keyword query interfaces over relational databases interpret keyword queries as Select-Project-Join SQL queries
with only conjunctions in the where clause [31, 42]. We denote the language used by the DBMS for the interpreted intents over the underlying database instance as E. The DBMS
2.1.3 User Strategy
observes
the set of queries user submits during the interacThe user strategy indicates the likelihood by which the
tion Qu . The interpretation of DBMS on queries that are not
user submits query q given that her intent is e. In practice,
in the set of submitted queries does not impact its interaction
a user has finitely many intents and submits finitely many
with the user. Furthermore, given a query, the DBMS exqueries over a database in a finite period of time. Hence,
plores finitely many alternative intents to interpret the query
we assume the sets of the user’s intents and queries are fiefficiently [21, 31, 14]. Thus, the DBMS strategy, denoted
nite. We index each user’s intent and each user’s query by
as D, is a mapping from Qu to a finite subset of E, denoted
1 ≤ i ≤ m and 1 ≤ j ≤ n, respectively. A user strategy,
as Ed . We index each intent in Ed by 1 ≤ ℓ ≤ o. Condenoted as U , is a m × n row-stochastic matrix from her set
sider the intents and queries in Table 2. The matrix on the
of intents to her set of queries. Consider again the intents
right hand side of Table 3(a) illustrates a DBMS strategy in
and queries in Table 2. The matrix on the left hand side of
which the DBMS interprets both q1 and q2 as e2 . The inTable 3(a) depicts a user strategy in which the user submits
tent language used by the DBMS in this example is the set
query q1 to express intents e1 , e2 , and e3 .
of conjunctive queries. DBMSs often implement their strategies using a real-valued function called scoring function over
Table 1: A database instance of relation Univ
pairs of queries and intents [45, 14, 21]. Given a query, the
Name
Abbreviation State Type
Rank DBMS selects the intent with the maximum score, executes
the intent on the database, and returns the resulting tuples. It
Missouri State University
MSU
MO
public 20
is shown, however, that such a deterministic mapping may
Mississippi State University MSU
MS
public 22
limit the effectiveness of interpreting queries in long-term
Murray State University
MSU
KY
public 14
interactions [30, 65, 62, 5]. Because the DBMS shows only
Michigan State University
MSU
MI
public 18
the result of the intent(s) with the highest score(s) to the user,
its learning remains largely biased toward the highly scored
Table 2: Queries and Intents
interpreted intents. To learn more about the user’s intents,
the DBMS has to show and get feedback on a sufficiently di2(a) Intents
verse set of intents. Of course, the DBMS should ensure that
this set of intents are relatively relevant to the query, otherIntent#
Intent
wise the user may become discouraged and give up querye1
ans(z) ← U niv(x, ‘M SU ’, ‘M S’, y, z)
ing. Because such a set of intents may be relatively large, it
e2
ans(z) ← U niv(x, ‘M SU ’, ‘M I’, y, z)
may not be useful to show all of them to the user during a
e3
ans(z) ← U niv(x, ‘M SU ’, ‘M O’, y, z)
single interaction. Hence, the DBMS may adopt a stochas2(b) Queries
tic strategy:it may randomly select and show the results of
intents such that the ones with higher scores are more often
Query#
Query
chosen [30, 65, 62, 5].
q1
‘MSU MI’
Researchers have shown that it is possible to find the stochasq2
‘MSU’
tic strategy that results in returning sufficiently many answers to keep users engaged and sufficiently diverse results
2.1.4 DBMS Strategy
to collect useful feedback and improve significantly the effectiveness of query answering [62, 30]. Hence, we redefine
The DBMS interprets the submitted queries to find the inthe DBMS strategy as an n × o row-stochastic matrix from
tents behind them. A DBMS may not exactly know the set
the set of queries in Qu to the set of interpreted intents Ed .
Consider again the intents and queries in Table 2. The matrix
on the right hand side of Table 3(b) depicts a DBMS strategy
in which the DBMS interprets query q2 as intent e1 and e2
each with probability 0.5.
2.1.5 Interaction & Reward
The data interaction game is a repeated game with identical interest between two players, the user and the database
system. At each round of the game, i.e., a single interaction,
the user selects the intent eu according to the prior probability distribution π. She then picks the query q according
to her strategy and submits it to the DBMS. The DBMS observes q and interprets q as intent ed based on its strategy,
executes ed on the underlying database, and returns its result
to the user. The user provides some feedback on the returned
tuples and informs the DBMS how relevant the result is to
her intent eu . The goal of both the user and the DBMS is
to have as many relevant tuples as possible in the returned
result. Hence, both the user and the DBMS receive some
reward according to the degree by which the returned result
matches the user’s intent eu . This reward is measured based
on the user feedback and using standard effectiveness metrics [45]. One example of such metrics is precision, which
is the ratio of the number of relevant and returned tuples
to the total number of returned tuples. In cases where the
user examines only the top-k returned answers, one may use
precision at k, p@k, that measure the precision of the topk returned tuples. At the end of each round, both the user
and the DBMS receive a reward equal to the value of the selected effectiveness metric for the returned result. Each pair
of the user and the DBMS strategy, (U ,D), is a strategy profile. Thus, the expected payoff for both players with strategy
profile (U ,D) in each interaction is
ur (U, D) =
m
X
i=1
πi
n
X
j=1
Uij
o
X
Djℓ r(ei , eℓ )
(1)
ℓ=1
where r is the effectiveness metric computed using the user’s
feedback on the result of the interpreted intent eℓ over the
underlying database.
2.1.6 Information
None of the players know the other player’s strategy. Let
the user and DBMS strategy at round t ∈ N of the game
be U (t) and D(t), respectively. With a slight abuse of notation, let ur (t) :=ur (U (t), D(t)). In round t ∈ N of the
game, the user has access to her sequence of intents (eu (i)),
queries (q(i)), and rewards (ur (i)) and the DBMS knows
the sequence of submitted queries (q(i)), interpreted intents
(ed (i)), and rewards (ur (i)), 1 ≤ i ≤ t, up to round t. We
note than whether and how the user and the DBMS will use
the aforementioned prices of information depend on their degrees of rationality, which we discuss later in the paper. The
prior probability distribution on intents, π is common knowledge between the players.
tain intent that has not received any positive feedback from
the user in the last round of the game.
Data Interaction Game: The data interaction game, G,
is represented as an 8-tuple:
(S, U, D, π, (eu (i)), (q(i)), (ed (i)), (ur (i)))
(2)
where 1 ≤ i ≤ t, in which S is the schema of the underlying
database.
Due to their lack of information, users may sometimes
submit queries that may not return any tuple over the underlying database [4]. For example, they may submit overspecified SQL queries whose execution over the underlying database produces an empty set. The DBMS typically
map these queries to relatively broad and relaxed intents
with non-empty result over the database [4, 67]. Thus, the
DBMS may be still able to solicit feedback for these types
of queries.
2.2 Open Problems
Traditionally, usable query interfaces, e.g., keyword query
interfaces, aim at improving users’ satisfaction by optimizing some effectiveness metrics, e.g., p@k, for their input
queries [14]. In our game-theoretic formalization, however,
the goal of the DBMS should be to guide the interaction to a
desired and stable state, i.e., equilibrium, in which, roughly
speaking, both players do not have any motivation to change
their strategies and they both get the maximum possible reward. There are three important questions regarding this
game.
• What are the desired and undesired equilibria of the game?
It is important to identify the non-optimal equilibria of
the game as the interaction may stabilize in these equilibria.
• What are the reasonable assumptions on the behavior and
the degree of rationality of the user?
• Given the answers to the previous two questions, what
strategy adaptation mechanism(s) the DBMS should use
to guide and converge the interaction to a desired equilibrium fast? At the first glance, it may seem that if
the DBMS adapts a reasonable learning mechanism, the
user’s adaptation can only help further the DBMS to reach
an optimal state as both players have identical interest.
Nevertheless, in some collaborative two-player games in
which both players adapt their strategies to improve their
payoff, the learning may not converge to any (desired)
equilibrium and cycle among several unstable states [54,
64]. More importantly, the DBMS should use an adaptation strategy that keeps users engaged [30]. In other
words, the adaptation mechanism may not significantly
reduce the payoff of the user for too many subsequent
sessions.
We present some preliminary results on the aforementioned
questions in the following sections.
3. EQUILIBRIUM ANALYSIS
2.1.7 Strategy Adaptation
Give the information available to each player, the player
may modify its strategy at the end of each round. For example, the DBMS may reduce the probability of returning cer-
3.1 Fixed User Strategy
In some settings, the strategy of a user may change in a
much slower time scale than that of the DBMS. In these
cases, it is reasonable to assume that the user’s strategy is
fixed. Hence, the game will reach a desirable state where the
DBMS adapts a strategy that maximizes the expected payoff.
Let a strategy profile be a pair of user and DBMS strategies.
D EFINITION 3.1. Given a strategy profile (U ,D), D is a
best response to U w.r.t. effectiveness measure r if we have
ur (U, D) ≥ ur (U, D′ ) for all the database strategies D′ .
A DBMS strategy D is a strict best response to U if the
inequality in Definition 3.1 becomes strict for all D′ 6= D.
E XAMPLE 3.1. Consider the database instance about universities that is shown in Table 1 and the intents, queries,
and the strategy profiles in Tables 2(a), 2(b), 3(a), and 3(b),
respectively. Given a uniform prior over the intents, the
DBMS strategy is a best response the user strategy w.r.t. w.r.t
precision and p@k in both strategy profiles 3(a) and 3(b).
D EFINITION 3.2. Given a strategy profile (U, D), an intent ei , and a query qj , the payoff of ei using qj is
ur (ei , qj ) =
o
X
Dj,ℓ r(ei , sℓ ).
ℓ=1
D EFINITION 3.3. The pool of intents for query qj in user
strategy U is the set of intents ei such that Ui,j > 0.
We denote the pool of intents of qj as P L(qj ). Our definition of pool of intent resembles the notion of pool of state in
signaling games [15, 18]. Each result sℓ such that Dj,ℓ > 0
may be returned in response to query qj . We call the set of
these results the reply to query qj .
D EFINITION 3.4. A best reply to queryPqj w.r.t. effectiveness measure r is a reply that maximizes ei ∈P L(qj ) πi Ui,j
ur (ei , qj ).
The following characterizes the best response to a strategy.
L EMMA 3.5. Given a strategy profile (U, D), D is a best
response to U w.r.t. effectiveness measure r if and only if D
maps every query to one of its best replies.
P ROOF. If each query is assigned to its best reply in D,
no improvement in the expected payoff is possible, thus D
is a best response for U . Let D be a best response for U
such that some query q is not mapped to its best reply in D.
Let rm ax be a best reply for q. We create a DBMS strategy
D′ 6= D such that all queries q ′ 6= q in D′ have the same
reply as they have in D and the reply of q is rmax . Clearly,
D′ has higher payoff than D for U . Thus, D is not a best
response.
The following corollary directly results from Lemma 3.5.
C OROLLARY 3.6. Given a strategy profile (U, D), D is
a strict best response to U w.r.t. effectiveness measure r if
and only if every query has one and only one best reply and
D maps each query to its best reply.
Given an intent e over database instance I, some effectiveness measures, such as precision, take their maximum for
other results in addition to e(I). For example, given intent
e, the precision of every non-empty result s ⊂ e(I) is equal
to the precision of e(I) for e. Hence, there are more than
one best reply for an intent w.r.t. precision. Thus, according
to Corollary 3.6, there is not any strict best response w.r.t.
precision.
3.2 Nash Equilibrium
In this section and Section 3.3, we analyze the equilibria
of the game where both user and DBMS may modify their
strategies. A Nash equilibrium for a game is a strategy profile where the DBMS and user will not do better by unilaterally deviating from their strategies.
D EFINITION 3.7. A strategy profile (U, D) is a Nash equilibrium w.r.t. a satisfaction function r if ur (U, D) ≥ ur (U ′ , D)
for all user strategy U ′ and ur (U, D) ≥ ur (U, D′ ) for all
database strategy D′ .
E XAMPLE 3.2. Consider again the database about universities that is shown in Table 1 and the intents, queries,
and the strategy profiles in Tables 2(a), 2(b), 3(a), and 3(b),
respectively. Both strategy profiles 3(a) and 3(b) are Nash
equilibria w.r.t precision and p@k. User and DBMS cannot
unilaterally change their strategies and receive a better payoff. If one modifies the strategy of the database in strategy
profile 3(b) and replaces the probability of executing and returning e1 and e3 given query q2 to ǫ and 1 − ǫ, 0 ≤ ǫ ≤ 1,
the resulting strategy profiles are all Nash equilibria.
Intuitively, the concept of Nash equilibrium captures the
fact that users may explore different ways of articulating and
interpreting intents, but they may not be able to look ahead
beyond the payoff of a single interaction when adapting their
strategies. Some users may be willing to lose some payoff in
the short-term to gain more payoff in the long run, therefore,
an interesting direction is to define and analyze less myopic
equilibria for the game [26].
If the interaction between user and DBMS reaches a Nash
equilibrium, they user do not have a strong incentive to change
her strategy. As a result the strategy of the DBMS and the expected payoff of the game will likely to remain unchanged.
Hence, in a Nash equilibrium the strategies of user and DBMS
are likely to be stable. Also, the payoff at a Nash equilibrium
reflects a potential eventual payoff for the user and DBMS
in their interaction. Query qj is a best query for intent ei if
qj ∈ arg maxqk ur (ei , qk ).
The following lemma characterizes the Nash equilibrium
of the game.
L EMMA 3.8. A strategy profile (U, D) is a Nash equilibrium w.r.t. effectiveness measure r if and only if
• for every query q, q is a best query for every intent e ∈
P L(q), and
• D is a best response to U .
P ROOF. Assume that (U, D) is a Nash equilibrium. Also,
assume qj is not a best query for ei ∈ P L(qj ). Let qj ′
be a best query for ei . We first consider the case where
′
= Uk,ℓ for
ur (ei , qj ′ ) > 0. We build strategy U ′ where Uk,ℓ
′
′
all entries (k, ℓ) 6= (i, j) and (k, ℓ) 6= (i, j ), Ui,j
= 0, and
′
′
′
Ui,j
′ = Ui,j . We have U 6= U and ur (U, D) < ur (U , D).
Hence, (U, D) is not a Nash equilibrium. Thus, we have
Ui,j = 0 and the first condition of the theorem holds. Now,
consider the case where ur (ei , qj ′ ) = 0. In this case, we
will also have ur (ei , qj ) = 0, which makes qj a best query
for ei . We prove the necessity of the second condition of the
theorem similarly. This concludes the proof for the necessity part of the theorem. Now, assume that both conditions
P ROOF. Let us assume that there is an intent ei and a
query qj such that 0 < Ui,j < 1. Since U is row stochastic,
there is a query qj ′ where 0 < Ui,j ′ < 1. Let ur (Ui,j , D) =
P
o
ℓ=1 Dj,ℓ r(ei , sℓ ). If ur (Ui,j , D) = ur (Ui,j ′ , D), we can
′
′
create a new user strategy U ′ where Ui,j
= 1 and Ui,j
′ = 0
′
and the values of other entries in U is the same as U . Note
3.3 Strict Nash Equilibrium
that the payoff of (U ,D) and (U ′ ,D) are equal and hence,
(U ,D) is not a strict Nash equilibrium.
A strict Nash equilibrium is a strategy profile in which the
If ur (Ui,j , D) 6= ur (Ui,j ′ , D), without loss of generalDBMS and user will do worse by unilaterally changing their
ity one can assume that ur (Ui,j , D) > ur (Ui,j ′ , D). We
equilibrium strategy.
construct a new user strategy U ′′ whose values for all en′′
D EFINITION 3.9. A strategy profile (U, D) is a strict Nash
tries except (i, j) and (i, j ′ ) are equal to U and Ui,j
= 1,
equilibrium w.r.t. effectiveness measure r if we have ur (U, D) > U ′′ ′ = 0. Because ur (U, D) < ur (U ′′ , D), (U ,D) is not a
i,j
ur (U, D′ ) for all DBMS strategies D′ 6= D and ur (U, D) >
strict Nash equilibrium. Hence, U must be a pure strategy.
′
′
ur (U , D) for all user strategies U 6= U .
Similarly, it can be shown that D should be a pure strategy.
If U is not onto, there is a query qj that is not mapped to
any intent in U . Hence, one may change the value in row j
Table 4: Queries and Intents
of D without changing the payoff of (U, D).
Assume that D is not one-to-one. Hence, there are queries
4(a) Intents
qi and qj and a result sℓ such that Di,ℓ = Dj,ℓ = 1. Because
(U, D) is a strict Nash, U is pure and we have either Ui,ℓ =
Intent# Intent
1 or Uj,ℓ = 1. Assume that Ui,ℓ = 1. We can construct
e3
ans(z) ← U niv(x, ‘M SU ’, ‘M O’, y, z)
strategy U ′ that have the same values as U for all entries
e4
ans(z) ← U niv(x, ‘M SU ’, y, ‘public’, z)
′
′
= 1. Since
except for (i, ℓ) and (j, ℓ) and Ui,ℓ
= 0, Uj,ℓ
e5
ans(z) ← U niv(x, ‘M SU ’, ‘KY ’, y, z)
′
the
payoffs
of
(U,
D)
and
(U
,
D)
are
equal,
(U,
D) is not a
4(b) Queries
strict Nash equilibrium.
Query# Query
Theorem 3.10 extends the Theorem 1 in [18] for our model.
q2
‘MSU’
In some settings, the user may knows and use fewer queries
q3
‘KY’
than intents, i.e., m > n. Because the DBMS strategy in
a strict Nash equilibrium is one-to-one, the DBMS strategy
does not map some of the results to any query. Hence, the
Table 5: Strict best strategy profile
DBMS will never return some results in a strict Nash equilibrium no matter what query is submitted. Interestingly, as
q2 q3
Example 3.1 suggests some of these results may be the ree3 e4 e5
e3 1 0
sults that perfectly satisfy some user’s intents. That is, given
q2 1 0 0
e4 1 0
intent ei over database instance I, the DBMS may never req3 0 0 1
e5 0 1
turn ei (I) in a strict Nash equilibrium. Using a proof similar
to the one of Lemma 3.8, we have the following properties
of strict Nash equilibria of a game. A strategy profile (U, D)
E XAMPLE 3.3. Consider the intents, queries, strategy prois a strict Nash equilibrium w.r.t. effectiveness measure r if
file, and database instance in Tables 4(a), 4(b), 5, and 1. The
and only if:
strategy profile is a strict Nash equilibrium w.r.t precision.
• Every intent e has a unique best query and the user stratHowever, the strategy profile in Example 3.2 is not a strict
egy maps e to its best query, i.e., e ∈ P L(qi ).
Nash equilibrium as one may modify the value of ǫ without
• D is the strict best response to U .
changing the payoff of the players.
of the theorem hold for strategies U and D. We can prove
that it is not possible to have strategies U ′′ and D′′ such that
ur (U, D) < ur (U ′′ , D) or ur (U, D) < ur (U, D′′ ) using a
similar method.
Next, we investigate the characteristics of strategies in a strict
Nash equilibria profile. Recall that a strategy is pure iff it has
only 1 or 0 values. A user strategy is onto if there is not any
query qj such that Ui,j = 0 for all intend i. A DBMS strategy is one-to-one if it does not map two queries to the same
result. In other words, there is not any result sell such that
Djℓ > 0 and Dj ′ ℓ > 0 where j 6= j ′ .
T HEOREM 3.10. If (U, D) is a strict Nash equilibrium
w.r.t. satisfaction function r, we have
• U is pure and onto.
• D is pure and one-to-one.
3.4 Number of Equilibria
A natural question is how many (strict) Nash equilibria
exist in a game. Theorem 3.10 guarantees that both user and
DBMS strategies in a strict Nash equilibrium are pure. Thus,
given that the sets of intents and queries are finite, there are
finitely many strict Nash equilibria in the game. We note that
each set of results is always finite. However, we will show
that if the sets of intents and queries in a game are finite, the
game has infinite Nash equilibria.
L EMMA 3.11. If a game has a non-strict Nash equilibrium. Then there is an infinitely many Nash equilibria.
P ROOF. The result follows from the fact that the payoff function (1) is a bilinear form of U and D, i.e. it is a
linear of D when U is fixed and a linear function of U ,
when D is fixed. If for D 6= D′ , (U, D) and (U, D′ ) are
Nash-equilibria, then ur (U, D) = ur (U, D′ ). Therefore
ur (U, αD + (1 − α)D′ ) = ur (U, D) for any α ∈ R. In
particular, for α ∈ [0, 1], if D, D′ are stochastic matrices,
αD + (1 − α)D′ will be a stochastic matrix and hence,
(U, αD + (1 − α)D′ ) is a Nash equilibrium as well. Similarly, if (U ′ , D) and (U, D) are Nash equilibria for U 6= U ′ ,
then ur (αU + (1 − α)U ′ , D) = ur (U, D) and (αU + (1 −
α)U ′ , D) is a Nash-equilibrium for any α ∈ [0, 1].
T HEOREM 3.12. Given a game with finitely many intents
and queries, if the game has a non-strict Nash equilibrium,
it has an infinite number of Nash equilibria.
P ROOF. Every finite game has always a mixed Nash equilibrium [57]. According to Theorem 3.10, a mixed Nash is
not a strict Nash equilibrium. Therefore, using Lemma 3.11,
the game will have infinitely many Nash equilibria.
3.5 Efficiency
In this section we discuss the efficiency of different equilibria. We refer to the value of the utility (payoff) in formula (1) at a strategy profile to the efficiency of the strategy.
Therefore, the most efficient strategy profile is naturally the
one that maximizes (1). We refer to an equilibria with maximum efficiency as an efficient equilibrium.
Thus far we have discussed two types of equilibria, Nash
and strict Nash, that once reached it is unlikely that either
player will deviate from its current strategy. In some cases
it may be possible to enter a state of equilibria where neither
player has any incentive to deviate, but that equilibria may
not be an efficient equilibrium.
The strategy profile in Table 3(b) provides the highest payoff for the user and DBMS given the intents and queries in
Tables 2(a) and 2(b) over the database in Table 1. However, some Nash equilibria may not provide high payoffs.
For instance, Table 3(a) depicts another strategy profile for
the set of intents and queries in Tables 2(a) and 2(b) over the
database in Table 1. In this strategy profile, the user has little knowledge about the database content and expresses all
of her intents using a single query q2 , which asks for the
ranking of universities whose abbreviations are MSU. Given
query q2 , the DBMS always returns the ranking of Michigan
State University. Obviously, the DBMS always returns the
non-relevant answers for the intents of finding the rankings
of Mississippi State University and Missouri State University. If all intents have equal prior probabilities, this strategy
profile is a Nash equilibrium. For example, the user will not
get a higher payoff by increasing their knowledge about the
database and using query q1 to express intent e2 . Clearly, the
payoff of this strategy profile is less than the strategy profile
in Table 3(b). Nevertheless, the user and the DBMS do not
have any incentive to leave this undesirable stable state once
reached and will likely stay in this state.
D EFINITION 3.13. A strategy profile (U ,D) is optimal w.r.t.
an effectiveness measure r if we have ur (U, D) ≥ u(U ′ , D′ )
for all DBMS strategies D′ and U ′
Since, the games discussed in this paper are games of identical interest, i.e. the payoff of the user and the DBMS are the
same, therefore, an optimal strategy (U, D) (w.r.t. an effectiveness measure r) is a Nash equilibrium.
L EMMA 3.14. A strategy (U ,D) is optimal if and only if
it is an efficient equilibrium.
P ROOF. Note that if (U, D) is optimal, then none of the
two players (i.e. the user and the DBMS) has a unilateral incentive to deviate. Therefore (U, D) is a Nash equilibrium.
On the other hand, since the payoff function (1) is a continuous function of U and D and the domain of row-stochastic
matrices is a compact space, therefore a maximizer (U, D)
of (1) exists and by the previous part it is a Nash equilibrium. Note that the efficiency of all strategies are bounded
by the efficiency of an optimal strategy and hence, any efficient equilibrium is optimal.
Similar to the analysis on efficiency of a Nash equilibria, there are strict Nash equilibria that are less efficient than
others. Strict Nash equilibria strategy profiles are unlikely
to deviate from the current strategy profile, since any unilateral deviation will result in a lower payoff. From this we can
say that strict Nash equilibria are also more stable than Nash
equilibria since unilateral deviation will always have a lower
payoff.
Table 6: Strategy Profile 1
6(a) User strategy
e1
e2
e3
q1
0
1
1
q2
1
0
0
6(b) Database strategy
q1
q2
e1
0
1
e2
0
0
e3
1
0
Table 7: Strategy Profile 2
7(a) User Strategy
e1
e2
e3
q1
0
0
1
q2
1
1
0
7(b) Database Strategy
q1
q2
e1
0
0
e2
0
1
e3
1
0
As an example of a strict Nash equilibrium that is not efficient, consider both strategy profiles illustrated in Tables 6
and 7. Note that the intents. queries, and results in this example are different from the ones in the previous examples.
For this illustration, we set the rewards to r(e1 , s1 ) = 1,
r(e2 , s2 ) = 2, r(e2 , s3 ) = 0.1, and r(e3 , s3 ) = 3 where all
other rewards are 0. Using our payoff function in Equation 1
we can calculate the total payoff for the strategy profile in
Table 6 as u(U, D) = 4.1. This strategy profile is a strict
Nash since any unilateral deviation by either player will result in a strictly worse payoff. Consider the strategy profile
in Table 7 with payoff u(U, D) = 5. This payoff is higher
than the payoff the strategy profile in Table 6 receives. It
is also not likely for the strategy profile with less payoff to
change either strategy to the ones in the strategy profile with
higher payoff as both are strict Nash.
4.
USER ADAPTATION MECHANISM
It is well established that humans show reinforcement behavior in learning [56, 47, 48]. Many lab studies with human
subjects conclude that one can model human learning using
reinforcement learning models [56, 47]. The exact reinforcement learning method used by a person, however, may vary
based on her capabilities and the task at hand. We have performed an empirical study of a real-world query log to figure out the reinforcement learning model(s) that best explain
the learning mechanism by which users adapt their strategies
during interaction with a data management system.
4.1 Query Workload
We have used a subsample of a Yahoo! query log for
our empirical study [63]. The Yahoo! query log consists of
queries submitted to a Yahoo! search engine over a period
of time in July 2010. Each record in the query log consists
of a time stamp, user cookie, query submitted, the 10 results
displayed to the user, and the positions of the user clicks. All
the record logs are anonymized such that each time stamp,
query, and returned result are saved as a unique identifier.
Table 8 illustrates an example of what the log record looks
like. Note, for the Results column, there are 10 results returned. Accompanying the query log is a set of relevance
judgment scores for each query and result pair. The relevance judgment scores determine user satisfaction with that
result. The score has the possible values of 0,1,2,3,4, with
0 meaning not relevant at all and 4 meaning the most relevant result. Table 9 illustrates an example of the relevance
judgment scores dataset. For example, if the user entered
QueryID = 00002ef d and one of results returned was
ResultID = 2722a07f , then the relevance of that result
would be three.
Table 8: Log Record Example
QueryID
00002efd
12fe75b7
CookieID
1deac14e
4b74d72d
TimeStampID
1279486689
1279546874
Results
2722a07f ... e1468bbf
dc381b38 ... df9a561f
Table 9: Relevance Judgment Score Example
QueryID
00002efd
12fe75b7
ResultID
2722a07f
dc381b38
Score
3
4
For our analysis we sorted the query log by the time stamp
attribute to simulate the time line of the users interaction
with the Yahoo! search engine. We determine the intent
behind each query by using the relevance judgment scores
for the results for each query. We consider the intent behind
each query to be the set of results, i.e., URLs, with non-zero
relevance scores.
4.2 Reinforcement Learning Methods
We have used and adapted six different reinforcement learning methods to model users’ strategy in interaction with data
systems. These models mainly vary based on 1) the degree
by which the user considers past interactions when computing future strategies, 2) how they update the user strategy,
and 3) the rate by which they update the user strategy. Some
models assume that the user leverages outcomes of her past
interactions when she updates her current strategy. Other
models allow the user to forget older interactions. These
methods also differ in how they use the value of reward to
update the user’s strategy. Some reinforce a behavior, e.g.,
using a certain query to convey an intent, after a successful
attempt by a fixed value independent of the amount of reward. Others use the value of received reward to reinforce a
behavior. Finally, a model may have some discounting factors to control the rate by which a behavior is reinforced. In
this study, we have used the value of NDCG for the returned
results of a query as the reward in each interaction. Since
NDCG models different levels of relevance, it provides a
more exact estimate of the true reward in an interaction than
other metrics that measure the quality of a ranked list, such
as precision at k.
4.2.1 Bush and Mosteller’s Model
Bush and Mosteller’s model increases the probability that
a user will choose a given query when searching for a specific intent by an amount proportional based on the reward
of using that query and the current probability of using this
query for the intent in the strategy [10]. It also decreases the
probabilities of queries not used in a successful interaction.
This method updates the probabilities of using queries for
the intent ei after an interaction using the following formulas.
Uij (t) + αBM · (1 − Uij (t)) qj = q(t) ∧ r ≥ 0
Uij (t+1) =
Uij (t) − β BM · Uij (t)
qj = q(t) ∧ r < 0
(3)
BM
Uij (t) − α
· Uij (t)
qj 6= q(t) ∧ r ≥ 0
Uij (t+1) =
BM
Uij (t) + β
· (1 − Uij (t) qj 6= q(t) ∧ r < 0
(4)
In the aforementioned formulas, αBM ∈ [0, 1] and β BM ∈
[0, 1] are parameters of the model, q(t) denotes the query
picked by the user at time t, and r is the reward of the interaction. If query qj is equal to the query chosen by the user
to represent intent ei , then Equation 3 is used. For all other
queries qj for the intent ei at time t, Equation 4 is used.
The probabilities of using queries for intents other than ei
remains unchanged. Since the value of NDCG is always
greater than zero, i.e., r > 0, the parameter β BM is never
used nor trained.
4.2.2 Cross’s Model
Cross’s model modifies the user’s strategy similar to Bush
and Mosteller’s model [16]. However, it uses the amount of
the received reward to update the user strategy. There are
two parameters in this model, αC and β C that influence the
rate of reinforcement
Uij (t) + R(r) · (1 − Uij (t)) qj = q(t)
Uij (t + 1) =
Uij (t) − R(r) · Uij (t)
qj 6= q(t)
(5)
R(r) = αC · r + β C
C
(6)
C
In the above formulas, α ∈ [0, 1] and β ∈ [0, 1] are the
parameters used compute the adjusted reward R(r) based
on the value of actual reward r. The parameter β C is a
static increment of the adjusted reward. Similar to Bush and
Mosteller’s model, the aforementioned formulas are used to
update the probabilities of using queries for the intent ei in
the current interaction. Other entries in the user’s strategy
are remained unchanged.
often assumes some minimum amount of reward is guaranteed when she queries the database. The model uses this
minimum amount to discount the received reward. We set
rmin to 0 in our analysis, representing that there is no expected reward in an interaction. Therefor the model uses the
total received reward to reinforce a strategy.
4.2.3 Roth and Erev’s Model
4.2.5 Win-Stay/Lose-Randomize
Roth and Erev’s model reinforces the probabilities directly
from the reward value r that is received when the user enters
query q(t) [51]. Its most important difference with other
models is that it explicitly accumulates all the rewards gained
by using a query to express an intent. Sij (t) in matrix S(t)
maintains the accumulated reward of using query qj to express intent ei over the course of the user and database system interactions up to round (time) t.
Sij (t) + r qj = q(t)
Sij (t + 1) =
(7)
Sij (t)
qj 6= q(t)
The Win-Stay/Lose-Randomize method uses only the most
recent interaction for an intent to determine the queries used
to express the intent in the future [7]. Assume that the user
conveys an intent ei by a query qj . If the reward of using qj
is above a specified threshold π the user will use qj to express ei in the future. Otherwise, if the the reward is below
the threshold, the user will simply randomize her strategy,
where for ei , all queries have an equal probability to be chosen.
Sij (t + 1)
Uij (t + 1) = P
n
Sij ′ (t + 1)
(8)
j′
Roth and Erev’s model increases the probability of using a
query to express an intent based on the accumulated rewards
of using that query over the long-term interaction of the user
and data management system. It does not explicitly penalize other queries. Of course, because the user strategy U is
row-stochastic, each query not used in a successful interaction will be implicitly penalized as when the probability of
a query increases, all others will decrease to keep U rowstochastic.
4.2.4 Roth and Erev’s Modified Model
Roth and Erev’s modified model is similar to the original
Roth and Erev’s model, but it has an additional parameter
that determines to what extent the user takes in to account
the outcomes of her past interactions with the system [20]. It
is reasonable to assume that the user may forget the results of
her much earlier interactions with the system. User’s memory is imperfect which means that over time the strategy may
change merely due to the forgetful nature of the user. This
is accounted for by the forget parameter σ ∈ [0, 1]. Matrix
S(t) has the same role it has for the Roth and Erev’s model.
Sij (t + 1) = (1 − σ) · Sij (t) + E(j, R(r))
R(r) · (1 − ǫ) qj = q(t)
E(j, R(r)) =
R(r) · (ǫ)
qj 6= q(t)
R(r) = r − rmin
Sij (t + 1)
Uij (t + 1) = P
n
Sij ′ (t + 1)
(9)
(10)
(11)
(12)
j′
In the aforementioned formulas, ǫ ∈ [0, 1] is a parameter
that weights the reward that the user receives, n is the maximum number of possible queries for a given intent ei , and
rmin is the minimum expected reward that the user wants to
receive. The intuition behind this parameter is that the user
4.2.6 Latest Reward
The Latest-Reward method reinforces the user strategy based
on the previous reward that the user has seen when querying
for an intent ei . All other queries have an equal probability to be chosen for a given intent. Let a user receive reward r ∈ [0, 1] by entering query qj to express intent ei .
The Latest-Reward method sets the probability of using qj
to convey ei in the user strategy, Ui j to r and distribute the
the remaining probability mass 1 − r evenly between other
entries related to intent ei , in Uik , where k 6= j.
4.2.7 Comparing the Methods
Next, we compare the aforementioned models in terms of
their use of past interaction and their update rules. Bush and
Mosteller’s, Cross’s, and both Roth and Erev’s models use
information from the past to compute the future strategies.
The Roth and Erev’s models use the information about the
past interactions more than others. Win-Stay/Lose-Randomize
and Latest-Reward models do not rely as much as the first
four methods on the outcomes of the previous interactions.
Cross’s, both Roth and Erev’s, and the Latest-Reward models use the value of the reward that is received after entering a
query to update the strategy. Bush and Mosteller’s and WinStay/Lose-Randomize models change their strategies based
on a fixed amount independent of the reward.
4.3 Training and Testing
Some models, e.g., Cross’s model, have some parameters
that need to be trained. We have used 5,000 records in the
query workload and found the optimal values for those parameters using a grid search and the sum of squared errors.
Each strategy has been initialized with an uniform distribution of probabilities, so that all queries are likely to be used
for a given intent at the initial strategy. Once we had the parameters estimated for each model, we let each model to run
over 300,000 records that follow the initial 5,000 records in
the query log to compute a user strategy. All models have
used the same set of queries and results. During the training
phase a total of 3,456 intents, 3,088 queries, and 253,952
users were covered. This means that queries and intents were
reused frequently and there were many returning users over
time. We have calculated the NDCG value for each interac-
tion using the available relevance judgment scores that accompanies Yahoo! query workload. Interestingly, we have
not observed any result with perfect or zero NDCG in the
query workload. At the end of this round, we get six user
strategies, one per reinforcement method.
We have evaluated the accuracy of the trained user strategies in predicting the future strategies of the users using the
interaction records for 2,000 unique intents in the query log
that follow the 300,000 records used in training. For each
intent, we have found its first log record that immediately
follows the records used in training and compared the predication of the strategy with the query actually used in this log
record to express the intent. During the testing phase a total of 2000 intents were used, 1806 queries, and 1999 users
were covered. To compare the prediction accuracies of the
strategies, we calculated the mean squared distance between
what a given strategy predicted and what the user actually
did.
4.4 Empirical Results
Method
Bush and
Mosteller
Cross
Roth and Erev
Roth and Erev
modified
WinStay/LoseRandomize
Latest-Reward
Mean
Squared
Distance
0.01031
Standard
Deviation
Parameters
0.07371
0.01035
0.07381
αBM
=
0.14
αC = 0.06
β C = 0.11
0.00836
0.00837
0.05721
0.05735
0.01400
0.05766
0.15205
0.19658
σ=0
ǫ = 0.18
π = 0.01
Table 10: The accuracies of learning algorithms
Table 10 shows the results from the tests that we performed
as well as the estimated parameters. A lower mean squared
distance implies that the model more accurately represents
the users’ learning method. Roth and Erev’s and Roth and
Erev’s modified models both perform the best out of all the
tested models. Since σ = 0 in Roth and Erev’s modified
model, this model does not forget the outcomes of the previous interactions. Thus, it behaves more similarly to the
original Roth and Erev’s model. Because Roth and Erev
models update the users strategies using the information of
the previous strategies and interactions, users use their previous strategies and the outcomes of their previous interactions with the system when they pick a query to express their
current intent. This result also indicates that the value of received reward should be considered when reinforcing a strategy.
Bush and Mosteller’s, Cross’s, and Win-Stay/Lose-Randomize
models perform worse than either of Roth and Erev’s models. Bush and Mosteller’s model has a relatively low value
of α. Therefore, the rate of reinforcement is quite slow as
the lower α is, the less a successful strategy is reinforced.
With an α of 0 for example, there would be no reinforce-
ment at all. Bush and Mosteller’s model also does not consider the reward when reinforcing and therefor cannot reinforce queries that get effective results more than others
that receive a smaller reward. Cross’s model suffers from
the same lack of reinforcement rate as Bush and Mosteller’s
but has an additional downfall. If the reward is extremely
low, almost zero, the query will still be reinforced as β is a
constant value independent of the reward. This means that
queries with higher reward will be reinforced more, but also
means that queries with an extremely low reward will still
be reinforced when they probably should be left alone. WinStay/Lose-Randomize does not provide an accurate predication because it does not consider the entire history of strategies that the user has used. It also does not explore the space
of possible queries to improve the effectiveness of the interaction. Hence, it seems that the users keep exploring possible queries to express an intent more effectively, although
they may already know a query that conveys the intent quite
successfully. Also, by only considering the previous reward,
Win-Stay/Lose-Randomize cannot make robust adjustments
and instead makes fixed changes in the model that are quite
drastic. Finally, Latest-Reward performs the worst when
compared to all models by an order of magnitude. This is
because not only does this method have not memory like
Win-Stay/Lose-Randomize, but it also reinforcing the strategy too drastically.
From our analysis it appears that users show a substantially intelligent behavior when adopting and modifying their
strategies. They leverage all most of their past interactions
and their outcomes, i.e., have an effective long-term memory. This is specially interesting as the results of previous
lab studies have indicated that mostly only proficient subjects rely on the accumulated rewards of the past interactions
and use Roth and Erev’s model for learning. Those studies
show that non-proficient users tend to use models that do not
leverage the information about the past interactions, such as
Cross’s model [11]. Also, the reward they receive directly
impacts how they reinforce their strategy and will dictate
what queries are used to represent intents in the future.
5. ADAPTATION MECHANISMS FOR DBMS
In many relevant applications, the user’s learning is happening in a much slower time-scale compared to the learning
of the DBMS. So, one can assume that the user’s strategy is
fixed compared to the time-scale of the DBMS adaptation.
Therefore, in this section, first, we consider the case that the
user is not adapting to the strategy of the DBMS and then,
we consider the case that the user’s strategy is adapting to the
DBMS’s strategy but perhaps on a slower time-scale. When
dealing with the game introduced in the previous sections,
many questions arise:
i. How can a DBMS learn or adapt to a user’s strategy?
ii. Mathematically, is a given learning rule effective?
iii. What would be the limiting behavior of a given learning
rule?
Here, we address the first and the second questions above.
Dealing with the third question is far beyond the page limits
of this paper. For simplicity of notation, we refer to intent ei
and result sℓ as intent i and ℓ, respectively, in the rest of the
paper. Hence, we have:
ur (U, D) =
m
X
i=1
πi
n
X
Uij
j=1
o
X
Djℓ riℓ ,
ℓ=1
where r : [m] × [o] → R+ is the effectiveness measure
between the intent i and the result, i.e., decoded intent ℓ.
5.1 Reinforcement Learning for an Arbitrary
Similarity Measure
As in [32], we consider Roth-Erev reinforcement learning
mechanism for adaptation of the DBMS adaption. For the
case that both the DBMS and the user adapt their strategies,
one can use the results in [32]. Let us discuss the DBMS
adaptation rule. The learning/adaptation rule happens over
discrete time t = 0, 1, 2, 3, . . . instances where t denotes
the tth interaction of the user and the DBMS. We refer to
t simply as the iteration of the learning rule. With this, the
reinforcement learning mechanism for the DBMS adaptation
is as follows:
a. Let R(0) > 0 be an n × o initial reward matrix whose
entries are strictly positive.
b. Let D(0) be the initial DBMS strategy with Djℓ (0) =
PoRjℓ (0)
Rjℓ (0) > 0 for all j ∈ [n] and ℓ ∈ [o].
ℓ=1
c. For iterations t = 1, 2, . . ., do
i. If the user’s query at time t is q(t), DBMS return a
result E(t) ∈ E with probability:
P (E(t) = i′ | q(t)) = Dq(t)i′ (t).
ii. User gives a reward rii′ given that i is the intent of
the user at time t. Note that the reward depends both
on the intent i at time t and the result i′ . Then, set
Rjℓ (t) + riℓ if j = q(t) and ℓ = i′
Rjℓ (t + 1) =
.
Rjℓ (t)
otherwise
(13)
iii. Update the DBMS strategy by
Rji (t + 1)
Dji (t + 1) = Po
,
ℓ=1 Rjℓ (t + 1)
(14)
for all j ∈ [n] and i ∈ [o].
In the above scheme R(t) is simply the reward matrix at
time t.
Few comments are in order regarding the above adaptation
rule:
- One can use available ranking functions, e.g. [12], for the
initial reward condition R(0) which possibly leads to an
intuitive initial point for the learning rule. One may normalize and convert the scores returned by these functions
to probability values.
- In step c.ii., if the DBMS has the knowledge of the user’s
intent after the interactions (e.g. through a click), the DBMS
sets Rji +1 for the known intent i. The mathematical analysis of both cases will be similar.
- In the initial step, as the DBMS uses a ranking function
to compute the probabilities, it may not materialize R and
D. As the game progresses, DBMS maintains the strategy and reward matrices with entries for only the observed
queries, their underlying intents, and their returned results.
Hence, the DBMS does not need to materialize R and D
for the sets of possible intents, queries, and results. DBMS
also does not need to know the set of user’s interns beforehand. Hence, the algorithm is practical for the cases where
the sought-for intents, submitted queries, and returned results are not very large. Moreover, R and D are generally
sparse. As queries and intents generally follow a power
law distribution [45], one may use sampling techniques to
use this algorithm in other settings. The rigorous theoretical and empirical analysis of applying such techniques are
interesting subjects of future work.
5.2 Analysis of the Learning Rule
In this section, we provide an analysis of the reinforcement mechanism provided above and will show that, statistically speaking, the adaptation rule leads to improvement
of the efficiency of the interaction. Note that since the user
gives feedback only on one tuple in the result, one can without the loss of generality assume that the cardinality of the
list k is 1.
For the analysis of the reinforcement learning mechanism
in Section 5 and for simplification, denote
u(t) := ur (U, D(t)) = ur (U, D(t)),
(15)
for an effectiveness measure r as ur is defined in (1).
We recall that a random process {X(t)} is a submartingale
[19] if it is absolutely integrable (i.e. E(|X(t)|) < ∞ for all
t) and
E(X(t + 1) | Ft ) ≥ X(t),
where Ft is the history or σ-algebra generated by X1 , . . . , Xt .
In other words, a process {X(t)} is a sub-martingale if the
expected value of X(t + 1) given X(t), X(t − 1), . . . , X(0),
is not strictly less than the value of Xt . Note that submartingales are nothing but the stochastic counterparts of monotonically increasing sequences. As in the case of bounded (from
above) monotonically increasing sequences, submartingales
pose the same property, i.e. any submartingale {X(t)} with
E(|X(t)|) < B for some B ∈ R+ and all t ≥ 0 is convergent almost surely. We refer the interested readers to [19] for
further information on this result (martingale convergence
theorem).
The main result in this section is that the sequence of
the utilities {u(t)} (which is indeed a stochastic process as
{D(t)} is a stochastic process) defined by (15) is a submartignale when the reinforcement learning rule in Section 5
is utilized. As a result the proposed reinforcement learning
rule stochastically improves the efficiency of communication between the DBMS and the user. More importantly, this
holds for an arbitrary reward/effectiveness measure r. This
is rather a very strong result as the algorithm is robust to the
choice of the reward mechanism.
To show this, we discuss an intermediate result. For simplicity of notation, we fix the time t and we use superscript +
to denote variables at time (t + 1) and drop the dependencies
at time t for variables depending on time t. Throughout the
rest of our discussions, we let {Ft } be the natural filtration
for the process {D(t)}, i.e. F is the σ-algebra generated by
D(0), . . . , D(t).
L EMMA 5.1. For any ℓ ∈ [m] and j ∈ [n], we have
E(u(t + 1 | Ft ) ≥ E(u(t) | Ft ) − βt ,
+
E(Djℓ
| Ft ) − Djℓ
= Djℓ ·
m
X
o
X
riℓ′
riℓ
Djℓ′
−
R̄j + ril ℓ′ =1
R̄j + riℓ′
πi Uij
i=1
Po
where R̄j =
ℓ′ =1
!
,
Rjℓ′ .
P ROOF. Fix ℓ ∈ [m] and j ∈ [n]. Let A be the event
that at the t’th iteration, we reinforce a pair (j, ℓ′ ) for some
+
ℓ′ ∈ [m]. Then on the complement Ac of A, Djℓ
(ω) =
Djℓ (ω). Let Ai,ℓ′ ⊆ A be the subset of A such that the
intent of the user is i and the pair (j, ℓ′ ) is reinforced. Note
that the collection of sets {Ai,ℓ′ } for i, ℓ′ ∈ [m], are pairwise
mutually exclusive and their union constitute the set A.
We note that
+
Djℓ
Using Lemma 5.1 and the above result, we show that up to
a summable disturbance, the proposed learning mechanism
is stochastically improving.
T HEOREM 5.3. Let {u(t)} be the sequence given by (15).
Then,
m
o
X
X
Rjℓ + ril
Rjℓ
=
1
+
1Ai,ℓ′
A
i,ℓ
R̄j + riℓ
′
R̄
+
r
j
iℓ
′
i=1
ℓ =1
ℓ′ 6=ℓ
+ Djℓ 1Ac .
for some
P∞non-negative random process {βt } that is summable
(i.e. t=0 β < ∞ almost surely). As a result {u(t)} converges almost surely.
P ROOF. Let u+ := u(t + 1), u := u(t),
uj := uj (U (t), D(t)) =
| Ft ) =
m
X
πi Uij Djℓ
i=1
+
m
X
i=1
πi Uij
X
Djℓ′
ℓ6=ℓ′
Rjℓ + riℓ
R̄j + riℓ
Pm
and also define R̄j := ℓ′ =1 Rjℓ′ . Note that uj is the efficiency of the jth signal/query.
Using the linearity of conditional expectation and Lemma 5.1,
we have:
o
m X
n
X
X
+
πi Uij
riℓ′ E(Djℓ
| Ft ) − Djℓ
E(u+ | Ft ) − u =
i=1 j=1
=
m X
n X
o
X
where p = P(A | F ). Note that Djℓ =
+
E(Djℓ
| Ft ) − Djℓ =
m
X
i=1
−
m
X
i=1
πi Uij
ℓ6=ℓ′
o
X
!!
ri′ ℓ
R̄j + ri′ ℓ
.
(17)
Pm
riℓ
i=1 πi Uij R̄j +riℓ .
E(u+ | Ft ) − u =
o
n
X
X
Djℓ yiℓ zjℓ −
ℓ=1
o
X
Djℓ yjℓ
o
X
.
Djℓ′ zjℓ′ (18)
ℓ′ =1
ℓ=1
!
Now, we express the above expression as
Rjℓ r
.
R̄j (R̄j + riℓ′ )
iℓ′
E(u+ | Ft ) − u = Vt + Ṽt
R
To show the main result, we will utilize the following result
in martingale theory.
T HEOREM 5.2. [50] A random process {Xt } converges
almost surely if Xt is bounded, i.e., E(|Xt |) < B for some
B ∈ R+ and all t ≥ 0 and
(16)
where
βt ≥ 0 is a summable sequence almost surely, i.e.,
P
β
< ∞ with probability 1.
t
t
Note that this result is a weaker form of the Robins-Siegmund
martingale convergence theorem in [50] but it will serve for
the purpose of our discussion.
(19)
where
Replacing R̄jlj with Djℓ and rearranging the terms in the
above expression, we get the result.
E(Xt+1 |Ft ) ≥ Xt − βt
i′ =1
ri′ ℓ′
−
Djℓ′
R̄j + ri′ ℓ′
ℓ′ =1
and hence,
Djℓ′
πi′ Ui′ j
Now, let yjℓ = i=1 πi Uij riℓ and zjℓ =
Then, we get from the above expression that
riℓ R̄j − Rjℓ
R̄j (R̄j + riℓ )
X
πi Uij Djℓ riℓ
i=1 j=1 ℓ=1
j=1
πi Uij Djℓ
ℓ=1
m
X
Pm
Rjℓ
+ (1 − p)Djℓ ,
R̄j + riℓ′
Rji
R̄j
πi Uij Djℓ riℓ(t) ,
i=1 ℓ=1
Therefore, we have
+
E(Djℓ
m X
o
X
and
n
o
X
1 X
2
Vt =
Djℓ yjℓ
−
R̄
j
j=1
ℓ=1
Ṽt =
o
n
X
X
j=1
Djℓ yjℓ
l=1
Djℓ′ z̃jℓ′ −
ℓ′ =1
ℓ=1
Further, z̃jℓ =
o
X
o
X
P
i=1
!2
Djℓ yjℓ ,
m
X
ℓ=1
Djℓ yjℓ z̃jℓ
!
.
(20)
2
riℓ
πi Uij R̄j (R̄j +riℓ ) .
We claim that Vt ≥ 0 for each t and {Ṽt } is a summable
sequence almost surely. Then, from (19) and Theorem 5.2,
we get that {ut } converges almost surely and it completes
the proof. Next, we validate our claims.
We first show that Vt ≥ P
0, ∀t. Note that D is a rowo
stochastic matrix and hence, ℓ=1 Djℓ = 1. Therefore, by
b. Let U (0) be the initial user’s strategy with
the Jensen’s inequality [19], we have:
o
X
Djℓ (yjℓ )2 ≥
ℓ=1
o
X
(Djℓ yjℓ )2 .
Uij (0) = Pn
ℓ=1
Hence, V ≥ 0.
We next claim that {Ṽt } is a summable sequence with
probability one. It can be observed from (20) that
Vt ≤
o
X
o2 n
j=1
R̄j2
.
Sij (0)
Sij ′ (0)
j ′ =1
(21)
since yjℓ ≤ 1, z̃jℓ ≤ R̄j−2 for each j ∈ [n], ℓ ∈ [m] and
D is a row-stochastic matrix. To prove the claim, it suffices to show that for each j ∈ [m], the sequence { R21(t) }
for all i ∈ [m] and j ∈ [n]. and let U (tk ) = U (tk − 1) =
· · · = U (tk−1 + 1) for all k.
c. For all k ≥ 1, do the following:
i. The nature picks a random intent t ∈ [m] with probability πi (independent of the earlier choices of the
nature) and the user picks a query j ∈ [n] with probability
P (q(tk ) = j | i(tk ) = i) = Uij (tk ).
j
is summable. Note that for each j ∈ [m] and for each t, we
have R̄j (t+1) = R̄j (t)+ǫt where ǫt ≥ ǫ > 0 with probability pt ≥ p > 0. Therefore, using the Borel-Cantelli Lemma
for adapted processes [19] we have { R21(t) } is summable
j
which concludes the proof.
The above result implies that the effectiveness of the DBMS,
stochastically speaking, increases as time progresses when
the learning rule in Section 5 is utilized. Not only that, but
this property does not depend on the choice of the effectiveness function (i.e. riℓ in this case). This is indeed a desirable
property for any adapting scheme for DBMS adaptation.
ii. The DBMS uses the current strategy D(tk ) and interpret the query by the intent i′ (t) = i′ with probability
U (i′ (tk ) = i′ | q(tk ) = j) = Dji′ (tk ).
iii. User gives a reward 1 if i = i′ and otherwise, gives
no rewards, i.e.
Sij (tk ) + 1 if j = q(tk ) and i(tk ) = i′ (tk )
+
Sij
=
Sij (tk )
otherwise
+
where Sij
= Sij (tk + 1).
iv. Update the user’s strategy by
Uij (tk + 1) = Pn
5.3 User Adaptation
Here, we consider the case that the user also adapts to the
DBMS’s strategy. When the user submits a query q and
the DBMS returns a result that fully satisfy the intent behind the query e, it is relatively more likely that the user
will use the query q to express e again during her interaction with the DBMS. On the other hand, if the DBMS returns a result that does not contain any tuple that is relevant to the e, it less likely that the user expresses e using
q in future. In fact, researchers have observed that users
show reinforcement learning behavior when interacting with
a DBMS over a period of time [11]. In particular, the authors in [11] have shown that some groups of users learned
to formulate queries with a model similar to Roth-Erev reinforcement learning. We define the similarity measure as
follows. For simplicity we assume that m = o and use the
following similarity measure:
riℓ =
1 if i = ℓ,
.
0 otherwise
In this case, we assume that the user adapts to the DBMS
strategy at time steps 0 < t1 < · · · < tk < · · · and in
those time-steps the DBMS is not adapting as there is no
reason to assume the synchronicity between the user and the
DBMS. The reinforcement learning mechanism for the user
is as follows:
a. Let S(0) > 0 be an m × n reward matrix whose entries
are strictly positive.
Uji (tk + 1)
,
Sij ′ (tk + 1)
(22)
j ′ =1
for all i ∈ [m] and j ∈ [n].
In the above scheme S(t) is the reward matrix at time t for
the user.
5.4 Analysis of User and DBMS Adaptation
In this section, we provide an analysis of the reinforcement mechanism provided above and will show that, statistically speaking, our proposed adaptation rule for DBMS,
even when the user adapts, leads to improvement of the efficiency of the interaction. With a slight abuse of notation,
let
u(t) := ur (U, D(t)) = ur (U (t), D(t)),
(23)
for an effectiveness measure r as ur is defined in (1).
L EMMA 5.4. Let t = tk for some k ∈ N. Then, for any
i ∈ [m] and j ∈ [n], we have
πi Uij
(Dji − ui (t))
S
+
1
iℓ
ℓ=1
E(Uij+ | Ft ) − Uij = Pn
(24)
where
ui (t) =
n
X
Uij (t)Dji (t).
j=1
P ROOF. Fix i ∈ [m], j ∈ [n] and k ∈ N. Let B be
the event that at the tk ’th iteration, user reinforces a pair
(i, ℓ) for some ℓ ∈ [n]. Then, on the complement B c of B,
Pij+ (ω) = Pij (ω). Let B1 ⊆ B be the subset of B such that
the pair (i, j) is reinforced and B2 = B \ B1 be the event
that some other pair (i, ℓ) is reinforced for ℓ 6= i.
We note that
Sij + 1
Sij
Uij+ = Pn
1B1 + Pn
1B2 + Uij 1B c .
ℓ=1 Siℓ + 1
ℓ=1 Siℓ + 1
P ROOF. Note from Theorem 5.3 and 5.5 that the sequence
{u(t)} satisfies all the conditions of Theorem 5.2. Hence,
proven.
5.5 Limiting Equilibria
One of the most important immediate directions to be explored is to study the limiting behavior of the adaptation
mechanism and identify its properties. Such questions are
Sij + 1
E(Uij+ | Fkt ) = πi Uij Dji Pn
partially addressed in [32] for the case that both the players
ℓ=1 Siℓ + 1
in a singling game adapt their strategy synchronously and
X
Sij
identically. In this case, in the limit, only certain equilibria
+ (1 − p)Uij ,
+
πi Uiℓ Dℓi Pn
′ + 1
can emerge. In particular, eventually, either an object, intent
S
′
iℓ
ℓ =1
ℓ6=j
in our model, will be assigned to many signals, queries in our
P
model, (synonyms) or many intents can be assigned to one
where p = U (B | Fkt ) = ℓ πi Uij Dji . Note that Uij =
S
ij
query (polysemy). But two intents will not be assigned to
Pn
and hence,
ℓ=1 Siℓ
two queries (see Theorem 2.3 in [32] for more details). The
authors in [32] follow a traditional language game approach
E(Uij+ | Ft ) − Uij =
! as explained in Section 6. Our immediate future research diX
1
rections on the study of the adaptation mechanism is to study
Pn
πi Uij Dji − πi Uij
Uiℓ Dℓi . the convergence properties of the proposed reinforcement al′
+
1
S
ℓ′ =1 iℓ
ℓ
gorithm in Section 5.1 and Section 5.3. In particular, when
the user is not adapting, does the strategy of the DBMS conwhich can be rewritten as in (24).
verge to some limiting strategy such as is the best response to
the user’s strategy? Also, when both the user and the DBMS
Using Lemma 5.1, we show that the process {u(t)} is a
adapt to each other, what equilibria will emerge? We believe
sub-martingale.
that answers to these questions will significantly contribute
to the proposed learning framework in the database systems
T HEOREM 5.5. Let t = tk for some k ∈ N. Then, we
and provide a novel theoretical perspective on the efficiency
have
of a query interface.
E(u(t + 1) | Ft ) − u(t) ≥ 0
(25)
Therefore, we have
6. RELATED WORK
where u(t) is given by (23).
+
P ROOF. Fix t = tk for some k ∈ N. Let u := u(t + 1),
u
:= u(t), ui := ui (U (t), D(t)) and also define S̃ i :=
Pm
ℓ′ =1 Siℓ′ + 1. Then, using the linearity of conditional expectation and Lemma 5.1, we have:
E(u+ | Ft ) − u =
m X
n
X
πi Dji E(Uij+ | Ft ) − Uij
i=1 j=1
=
m X
n
X
i=1 j=1
=
m
X
π2
i=1
i
S̃ i
πi Uij
Dji − ui
ℓ′ =1 Sjℓ′ + 1
πi Dji Pm
n
X
j=1
Uij (Dji )2 − (ui )2 .
(26)
Pm
Note that U is a row-stochastic matrix and hence, i=1 Uij =
1. Therefore, by the Jensen’s inequality [19], we have:
2
n
n
X
X
Uij (Dji )2 ≥
Dji Uij = (ui )2 .
j=1
j=1
Replacing this in the right-hand-side of (26), we conclude
that E(u+ | Ft ) − u ≥ 0 and hence, the sequence {u(t)} is
a submartingale.
C OROLLARY 5.6. The sequence {u(t)} given by (15) converges almost surely.
Signaling games model communication between two or
more agents and have been widely used in economics, sociology, biology, and linguistics [39, 15, 49, 18]. Generally
speaking, in a signaling game a player observes the current
state of the world and informs the other player(s) by sending
a signal. The other player interprets the signal and makes a
decision and/or performs an action that affect the payoff of
both players. A signaling game may not be cooperative in
which the interests of players do not coincide [15].
Our framework extends a particular category of signaling
games called language games [61, 49, 18] and is closely
related to learning in signaling games [32, 59, 23]. These
games have been used to model the evolution of a population’s language in a shared environment. In a language
game, the strategy of each player is a stochastic mapping
between a set of signals and a set of states. Each player
observes its internal state, picks a signal according its strategy, and sends the signal to inform other player(s) about its
state. If the other player(s) interpret the correct state from
the sent signal, the communication is successful and both
players will be rewarded. Our framework, however, differs
from language games in several fundamental aspects. First,
in a language game every player signals, but only one of our
players, i.e., user, sends signals. Second, language games
model states as an unstructured set of objects. However, each
user’s intent in our framework is a set of tuples and different intents may intersect. Third, we use widely used similarity functions between the desired and returned answers
for a query to measure the degree of success of answering a
query. Fourth, the signals in language games do not posses
any particular meaning and can be assigned to every state. A
database query, however, restricts its possible answers. Finally, there is not any work on language games on analyzing
the dynamics of reinforcement learning where players learn
in different time scales.
Game theoretic approaches have been used in various areas of computer science, such as distributed systems, planning, security, and data mining [3, 38, 27, 55, 53, 28, 43].
Researchers have also leveraged economical models to build
query interfaces that return desired results to the users using the fewest possible interactions [66]. In particular, researchers have recently applied game theoretic approaches
to model the actions taken by users and document retrieval
systems in a single session [41]. They propose a framework to find out whether the user likes to continue exploring the current topic or move to another topic. We, however, explore the development of common representations
of intents between the user and DMBS. We also investigate
the querying and interactions that may span over multiple
queries and sessions. Moreover, we analyze equilibria of
the game and convergence rates of some strategy adaptation
methods for the user and DBMS. Finally, we focus on structured rather than unstructured data. Avestani et al. have used
signaling games to create a shared lexicon between multiple
autonomous systems [6]. We, however, focus on modeling
users’ information needs and development of mutual understanding between users and the DBMS. Moreover, as opposed to the autonomous systems, a DBMS and user may update their information about the interaction in different time
scales. We also provide a rigorous analysis of the game equilibria and propose novel strategy adaptation mechanisms for
the DBMS.
We have proposed the possibility of using signaling games
to model the interaction between users and DBMS and provided an initial result in a short paper in [58]. The current paper precisely formalizes and significantly develops
our framework, provides new results, and extends our previous result for practical settings. First, we precisely define
the concepts in our framework, e.g., intent and query, and
their relationships. Second, we define the equilibria of the
game, analyze their properties, and show some equilibria are
desirable. Third, our previous work models a database as a
set of tuples and assumes that the desired answer for each
query is a single tuple. These assumptions do not generally
hold in practice. In this paper, we analyze both the equilibria
and learning algorithms for the game where the answers to
a query are a set or ranked lists of tuples. Usable query interfaces usually only partially satisfy the user’s information
needs. However, authors in [58] assume that a result either
fully satisfies the user’s information need or does not contain
information relevant to the submitted intent. In this paper,
we model the cases where results partially satisfy user’s intents. We show that the ability to return results that partially
match the desired answers of a query significantly impacts
the equilibria of the game and the algorithms for updating
the strategy of the DBMS. Finally, we rigorously analyze
our proposed algorithm for updating DBMS strategy where
user also changes her strategy in a possibly different time
scale than the DBMS.
7. CONCLUSION & FUTURE WORK
Many users query and explore databases interactivity and
over a relatively long period of time. We modeled the interaction between the user and the DBMS as a signaling
game, where the players start with different mappings between queries and desired results, and would like to reach a
mutual understanding. We formally analyzed different equilibria of this game. We also showed that some equilibria
may not have an optimal payoff. We proposed an adaptation
mechanism for the DBMS to learn the querying strategy of
the user and proved that this mechanism increases the expected payoff for both user and the DBMS in average and
converges almost surely. This case was extended for when
the user changes her strategy and we provided an adaptation
mechanism for the DBMS that converges almost surely. Interactive querying of databases has recently received a lot of
attention [34]. We believe that our proposed framework will
provide a basis for deep exploration of some interesting and
important problems in this context. In particular, we plan to
explore the properties of Nash and efficient Nash equilibria
in this game. We also plan to analyze other types of equilibria ans stable states for the game. Moreover, we would
like to design and explore other adaptation algorithms for
the DBMS that converge to efficient equilibria and are robust to various user’s adaptation mechanisms. The equilibria
discussed in this paper were defined for a database instance.
We believe that one may find some equilibria that share important properties over all databases instances of the same
schema. This will open the door to explore the properties of
the game in database schema level.
8. REFERENCES
[1] S. Abiteboul, R. Hull, and V. Vianu. Foundations of
Databases: The Logical Level. Addison-Wesley, 1994.
[2] A. Abouzied, D. Angluin, C. H. Papadimitriou, J. M.
Hellerstein, and A. Silberschatz. Learning and
verifying quantified boolean queries by example. In
PODS, 2013.
[3] I. Abraham, D. Dolev, R. Gonen, and J. Halpern.
Distributed computing meets game theory: robust
mechanisms for rational secret sharing and multiparty
computation. In PODC, 2006.
[4] S. Agrawal, S. Chaudhuri, G. Das, and A. Gionis.
Automated ranking of database query results. In
CIDR, 2003.
[5] P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time
analysis of the multiarmed bandit problem. Machine
learning, 47(2-3):235–256, 2002.
[6] P. Avesani and M. Cova. Shared lexicon for distributed
annotations on the Web. In WWW, 2005.
[7] J. A. Barrett and K. Zollman. The role of forgetting in
the evolution and learning of language. Journal of
Experimental and Theoretical Artificial Intelligence,
21(4):293–309, 2008.
[8] T. Beckers et al. Report on INEX 2009. SIGIR Forum,
44(1), 2010.
[9] A. Bonifati, R. Ciucanu, and S. Staworko. Learning
join queries from user examples. TODS, 40(4), 2015.
[10] R. R. Bush and F. Mosteller. A stochastic model with
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
applications to learning. The Annals of Mathematical
Statistics, pages 559–585, 1953.
Y. Cen, L. Gan, and C. Bai. Reinforcement learning in
information searching. Information Research: An
International Electronic Journal, 18(1):n1, 2013.
S. Chaudhuri, G. Das, V. Hristidis, and G. Weikum.
Probabilistic information retrieval approach for
ranking of database query results. TODS, 31(3), 2006.
Y. Chen, W. Wang, Z. Liu, and X. Lin. Keyword
search on structured and semi-structured data. In
SIGMOD, 2009.
Y. Chen, W. Wang, Z. Liu, and X. Lin. Keyword
search on structured and semi-structured data. In
SIGMOD, 2009.
I. Cho and D. Kreps. Signaling games and stable
equilibria. Quarterly Journal of Economics, 102,
1987.
J. G. Cross. A stochastic learning model of economic
behavior. The Quarterly Journal of Economics,
87(2):239–266, 1973.
K. Dimitriadou, O. Papaemmanouil, and Y. Diao.
Explore-by-example: An automatic query steering
framework for interactive data exploration. In
SIGMOD, 2014.
M. C. Donaldson, M. Lachmannb, and C. T.
Bergstroma. The evolution of functionally referential
meaning in a structured world. Journal of
Mathematical Biology, 246, 2007.
R. Durrett. Probability: theory and examples.
Cambridge university press, 2010.
I. Erev and A. E. Roth. On the Need for Low
Rationality, Gognitive Game Theory: Reinforcement
Learning in Experimental Games with Unique, Mixed
Strategy Equilibria. 1995.
R. Fagin, B. Kimelfeld, Y. Li, S. Raghavan, and
S. Vaithyanathan. Understanding Queries in a Search
Database System. In PODS, 2010.
G. H. L. Fletcher, J. V. D. Bussche, D. V. Gucht, and
S. Vansummeren. Towards a theory of search queries.
In ICDT, 2009.
M. J. Fox, B. Touri, and J. S. Shamma. Dynamics in
atomic signaling games. Journal of theoretical
biology, 376, 2015.
D. Fudenberg and D. Levine. The Theory of Learning
in Games. MIT Press, 1998.
N. Fuhr and T. Rolleke. A probabilistic relational
algebra for the integration of information retrieval and
database systems. TOIS, 15, 1997.
A. Ghosh and S. Sen. Learning toms: Towards
non-myopic equilibria. In AAAI, 2004.
G. D. Giacomo, , A. D. Stasio, A. Murano, and
S. Rubin. Imperfect information games and
generalized planning. In IJCAI, 2016.
G. Gottlob, N. Leone, and F. Scarcello. Robbers,
marshals, and guards: Game theoretic and logical
characterizations of hypertree width. In PODS, 2001.
L. A. Granka, T. Joachims, and G. Gay. Eye-tracking
analysis of user behavior in www search. In SIGIR,
2004.
[30] K. Hofmann, S. Whiteson, and M. de Rijke. Balancing
exploration and exploitation in listwise and pairwise
online learning to rank for information retrieval.
Information Retrieval, 16(1), 2013.
[31] V. Hristidis, L. Gravano, and Y. Papakonstantinou.
Efficient IR-Style Keyword Search over Relational
Databases. In VLDB 2003.
[32] Y. Hu, B. Skyrms, and P. Tarrès. Reinforcement
learning in signaling game. arXiv preprint
arXiv:1103.5818, 2011.
[33] J. Huang, R. White, and G. Buscher. User see, user
point: Gaze and cursor alignment in web search. In
CHI, 2012.
[34] S. Idreos, O. Papaemmanouil, and S. Chaudhuri.
Overview of data exploration techniques. In SIGMOD,
2015.
[35] H. V. Jagadish, A. Chapman, A. Elkiss,
M. Jayapandian, Y. Li, A. Nandi, and C. Yu. Making
database systems usable. In SIGMOD, 2007.
[36] N. Khoussainova, Y. Kwon, M. Balazinska, and
D. Suciu. Snipsuggest: Context-aware autocompletion
for sql. PVLDB, 4(1), 2010.
[37] B. Kimelfeld and Y. Sagiv. Finding and
Approximating Top-k Answers in Keyword Proximity
Search. In PODS, 2005.
[38] D. Koller and A. Pfeffer. Generating and solving
imperfect information games. In IJCAI, 1995.
[39] D. Lewis. Convention. Cambridge: Harvard University
Press, 1969.
[40] H. Li, C.-Y. Chan, and D. Maier. Query from
examples: An iterative, data-driven approach to query
construction. PVLDB, 8(13), 2015.
[41] J. Luo, S. Zhang, and H. Yang. Win-win search:
Dual-agent stochastic game in session search. In
SIGIR, 2014.
[42] Y. Luo, X. Lin, W. Wang, and X. Zhou. SPARK:
Top-k Keyword Query in Relational Databases. In
SIGMOD 2007.
[43] Q. Ma, S. Muthukrishnan, B. Thompson, and
G. Cormode. Modeling collaboration in academia: A
game theoretic approach. In BigScholar, 2014.
[44] D. Maier, D. Rozenshtein, S. Salveter, J. Stein, and
D. S. Warren. Toward logical data independence: A
relational query language without relations. In
SIGMOD, 1982.
[45] C. Manning, P. Raghavan, and H. Schutze. An
Introduction to Information Retrieval. Cambridge
University Press, 2008.
[46] A. Nandi and H. V. Jagadish. Guided interaction:
Rethinking the query-result paradigm. PVLDB, 102,
2011.
[47] Y. Niv. The neuroscience of reinforcement learning. In
ICML, 2009.
[48] Y. Niv. Reinforcement learning in the brain. The
Journal of Mathematical Psychology, 53(3):139–154,
2009.
[49] M. A. Nowak and D. C. Krakauer. The evolution of
language. PNAS, 96(14), 1999.
[50] H. Robbins and D. Siegmund. A convergence theorem
for non negative almost supermartingales and some
applications. In Herbert Robbins Selected Papers.
Springer, 1985.
[51] A. E. Roth and I. Erev. Learning in extensive-form
games: Experimental data and simple dynamic models
in the intermediate term. Games and economic
behavior, 8(1):164–212, 1995.
[52] A. D. Sarmaa, A. Parameswaran, H. Garcia-Molina,
and J. Widom. Synthesizing view definitions from
data. In ICDT, 2010.
[53] M. Schuster and T. Schwentick. Games for active
XML revisited. In ICDT, 2015.
[54] L. S. Shapley et al. Some topics in two-person games.
Advances in game theory, 52(1-29):1–2, 1964.
[55] Y. Shoham. Computer science and game theory.
Commun. ACM, 51(8), 2008.
[56] H. Shteingart and Y. Loewenstein. Reinforcement
learning and human behavior. Current Opinion in
Neurobiology, 25:93–98, 04/2014 2014.
[57] S. Tadelis. Game Theory: An Introduction. Princeton
University Press, 2013.
[58] A. Termehchy and B. Touri. A signaling game
approach to database querying and interaction. In
ICTIR, 2015.
[59] B. Touri and C. Langbort. Language evolution in a
noisy environment. In (ACC). IEEE, 2013.
[60] Q. Tran, C. Chan, and S. Parthasarathy. Query by
output. In SIGMOD, 2009.
[61] P. Trapa and M. Nowak. Nash equilibria for an
evolutionary language game. Journal of Mathematical
Biology, 41, 2000.
[62] A. Vorobev, D. Lefortier, G. Gusev, and P. Serdyukov.
Gathering additional feedback on search results by
multi-armed bandits with respect to production
ranking. In WWW, 2015.
[63] Yahoo! Yahoo! webscope dataset anonymized Yahoo!
search logs with relevance judgments version 1.0.
http://labs.yahoo.com/Academic_Relations,
2011. [Online; accessed 5-January-2017].
[64] H. P. Young. Strategic learning and its limits. OUP
Oxford, 2004.
[65] Y. Yue, J. Broder, R. Kleinberg, and T. Joachims. The
k-armed dueling bandits problem. J. Comput. Syst.
Sci., 78(5), 2012.
[66] C. Zhai. Towards a game-theoretic framework for
information retrieval. In SIGIR, 2015.
[67] X. Zhou, J. Gaugaz, W.-T. Balke, and W. Nejdl. Query
relaxation using malleable schemas. In SIGMOD,
2007.
| 2cs.AI
|
A note on radial basis function computing
W. Chen* and W. He**
*
Department of Mechanical System Engineering, Shinshu University, Wakasato 4-17-1,
Nagano City, Nagano 380-8533, Japan ([email protected])
**
Department of Electrical Engineering, Jiangsu University of Science and Technology,
Zhenjiang City, Jiangsu Province 212013, P. R. China.
Key words: radial basis function, boundary knot method, nonlinear problem, time-space distance function,
wavelet RBF, centrosymmetric matrix, ill-conditioning.
1. Introduction
This note carries three purposes involving
our latest advances on the radial basis function
(RBF) approach. First, we will introduce a new
scheme employing the boundary knot method
(BKM) [1] to nonlinear convection-diffusion
problem. It is stressed that the new scheme
directly results in a linear BKM formulation of
nonlinear problems by using response
point-dependent RBFs, which can be solved by
any linear solver. Then we only need to solve a
single nonlinear algebraic equation for desirable
unknown at some inner node of interest. The
numerical results demonstrate high accuracy
and efficiency of this nonlinear BKM strategy.
Second, we extend the concepts of distance
function, which include time-space and variable
transformation distance functions. Finally, we
demonstrate that if the nodes are symmetrically
placed, the RBF coefficient matrices have either
centrosymmetric or skew centrosymmetric
structures. The factorization features of such
matrices lead to a considerable reduction in the
RBF computing effort. A simple approach is
also presented to reduce the ill-conditioning of
RBF interpolation matrices in general cases.
2. BKM linear formulation of nonlinear
problems
Chen and Tanaka [1] found that if only
boundary knots are used, the BKM can
formulate linear analogization of nonlinear
differential equations with linear boundary
conditions.
Consider
the
Burger-like
convection-diffusion equation [2]
∇ u − ux u = 0
2
(1)
with inhomogeneous boundary condition
u = −2 / x
(2)
Eq. (2) is also a particular solution of Eq. (1).
The geometry of the problem is an ellipse with
semi-major axis of length 2 and semi-minor
axis of length 1. Note that the origin of the
Cartesian co-ordinates system is dislocated to
the node (3,0) to circumvent singularity at x=0.
Using the scheme given in [1], Eq. (1) is
restated as
∇ u + u = u + ux u .
2
(3)
When only the boundary nodes are employed,
the resulting BKM formulation will be a linear
algebraic equation in terms of 2D Helmhotz
non-singular general solution. For details of
such BKM procedure see [1]. The numerical
solutions of this normal BKM procedure have
the average relative errors 3.9e-2 for N=5,
1.1e-1 for N=9, 1.4e-1 for N=13 at some inner
nodes, where N is the number of boundary
nodes used. It is noted that the performances are
unstable and solutions inaccurate. If we use
interior points in the BKM, the accuracy and
stability will be improved greatly at the expense
of sacrificing linear formulation as in the dual
reciprocity BEM (DRM) [2]. It is note that the
convection term rather than nonlinear
constitution here causes the deficiency of the
BKM solutions if not using inner points.
In the DRM, it is reported that the use of the
fundamental solution of convection diffusion
equation can significantly improve the solution
accuracy of transient convection diffusion
problem with only one inner point. This
suggests us that a non-singular general solution
of convection-diffusion operator may be much
more suitable for this problem. By analogy with
the
fundamental
solution
of
the
convection-diffusion equation, the response
knot-dependent non-singular general solution of
Eq. (1) is given by
u(r, x ) = e
u
φ
r ,
2
−u (x −x k ) 2
(4)
where radial function is the zero order Bessel
Jo or modified Bessel function Io of the first
kind dependent on the sign of flow velocity u.
In this case it is the latter. The RBF
approximation is given by
ui =
N
∑α e
−ui ( x i −x k ) 2
k
k=1
ui
I0
rik .
2
(5)
Note that the above RBF representation differs
from the normal one in that we here use the
response point-dependent RBFs even if we do
not know the value of u. In terms of the BKM
using only boundary knots, Eq. (1) is
analogized by substituting boundary conditions
into Eq. (5), namely,
N
∑ α e(
k
k=1
xi − x k ) x i
(
)
I 0 rik 2 x i = −2 / xi .
(6)
The formulation (6) is a set of simultaneous
linear algebraic equations and can be solved
easily by any linear solver.
Then we can evaluate the value of u at any
inner nodes of interest through the solution of a
single nonlinear equation (5). Note that the RBF
expansion coefficients in Eq. (5) are now
already evaluated from Eq. (6) and the only one
unknown is the value of u at a specified single
inner knot. This study used simple bisection
method to handle such a single nonlinear
equation. There are no concerns here relating to
the expensive repeated evaluation and inverse
of Jacobian matrix, stability issue and the
careful guess of initial solutions.
In summary, the present nonlinear BKM
scheme can be viewed as a two-step procedure.
First, the linear BKM formulation of nonlinear
problems is produced using response
node-dependent RBFs. Then, the second step is
to calculate the solution at any inner node of
interest through solving a single nonlinear
algebraic equation.
Table 1. Relative errors for BKM linear
formulation of Burger-like equation
x
4.5
4.2
3.6
3.0
2.4
1.8
3.9
3.3
3.0
2.7
2.1
y DRM(33) BKM (9) BKM (11)
0.0
2.3e-3
2.8e-3
2.5e-3
-0.35 2.1e-3
2.3e-3
2.9e-3
-0.45 5.4e-3
4.4e-3
6.2e-3
-0.45 4.5e-3
1.0e-2
9.2e-3
-0.45 1.2e-3
1.2e-2
5.7e-3
-0.35 9.0e-4
7.0e-3
3.2e-3
0.0
3.9e-3
4.1e-3
5.5e-3
0.0
3.3e-3
9.1e-3
1.0e-2
0.0
4.5e-3
1.2e-2
1.1e-2
0.0
2.7e-3
1.4e-2
1.1e-2
0.0
3.2e-3
1.1e-2
4.9e-3
Table 1 lists the BKM results compared with
the solutions of the dual reciprocity BEM [2].
Average relative errors of the present BKM for
N=9, 11 are respectively 8.5e-3, 7.5e-3.
Compared with the previous BKM procedure,
the solution accuracy is significantly improved
while still keeping the linear BKM formulation
of nonlinear problems. However, it is noted that
the solution accuracy is still not always
improved with incremental number of boundary
knots. For example, average relative errors for
N=13, 15, 17, 19, 21 are respectively 8.3e-3,
8.3e-3, 8.8e-3, 8.9e-3, 1.9e-2. Namely, the
highest average accuracy is achieved for N=11.
As in the other global collocation techniques,
this is due to the ill-conditioning system matrix
for large number of knots. Anyway overall
solution procedure is rather stable. The
accuracy and efficiency of the present BKM
scheme are very encouraging.
On the other hand, the solutions of the DRM
[2] were obtained with 16 boundary nodes and
17 inner points. Therefore, it is not surprising
that the DRM solutions [2] are slightly more
accurate than the present BKM ones. It is also
noted that the DRM formulation is a set of
simultaneous nonlinear algebraic equations.
The programming, computing effort and storage
requirements in the DRM are much higher than
the BKM. The present case only involves the
Dirichlet conditions. For more complicated
boundary conditions, the resulting BKM
formulation may not be a set of linear algebraic
equations even if we only employ boundary
knots. By any measure, however, the size of the
BKM analogous equations will be much smaller
than that of the DRM.
The essential idea behind this work may be
extended to the BEM, DRM, method of
fundamental solution, and multiple reciprocity
BEM for nonlinear problems. Namely, as in
handling linear varying parameter problems, the
response knot-dependent fundamental solution
may be employed in the nonlinear computing
even if some parameters of fundamental
solution are unknown.
3. Redefinition of distance functions
3.1. Time-space distance
Golberg and Chen [3] summarized the
Euclidean and geodisc distance function as the
two kinds of distance functions used in the RBF.
Chen and Tanaka [1] introduced a time-space
distance function to eliminate time dependence
within the framework of the RBF numerical
schemes. This study further develops this work
in combination with the Green RBF [1] to
create efficient operator-dependent RBFs. Note
that the Green RBF was named after the general
solution RBF (GS-RBF) in [1]. Here we feel it
is more proper to call it as Green RBF since it
uses both the fundamental solution of the
related operator and inhomogeneous terms
based on the known Green second identity.
First, let us consider the equation governing
wave propagation
∇u=
2
1
u + f (p, t) .
c 2 tt
(7)
s = ict ,
(8)
Let
where p denotes multidimensional variable and
i means unit imaginary number, we have
∇ 2u + uss = f (p,t ) .
(9)
By analogy with the Euclidean definition of
distance variable, the time-space distance
function is defined
rj = rpj + (s − s j ) = rpj − c (t − tj ) , (10)
2
2
2
2
2
where rpj denotes the normal spatial Euclidean
distance function. The above definition can lead
to complex distance variable due to the
presence of minus operation. It is safe to use
r j = r pj2 + c 2 ∆t 2j .
(
(11a)
)
r j = c 2 ∆t 2j − rpj2 H c∆t j − rpj2 .
(11b)
r j = rpj − c∆t j & r j = rpj + c∆t j . (11c)
The definition (11a,b,c) of distance function
differ from the standard radial distance function
in that the time variable is handled equally as
the space variables. Note that distances (11a)
are respectively used in two RBFs for one
problem. In general, hyperbolic and elliptic
equations have solutions whose arguments have
the form p+at and p+ibt respectively, where a
and b are real. Namely,
u( p, t) = f (ζ )
(12)
where is some linear combination of p and t.
This provides some theoretical support to use
time-space distance functions (10) and (11).
However, the above situations do not hold for
Hankel transforms of f(r) and g(r), respectively.
In terms of the Green RBF [1], we have
parabolic-type diffusion equation
∇u=
2
1
u + f ( p, t) .
k t
(13)
Its transient fundamental solution is well known
u =
*
(t
1
j
−t
)
d2
r
H tj − t , (14)
exp −
4k t j − t
(
2
p
)
(
)
where d is the space dimensionality, H is the
Heaviside function. We can define the
corresponding time-space RBF (TS-RBF) as
φ (rp ,t, t j ) = h(rp ,t, t j )u (r, t,t j ) ,
(15)
Here h is chosen according to problem feature.
It is stressed that in this case the response and
source nodes must be totally staggered to avoid
singularity in time dimension.
On the other hand, Chen et al. [4] proposed
time-space non-singular general solution
*
( )
φ rp ,
− k t− t j
()
The Green-RBF is constructed based on the
canonical form of some operators such as the
known Laplace or Helmholtz operators.
However, many engineering problems do not
possess such standard form operators. This
section will show that some cares may be taken
to handle these problems.
The general second order partial differential
system with varying coefficient can be stated as
(16)
for diffusion problem and
[ ( ( ))+ Dsin(c(t − t ))]φ (r ) (17)
j
p
for wave problems, where (rp) is the zero
order Bessel function of the first kind for 2D
problems and sin(rp)/rp for 3D problems. For
example, consider the free symmetrical
vibration of a very large membrane governed
by the equation
∂ 2 z 1 ∂z 1 ∂ 2 z
+
=
∂r 2 r ∂r c 2 ∂t 2
1
c
∞
∫ ξf (ξ )cos(ξct )J (ξr )dξ +
∞
∫ g (ξ )sin(ξct )J (ξr)dξ
0
ξ = f1 (x, y) , η = f2 (x, y ) .
(19)
0
r=
(22)
(ξ − ξ ) + (η − η )
2
2
j
j
.
(23)
Substituting Eq. (22) into Eq. (23), we define
the distance function in terms of the original
independent variables of x and y as
r=
[f (x, y) − f (x , y )]
+[f (x, y ) − f (x , y )] .
2
1
1
j
j
2
2
where upper-dashed f and g are the zero-order
(21)
The corresponding distance function of the
Euclidean norm is given by
0
0
∂2u
∂2u
∂2u
+ T 2 = 0,
2 + S
∂x
∂xy
∂y
where R, S and T are continuous functions of x
and y. We can translate it into the canonical
Laplacian by a suitable change of independent
variables [5]
(18)
with z=f(r), ∂z ∂t = g(r ) when t=0. We have the
solution [5]
z(r, t) =
0
Substituting the non-singular general solutions
(16), (17) and (20) into Eq. (15) produces the
TS-RBF without singularity.
The time-space distance function and
corresponding TS-RBF are also expected to be
applicable to transient data processing such as
motion picture and movie.
R
u* = C cos c t − t j
j
3.2 Varying parameter problems
*
u = Ae
[ ( ( ))+ Bsin(c(t − t ))]J (r ). (20)
z(r, t) = Acos c t − t j
2
j
j
(24)
In the following we illustrate one special
case. Let us consider [6]
ym
∂ 2u ∂ 2 u
+
= 0 , (y ≥ 0, m
∂x 2 ∂y 2
−2) ,
(25)
its general solution is
−2 β
( ),
u = r2 w r
2
(26)
pre-wavelet RBFs with rj2 + c 2j instead of r
into the RBFs, where cj is dilution parameters.
For example, numerical experiments with
pre-wavelet TPS rj2 m ln rj2 + c 2j
manifests
spectral convergence as in the multiquadratic
(MQ). This work can be generalized by
u=
N
∑ c φ (λ r
k
k k
+ dk ) ,
(31)
k=1
where =m/2(m+2), w is the hypergoemetric
functions,
r=
r1 =
(
r2 =
(x − x )
x − xj
)
2
2
j
+
r1
,
r2
4
(m + 2 )2
(27a)
2
m+ 2
m+2 2
2
y
−
y
j
, (27b)
2
m+ 2
m+2 2
4
+
+ y j 2 . (27c)
2 y
m
+
2
(
)
where
and d are respectively dilate and
location coefficients of the wavelet. is the
RBF which can be here seen as a wavelet parent
function. Such wavelet-like RBF may be
especially attractive for adaptable handling
geometry singularity and localized shock-like
solutions due to its inherent multiscale feature
combined with a spatial localization.
Fasshauer and Schumaker [7] summarized
some wavelets using sphere RBFs. It will be
beneficial to pay more attentions on this aspect,
especially for creating orthonormal wavelet-like
RBF.
Let
ξ = x,
we have
η=
m+2
2
y 2 ,
m +2
∂2u ∂2u
+
= 0.
∂ξ 2 ∂η 2
(28)
(29)
Substituting Eq. (28) into Eq. (23) also yields
the distance function (27b). Similar situation
holds for another example
∂2u
∂2u
∂u
+α
= 0 , (y ≥ 0) ,
2 + y
∂x
∂y 2
∂y
(30)
where is a coefficient.
The above analyses show that for one certain
problem, we can use multiple different
definitions of distance variable simultaneously
such as Eqs. (27a,b,c). The use of RBF should
fully consider the features of the targeted
problems.
3.3. Wavelet RBF
Chen
and
4. Coefficient matrix structures and
ill-conditioning
It is well known that the RBF interpolation
matrix has symmetric structure irrespective of
the geometry and node placements. This study
will show that the RBF matrix also carries the
centrosymmetric structure if the nodes are
symmetrically placed. Let us consider a
d-dimension space problem. The Euclidean
distance norm is defined as
rij =
∑ (x( ) − x( ) )
d
k
i
[1]
2
.
(32)
k=1
The symmetric node spacing is understood as
k)
)
(k )
xi(k ) + x (Nk+1−i
= x (j k ) + x (N+1−
,
j = c
(33)
where c(k) are constants. Thus, it is obvious
rij = rN+ 1− i,N+1− j .
Tanaka
k
j
(34)
constructed
Based on Eq. (34), one can easily verify that the
RBF coefficient matrices have centrosymmetric
structure as shown in Eq. (34) for even order
derivative and skew centrosymmetric structure
as shown below
aij = −aN +1−i,N +1− j
(35)
for odd order derivative, where a denotes an
entry of a matrix. Therefore, if the nodes are
symmetrically placed, the RBF coefficient
matrices for derivatives have either symmetric
centrosymmetric
or
skew
symmetric
centrosymmetric structures.
Centrosymmetric matrices can easily be
decomposed into two half-sized matrices. Such
factorization merit leads to a considerable
reduction in computing effort for determinant,
inversion and eigenvalues. For more related
details see [8-9]. On the other hand, even if the
total node spacing is not symmetric, such
decomposition processing can still effectively
reduce the ill-conditioning of the resulting RBF
matrices by preconditioning
I
A=
I
−J I
A
J −J
I
J
(36)
where A means a RBF matrix of even order, J is
contra-identity matrix. A similar but distinct
preconditioning transform matrix exists for odd
order matrix.
In addition, we observe that as in the
traditional collocation method [10], very large
entries of RBF derivative matrices appear in the
upper and lower two rows and middle columns
and largely account for the ill-conditioning of
large-size RBF system. Accordingly some
elementary matrix transformations can simply
significantly reduce the RBF ill-conditioning
with very little effort.
References
1
2
Chen, W. and Tanaka, M., New Insights in
Boundary-only and Domain-type RBF
Methods, Inter. J. Nonlinear Sci. Numer.
Simulation, 1(2000), 145-152.
Partridge, P.W., Brebbia, C.A. and Wrobel,
L.W., The Dual Reciprocity Boundary
Element Method, Comput. Mech. Publ.,
1992.
Golberg, M.A. and Chen, C.S., A
bibliography on radial basis function
approximation,
Boundary
Element
Communications, 7(1996), 155-163.
4 Chen, W. and Tanaka, M., A meshless,
exponential convergence, integration-free,
and boundary-only RBF technique, Comput.
Math. Appl. (submitted), 2000.
5 Sneddon, I., Elements of Partial
Differential Equations, McGRAW-Hill, Inc,
Tokyo, 1957.
6 Cmnphob, M.M., Examples of Partial
Differential Equations, (Japanese version
translated by M. Sayigo), Morikata Press,
Tokyo, 1982.
7 Fasshauer, G. and Schumaker, L., Scattered
Data Fitting on the Sphere. In Mathematical
Methods for Curves and Surfaces II, (M.
Daehlen, T. Lyche, and L. L. Schumaker
eds.), Vanderbilt University Press, 1998,
117-166.
8 Cantoni, A. and butler, P., Eigenvalues and
eigenvectors of symmetric centrosymmetric
matrices, Linear Algebra Appl. 13(1976),
275-288.
9 Chen, W., Wang, X. and Zhong, T., The
structure of weighting coefficient matrices
of harmonic differential quadrature and its
applications, Commun. Numer. Methods
Engng., 12(1996), 455-460.
10 Beruer, K. and Everson, R., On the errors
incurred calculating derivatives using
Chebyshev polynomials, J. Comput. Phys.,
99(1992), 56-67.
3
| 5cs.CE
|
arXiv:1605.01079v1 [cs.DS] 3 May 2016
Newer method of string comparison: the Modified
Moving Contracting Window Pattern Algorithm
Tiago Tresoldi
Universidade Federal do Rio Grande (FURG)
[email protected]
May 5, 2016
Abstract
This paper presents a new algorithm, the Modified Moving Contracting Window Pattern Algorithm (CMCWPM), for the calculation of field
similarity. It strongly relies on previous work by Yang et al. (2001), correcting previous work in which characters marked as inaccessible for further pattern matching were not treated as boundaries between subfields,
occasionally leading to higher than expected scores of field similarity. A
reference Python implementation is provided.
Keywords: field similarity, string similarity, string comparison.
1
Introduction
Assessing string similarity (or, as commonly referred in the literature, “field similarity”), is a recurrent problem in computer science, particularly when dealing
with natural language and genetic data. A review of relevant literature on the
topic is presented by Yang et al. [13], which stresses the importance of these
methods in applications such as data searching and cleansing, Web searching,
computational biology, and data compression.
These areas of research were considered by the authors before introducing
their method for field comparison, the Moving Contracting Window Pattern
Algorithm (MCWPA). MCWPA is fundamentally different from the general
methods for field comparison based in measures of “edit distance” (such as the
one proposed by Wagner et Fischer, [14]), which expand initial work by Vladimir
Levenshtein [8]. These methods, commonly referred to under the single label of
“Levenshtein Distance”, are generally defined as counts of the minimum number
of single character “edits” required to mutate a field FX into a field FY , where
“edit” is defined as either an insertion, a deletion, or a substitution of a single
character; in general, no difference in weight for the various types of “edit” is
specified. The method for this computation of field similarity is closely related
to pairwise field alignment and usually implemented after the aforementioned
Wagner-Fischer algorithm or through dynamic programming approaches such
as the one proposed by Guseld [4].
The algorithm proposed by Yang et al. extends and generalizes an alternative “token–based” approach by Lee et al. [7], developed in the context of
1
data cleansing. In their paper, the authors present a pseudo-code for the algorithm, claiming that their solution “not only achieve[s] higher accuracy but also
gain[s] the time complexity O(knm) (k < 0.75) for worst case”, comparing the
accuracy of their proposal with the one of the method by Lee et al. and concluding that “[t]heoretical analysis, concrete examples and experimental result
show that [the proposed] algorithms can significantly improve the accuracy and
time complexity of the calculation of Field Similarity”.
In the course of a research conducted around 2005 with extensive usage of the
Natural Language Toolkit (NLTK), a Python library and framework for Natural
Language Processing by Bird et al. [2], the author of the current paper needed
to perform hundreds of field comparisons for sorting lists of fields according to
their similarity to a number of reference field, usually short strings containing
natural language data. Levenshtein distance, the most recommended method,
proved slow when computed without dynamic methods and, more importantly,
was found to be unsuitable for a considerable number of cases, as its scores of
similarity, adjusted to ratios between 0.0 and 1.0, failed to match the magnitude of similarity that most speakers of the natural languages in study would
expect or report. A theoretical investigation suggested that the obstacle was
an intrinsic limitation of the algorithm itself given by its focus in general field
comparison (i.e., with no a priori assumptions on the entropy of both fields),
and was triggered by idiosyncrasies of the morphology and the orthography of
the languages in analysis. While the difficulties could in part be circumvented
with a combination of orthographic, phonological and, exceptionally, morphological mappings, the decision rested in adopting new methods. A bibliographic
research suggested the paper by Yang et al., and, while for our purposes the
new algorithm performed better than the edit distance method, its results were
occasionally unexpected and, eventually, worse than Levenshtein distance for
some corner cases. The author wrote a revised version that partially solved the
deviations, and the Python module which implemented them was eventually
included among the “contributions” to NLTK1 .
When needing to perform a similar task almost a decade after that first
revision, the author decided to write a new version which fully and correctly
implemented the revised method, presenting the algorithm and its implementation in this paper.
2
Background
A brief but throughly description of the “token-based” approach for the computation of field similarity proposed by Lee et al. is given in the second section
of Yang et al., from which the outline of the current section is developed.
Let a field X of length n be composed of tokens (such as “words”) TX1 , TX2 ,
. . . , TXn and the corresponding field Y of length m be composed of tokens TY1 ,
TY2 , . . . , TYm . Each token TXi , where 1 ≤ i ≤ n, is compared with each token
TYj , where 1 ≤ j ≤ m. Let DoSX1 , DoSX2 , . . ., DoSXn , DoSY1 , DoSY2 , . . .,
DoSYm be the maximum degree of similarity for tokens OX1 , OX2 , . . . , OXn ,
OY1 , OY2 , . . . , OYm , respectively. The Field Similarity between FX and FY is
computed as follows:
1 The module has apparently been removed from newer versions of NLTK, but can be easily
found in public forks based on older versions of the toolkit.
2
Pn
DoSXi +
Pm
j=1 DoSYj
(1)
n+m
The algorithm proposed by Yang et al. generalizes the “tokens” employed
by Lee et al., essentially words in natural languages, into “window patterns”,
which are defined as subfields of minimal length equal to 1. As in the first
example given in their paper, for the string "abcde", considering a window of
size 3 sliding from left to right, the series of patterns obtained is composed of
"abc", "bcd", and "cde". The field similarity in MCWPA is given by the sum
of the squares of the number of the same characters, or minimal units, between
fields FX and FY , which is defined as the cumulative sum of the square of
combined length of minimal units matched in both fields, i.e. twice the length
of the pattern; the sum is accumulated while marking already matched subfields
as inaccessible for further comparisons.
Thus, in MCWPA, let a field FX of n characters and a field FY of m characters; the field similarity between the two fields, which “approximately reflects
the ratio of the total number of the common characters in two fields to the total number of characters in two fields”, where SSNC represents the Sum of the
Square of the Number of Same Characters between FX and FY , is computed as
follows:
s
SSN C
(2)
SIMF(X,Y ) =
(n + m)2
SIMF(X,Y ) =
i=1
The algorithm is described in depth by Yang et al., with a number of examples and graphical representations of the inner workings of the sliding window
approach.
3
Changes to MCWPA
The author of the current paper first implemented the MCWPA algorithm in
Python following the pseudo-code given in Figure 1 in Yang et al.. While the
authors did not offer actual code or reference values to test implementations
of theirs algorithms, all the examples could be matched, suggesting that the
implementation was correct.
When testing the implementation in production code, however, it was verified that for some corner cases the results returned were unsuitable, with scores
generally higher than what was expected by human reviewers. The author also
experimented with some random strings used in imitation of genetic data, generated by an ad hoc weighted random function according to a table of DNA codon
frequencies for the human genome; for a restricted number of test samples the
results were considered equally unsatisfactory.
An investigation of the algorithm did not prove sufficient for identifying any
theoretical limit as a source for the unsatisfactory scores. After implementing
the algorithm in multiple and different ways, by a trial and error methodology
an hypothesis was developed that the problem resided in a simplification of the
original implementation, which can be found in the pseudo-code itself and might
have been intentional, as MCWPA is less computationally expansive than the
3
revised method here proposed and the limitation affected a small number of
cases.
In detail, while the theoretical description of the paper and the pseudo-code
correctly call for marking characters in FX and FY as “inaccessible” after a
given pattern matching, the implementation was apparently not marking inaccessible characters as a boundary for future pattern matchings, thus allowing
new windows to “jump over them”. The limitation might have been introduced
when adapting the method from a token based to a character based approach,
as the implementation in Lee et al. doesn’t seem to allow non contiguous tokens
to be matched. This hypothesis cannot be verified without access to the original
source code.
To illustrate the difference of the algorithm here proposed, the implementation yielding results considered wrong would, when matching the strings
"A123B" and "123AB", first match the pattern of characters "123", but after the deletion of this pattern in both fields it would not treat the residual
characters as groups of non-overlapping and non-contiguous subfields ("A" and
"B" for the first field, "" and "AB" for the second), but as two identical "AB"
strings. When reducing the length of the window from 3 to 2, the implementation would incorrectly find a match of a substring of length 2, when, from the
theoretical stand point of the algorithm, it would be supposed to identify two
different matches of length 1, with a lower final score.
The Python implementation presented in the following section solves this
problem by replacing the operation of string concatenation of the first version by
operations on lists, introducing the concept of “sub-fields”, i.e., non-overlapping
and non-contiguous factors resulting after the removal of a specific factor (the
pattern being matched) from a starting field or subfields. When the algorithm
matches a pattern, it returns two subfields, i.e., the characters that precede and
the characters that follow the matched pattern, if any. As the subfields might
be empty, when the match includes the first or the last character in the string,
a check is performed to filter out such empty subfields from the list of subfields.
As stated, the distortion was only found in corner cases, and actual scores
were higher than the expect only to a reduced limit. However, as bibliographical research did not find any mention or correction to Yang et al., even though
their paper has a considerable number of citations, the author found it useful
to publish this corrected implementation under the name of MMCWPA (Modified Moving Contracting Window Pattern Algorithm). The author wishes to
publicity note that his implementation distributed with old versions of NLTK
at the time of writing is still affected by the problem describe above, and should
be replaced whenever possible by the one presented here.
4
Python implementation
A stand-alone Python implementation for the algorithm is presented in this
section. As per the terms of the “MIT License”, permission is hereby granted,
free of charge, to any person obtaining a copy of this software and associated
documentation files (the “Software”), to deal in the Software without restriction,
including without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Software, and to permit persons
to whom the Software is furnished to do so, subject to the following conditions:
4
• The above copyright notice and this permission notice shall be included
in all copies or substantial portions of the Software.
• The software is provided “as is”, without warranty of any kind, express
or implied, including but not limited to the warranties of merchantability,
fitness for a particular purpose and noninfringement. In no event shall the
author be liable for any claim, damages or other liability, whether in an
action of contract, tort or otherwise, arising from, out of or in connection
with the software or the use or other dealings in the software.
Listing 1: CMCWPA
def mmcwpa( f x , f y , s s n c ) :
”””
An i m p l e m e n t a t i o n o f t h e M o d i f i e d Moving C o n t r a c t i n g
Window P a t t e r n A l g o r i t h m (MMCWPA) t o c a l c u l a t e s t r i n g
s i m i l a r i t y , r e t u r n s a l i s t o f non−o v e r l a p p i n g ,
non−c o n t i g u o u s f i e l d s Fx , a l i s t o f non−o v e r l a p p i n g ,
non−c o n t i g u o u s f i e l d s Fy , and a number i n d i c a t i n g t h e
Sum o f t h e S qu are o f t h e Number o f t h e same
C h a r a c t e r s . This f u n c t i o n i s i n t e n d e d t o be ” p r i v a t e ” ,
c a l l e d from t h e ” p u b l i c ” s t r i n g c o m p ( ) f u n c t i o n b e l o w .
@param f x : A C{ l i s t } o f C{ s t r i n g s } .
@param f y : A C{ l i s t } o f C{ s t r i n g s } .
@param s s n c : A C{ f l o a t } .
@return : A C{ l i s t } o f C{ s t r i n g s } w i t h
non−o v e r l a p p i n g , non−c o n t i g u o u s s u b f i e l d s
f o r Fx , a C{ l i s t } o f C{ s t r i n g s } w i t h
non−o v e r l a p p i n g , non−c o n t i g u o u s s u b f i e l d s
f o r Fy , and a C{ f l o a t } w i t h t h e v a l u e o f
t h e SSNC c o l l e c t e d so f a r .
@rtype : C{ l i s t } o f C{ s t r i n g s } , C{ l i s t } o f C{ s t r i n g s } ,
C{ f l o a t }
”””
# t h e b o o l e a n v a l u e i n d i c a t i n g i f a t o t a l or p a r t i a l
# match was fou n d bet ween s u b f i e l d s o f Fx and Fy ; when
# a match i s found , t h e v a r i a b l e i s u sed t o c a s c a d e
# out of the loops of the fu n ct ion
match = F a l s e
# t h e v a r i a b l e s where t o s t o r e t h e new c o l l e c t i o n s o f
# s u b f i e l d s , i f any match i s fou n d ; i f t h e s e v a l u e s
# are n ot chan ged and t h e empty l i s t s are r e t u r n e d ,
# s t r i n g c o m p ( ) w i l l b r e a k t h e l o o p o f comparison ,
# c a l c u l a t e t h e s i m i l a r i t y r a t i o and r e t u r n i t s v a l u e
new f x , new f y = [ ] , [ ]
5
# s e a c h p a t t e r n s i n a l l s u b f i e l d s o f Fx ; t h e i n d e x o f
# t h e s u b f i e l d i n t h e l i s t i s u sed f o r u p g r a d i n g t h e
# l i s t , i f a p a t t e r n i s fou n d
for i d x x , s f x in enumerate ( f x ) :
# ‘ l e n g t h ‘ s t o r e s t h e l e n g t h o f t h e s l i d i n g window ,
# from f u l l l e n g t h t o a s i n g l e c h a r a c t e r
for l e n g t h in r a ng e ( l e n ( s f x ) , 0 , −1):
# ‘ i ‘ s t o r e s the s t a r t i n g index of the s l i d i n g
# window i n Fx
for i in r a ng e ( l e n ( s f x )− l e n g t h + 1 ):
# e x t r a c t t h e p a t t e r n f o r mat chin g
p a t t e r n = s f x [ i : i+l e n g t h ]
# l o o k f o r t h e p a t t e r n i n Fy
for i d x y , s f y in enumerate ( f y ) :
# ‘ j ‘ s t o r e s t h e s t a r t i n g i n d e x i n Fy ; t h e
# Python f i n d ( ) f u n c t i o n r e t u r n s −1 i f t h e r e
# i s no match
j = s f y . find ( pattern )
i f j > −1:
# t h e p a t t e r n was fou n d ; s e t ‘ n ew fx ‘ and
# ‘ n ew fy ‘ t o v e r s i o n o f ‘ f x ‘ and ‘ fy ‘ w i t h
# t h e p a t t e r n s removed , u p d a t e t h e SSNC and
# s e t ‘ match ‘ as True , i n o r d e r t o c a s c a d e
# out of the loops
tmp
tmp
new
new
x
y
f
f
=
=
x
y
[ sf
[ sf
= f
= f
x
y
x
y
[:
[:
[:
[:
i],
j],
idx
idx
sf
sf
x]
y]
x [ i+l e n g t h
y [ j+l e n g t h
+ tmp x +
+ tmp y +
:]]
:]]
f x [ idx x +1:]
f y [ idx y +1:]
s s n c += ( 2 ∗ l e n g t h )∗ ∗ 2
match = True
break
# i f t h e c u r r e n t p a t t e r n was found , end s e a r c h
i f match :
break
# i f a match was found , end s l i d i n g window
i f match :
break
# i f a match was found , end Fx s u b f i e l d en u merat ion
i f match :
break
# remove any empty s u b f i e l d s due t o p a t t e r n removal
6
n e w f x = [ s f for s f in n e w f x i f s f ]
n e w f y = [ s f for s f in n e w f y i f s f ]
return n e w f x , n e w f y , s s n c
def s t r i n g c o m p ( s t r x , s t r y ) :
len x , len y = len ( s tr x ) , len ( s tr y )
f x , f y = [ str x ] , [ str y ]
ssnc = 0.0
while True :
f x , f y , s s n c = mmcwpa( f x , f y , s s n c )
i f l e n ( f x ) == 0 or l e n ( f y ) == 0 :
break
return ( s s n c / ( ( l e n x+l e n y ) ∗ ∗ 2 . ) ) ∗ ∗ 0 . 5
5
Evaluation
Table 1 provides some reference scores as returned by the Python implementation given in the previous section. Besides some new examples, we reproduce all
the string comparison employed by Yang et al. when presenting the MCWPA.
FX
Table 1: Evaluation for MMCWPA
FY
SIMF(X,Y )
abc
abcdef
Austria
Python
a123b
129 Industry Park
abc de
de abc
abc de
Fu Hui
Fu Hui
abcdefgh ijklmnpo
akabc axyz mo
abcdefagha
Gao Hua Ming
zeng zeng
def
abcdef
Australia
python
ab123
129 Indisttry Park
abc k de
de abc
de abc
Mr Fu Hui
Fu Mr Hui
abcdefgh ijklmnwo
aabc axyz muo
aijklamabc
Gao Ming Hua
zeng hong
7
0.0000
1.0000
0.6731
0.8333
0.6982
0.6101
0.6388
1.0000
0.6236
0.8000
0.5962
0.8843
0.7768
0.3316
0.5892
0.5983
6
Conclusion
This paper presented a new algorithm, Modified Moving Contracting Window
Pattern Algorithm (MMCWPA) for the calculation of field similarity, strongly
relying on previous work by Yang et al. (which are in no way associated with
this work). As for MCWPA, theoretical analysis, concrete examples, and experimental results indicate that MMCWPA improves the accuracy and efficiency of
the calculation of field similarity and should be considered alongside with other
field metrics, particularly when dealing with short strings representing natural
language.
References
[1] S. Alstrup, G. S. Brodal and T. Rauhe, “Pattern matching in dynamic
texts”, In Proceedings of the 11th Annual Symposium on Discrete Algorithms, pages 819–828, 2000.
[2] S. Bird, E. Loper, and E. Klein, “Natural Language Processing with
Python”, O’Reilly Media Inc., 2009.
[3] Z. Galil, “Open problems in stringology”, In Combinatorial Algorithms on
Words, pages 1–8. Springer, 1985.
[4] D. Guseld, “Algorithms on Strings, Trees and Sequences”, in Computer
Science and Computational Biology, CUP, 1997.
[5] R. M. Karp, R. E. Miller, and A. L. Rosenberg, “Rapid identification of repeated patterns in strings, trees and arrays”, In 4th Symposium on Theory
of Computing, pages 125–136, 1972.
[6] R. M. Karp and M. O. Rabin, “Efficient randomized pattern-matching
algorithms”, IBM Journal of Research and Development, 31(2):249–260,
March 1987.
[7] M. L. Lee, H. Lu, T. W. Ling, and Y. T. Ko, “Cleansing data for mining and warehousing”, in Proceedings of the 10th International Conference
on Database and Expert Systems Applications (DEXA99), pages 751–760,
August 1999.
[8] V. I. Levenshtein, “Binary codes capable of correcting deletions, insertions,
and reversals”, Soviet Physics Doklady 10(8):707–710, February 1966.
[9] K. Mehlhorn, R. Sundar and C. Uhrig, “Maintaining dynamic sequences
under equality tests in polylogarithmic time”, Algorithmica, 17(2):183–198,
February 1997.
[10] A. E. Monge and C. P. Elkan, “The field matching problem: Algorithms
and applications”, Proc. of the 2nd Int. Conference on Knowledge Discovery
and Data Mining, pages 267–270, 1996.
[11] S. Muthukrishnan and S. C. Sahinalp, “Approximate nearest neighbors
and sequences comparison with block operations”, In 32nd Symposium on
Theory of Computing, 2000.
8
[12] S. C. Sahinalp and U. Vishkin, “Simmetry breaking for suffix tree construction”. In 26th Symposium on Theory of Computing, pages 300–309,
1994.
[13] Q. X. Yang, S. S. Yuan, L. Chun, L. Zhao, S. Peng, “Faster Algorithm of
String Comparison”, eprint arXiv:cs/0112022, December 2001.
[14] R. Wagner and M. Fischer, “The String to String Correction Problem”,
JACM 21, pages 168–173, 1974.
9
| 8cs.DS
|
arXiv:1610.06026v1 [math.ST] 19 Oct 2016
Least squares estimation in the monotone single
index model
Fadoua Balabdaoui, Cécile Durot and Hanna Jankowski
Université Paris-Dauphine
PSL Research University
CNRS, CEREMADE
75016 Paris, France
e-mail: [email protected]
Seminar für Statistik
ETH Zürich
8092, Zürich, Schweiz
e-mail: [email protected]
Université Paris Nanterre
Modal’x
200 avenue de la république
92001 Nanterre Cedex, France
e-mail: [email protected]
Department of Mathematics and Statistics
York University
4700 Keele Street
Toronto ON, M3J 1P3, Canada
e-mail: [email protected]
Abstract: . We study the monotone single index model where a real response variable Y is
linked to a d-dimensional covariate X through the relationship ErY |Xs “ Ψ0 pαT
0 Xq almost
surely. Both the ridge function, Ψ0 , and the index parameter, α0 , are unknown and the
ridge function is assumed to be monotone on its interval of support. Under some regularity
conditions, without imposing a particular distribution on the regression error, we show
the n´1{3 rate of convergence in the ℓ2 -norm for the least squares estimator of the bundled
function ψ0 pαT
0 ¨q, and also that of the ridge function and the index separately. Furthermore,
we show that the least squares estimator is nearly parametrically rate-adaptive to piecewise
constant ridge functions.
AMS 2000 subject classifications: Primary 62G08, 62G20; secondary 62J12.
Keywords and phrases: least squares, maximum likelihood, monotone, semi-parametric,
shape-constraints, single-index model.
1. Introduction
1.1. The generalized linear model and the single index model
The generalized linear model is widely used in econometrics and biometrics as a standard tool
in parametric regression analysis, see e.g. Dobson and Barnett (2008), and can be described as
follows. Let n be the size of a sample of individuals randomly selected from some population. For
each individual i “ 1, . . . , n, a response variable and a d-dimensional vector of covariates have
been measured. The observations are assumed to be independent copies of a pair pX, Y q such
that
EpY |Xq “ Ψ0 pαT0 Xq
(1.1)
almost surely with an unknown index α0 P Rd zt0u and a monotone ridge function Ψ0 , also
called the inverse link function. Here, α0 is unknown whereas Ψ0 is assumed to be known. In
1
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
2
the generalized linear model, the conditional density of Y given X “ x with respect to a given
dominating measure (typically either Lebesgue measure or a counting measure) is assumed to
take the form
"
*
yℓpµpxqq ´ B ˝ ℓpµpxqq
y ÞÑ hpy, φq exp
(1.2)
φ
where h is a normalizing function, µpxq is the mean, φ ą 0 is a possibly unknown dispersion
parameter, ℓ is a given real valued function with first derivative ℓ1 ą 0, and inverse ℓ´1 “ B 1 .
The generalized linear model includes of course the Gaussian linear model, which is arguably the
most popular parametric regression model. In this model, the conditional distribution function of
Y given X is Gaussian with an unknown variance σ 2 independent of X and expectation EpY |Xq “
αT0 X for an unknown vector α0 . Here, φ “ σ 2 and both ℓ and Ψ0 are the identity function on R.
The generalized linear model covers a number of other popular parametric regression models such
as the Poisson model, and logit and probit models for Binomial observations. Although popular,
parametric models lack the flexibility offered by more robust approaches.
The single index model can be viewed as an attempt at extending the generalized linear one
with the aim of gaining more flexibility. It is widely used, for instance in econometrics, as a
compromise between too restrictive parametric assumptions and a fully non parametric setting.
Fully non-parametric models can suffer from the “curse of dimensionality” in high-dimensional
problems. In contrast, the single index model assumes that the conditional expectation of Y
given X depends only on the linear combination αT0 X (the linear predictor) of the components
of X. This means that similar as above, the conditional expectation of Y given X takes the form
EpY |Xq “ Ψ0 pαT0 Xq almost surely, but now, the ridge function Ψ0 is unknown. Furthermore, it is
no longer assumed that the conditional distribution of Y given X belongs to a given exponential
family, making the model even more flexible, and robust to assumption misspecification.
Standard methods for estimating α0 and Ψ0 rely on smoothness assumptions on Ψ0 , and hence
involve a smoothing parameter which has to be carefully chosen, see e.g. Härdle et al. (1993),
Chiou and Müller (2004), Hristache et al. (2001) and references therein. Note that α0 and Ψ0
are not identifiable if left unrestricted. To see this, let }α0 } denote the Euclidean norm of α0 ,
β0 “ α0 {}α0 } and Φ0 ptq “ Ψ0 p}α0 }tq for all t. Then, we have EpY |Xq “ Ψ0 pαT0 Xq “ Φ0 pβ0T Xq
almost surely. We obtain the same equalities by considering β0 “ ´α0 and Φ0 ptq “ Ψ0 p´tq for
all t. This issue could be resolved by assuming e.g. that }α0 } “ 1 and the first non-null entry
of α0 is positive. Additional constraints may be also imposed on Ψ0 and the distribution of the
covariate X in order to ensure identifiability.
1.2. The monotone single index model
In this paper, we are interested in the single index model where the unknown ridge function is
assumed to be monotone. This is motivated by the fact that monotonicity appears naturally in
various applications, which is one of the reasons behind the popularity of the generalized linear
model. We follow a non-parametric estimation approach which does not require a smoothing
parameter. Precisely, we assume (1.1) where α0 P Rd zt0u is such that }α0 } “ 1, and Ψ0 is assumed
to be non-decreasing. The latter assumption can be made without loss of generality since changing
α0 into ´α0 changes automatically the monotonicity direction of Ψ0 . Furthermore, we do no make
any specific assumption about the conditional distribution of Y given X, thereby avoiding model
misspecification. Throughout this paper, we will refer to this model as the monotone single index
model, a term that has been already used in the literature.
The monotone single index model, with the additional assumption that Y ´EpY |Xq is independent of X, has been considered by Foster et al. (2013), where an estimator for α0 was proposed
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
3
based on combining isotonic regression with a smoothing method (which involves a tuning parameter), and also by Kakade et al. (2011), where an algorithm for simultaneously estimating the
index and the ridge function is provided under the assumption that the ridge function is Lipschitz
(the Lipschitz constant is a parameter of the algorithm). It enters the setting of Han (1987) with
(using the notation of that paper) F px, uq “ f pxq ` u and Dptq “ t with f a monotone function.
In the general setting they consider, Han (1987) prove consistency of a non-parametric estimator
for the index that does not require any tuning parameter. The monotone single index model
is also closely related to the model considered by Chen and Samworth (2016), who, in contrast
to the approach followed here, assume that the conditional distribution of the response variable
given the predictor vector belongs to a given exponential family. Another difference is that the
authors consider a more general assumption on the conditional expectation. Indeed, they allow
the conditional expectation of Y given X to depend on several linear predictors: with ℓ taken from
(1.2), ℓpEpY |Xqq can be written as a sum of ridge functions of the linear predictors, where each
ridge function satisfies a different shape constraint such as monotonicity, convexity, or linearity.
The authors show consistency for a slightly modified maximum likelihood estimator, which is
obtained by maximizing the likelihood over the closure of the set of all possible parameters.
We also point out that the monotone single index model covers the problem of current status
regression, in which the response Y ě 0 is not completely observed. Instead, independent copies
of pX, C, ∆q are observed where X P Rd is the predictor, C ě 0 is an observed censoring time
independent of Y , and ∆ “ 1tY ďCu Although non-observed, Y is assumed to satisfy the linear
regression model Y “ αT0 X ` ε where α0 P Rd and ε is independent of pC, Xq with unknown
distribution function F0 . Let X̃ denote the random vector in Rd`1 such that X̃ T “ pC, X T q, α̃0
the vector in Rd`1 such that α̃T0 “ p1, ´αT0 q and Ỹ “ ∆. Then, we have EpỸ |X̃q “ F0 pα̃T0 X̃q
where F0 is non-decreasing (since it is a distribution function). Here, the conditional distribution
of Ỹ given X̃ is Bernoulli, with density function (with respect to the counting measure on t0, 1u)
given by (1.2) where ℓ is the logit function ℓpµq “ logpµ{p1 ´ µqq for µ P p0, 1q. Note that in the
particular case where the censoring time C ” 0 is degenerate has been widely used in econometrics
and is usually referred to as the binary choice model. The maximum likelihood estimator (MLE)
of α0 was proved to be consistent by Cosslett (1983), and Murphy et al. (1999) prove that the rate
of convergence is Opn´1{3 q in the one-dimensional case
? (that is, when d “ 1). The latter paper
also proves that an appropriately penalized MLE is n-consistent but the considered estimator
?
is difficult to implement. Groeneboom and Hendrickx (2016) consider several alternative nconsistent estimators based on a truncated likelihood, where the truncation parameter has to be
chosen by the practitioner.
1.3. Contents of the paper
In the monotone single index model, we consider the least squares estimator which aims at estimating both the index and ridge function without choosing a tuning parameter. We give a
characterization of the least-squares estimator (LSE) of pΨ0 , α0 q under the monotonicity constraint using a profiling approach. Furthermore, letting g0 pxq “ EpY |X “ xq “ Ψ0 pαT0 xq the
regression function, we prove that under appropriate conditions the LSE of g0 converges at a
n1{3 -rate in the L2 -sense. Then, we consider the LSE of α0 and Ψ0 separately and prove their
n1{3 -consistency. The n1{3 -rate of convergence that we obtain for the LSE of the parametric index,
α0 , may be due to our strategy of proof, as we derive this rate from the n1{3 -rate of the LSE of
g0 . Thus, it is not excluded that sharper rates could be obtained using alternative methods. This
is however out of the scope of this work. For the sake of completeness, we devote a small section
to the maximum likelihood estimators (MLE) under the assumption that the conditional distribution of Y given X belongs to a given exponential family. Here, we maximize the likelihood over
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
4
the set of all possible parameters in contrast to the approach followed by Chen and Samworth
(2016), where maximization is performed over the closure of that set. We prove existence of the
MLE and give a characterization thereof but we chose not to pursue further investigations on this
estimator: we prefer using a more robust estimator like the LSE, which does not require specific
assumptions on the conditional distribution of the response variable given the predictor.
Furthermore, we also study the adaptive properties of our LSE approach. Shape-constrained
methods are well-known to be naturally adaptive. See, for example, van de Geer (1993, 2000);
Chatterjee et al. (2015); Han and Wellner (2016). Following Chatterjee et al. (2015), we study
the global convergence rate when the truth Ψ0 is piecewise constant. In Theorem 4.2, we show
that the rate of convergence of g0 in this case is nearly parametric.
The paper is organized as follows. In Sections 2 and 3, we show existence of the LSE and MLE
of pΨ0 , α0 q and give a characterization of the estimators. Our main result is given in Theorem 4.1
of Section 4, where we establish the n1{3 -convergence rate of the LSE of the bundled regression
function g0 p¨q “ Ψ0 pαT0 ¨q. In Section 5, we show under some specified assumptions that the LSE
of α0 and Ψ0 converge separately at the same rate in the Euclidean norm on Rd and the L2 -norm
on the set of real valued functions respectively, provided that we restrict integration to a bounded
subset of the domain of Ψ0 . Section 4 also gives the adaptive rates of convergence for the LSE.
The proof of the main Theorem 4.1 is given in Section 6. The proof of Theorem 4.2 as well as
some additional proofs are given in an Appendix in the Supplementary Material.
2. Existence and characterization of the least squares estimator
Assume that we observe an i.i.d. n-sample from pX, Y q such that EpY |Xq “ Ψ0 pαT0 Xq almost
surely, where both the index α0 and the monotone ridge function Ψ0 are unknown. For identifiability purposes, as explained in the introduction, α0 is assumed to belong to the d-dimensional
unit Euclidean sphere denoted by Sd´1 , whereas Ψ0 is assumed to be non-decreasing on its domain, which contains the range of the linear predictor
αT0 X. Assume that we have m distinct
´1 řni
d
data points X1 , . . . , Xm in R and denote Yi “ ni
j“1 Yij for all i “ 1, . . . , m, where ni is the
number of observationsřat the data point Xi and Yi1 , Yi2 , . . . , Yini are the corresponding response
variables. Hence, n “ m
i“1 ni .
The goal is to find the LSE of pΨ0 , α0 q, the minimizer of the least-squares criterion
hn pΨ, αq “
m
ÿ
i“1
(2
ni Yi ´ ΨpαT Xi q
over M ˆ Sd´1 , where M is the class of all non-decreasing functions on R. Using the profiled
least-squares approach, we first consider the problem of minimizing Ψ ÞÑ hn pΨ, αq over M for
p α in the criterion. Finally, we minimize
a fixed α P Sd´1 , then plug the obtained minimizer Ψ
n
α
p
α ÞÑ hn pΨn , αq over α P Sd´1 . Denoting the minimizer by α
pn , the minimum of hn pΨ, αq is then
p n q where Ψ
pn “ Ψ
p αp n , whence pp
p
achieved at pp
αn , Ψ
α
,
Ψ
q
is
the
(non unique) LSE. In the following,
n
n
n
we discuss precisely the way in which the LSE is non-unique, the algorithm to compute it, as
well as existence.
First consider the problem of minimizing Ψ ÞÑ hn pΨ, αq over M for a fixed α P Sd´1 . For this
task, order the observations based on the values of the linear predictors αT Xi ’s corresponding
to α. Ties may occur in the sense that we may have αT Xi “ αT Xj for i ‰ j, so we denote
α
the distinct ordered statistics of
by mα the number of distinct values, and by Z1α ă . . . ă Zm
α
T
T
α X1 , . . . , α Xm . Hence, we have mα ď m with an equality in the case of no ties. Then, with
nα
k “
m
ÿ
i“1
ni IαT Xi “Zkα
and Ykα “
m
1 ÿ
ni Yi IαT Xi “Zkα ,
nα
k i“1
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
we have
hn pΨ, αq “
m
ÿα
k“1
2
α
α
nα
k tYk ´ ΨpZk qu `
m
ÿ
i“1
ni Yi2 ´
m
ÿα
k“1
α 2
nα
k pYk q .
5
(2.3)
This means that minimizing hn pΨ, αq with respect to Ψ amounts to minimizing
Ψ ÞÑ
m
ÿα
k“1
α
α
nα
k tYk ´ ΨpZk qu
2
ř mα α α
2
nk tYk ´ ηk u over the set of all real numbers
over M, which in turn amounts to minimizing k“1
α
η1 ď ¨ ¨ ¨ ď ηmα , where we set ηk “ ΨpZk q. It is a standard fact in isotonic regression (see e.g.
Theorem 1.1 in Barlow et al. (1972)) that the minimum is achieved and that pη1 , . . . , ηmα q is a
pα
pα
minimizer if, and only if ηk “ dpα
k for all k “ 1, . . . , mα , where d1 ď . . . ď dmα are equal to the
left derivatives of the greatest convex minorant of the cumulative sum diagram defined by the
set of points
#
˜
¸
+
i
i
ÿ
ÿ
α
α α
nj ,
p0, 0q,
nj Yj , i “ 1, . . . , mα .
(2.4)
j“1
j“1
p α if and only if
Hence, the minimum of Ψ ÞÑ hn pΨ, αq over M is achieved at a non-decreasing Ψ
n
for all i “ 1, . . . , mα ,
α
pα
pα
(2.5)
Ψ
n pZi q “ di .
The solution is not unique, but is uniquely defined at the points Ziα .
pα
Next, we search for α
pn that minimizes p
hn pαq :“ hn pΨ
n , αq over α P Sd´1 . The following
proposition, which is proved in Subsection 8.1 of the Supplement, shows that the optimization
problem can be reduced to minimizing the above function over S X , the set of all α P Sd´1 such
that αT Xi ‰ αT Xj for all i ‰ j.
Proposition 2.1. We have
pα
pα
inf hn pΨ
n , αq “ inf hn pΨn , αq.
αPSd´1
αPS X
Now for all α P S X , we have mα “ m and there exists an ordering (i.e. a permutation on
t1, . . . , mu) πα induced by α such that αT Xπα p1q ă ¨ ¨ ¨ ă αT Xπα pmq , and we have
p α , αq “
hn pΨ
n
m
ÿ
k“1
!
)2
α
pα
nα
k Yk ´ dk
(2.6)
for all α P S X . This depends only on the ordering induced by α (and on the observations). As the
number of inducible orderings is finite, minimizing p
hn over S X hence amounts at minimizing p
hn
over a finite set, so that the minimum is achieved. This means that there exists a (non-unique)
α
pn that minimizes p
hn pαq over α P Sd´1 , and such that mαp n “ m.
p n q. We point
We conclude the section by a short comment on how to implement the LSE pp
αn , Ψ
out that from the above characterization, it follows that the LSE is not unique. Let P X be the
set of all inducible orderings, that is the set of all permutations π on t1, . . . , mu such that there
exist α P Sd´1 that induces π, i.e. that satisfies αT Xπp1q ă ¨ ¨ ¨ ă αT Xπpmq . For each π P P X ,
compute the left derivatives dπ1 ď . . . ď dπm of the greatest convex minorant of the cumulative
sum diagram defined by the set of points
¸
+
#
˜
i
i
ÿ
ÿ
nπpjq ,
nπpjq Yπpjq , i “ 1, . . . , m
p0, 0q,
j“1
j“1
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
6
using the pool adjacent violator algorithm (PAVA) or a similar device, see Barlow et al. (1972) for
(
ř
π 2
a precise description of the algorithm. Then choose π
pn that minimizes m
k“1 nπpkq Yπpkq ´ dk
over π P P X . The above characterization shows that α
pn is any index in Sd´1 that induces π
pn ,
p n is any non-decreasing function that satisfies Ψ
p n pp
and Ψ
αTn Xπp n piq q “ dπip n for all i.
The difficulty with the above line of implementation is that it requires that the set of all
inducible orderings be computable. One can check that the number of inducible orderings is on
the order of m2pd´1q , see Cover (1967). Also, given an inducible ordering, it is not easy to compute
the set of corresponding inducing indices in Sd´1 in d ą 2. As we are not aware of an efficient
algorithm to solve those problems, we suggest instead of using inducible orderings to estimate
α
pn using a stochastic search as done in Chen and Samworth (2016).
3. Existence and characterization of the maximum likelihood estimator
Similar to Section 2, we observe an i.i.d. n-sample from pX, Y q such that EpY |Xq “ Ψ0 pαT0 Xq
almost surely, where both the index α0 P Sd´1 and the non-decreasing ridge function Ψ0 are unknown. We are now interested in the maximum likelihood estimator (MLE) of pΨ0 , α0 q assuming
that with respect to some dominating measure, the conditional distribution of Y given X “ x
has a density of the form (1.2) where µpxq is the mean that belongs to a given interval pa, bq Ă R,
ℓ is a real valued function with a strictly positive first derivative on pa, bq, φ ą 0 is a dispersion
parameter, and h is a normalizing function. Here, a and/or b could be infinite.
If P pY “ a|Xq ą 0, we assume that aℓpηq ´ B ˝ ℓpηq converges to a finite limit as η Ó a.
Similarly, if P pY “ b|Xq ą 0, we assume that bℓpηq ´ B ˝ ℓpηq converges to a finite limit as
η Ò b. Note that these additional assumptions are satisfied in popular discrete models such as the
binomial model (where conditionally on X “ x, kY is a binomial variable with parameters k and
kµpxq) and the Poisson model (where conditionally on X “ x, Y has a Poisson distribution).
Assume, furthermore, that Y P ra, bs.
The goal is to find the MLE of pΨ0 , α0 q, if it exists. Using the same notation as in Section 2,
this amounts to minimizing
ℓn pΨ, αq
:“ ´
m
ÿ
i“1
ni Yi ℓrΨpαT Xi qs ´ B ˝ ℓrΨpαT Xi qs
(
over MˆSd´1 . Using the profiled likelihood approach, we first minimize Ψ ÞÑ ℓn pΨ, αq over M for
arbitrary fixed α P Sd´1 . An interesting fact is that the minimum is achieved at those functions
Ψ that minimize the least squares criterion hn p . , αq (and only at these ones). Considering the
characterization that we obtain for the minimizers of the least squares criterion, this can be
formulated as follows (see Section 8.2 in the Supplement for a proof).
Proposition 3.1. Under the above assumptions, for all fixed α P Sd´1 , ℓn pΨ, αq attains its
p α P M is a minimizer if, and only if, it satisfies
minimum over Ψ P M. Moreover, a function Ψ
n
(2.5) for all i “ 1, . . . , mα .
pα
Next, we search for a minimizer of ℓpn pαq :“ ℓn pΨ
n , αq over Sd´1 . Similar to Proposition 2.1,
p
it can be proved that the minimum of ℓn over Sd´1 is equal to the minimum over S X . Then,
similarly to the case of the least squares criterion, one concludes that the minimum can be
reduced to the minimum over a finite set, and hence is achieved. Denoting by α̃n a minimizer,
p α̃n , α̃n q is a MLE of pΨ0 , α0 q. Similar to the LSE, it is not unique and can
we conclude that pΨ
n
be computed by combining PAVA with a stochastic search.
The above results for the MLE compare to results obtained by Chen and Samworth (2016) in
a more general model. Specializing the model they consider to the current model, their results
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
7
show existence and give a characterization, for estimators of f0 “ ℓ ˝ Ψ0 and α0 , where ℓ is taken
from (1.2). Because ℓ is strictly increasing, estimating f0 is equivalent to estimating Ψ0 . When
estimating f0 , one may encounter estimators which take infinite values: whereas the MLE Ψ̃n
of Ψ0 always takes finite values (as can be seen from the preceding characterization), it may
happen that the MLE ℓ ˝ Ψ̃n of f0 takes infinite values. To give a simple example, assume that
the conditional distribution of Y given X is Bernoulli, and that the observations are such that
the corresponding MLE of Ψ0 satisfies inf t Ψ̃n ptq “ 0. Since in this case ℓ is the logit function,
we have ℓpηq Ñ ´8 as η Ñ 0 so that the MLE of f0 takes an infinite value in this situation.
To handle this technicality, Chen and Samworth (2016) must estimate pf0 , α0 q by maximizing a
criterion over the closure of the set of all possible parameters. This may offer advantages in the
complex model they consider, but for our setting, it is more natural to estimate pΨ0 , α0 q instead
of pf0 , α0 q.
We also point out that the LSE is a natural alternative to the MLE. Although for any fixed
α, the minimizers of hn p . , αq coincide with the minimizers of ℓn p . , αq, the MLE and the LSE of
pΨ0 , α0 q do not necessarily coincide, except of course in the Gaussian case. On the other hand,
the MLE is appropriate only if the assumption of an exponential family of distributions is valid.
However, the LSE is appropriate independently of any assumption on the conditional distribution
of Y given X. Hence, the LSE is more robust, and this is the estimator we focus on in the rest of
this paper. In particular, we no longer assume that the conditional distributions come from an
exponential family.
4. Convergence of the LSE for the regression function
We consider the same setting as in Section 2, however, we now also assume that X has a continuous distribution. This means that we observe an i.i.d. n-sample pXi , Yi q, i “ 1, . . . , n from a pair
pX, Y q where with probability one, all the Xi ’s are different from each other (hence, in the notation of Section 2, n “ m and ni “ 1 for all i “ 1, . . . , n). It is assumed that EpY |Xq “ Ψ0 pαT0 Xq
almost surely, where both the index α0 and the monotone ridge function Ψ0 are unknown, so the
regression function is defined by
g0 pxq “ Ψ0 pαT0 xq
(4.7)
for (almost-) all x in the support of X and its least-squares estimator (LSE) is given by
p n pp
gpn pxq “ Ψ
αTn xq
(4.8)
p n, α
for (almost-) all x in the support of X, where pΨ
pn q is the LSE of pΨ0 , α0 q studied in Section
p n, α
2: pΨ
pn q minimizes
n
ÿ
(2
hn pΨ, αq “
Yi ´ ΨpαT Xi q
i“1
p n that is left-hand continuous
over M ˆ Sd´1 . For convenience, we consider below a solution Ψ
and piecewise constant, and that can jump only at the points α
pTn Xi (but not necessarily all), for
i “ 1, . . . , n. Hence, ψpn is uniquely defined whereas α
pn is not unique (p
αn denotes an arbitrary
minimizer of p
hn , in the notation of Section 2). In this section, we are interested in the consistency
and rate of convergence of gpn in the L2 -sense.
4.1. The rate of convergence
We begin with some notation and assumptions. Let X be the support of the random vector of
covariate X. Let P be the joint distribution of pX, Y q, Px the conditional distribution of Y given
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
8
X “ x, and Q the marginal distribution of X. Theorem 4.1 below will be established under the
following assumptions.
A1. X is a bounded convex set,
A2. there exists a constant K0 ą 0 such that |g0 pxq| ď K0 for all x P X ,
A3. there exist constants a0 ą 0 and M ą 0 such that for all integers s ě 2 and x P X
ż
|y|s dPx pyq ď a0 s!M s´2 ,
(4.9)
A4. the distribution Q admits a density q with respect to Lebesgue measure and there exist
constants q ą 0, q̄ ą 0 such that for all x P X ,
q ď qpxq ď q̄
(4.10)
Assumption A1 ensures that for all α P Sd´1 , the set tαT x, x P X u, which is the support of
the linear predictor αT X corresponding to α, is convex (i.e. is an interval). Hence, we consider
functions of the form ΨpαT Xq where α P Sd´1 and Ψ is a non-decreasing function on its interval
of support tαT x, x P X u.
Assumption A3 ensures that conditionally on X “ x, the response variable Y is uniformly
integrable in x. This assumption is clearly satisfied if Y is a bounded random variable. It is
also satisfied if the conditional distribution of Y given X belongs to an exponential familly, see
Proposition 8.1 in Appendix 8.6 for more details.
Assumption A4 makes L2 -distances with respect to the probability measure Q and Lebesgue
measure equivalent. Also, certain arguments in the entropy calculations below can be given in
slightly simpler form using Lebesgue measure. Therefore, we have opted for the L2 -distance with
respect to this measure while computing bounds for the estimation error associated with the
LSE.
Theorem 4.1. With g0 and p
gn defined by (4.7) and (4.8) respectively, under assumptions A1-A4
above we have
ˆż
˙1{2
2
“ Op pn´1{3 q.
pp
gn pxq ´ g0 pxqq dx
X
On the other hand, when the truth is piecewise constant, we can show that the rate of convergence is much faster.
Theorem 4.2. Under assumptions A1-A4 and assuming that Ψ0 an increasing piecewise constant
function where κ denotes the number of unique values taken by Ψ0 , we have
ˆż
˙1{2
´
¯
“ Op pn{κq´1{2 plog nq3{2 .
pp
gn pxq ´ g0 pxqq dx
2
X
The proofs of Theorems 4.1 and 4.2 are similar to each other. For this reason, we present in
the next Section the main steps and technical difficulties encountered in proving Theorem 4.1,
while a complete proof of Theorem 4.2 is given in Appendix A (Section 7 in the Supplement).
Remark 4.3. To our best knowledge, the first adaptive results for the Grenander estimator
appear in Groeneboom and Pyke (1983), where the asymptotic distribution is obtained exactly.
Later, van de Geer (1993) showed how the same rate of convergence can be obtained via empirical
process arguments using local entropy bounds. Specficially, in van de Geer (1993) it is shown that
the MLE of a decreasing density converges at rate n´1{2 plog nq1{2 in Hellinger distance when
the true density is uniform. More recently, risk bounds were developed for isotonic regression by
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
9
Chatterjee et al. (2015), who obtain the same rate in the L2 norm. Similar (nearly parametric)
results were found in Han and Wellner (2016) for the multivariate convex bounded LSE when
the truth is a polyhedral functions supported on a polytopal region for d ď 4.
4.2. The main steps of the proof of Theorem 4.1
We will consider monotone functions with various intervals of support with the aim of comparing
them in a specific sense. For more convenience, we will extend each considered function outside
its actual support by taking the extension to be constant to the left and right of the end points of
that support. Depending on the context at hand, we will either consider extension to the whole
real line, or an extension to a bounded interval that will contain the support of all monotone
functions involved.
By definition of the LSE, gpn maximizes the criterion
Mn g
n
1 ÿ!
gpXi q2 )
Yi gpXi q ´
n i“1
2
:“
(4.11)
over the set of all functions g of the form gpxq “ ΨpαT xq, x P X with α P Sd´1 and Ψ P M. It
would have been easier to prove Theorem 4.1 using standard results from empirical process theory
if the LSE were known to be bounded in probability by some constant which is independent of
n. Unfortunately, we do not know whether this holds true. Instead, the following lemma can be
established (see Subection 6.1 for a proof).
Lemma 4.4. We have
min Yk ď gpn pxq ď max Yk
1ďkďn
1ďkďn
(4.12)
for all x P X . Moreover, under assumptions A2 and A3 we have
sup |p
gn pxq|
xPX
ď
max |Yk | “ Op plog nq.
1ďkďn
Note that under the more restrictive assumption that Y is a bounded random variable (so that
maxk |Yk | is bounded), we obtain that gpn is also bounded. This is the case for instance in the
current status model which, as explained in the introduction, is a special case of the model we
consider with Y P t0, 1u. For this reason, the arguments developed in Groeneboom and Hendrickx
(2016) in the current status model cannot be directly adapted to our setting.
gn maximizes Mn g over
Now it follows from Lemma 4.4 that, with arbitrarily large probability, p
the set of all functions g that are bounded in absolute value by C log n for some appropriately
chosen C ą 0, and take the form gpxq “ ΨpαT xq, x P X with pα, Ψq P Sd´1 ˆM. Denote by Pn the
empirical distribution corresponding to pX1 , Y1 q, . . . , pXn , Yn q, and let fpn px, yq “ yp
gn pxq´p
gn2 pxq{2
for x P X and y P R. Since Mn g “ Pn f with
f px, yq “ ygpxq ´ g 2 pxq{2
(4.13)
for all x P X , y P R, this means that, with arbitrarily large probability, fpn maximizes Pn f over
the set of all functions f of the form (4.13) for some function g that is bounded in absolute
value by C log n and takes the form gpxq “ ΨpαT xq, x P X with pα, Ψq P Sd´1 ˆ M. Hence,
classical arguments for maximizers of the empirical process over a class of functions (where g
can be assumed to be bounded by C log n) can be used to compute the rate of convergence of
the estimator. This requires bounds for the entropy of the class of functions f of the form (4.13)
together with a basic inequality that makes the connection between the mean of Mn pg ´ g0 q and
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
10
a distance between g and g0 . The entropy bounds are given in Section 4.3 below whereas the
basic inequality is given in the following lemma.ş For each bounded function g : X Ñ R, we define
Mg “ Pf where f is given by (4.13) and Pf “ f dP, which means that Mg is the expected value
of Mn g:
ż
!
g 2 pxq )
Mg “
ygpxq ´
dPpx, yq.
(4.14)
2
X ˆR
Lemma 4.5. Let g : X Ñ R be a bounded funtion. Under the assumption A4 we have Mpg´g0 q ď
´qD2 pg, g0 q{2 where
˙1{2
ˆż
.
(4.15)
pgpxq ´ g0 pxqq2 dx
Dpg, g0 q “
X
If classical arguments for maximizers over a class of functions are used based on the previous
basic inequality and Lemma 4.4 (which allows to restrict attention to functions g that are bounded
by C log n, for some large C ą 0 that does not depend on n), the obtained rate of convergence
would be inflated by a logarithmic factor:
Theorem 4.6. Under assumptions A1-A4, we have Dpp
gn , g0 q “ Op pn´1{3 plog nq5{3 q.
It may seem superfluous to add Theorem 4.6. However, the obtained unrefined rate of convergence will be used to get rid of the additional logarithmic factor. To explain how this works, set
v “ Cn´1{3 log n and K “ C log n for some constant C ą 0 to be chosen appropriately. Lemma
4.4 and Theorem 4.6 are used to show that with a probability that can be made arbitrarily large
by choice of C, the LSE p
gn is bounded by K while the L2 -distance between p
gn and g0 is smaller
that v. Hence, with arbitrarily large probability, the LSE maximizes Mn g over the set GKv of all
functions g of the form gpxq “ ΨpαT xq with α P Sd´1 and Ψ P M such that |gpxq| ď K for all
x P X and
Dpg, g0 q ď v.
(4.16)
Hence, although optimization cannot be restricted to a set of functions that are uniformly
bounded in n, we can work with a class of of functions that are bounded in the L2 -norm. The
merit of the latter is that under A2, a function g P GKv can be shown to exceed 2K0 only on a
subset of X with Lebesgue measure not larger than pv{K0 q2 . The fact that considered functions g
are bounded by 2K0 except on such a small region will balance out the large values that g might
have on the same region, and this will prove to be very advantageous in computing the final
entropy of the original class of functions. To estimate this entropy, each function gp.q “ ΨpαT .q
will be decomposed as follows:
g “ pg ´ ḡq ` ḡ
where ḡ is the truncated version of g defined
$
’
&gpxq
ḡpxq “ 2K0
’
%
´2K0
(4.17)
by
if |gpxq| ď 2K0
if gpxq ą 2K0
if gpxq ă ´2K0 .
(4.18)
The set of all possible ḡ forms now a class of bounded functions on which standard arguments
from empirical processes theory apply. On the other hand, the differences g ´ ḡ form a set of
functions whose supremum norm increases with n and which, by the discussion above, take the
value zero except on regions of a very small size. Those two classes of functions will be treated
with different arguments. Below are some entropy results required in the proof of Theorems 4.6
and 4.1. The complete proofs are given in Subsections 6.6 and 6.7 whereas the proofs of the
lemmas are given in Subsections 6.3 to 6.5.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
11
4.3. Entropy results
We begin with some notation. For any class of functions F equipped with a norm } ¨ }, and ε ą 0,
we denote by HB pε, F , } ¨ }q the corresponding bracketing entropy:
HB pε, F , } ¨ }q “ log NB pε, F , } ¨ }q
U
L
where NB pε, F , } ¨}q “ N is the smallest number of pairs of functions pf1U , f1L q, . . . , pfN
, fN
q such
L
U
L
that all }fj ´ fj } ď ε and for each f P F , there exists a j P t1, . . . , N u such that fj ď f ď fjU .
Moreover, assuming that X has a bounded support X , we set R “ supxPX }x} where } . } denotes
the Euclidean norm in Rd . We then have
|αT x| ď R
f or all x P X
(4.19)
for all α P Sd´1 , using the Cauchy-Schwarz inequality. Hence, we will consider monotone functions
Ψ defined on a sub-interval of r´R, Rs; we will extend these functions to the entire interval
r´R, Rs by taking the extension to be constant to the left and right of the end points of the
support. In the rest of the paper, we will use the following notation
– } ¨ }2,I is the L2 norm on I, where I can denote either an interval of R or a convex subset
of Rd , with respect to the corresponding Lebesgue measure,
– MIK is the class of all nondecreasing functions on some interval I that are bounded in
absolute value by K,
– GIK is the class of functions gpxq “ ΨpαT xq, x P X where α P Sd´1 , Ψ P MIK ,
– FIK is the class of functions f of the form (4.13), x P X , y P R, and g P GIK ,
– GIKv is the class of functions g P GIK satisfying the condition (4.16),
– FIKv is the class of functions f of the form (4.13), x P X , y P R, and g P GIKv ,
– G IKv is the class of functions g ´ ḡ , where g P GIKv and ḡ is given in (4.18),
– F IKv is the class of functions f ´ f¯, where f takes the form (4.13) for some g P GIKv and
f¯px, yq “ yḡpxq ´ ḡ 2 pxq{2, x P X , y P R.
Our starting point is the two following results.
Lemma 4.7. There exists a universal constant A ą 0 such that
?
A b´aK
HB pε, Mra,bsK , } . }2,ra,bs q ď
ε
for all ´8 ă a ă b ă 8, ε ą 0 and K ą 0.
The bound is an easy consequence of Theorem 2.7.5 in van der Vaart and Wellner (1996), and
hence its proof is omitted. The next result, which follows from Lemma 22 of Feige and Schechtman
(2002), gives a bound on the minimal number of subsets with diameter at most ε, say, into
which Sd´1 can be divided. Here, the diameter of some given subset A Ă Sd´1 , is given by
suppx,yqPA2 }x ´ y}.
Lemma 4.8. Fix ε P p0, π{2q, and let N pε, Sd´1 q be the number of subsets of equal size with
diameter at most ε into which Sd´1 can be partitioned. Then, there exists a universal constant
A ą 0, such that N pε, Sd´1 q ď pA{εqd .
In what follows, we assume that the assumptions A1 -A4 are fulfiled. The next step is to use
the results above to construct ε-brackets for the classes Gr´R,RsK , Fr´R,RsK , G r´R,RsKv and
F r´R,RsKv . We begin with the classes Gr´R,RsK and Fr´R,RsK .
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
12
Lemma 4.9. Let ε ą 0 and K ą ε. Then, there exists A1 ą 0 depending only on d and R such
that
´
¯ A K
1
.
HB ε, Gr´R,RsK , } . }2,X ď
ε
Moreover, if K ą 1 then there exists A2 ą 0 depending only on d, R, a0 and q̄ such that
´
¯ A K2
2
HB ε, Fr´R,RsK , } . }P ď
ε
ş
where } . }P is the L2 -norm corresponding to P: }f }2P “ X ˆR f 2 px, yqdPpx, yq.
Lemma 4.10. Let ε ą 0 and v ą 0. There exist constants A1 ą 0 and A2 ą 0 depending only
on K0 , d and X such that
ˆ
˙
´
¯
A1 K 2
A1 Kv
HB ε, Ḡr´R,RsKv , } . }2,X ď
` d log
ε
ε2
for all K ą ε such that Kv ą A2 ε. Moreover, there exists A1 ą 0 depending only on K0 , d, Q
and a0 such that
ˆ
˙
´
¯ A K 2v
A1 K 4
1
` 2d log
HB ε, F̄r´R,RsKv , } . }P ď
ε
ε2
for all K ą 1 _ ε such that Kv ą A2 ε.
The next lemma will be needed to give entropy bounds in the Bernstein norm. We recall that
the Bernstein norm of some function f with respect to P is defined by
´
¯1{2
.
}f }B,P “ 2Ppe|f | ´ 1 ´ |f |q
With C̃ ą 0 a constant appropriately chosen, Lemma 4.11 below derives form Lemma 4.9 and
4.10 upper bounds on the bracketing number of the classes
!
)
Frr´R,RsKv :“ pf ´ f0 qC̃ ´1 , f P Fr´R,RsKv
(4.20)
and
!
)
´1
Ă̄
F
:“
f
C̃
,
f
P
F̄
r´R,RsKv
r´R,RsKv
(4.21)
with respect to the Bernstein norm. This will enable us to use Lemma 3.4.3 of van der Vaart and Wellner
(1996) which does not require the class of functions of interest to be bounded.
Lemma 4.11. Let ε ą 0 and v ą 0. Let M be the same constant from the moment condition
(4.9) of Assumption A3. Let C̃ “ 4M K 2 such that K ě p2K0 q _ 2. Then, there exist constants
A1 ą 0 and A2 ą 0 depending only d, R, a0 , M , q̄ such that
´
¯ A
1
HB ε, Frr´R,RsKv , } ¨ }B,P ď
, and }f˜}B,P ď A2 v
ε
for all f˜ P Frr´R,RsKv . Similarly, there exist constants A1 ą 0 and A2 ą 0 depending only d, R,
a0 , M , q̄ and K0 such that
ˆ ˙
´
¯ A v
A1
1
Ă̄
HB ε, Fr´R,RsKv , } ¨ }B,P ď
` 2d log
, and }f˜}B,P ď A2 v
ε
ε2
Ă̄
for all f˜ P F
r´R,RsKv , provided that Kv ą A2 ε and K ą ε.
Note that the condition ε ă 2 guarantees that K ą ε since K ě 2. Also, we point out that the
constants A1 and A2 many not be the same ones as in Lemma 4.10 but we can always increase
their respective values so that they are suitable for Lemma 4.11.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
13
5. Convergence of the separated estimators
p n to Ψ0 . Moreover, we are
pn to α0 and Ψ
We now derive from Theorem 4.1 convergence of α
interested in the rate of convergence of the two estimators. Convergence can happen only under
uniqueness of the limit so first we prove identifiability of Ψ0 and α0 under appropriate conditions.
5.1. Identifiability of the separated parameters
Let pX, Y q be a pair of random variables, where X takes values in Rd and Y is an integrable real
valued random variable such that (1.1) holds for some α0 P Sd´1 and Ψ0 P M. Identifiability
of the parameter pα0 , Ψ0 q means here that if we can find β in Sd´1 , and h in M such that
Ψ0 pαT0 Xq “ hpβ T Xq a.s. then β “ α0 and h “ Ψ0 on Cα0 “ Cβ , where for all α P Sd´1 we
set Cα “ tαT x, x P X u with X being the support of X. Although identifiability can be derived
from Lin and Kulasekera (2007) when assuming that Ψ0 is non-constant and continuous, for
completeness we state below identifiability under a slightly less restrictive assumption, namely
left– (or right–) continuity instead of continuity. A proof can be found in Section 8.3 in the
Supplement. In the sequel, for all α P Sd´1 , we denote by Mα the set of all non-decreasing
functions on the interval Cα .
Proposition 5.1. Assume A1 and that (1.1) holds for some α0 P Sd and some Ψ0 P Mα0 that
is not constant on Cα0 , and either left- or right-continuous on Cα0 with no discontinuity point at
the boundaries of Cα0 . Then, pΨ0 , α0 q is uniquely defined.
5.2. Convergence of the separated estimators
p n q where as before 4, Ψ
p n denotes the left-continuous
We begin by establishing consistency of pp
αn , Ψ
LSE of Ψ0 extended to R, see Section 8.4 in the Supplement for a proof.
Theorem 5.2. Assume A1-A4. Assume moreover that Ψ0 is non-constant and continuous on
Cα0 . We then have
}p
αn ´ α0 } “ op p1q
(5.22)
p n ptq ´ Ψ0 ptq| “ op p1q
sup |Ψ
(5.23)
with } ¨ } the Euclidean norm and
tPI
for all compact intervals I Ă R such that K´ ă Ψ0 ptq ă K` for all t P I. Here, K` and K´
denote the largest and smallest values of Ψ0 on Cα0 .
p n . To show that both α
p n inherit
Next, we establish rates of convergence for α
pn and Ψ
pn and Ψ
1{3
the n rate of convergence from the joint convergence established for the full estimator gpn p . q “
p n pp
Ψ
αTn ¨q, some additional assumptions are needed.
A5. There exists an interior point z0 P Cα0 such that Ψ0 is continuously differentiable in the
neighborhood of z0 , with Ψ10 pz0 q ą 0.
A6. The density of X, q, is continuous on X .
Let c “ inf Cα0 and c “ sup Cα0 . Our main result here is the following. It is proved in Section
8.4 in the Supplement.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
14
αn ´ α0 } “ Op pn´1{3 q. If
Corollary 5.3. Assume A1-A6 and Ψ0 is continuous on Cα0 . Then, }p
moreover, Ψ0 has a derivative bounded from above on Cα0 , then
¸1{2
˜ż
c´vn
2
p
“ Op pn´1{3 q
(5.24)
pΨn ptq ´ Ψ0 ptqq dt
c`vn
for all sequences vn such that n1{3 vn Ñ 8 and c ` vn ď c ´ vn .
6. Proof of Theorem 4.1
We will extend any monotone function defined on a given interval either on the whole real line or
a bounded interval that contains the support of all the monotone functions involved. This can be
done by taking the extension to be constant to the left and right of the end points of the original
support.
6.1. Proof of Lemma 4.4
pα
We use the same notation as in Section 2. Because for all fixed α, Ψ
n satisfies (2.5), where
α
α
p
p
d1 ď . . . ď dmα are the left derivatives of the greatest convex minorant of the cumulative sum
diagram defined by (2.4), we have
ř mα
řk
α α
α α
i“mα ´k ni Yi
α
T
i“1 ni Yi
p
ř
min
ď Ψn pα Xj q ď max
řk
m
α
α
α
1ďkďmα
0ďkďmα ´1
i“mα ´k ni
i“1 ni
T
pα
for all j “ 1, . . . , n. Therefore, min1ďiďn Yi ď Ψ
n pα Xj q ď max1ďiďn Yi for all α P Sd´1 and
T
pα
j “ 1, . . . , n. The inequalities (4.12) follow since by definition, gpn pxq takes the form Ψ
n pα xq for
all x, where α is replaced by the LSE α
pn P Sd´1 .
Now we prove that under the assumptions A2 and A3,
max |Yi | “ Op plog nq.
(6.25)
1ďiďn
For an integer s ě 2, it follows from the convexity of the function z ÞÑ |z|s on R that
´
¯
Er|Y ´ g0 pxq|s |X “ xs ď 2s´1 Er|Y |s |X “ xs ` |g0 pxq|s
´
¯
ď 2s´1 s!a0 M s´2 ` K0s
ď s!b0 pM 1 qs´2
with b0 “ 2pa0 `K02 q and M 1 “ 2pM _K0 q. Now, using Lemma 2.2.11 of van der Vaart and Wellner
(1996) with n “ 1 and Ỹ “ Y ´ g0 pxq and after integrating out with respect to dQ, we obtain
´
¯
´
¯
´t2
P |Ỹ | ą t ď 2 exp
2p2b0 ` M 1 tq
for all t ą 0. Hence, with t “ C log n such that K0 ă C log n{2 we have that
P
´
max |Yi | ą C log n
1ďiďn
¯
ď
ď
n
ÿ
i“1
´
¯
P |Ỹ1 | ą C log n{2
2n exp
ˆ
´C 2 log n
4 p4b0 plog nq´1 ` M 1 Cq
˙
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
15
which converges to 0 as n Ñ 8 provided that C is sufficiently large so that 8M 1 C ă C 2 . Lemma
4.4 follows by combining (4.12) and (6.25).
6.2. Proof of Lemma 4.5
We have
ż "
Mpg ´ g0 q “
X
g0 pxq2
gpxq2
`
g0 pxqpgpxq ´ g0 pxqq ´
2
2
*
dQpxq,
using the fact that ErY |X “ xs “ g0 pxq. Hence
ż
ż
q
1
Mpg ´ g0 q “ ´
pgpxq ´ g0 pxqq2 dQpxq ď ´
pgpxq ´ g0 pxqq2 dx
2 X
2 X
using assumption A4. This completes the proof of Lemma 4.5.
6.3. Proof of Lemma 4.9
Let εα “ ε2 K ´2 P p0, 1q. By Lemma 4.8, Sd´1 can be covered by N neighborhoods with diameter
d
at most εα where N ď pAε´1
α q with A ą 0 a universal constant. Let tα1 , . . . , αN u denote
elements of each of these neighborhoods. Now, consider an arbitrary g P Gr´R,RsK . Then, gpxq “
ΨpαT xq, x P X , for some Ψ P Mr´R,RsK and α P Sd´1 . We can find i P t1, . . . , N u such that
}α ´ αi } ď εα . Then, extending Ψ to r´2R, 2Rs and using the monotonicity of Ψ together with
the Cauchy-Schwarz inequality we can write for all x P X that
gpxq
“ ΨpαTi x ` pα ´ αi qT xq ď ΨpαTi x ` εα Rq
and gpxq ě ΨpαTi x ´ εα Rq. Note that both αTi x?
´ εα R and αTi x ` εα R belong to r´2R, 2Rs.
1
Then, Lemma 4.7 implies that with N “ expp2A RKε´1 q, at the cost of increasing A, we can
U
find brackets rΨL
j , Ψj s covering the class of functions Mr´2R,2RsK such that
ż 2R
L
2
2
pΨU
j ptq ´ Ψj ptqq dt ď ε .
´2R
1
for j “ 1, . . . , N . Now returning to g, and using that Ψ P Mr´2R,2RsK , we can see that
T
U
T
ΨL
j pαi x ´ εα Rq ď gpxq ď Ψj pαi x ` εα Rq
for some j “ 1, . . . , N 1 and all x P X . We will show that there exists B ą 0 depending only on
T
U
T
d, R such that the new bracket rΨL
j pαi x ´ εα Rq, Ψj pαi x ` εα Rqs, x P X satisfies
¸1{2
˜ż
¯2
´
L
T
T
ď Bε.
(6.26)
ΨU
j pαi x ` εα Rq ´ Ψj pαi x ´ εα Rq dx
X
It follows from the Minkowski inequality that the left-hand side in (6.26) is at most
˜ż
¸1{2
´
¯2
T
ΨpαTi x ´ εα Rq ´ ΨL
j pαi x ´ εα Rq dx
X
`
˜ż
X
`
´
¸1{2
¯2
T
T
ΨU
j pαi x ` εα Rq ´ Ψpαi x ` εα Rq dx
˜ż
X
´
ΨpαTi x
` εα Rq ´
ΨpαTi x
¸1{2
¯2
.
´ εα Rq dx
(6.27)
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
16
Using the?fact that for any α P Sd´1 there exists at least an index j P t1, . . . , du such that
|αj | ě 1{ d, we can assume without loss of generality that the first component of αi satisfies
this property. Hence, using the change of variable t1 “ αTi x ´ εα R and tj “ xj , j “ 2, . . . , d and
that fact that t1 P r´2R, 2Rs, tj P r´R, Rs, j “ 2, . . . , d we get
ż ´
¯2
T
ΨpαTi x ´ εα Rq ´ ΨL
j pαi x ´ εα Rq dx
X
ď
ď
ż 2R ´
¯2
?
d´1
dp2Rq
Ψptq ´ ΨL
j ptq dt
´2R
?
d´1 2
dp2Rq ε
and a similar bound is found for the square of the second integral in (6.27). Hence, the left-hand
side in (6.26) is less than or equal to
˜ż
2pd`1q{2 d1{4 Rpd´1q{2 ε `
X
´
¸1{2
¯2
.
ΨpαTi x ` εα Rq ´ ΨpαTi x ´ εα Rq dx
Now using the change of variable t1 “ αTi x ` εα R and tj “ xj , j “ 2, . . . , d and assuming
again
?
without loss of generality that the first component of αi is greater or equal than 1{ d we can
write
ż ´
ż 2R ´
¯2
¯2
ΨpαTi x ` εα Rq ´ ΨpαTi x ´ εα Rq dx ď d1{2 p2Rqd´1
Ψptq ´ Ψpt ´ 2εα Rq dt.
´2R
X
By monotonicity of Ψ (which can be extended to the whole real line) and the fact that it is
bounded in absolute value by K, this is at most
1{2
2Kd
d´1
p2Rq
ż 2R ´
´2R
1{2
“ 2Kd
¯
Ψptq ´ Ψpt ´ 2εα Rq dt
˜ż
2R
d´1
p2Rq
2R´2εα R
Ψptqdt ´
ż ´2R
Ψptqdt
´2R´2εα R
¸
ď 2d`2 d1{2 Rd ε2 ,
using that εα “ ε2 {K 2 . Hence,
˜ż
X
¸1{2
´
¯2
T
T
ď 2pd`2q{2 d1{4 Rd{2 ε.
Ψpαi x ` εα Rq ´ Ψpαi x ´ εα Rq dx
(6.28)
Combining
the inequalities above, we get the claimed inequality in (6.26) with B “ 2pd`1q{2 d1{4 Rpd´1q{2 p1`
?
2Rq. It follows that
HB pBε, Gr´R,RsK , } . }2,X q ď
ď
ď
log N ` log N 1
?
ˆ
˙
AK 2
2A RK
d log
`
ε2
ε
´
? ¯
´1
1{2
2ε
dA ` A R K
?
since log x ď 2 x for all x ą 0. The first assertion of Lemma 4.9 follows.
To prove the second assertion, we need to build brackets for the class of functions px, yq ÞÑ
ygpxq, x P X , y P R, and then for the class of functions g 2 , with g P Gr´R,RsK . In the following,
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
17
we denote the former class by G1 and the latter by G2 with which we begin. Note that g 2 pxq “
spxq “ Ψ2 pαT xq “ hpαT xq for some function h that is either monotone non-decreasing, monotone
non-increasing or U -shaped depending on the sign of Ψ. Hence, the function h can be always
decomposed into the difference of two monotone functions on r´R, Rs that are bounded by K 2 .
If K 2 ą ε (which holds for all ε ą 0 and K ą ε such that K ą 1), we can use similar arguments
as above to conclude that there exist a constant B0 ą 0 depending only on d and R such that
´
¯ B K2
0
HB ε, G2 , } . }2,X ď
.
ε
Using the fact that any element s P G2 satisfies
ż
ż
ż ż
s2 pxqdx,
s2 pxqdQpxq ď q̄
s2 pxqdPpx, yq “
R
X
X
X
where q̄ is the upper bound of the density of Q in (4.10), it follows that
HB
´?
¯ B K2
0
.
q̄ε, G2 , } . }P ď
ε
(6.29)
Now we turn to G1 . With N “ NB pε, Gr´R,RsK , } . }2,X q, we will denote by tpgiL , giU q, i P t1, . . . , N uu
a cover of ε-brackets for Gr´R,RsK . For all i “ 1, . . . , N , define
#
#
U
ygiL pxq if y ě 0
yg
pxq
if
y
ě
0
i
L
.
(6.30)
,
k
px,
yq
“
kiU px, yq “
i
ygiU pxq if y ď 0
ygiL pxq if y ď 0
Now, take g P Gr´R,RsK and let i P t1, . . . , N u such that giL ď g ď giU . Then, we have kiL px, yq ď
ygpxq ď kiU px, yq so that tpkiL , kiU q, i P t1, . . . , N uu form a bracketing cover for G1 . We will now
compute its size. We have that
ż
ż
´
¯2
´
y 2 ˆ giU pxq ´ giL pxqq2 dPpx, yq
kiU px, yq ´ kiL px, yq dPpx, yq “
RˆX
RˆX
ż ´
¯2
ď 2a0
giU pxq ´ giL pxq dQpxq
żX ´
¯2
ď 2a0 q̄
giU pxq ´ giL pxq dx ď 2a0 q̄ ε2
X
where a0 is taken from (4.9). Hence,
¯
´a
ď
HB
2a0 q̄ε, G1 , } . }P
ď
´
¯
HB ε, Gr´R,RsK , } . }2,X
A1 K
,
ε
(6.31)
using the first assertion of the lemma. But for all ε ą 0, we have
HB pε, Fr´R,RsK , } . }P q ď HB pε{2, G1 , } . }P q ` HB pε, G2 , } . }P q
(6.32)
and hence we obtain the second assertion of the lemma. This completes the proof of Lemma
4.9.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
18
6.4. Proof of Lemma 4.10
Let g P Gr´R,RsKv such that gpxq “ ΨpαT xq. Here, the support of Ψ is the interval tαT x, x P X u,
the boundaries of which we denote by a and b. We shall show below that |Ψptq| ď 2K0 except
on a region of small size. To do that, we shall use the condition (4.16) together with the triangle
inequality to get
ż
ż
1t|ΨpαT xq´Ψ0 pαT0 xq|ąK0 u dx
1t|ΨpαT xq|ą2K0 u dx ď
X
X
ď
ż ˆ
X
ΨpαT xq ´ g0 pxq
K0
˙2
dx ď v 2 K0´2 .
All components of αj are smaller than or equal to one in absolute value, so with the change of
variable t1 “ αT x, t2 “ x2 , . . . , td “ xd , we have that
ż
żb
1t|ΨpαT xq|ą2K0 u dx ě D1
1t|Ψptq|ą2K0 u dt
X
a
ş
where D1 “ Xd´1 dx2 . . . dxd ą 0 with Xd´1 “ tpx2 , . . . , xd q, x “ px1 , . . . , xd q P X u. Combining
the two preceding displays, we conclude that
żb
1t|Ψptq|ą2K0 u dt ď D2 v 2 .
a
where D2 “ D1´1 K0´2 . By monotonicity of Ψ, this means that |Ψptq| ď 2K0 for all t in the interval
ra ` D2 v 2 , b ´ D2 v 2 s. Now, note that gpxq ´ ḡpxq takes the form of hpαT xq where h is a monotone
function on ra, bs that is bounded by K and such that
$
’
if |Ψptq| ď 2K0
&0
hptq “ Ψptq ` 2K0 if Ψptq ă ´2K0
’
%
Ψptq ´ 2K0 if Ψptq ą 2K0 .
Hence, we can have hptq ‰ 0 only for t P ra, a ` D2 v 2 s Y rb ´ D2 v 2 , bs. Consider tα1 , . . . , αN u the
d
grid providing a εα -cover for Sd´1 with εα “ ε2 {K 2 and N ď Apε´1
α q , see Lemma 4.8. Similar
to the proof of Lemma 4.9, with αi such that }α ´ αi } ď εα , we then have
hpαTi x ´ εα Rq ď gpxq ´ ḡpxq ď hpαTi x ` εα Rq
(6.33)
for all x P X , where h has been extended to the larger interval rai , bi s where
ai “ inftαTi x ´ εα R, x P X u
and bi “ suptαTi x ` εα R, x P X u.
Note that |a ´ ai | ď 2εα R and |b ´ bi | ď 2εα R. From what precedes, with the extended version
of h we can have hptq ‰ 0 only for t P Ii,1 Y Ii,2 where
”
ı
ı
Ii,1 “ ai , ai ` D2 v 2 ` 2εα R and Ii,2 “ rbi ´ D2 v 2 ´ 2εα R, bi .
Hence, we only need to construct brackets for monotone functions that are bounded by K
and defined on Ii,1 and Ii,2 . This can be done by using Lemma 4.7 for MIi,1 K and MIi,2 K .
These ?
classes admit Ni ε-brackets for the L2 norm on the corresponding intervals where Ni ď
exppA 2D2 Kv{εq (at the cost of increasing the value of A), provided that ε ą 0 satisfies 2εα R ă
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
19
1{2
U
D2 v 2 ; i.e., ε ă pD2 {p2Rqq Kv. Based on those brakets, we can build brackets rhL
j , hj s, j “
2
L
U
1, . . . , Ni such that hj ď h ď hj for some j on rai , bi s and
ż bi
ai
|hL
j ptq
´
2
hU
j ptq| dt
“
ď
ż ai `2D2 v2
ai
2
|hL
j ptq
´
2
hU
j ptq| dt
`
2ε .
ż bi
bi ´2D2 v 2
U
2
|hL
j ptq ´ hj ptq| dt
(6.34)
Note that we have omitted writing the dependence on i for the functions in the brackets. By
(6.33) we have
T
U
T
hL
j pαi x ´ εα Rq ď gpxq ´ ḡpxq ď hj pαi x ` εα Rq
for all x P X and it remains to compute the size of the obtained brackets. By the Minkowski
inequality we have
ˆż ´
X
¯2 ˙1{2
T
L
T
dx
hU
pα
x
`
ε
Rq
´
h
pα
x
´
ε
Rq
α
α
j
i
j
i
ˆż ´
¯2 ˙1{2
U
T
U
T
hj pαi x ` εα Rq ´ hj pαi x ´ εα Rq dx
ď
X
ˆż ´
¯2 ˙1{2
T
L
T
U
`
hj pαi x ´ εα Rq ´ hj pαi x ´ εα Rq dx
X
ˆ
ż bi ´2εα R ´
¯2 ˙1{2
U
dx
ď d1{2 D1
pt
`
2ε
Rq
´
h
ptq
hU
α
j
j
ai
ˆ
ż bi ´
¯2 ˙1{2
L
` d1{2 D1
dx
hU
ptq
´
h
ptq
j
j
ai
by using the change of variable t1 “ αTi x ´ εα R, t2 “ x2 , . . . , td “ xd and using that at least one
U
component in αi is larger than d´1{2 . Since hL
j and hj can be choosen monotone and bounded
in absolute value by K we conclude from (6.34) that
ˆż ´
X
ď
ď
T
hU
j pαi x
ˆ
ˆ
` εα Rq ´
2Kd1{2 D1
T
hL
j pαi x
¯2 ˙1{2
´ εα Rq dx
˙1{2
ż bi ´2εα R ´
¯ ˙1{2 ˆ
1{2
U
`
2d
D
ε
hU
pt
`
2ε
Rq
´
h
ptq
dx
1
α
j
j
ai
8εα RK 2 d1{2 D1
˙1{2
˙1{2
ˆ
` 2d1{2 D1
ε.
Since by definition, εα “ ε2 {K 2 , this means that we have Cε-brackets where C depends on D1 ,
d and R only. Hence,
´
¯
HB Cε, Ḡr´R,RsKv , } ¨ }2,X
ď 2 logpNi q ` logpN q
1{2
ď
23{2 AD2 Kv
AK 2
` d log 2
ε
ε
and the first assertion of the lemma follows.
To prove the second assertion, recall that pf ´ f¯qpx, yq “ ypgpxq ´ ḡpxqq ´ 21 pg 2 pxq ´ ḡ 2 pxqq.
We need then to build brackets for the functions px, yq ÞÑ ypgpxq ´ ḡpxqq, and those for the
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
20
functions x ÞÑ g 2 pxq ´ ḡ 2 pxq. The construction of the brackets goes along the same line as the
construction of the brackets for the classes G1 and G2 in the proof of Lemma 4.9 above, where
for the latter class, we use the fact that all functions in the class take the form hpαT xq where
h is either monotone or U-shaped, and vanishes when |gpαT xq| ď 2K0 , which implies that the
function vanishes except on at most two intervals of maximal length D2 v 2 .
Hence, we can find A1 ą 0 and A2 ą 0 depending on d, K0 , X , Q and a0 such that
´
¯ A K 2v
A1 K 4
1
HB ε, F̄r´R,RsKv , } . }P ď
` 2d log
ε
ε2
for K ą 1 _ ε such that K 2 v ą A2 ε. This completes the proof of Lemma 4.10.
6.5. Proof of Lemma 4.11
We start by noting that entropy bound on the class Frr´R,RsKv is smaller than the entropy bound
for the class Frr´R,RsK as a consequence of inclusion of the former class in the latter. We will
now show that the upper bound with respect to the Bernstein norm for the class Frr´R,RsKv is of
the claimed order. Let N1 “ NB pε{2, G1 , } . }P q and N2 “ NB pε, G2 , } . }P q, where G1 is the class of
functions tpx, yq ÞÑ ygpxq, g P Gr´R,RsKv u and G2 that of tx ÞÑ g 2 pxq, g P Gr´R,RsKv u. Consider
U
the same brackets rkjL , kjU s, j “ 1, . . . , N1 covering G1 and rsL
i , si s, i “ 1, . . . , N2 covering G2
L
U
as in the proof of Lemma 4.9: kj and kj satisfy (6.30) with appropriate gjL and gjU . Note that
giL and giU can be always taken to be bounded by K, because otherwise we can take instead
U
giL _ p´Kq and giU ^ K. The same thing holds for sL
i and si which can be taken to be bounded
2
by K . Let f P Fr´R,RsKv and C̃ ą 0 a fixed constant to be chosen later. Then, there exists
L
U
pi, jq P t1, . . . , N1 u ˆ t1, . . . , N2 u such that fi,j
ď f ď fi,j
and
#
ygjL pxq ´ 21 sU
i pxq,
L
fi,j
px, yq “
ygjU pxq ´ 21 sU
i pxq,
if y ě 0
if y ă 0
U
; fi,j
px, yq “
#
ygjU pxq ´ 21 sL
i pxq,
ygjL pxq ´ 12 sL
i pxq,
if y ě 0
if y ă 0.
k
Now note that
for all k ě 2, we can write
ş a kgiven function h such that h is P-integrable
ř8 for
2
}h}B,P “ 2 k“2 p |h| dPq{k!. Hence, by convexity of x ÞÑ xk for k ě 2, we have
›
›
"
ż
8
ˇk *
ˇk
ˇ
› f U ´ f L ›2
ÿ
2k´1
1 ˇˇ
ˇ
ˇ
ˇ
› i,j
i,j ›
L
ď2
pxq
´
s
pxq
|y|k ˇgjU pxq ´ gjL pxqˇ ` k ˇsU
›
ˇ dPpx, yq.
›
i
i
k
›
›
2
C̃
k!
C̃
X
ˆR
k“2
B,P
Integrating first with respect to y using the assumption A3 yields that the right-hand side is at
most
*
"
ż
ż
8
ÿ
2k´1
p2K 2 qpk´2q
U
L 2
2
|s
´
s
|
dQ
a0 M k´2 k! ˆ p2Kqk´2
|gjU ´ gjL |2 dQ `
j
j
k
2k
X
X
k“2 k!C̃
*
"
8
ÿ
p2K 2 qpk´2q 2
2k´1
q̄ε
,
ď 2A
a0 M k´2 k! ˆ p2Kqk´2 q̄ε2 `
k
2k
k“2 k!C̃
where A ą 0 depends only on d, R, a0 and q̄. Hence,
¸
˜
˙k´2
˙k´2
8 ˆ
8 ˆ
›
›2
ÿ
4M K
1
1 ÿ 2K 2
2Aq̄
› U
L
´1 ›
ε2 ,
2a0
`
ď
›pfi,j ´ fi,j qC̃ ›
2
4
pk
´
2q!
B,P
C̃
C̃
C̃
k“2
k“2
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
21
using the fact that k! ě 2pk ´ 2q! for all k ě 2. We conclude that
›
›2
› U
›
L
qC̃ ´1 ›
›pfi,j ´ fi,j
B,P
ď
´
¯
2
1
2Aq̄
p2a0 _ 1{4q
` e2K {C̃ ε2 .
C̃ 2
1 ´ 4M K{C̃
since K ě 2, the choice C̃ “ 4M K 2 yields
U
L
}pfi,j
´ fi,j
qC̃ ´1 }2B,P
ď
ď
¯´ K
¯
´1
2Aq̄ ´
2a0 _ 1{4
` ep2Mq
ε2
K ´1
C̃ 2
B 2 K ´4 ε2
(6.35)
where B depends on d, R, q̄, M, a0 only. Using (6.29) and (6.31), this means that there exists A2
depending on q̄, a0 , d, R only such that
HB
´ Bε
K2
, Frr´R,RsKv , } . }B,P
¯
ď
log N1 ` log N2
ď
A2 K 2
.
ε
(6.36)
This in turn implies that
´
¯
HB ε, Frr´R,RsKv , } . }B,P
ď
log N1 ` log N2
ď
A1
A2 BK 2
“
εK 2
ε
as claimed in the statement of the lemma.
To show the second claim, we will use again the series expansion of the Bernstein norm.
Similar as above, using that g0 is bounded by K0 ď K, and that for arbitrary f˜ “ pf ´ f0 qC̃ ´1 P
Frr´R,RsKv , the corresponding g P Gr´R,RsKv satisfies (4.16), we obtain
}f˜}2B,P
"
*
ż
ż
p2K 2 qpk´2q
2k
2
2 2
a0 M k´2 k!p2Kqk´2 pg ´ g0 q2 dQ `
pg
´
g
q
dQ
0
k
2k
X
X
k“2 k!C̃
*
"
ż
ż
8
ÿ
2k
2p2K 2 qk´1
2
k´2
k´2
2
ď
pg ´ g0 q dQ
a0 M
k!p2Kq
pg ´ g0 q dQ `
k
2k
X
X
k“2 k!C̃
ˆ
˙k´2
˙k´2 ˙
8 ˆ
8 ˆ
4a0 ÿ 4M K
1
p4Kq2 ÿ 4K 2
q̄v 2
`
ď
2
2
k!
C̃ k“2
C̃
C̃
C̃
k“2
ˆ
˙
a0
1
ď
(6.37)
`
e1{M q̄v 2 ,
2M 2 K 4
4M 2 K 2
ď
8
ÿ
using that K ě 2. The second claim follows.
Using the same arguments as above in combination with the entropy bound for F̄r´R,RsKv
obtained in Lemma 4.10 and which remains the same if the class Fr´R,RsKv is shifted by a fixed
function (here f0 ) we can show that
ˆ ˙
´
¯ A v
A1
1
Ă̄
HB ε, F
}
.
}
ď
`
2d
log
B,P
r´R,Rsv
ε
ε2
at the cost of increasing the constant A1 . To show the second assertion for the elements of
Ă̄
r
F
r´R,RsKv , we can use again the same arguments as for the class Fr´R,RsKv . Indeed, the condition
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
22
K ě 2K0 _ 2 implies that maxp|g|, |ḡ|q ď K since |ḡ| ď 2K0 . Moreover, with C̃ “ 4M K 2 we get
Ă̄
for any element f˜ P F
r´R,RsKv that
˙ ż ´
ˆ
¯2
1 1{p2Mq
a0
`
e
dx
q̄
gpxq
´
ḡpxq
}f˜}2B,P ď
32M 2 4M 2
X
˙
ˆ
1 1{p2Mq
a0
`
e
q̄v 2
(6.38)
ď 4
32M 2 4M 2
where in the last line we used the fact that by convexity of x ÞÑ x2 ,
ż
ż
ż ´
¯2
gpxq ´ ḡpxq dx ď 2 pgpxq ´ g0 pxqq2 dx ` 2 pḡpxq ´ g0 pxqq2 dx,
X
X
żX
2
ď 4 pgpxq ´ g0 pxqq dx,
(6.39)
X
as pḡ ´ g0 q2 ď pg ´ g0 q2 by definition of ḡ. This completes the proof of Lemma 4.11.
l
6.6. Proof of Theorem 4.6
In the sequel, we assume that A1-A4 hold. Below, we give a uniform bound for the centered
process Mn ´ M, with Mn and M defined by (4.11) and (4.14) respectively. In the sequel, the
notation À means “is bounded up to a constant”, where the constant depends on d or K0 and
other fixed values used in the assumptions A1-A4, such as a0 , M, or R. Moreover, the capital E
denotes outer expectation in cases when we consider expectation of a random variable which we
have not proved to be measurable.
Proposition 6.1. Let K ą 1. Then, for all v P p0, 2Ks we have
„
ˇ
ˇ
?
ˇ
ˇ
nE
sup
ˇpMn ´ Mqpg ´ g0 qˇ À φn pvq
(6.40)
gPGr´R,RsKv
where φn pvq “ v 1{2 K 5{2 p1 ` K 1{2 v ´3{2 n´1{2 q.
Proof. Note that with K̃ “ 2K ` p2a0 q1{2 ` 2K0 ě K we have
„
„
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
E
sup
sup
ˇpMn ´ Mqpg ´ g0 qˇ ď E
ˇpMn ´ Mqpg ´ g0 qˇ ,
gPGr´R,RsKv
gPGr´R,RsK̃v
so possibly replacing K by K̃, we assume without loss of generality that K ě p2a0 q1{2 , K ě 2
and K ě 2K0 . Define for η ą 0 and fixed v P p0, 2Ks
żηc
´
¯
1 ` HB ε, Frr´R,RsKv , } . }B,P dε
Jpηq “
0
where we recall that } . }B,P is the Bernstein norm, and Frr´R,RsKv is defined in (4.20) with
C̃ “ 4M K 2. By Lemma 4.11, there exists a constant A2 ą 0 depending only on d,R, a0 ,
M and q̄ such that }f˜}B,P ď A2 v for all f˜ P Frr´R,RsKv . It follows from Lemma 3.4.3 of
van der Vaart and Wellner (1996) (using the notation of that book) that
˜
`
˘¸
`
˘
J A2 v
Er}Gn }Frr´R,RsKv s À J A2 v 1 ` 2 2 ?
A2 v n
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
where by Lemma 4.11 and the inequality
żηc
Jpηq ď
0
23
?
?
?
u ` v ď u ` v for u, v ě 0 we have that
1`
A1
1{2
dε ď η ` 2A1 η 1{2
ε
for all η ą 0. Note that v P p0, 2Ks implies that v ď v 1{2 21{2 K 1{2 , and hence
`
˘ `
1{2 1{2 ˘
J A2 v ď A2 21{2 K 1{2 ` 2A1 A2 v 1{2 ď A3 K 1{2 v 1{2
1{2 1{2
using that K ě 1, where A3 “ A2 21{2 ` 2A1 A2 . Hence, by definition of Frr´R,RsKv ,
„
ˇ
ˇ
?
ˇ
ˇ
nE
sup
“ Er}Gn }Fr´R,RsKv s
ˇpMn ´ Mqpg ´ g0 qˇ
gPGr´R,RsKv
which completes the proof of Proposition 6.1.
“ 4M K 2 Er}Gn }Frr´R,RsKv s À φn pvq
Now we are ready to give the proof of Theorem 4.6.
Proof of Theorem 4.6. In the sequel, we consider K “ C log n for some C ą 0 that does not
depend on n, and v P p0, 2Ks. It follows from Proposition 6.1 that for all n sufficiently large, we
have (6.40) where
¯
? ´
φn pvq “ plog nq5{2 v 1 ` plog nq1{2 v ´3{2 n´1{2 .
Since with D taken from (4.15), we have Dpg, g0 q ď }g}8 ` }g0 }8 ď 2K for sufficiently large
n and all g P Gr´R,RsK , the above inequality holds for all v ą 0. Furthermore, gpn maximizes
Mn g over the set of all functions g of the form gpxq “ ΨpαT xq, x P X with α P Sd´1 and
Ψ a non-decreasing function on r´R, Rs, and it follows from Lemma 4.4 that with arbitrarily
large probablity by choice of C, p
gn maximizes Mn g over the restricted set Gr´R,RsK . Hence,
we can use Lemma 4.5 and Theorem 3.2.5 in van der Vaart and Wellner (1996), with α “ 1{2
and rn „ n1{3 plog nq´5{3 , to conclude that Dpp
gn , g0 q “ Op˚ pn´1{3 plog nq5{3 q, which completes the
proof of Theorem 4.6.
6.7. Proof of Theorem 4.1
Still assuming that A1-A4 hold, we give a second uniform bound for Mn ´M. The bound is sharper
than the one obtained in Proposition 6.1 for the case where all functions g in the considered class
of functions satisfy (4.16) for some v ď plog nq2 n´1{3 . As before, the notation À means “is
bounded up to a constant”.
Proposition 6.2. Let K “ C log n for some fixed C ą 0, and v P p0, plog nq2 n´1{3 s. Then for n
large enough we have that
„
ˇ
ˇ φ pvq
ˇ
ˇ
n
E
sup
pM
´
Mqpg
´
g
q
ˇ n
0 ˇ À ?
n
gPGr´R,RsKv
`
˘
where φn pvq “ Av 1{2 1 ` v ´3{2 n´1{2 .
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
24
Proof. Assume n large enough so that K ě p2K0 q _ 2 and use (4.17) to write
„
ˇ
ˇ
ˇ
ˇ
E
sup
´
Mqpg
´
g
q
ˇpMn
0 ˇ
gPGr´R,RsKv
ďE
Hence,
?
nE
„
„
sup
gPGr´R,RsKv
sup
gPGr´R,RsKv
„
ˇ
ˇ
ˇ
ˇ
ˇpMn ´ Mqpg ´ ḡqˇ ` E
sup
gPGr´R,RsKv
ˇ
ˇ
ˇ
ˇ
ˇpMn ´ Mqpḡ ´ g0 qˇ .
ˇ
ˇ
ˇ
ˇ
ˇpMn ´ Mqpg ´ g0 qˇ ď Er}Gn }F1 s ` Er}Gn }F2 s
(6.41)
where F1 “ F̄r´R,RsKv and F2 is the class of functions f¯ ´ f0 such that f¯ P Fr´R,Rsp2K0 qv . To
give a bound for the first term on the right-hand side, consider v 1 “ plog nq2 n´1{6 . It follows from
Lemma 4.11 that for all ε P p0, A2 vs, we have
ˆ ˙
´
¯
v1
A1
A1 v 1
Ă̄
1
ď A1 p1 ` 2dq
`
2d
log
(6.42)
HB ε, F
,
}
¨
}
ď
B,P
r´R,RsKv
2
ε
ε
ε
provided that n is sufficiently large, where we used the fact that logpxq ď 4x1{4 for all x ą 0 for
Ă̄
Ă̄
the second inequality. Since the class Fr1 :“ F
r´R,RsKv is included in Fr´R,RsKv 1 , its ε-bracketing
entropy can
also bounded above by (6.42) for all ε P p0, A2 vs. Using again the inequality
? be ?
?
x ` y ď x ` y for all x, y ě 0 we can write
ż A2 v c
¯
´
`
˘
1 ` HB ε, Fr1 , } ¨ }B,P dε
J1 A2 v
:“
0
ď
´
¯1{2
A2 v ` 2 A1 p1 ` 2dqv 1
pA2 vq1{2 ď A3 pv 1 vq1{2
using that v ă v 1 and K ą 1, where A3 ą 0 is a constant depending on the same parameters
as A1 and A2 . Lemma 4.11 implies that }f˜}B,P ď A2 v for all f˜ P Fr1 . Invoking Lemma 3.4.3 of
van der Vaart and Wellner (1996) allows us to write that
˜
`
˘¸
`
˘
J1 A2 v
Er}Gn }Fr1 s À J1 A2 v 1 ` 2 2 ?
A2 v n
ˆ
˙
pv 1 q1{2
1 1{2
ď A3 pv vq
1 ` 3{2 ?
v
n
at the cost of increasing A3 . Now, using the definition of Fr1 , we have that
ˆ
˙
1
2
1{2
Er}Gn }F1 s “ 4M K Er}Gn }Fr1 s À v
1 ` 3{2 ?
.
v
n
This gives a bound for the first term on the right hand side of (6.41).
To deal with the second term, we apply Lemma 4.11 to the class Fr2 “ Frr´R,Rsp2K0 qv with
K “ 2K0 . Here, C̃ “ 4M K02 is independent of n, and J2 pA2 vq ď A3 v 1{2 for some A3 ą 0 that
does not depend on n, where J2 is defined in the same manner as J1 with Fr1 replaced by Fr2 . By
Lemma 3.4.3 of van der Vaart and Wellner (1996), we conclude that
˙
ˆ
1
2
1{2
.
Er}Gn }F2 s “ 16M K0 Er}Gn }Fr2 s À v
1 ` 3{2 ?
v
n
Combining the calculations developed for both classes together with (6.41) gives the claimed
form of the entropy bound.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
25
Proof of Theorem 4.1. Theorem 4.6 implies that with a probability that can be made arbitrarily
large, the LSE gpn belongs to Gr´R,RsKv with K “ C log n and v “ plog nq2 n´1{3 for some
C ą 0 that does not depend on n. The result follows now by applying again Theorem 3.2.5 of
van der Vaart and Wellner (1996) with α “ 1{2 and rn „ n1{3 .
Acknowledgements
Part of this work was done while the third author was at the University Paris Nanterre (laboratory
Modal’x) as an invited visiting researcher. The third author thanks Paris Nanterre for the support.
References
Barlow, R. E., Bartholomew, D. J., Bremner, J. M. and Brunk, H. D. (1972). Statistical inference under order restrictions. The theory and application of isotonic regression.
John Wiley & Sons, London-New York-Sydney. Wiley Series in Probability and Mathematical
Statistics.
Chatterjee, S., Guntuboyina, A. and Sen, B. (2015). On risk bounds in isotonic and other
shape restricted regression problems. Ann. Statist. 43 1774–1800.
Chen, Y. and Samworth, R. (2016). Generalised additive and index models with shape constraints. Journal of the Royal Statistical Society: Series B 78 729–754.
Chiou, J.-M. and Müller, H.-G. (2004). Quasi-likelihood regression with multiple indices and
smooth link and variance functions. Scandinavian Journal of Statistics 31 367–386.
Cosslett, S. R. (1983). Distribution-free maximum likelihood estimator of the binary choice
model. Econometrica: Journal of the Econometric Society 765–782.
Cover, T. M. (1967). The number of linearly inducible orderings of points in d-space. SIAM
Journal on Applied Mathematics 15 434–439.
Dobson, A. J. and Barnett, A. G. (2008). An introduction to generalized linear models. 3rd
ed. Texts in Statistical Science Series, CRC Press, Boca Raton, FL.
Feige, U. and Schechtman, G. (2002). On the optimality of the random hyperplane rounding
technique for MAX CUT. Random Structures Algorithms 20 403–440. Probabilistic methods
in combinatorial optimization.
Foster, J. C., Taylor, J. M. and Nan, B. (2013). Variable selection in monotone single-index
models via the adaptive lasso. Statistics in Medicine 32 3944–3954.
Groeneboom, P. and Hendrickx, K. (2016). Current status linear regression. Available on
arXiv: http://arxiv.org/abs/1601.00202 .
Groeneboom, P. and Pyke, R. (1983). Asymptotic normality of statistics based on the convex
minorants of empirical distribution functions. Ann. Probab. 11 328–345.
Han, A. K. (1987). Non-parametric analysis of a generalized regression model: the maximum
rank correlation estimator. Journal of Econometrics 35 303–316.
Han, Q. and Wellner, J. A. (2016). Multivariate convex regression: global risk bounds and
adaptation. Available on arXiv: http://arxiv.org/abs/1601.06844 .
Härdle, W., Hall, P. and Ichimura, H. (1993). Optimal smoothing in single-index models.
The Annals of Statistics 21 157–178.
Hristache, M., Juditsky, A. and Spokoiny, V. (2001). Direct estimation of the index
coefficient in a single-index model. Annals of Statistics 595–623.
Kakade, S. M., Kanade, V., Shamir, O. and Kalai, A. (2011). Efficient learning of generalized linear and single index models with isotonic regression. In Advances in Neural Information
Processing Systems.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
26
Lin, W. and Kulasekera, K. (2007). Identifiability of single-index models and additive-index
models. Biometrika 94 496–501.
Murphy, S., Van der Vaart, A. and Wellner, J. (1999). Current status regression. Mathematical Methods of Statistics 8 407–425.
van de Geer, S. (1993). Hellinger-consistency of certain nonparametric maximum likelihood
estimators. Ann. Statist. 21 14–44.
van de Geer, S. A. (2000). Empirical Processes in M-estimation, vol. 6. Cambridge University
Press.
van der Vaart, A. W. (1998). Asymptotic Statistics, vol. 3 of Cambridge Series in Statistical
and Probabilistic Mathematics. Cambridge University Press, Cambridge.
van der Vaart, A. W. and Wellner, J. A. (1996). Weak convergence and empirical processes.
Springer Series in Statistics, Springer-Verlag, New York. With applications to statistics.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
27
Supplement to ”Least squares estimation in the monotone
single index model”
7. Appendix A: All proofs for Adaptive Rates of Convergence
We use the same notation as in Section 4.3, with the following additions. We also assume throughout A1-A4.
– Let MIK denote the class of monotone increasing functions Ψ : I ÞÑ r´K, Ks, as before.
Furthermore, suppose that Ψ0 : I ÞÑ r´K, Ks denote any fixed increasing function, and let
0
MΨ
IK pδq denote the class of monotone increasing functions Ψ : I ÞÑ r´K, Ks, where I is a
connected interval and such that }Ψ ´ Ψ0 }2,I ď δ.
– The notation À means “is bounded up to a constant”, as in van der Vaart and Wellner (1996,
page 289). Here, the constant depends on d or K0 and other fixed values used in the assumptions
A1-A4, such as a0 , M, or R.
7.1. Entropy bounds
When the class of monotone functions is known to be close to a constant, we obtain the following
closer bound. This follows from van de Geer (1993, Lemma 4.7).
Lemma 7.1. Suppose that Ψ0 “ 0. Then, there exists a constant A ą 0 such that
ˆ
˙
ˆ ˙
2KL1{2
δ
Ψ0
log
HB pε, MIK pδq, } ¨ }2,I q ď A
ε
ε
for all ε ą 0, where L denotes the length of the interval I.
Corollary 7.2. There exists a universal constant A ą 0 such that
ˆ
˙
ˆ ˙
2KpκLq1{2
δ
0
HB pε, MΨ
log
pδq,
}
¨
}
q
ď
Aκ
2,I
IK
ε
ε
for all δ ě ε ą 0, where L denotes the length of the interval I and Ψ0 : I ÞÑ r´K, Ks is an
increasing piecewise constant function. Here, κ denotes the number of unique values taken by Ψ0 .
Lemma 7.3. Assume that Ψ0 in (1.1) is an increasing piecewise constant function with κ unique
values. Then, for any 0 ă ε ď δ ă K ´1 and K ě 3 we have
ˆ ˙
ˆ ? ˙
δ
K κ
HB pε, Gr´R,RsKδ , } ¨ }2,X q À κ
log
,
ε
ε
?
for δ ď 1{p4CR,d R ` 1q where Gr´R,RsKδ is defined as in Section 4.3.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
28
The above results quantify the effect of the finite-dimensional index parameter on the entropy
of the class Gr´R,RsKδ . Namely, its size is negligible when compared to the size of the monotone
class of functions. This is, of course, an expected result, as in a semi-parametric model the
non-parametric portion should “dominate” the rate of convergence.
Finally, we obtain additional results for an appropriate set of functions using the Bernstein
norm. We choose to work with the Bernstein norm } ¨ }B,P because it provides slightly tighter
results in terms of the upper bound K, allowing us to use van der Vaart and Wellner (1996,
Lemma 3.4.3) over van der Vaart and Wellner (1996, Lemma 3.4.2).
Lemma 7.4. There exists a constant c depending on a0 , M, q̄ such that, for K ą 1{6,
ˆ
˙
˙
ˆ
(
1
Kε
HB ε,
ˆ Fr´R,RsKδ ´ f0 , } ¨ }B,P
, Gr´R,RsKpKδ{cq , } ¨ }2,X .
ď HB
12M K 2
c
and
›
›
›
`
˘›2
1
›
›
f
´
f
0 ›
› 12M K 2
B,P
ď c2 K ´2
ż
X
pgpxq ´ g0 pxqq2 dx.
(7.43)
Proof of Lemma 7.4. Recall that f “ yg ´ g 2 {2. Therefore, f ´ f0 “ ypg ´ g0 q ´ pg 2 ´ g02 q{2.
Consider some ε1 , δ1 to be chosen later. Suppose that g P GIKδ1 where I “ r´K, Ks. Then, from
the definition of the Bernstein norm, we have for C “ 12M K 2 ,
›
›2
›
›
›pf ´ f0 qC ´1 ›
B,P
"
8
ˇk *
ˇ
ˇk
k ż
ÿ
1 ˇˇ 2
2
ˇ
ˇ
2
kˇ
pxq
´
g
pxq
|y|
´
g
pxq
`
ď
ˇg
ˇ dPpx, yq
ˇgpxq
ˇ
0
0
k
k
k!C
2
X ˆR
k“2
"
ż ´
8
¯2
ÿ
2k
k´2
k´2
ď
a
M
k!
ˆ
p2Kq
dQpxq
gpxq
´
g
pxq
0
0
k!C k
X
k“2
*
ż
p2K 2 qpk´2q
2
2
pgpxq ` g0 pxqq pgpxq ´ g0 pxqq dx
`
2k
X
"
*ż ´
8
¯2
ÿ
2k
p2K 2 qpk´2q ˆ p2Kq2
k´2
k´2
ď q
dx
a
M
k!
ˆ
p2Kq
`
gpxq
´
g
pxq
0
0
k!C k
2k
X
k“2
˙
"
*ż ´
ˆ
¯2
1
1
2 1{p6Mq
`
4K
e
dx
4a
gpxq
´
g
pxq
ď q
0
0
p12M K 2q2
1 ´ 1{6K
X
ż ´
¯2
ď c21 K ´2
gpxq ´ g0 pxq dx ď c21 K ´2 δ12 .
X
This also proves (7.43).
giU u, such that }giL ´ giU }2,X ď ε1 , with
Next, consider brackets for the ´class g P GIKδ1 , tgiL , ¯
i “ 1, . . . , N1 , where log N1 “ HB ε1 , g P GIKδ1 , } ¨ }2,X . We now use this class of functions to
form further brackets. We make use of the fact that 2K ě giL ` K ě 0 throughout. First, note
that if giL ď g ď giU then
pgiL ` Kq2 ´ 2KgiU ´ K 2 ď g 2
ď pgiU ` Kq2 ´ 2KgiL ´ K 2 .
Therefore, defining
fiL px, yq
#
(
ygiL pxq ´ 21 pgiU q2 ` 2KpgiU ´ giL q pxq,
(
“
ygiU pxq ´ 12 pgiU q2 ` 2KpgiU ´ giL q pxq,
if y ě 0
if y ă 0
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
and
fiU px, yq
“
#
(
ygiU pxq ´ 21 pgiL q2 ` 2KpgiL ´ giU q pxq,
(
ygiL pxq ´ 12 pgiL q2 ` 2KpgiL ´ giU q pxq,
29
if y ě 0
if y ă 0
we have that fiL ď f ď fiU and fiL ´ f0 ď f ´ f0 ď fiU ´ f0 (noting that ppfiU ´ f0 q ´ pfiL ´ f0 qq “
pfiU ´ fiL q). It remains to check the size of the brackets.
#
(
y ` 2K ` 21 pgiU ` giL q pgiU ´ giL q, if y ě 0,
U
L
(
fi ´ fi “
y ´ 2K ` 12 pgiU ` giL q pgiL ´ giU q, if y ă 0.
ř8 ´ ş k ¯
Recall that }h}2B,P “ 2 k“2
|h| dP {k!. Hence, choosing C “ 12M K 2, we find
}pfiU ´ fiL q{C}2B,P
#
ď
“
Kě1
ď
˙k´2
˙k´2 + ż
8 ˆ
8 ˆ
1
36K 2 ÿ 12K 2
4 ÿ 2M K
`
|giU ´ giL |2 dx
q a0 2
C k“2
C
C 2 k“2
C
k!
*ż
"
1
36K 2 1{M
4
`
e
|g U ´ g L |2 dx
q a0 2
C 1 ´ 1{6K 2
C2
"
*
4
1
36
1{M
q a0
e
`
K ´2 ε21 ď c22 K ´2 ε21 .
144M 2 1 ´ 1{6 144M 2
Finally, take c2 “ maxtc21 , c22 u. We therefore choose ε21 “ K 2 ε2 {c2 and δ12 “ K 2 δ 2 {c2 .
7.2. Obtaining the rate result
As in the proof of Theorems 4.1 and 4.6, we make use throughout of the basic inequality given
in Lemma 4.5. This is the inequality which allows us to make use of the empirical process results
of van der Vaart and Wellner (1996).
Theorem 7.5. Under assumptions A1-A4 and assuming that Ψ0 an increasing piecewise constant
function and κ denotes the number of unique values taken by Ψ0 , we have
˙1{2
ˆ´ ¯
˙
ˆż
n ´1{2
2
“ Op
plog nq3{2 .
pp
gn pxq ´ g0 pxqq dx
κ
X
The key ingredient in the proof is the bound on the class of functions that p
gn ´ g0 are known
to lie in. It is obtained from empirical process theory (van der Vaart and Wellner, 1996, Lemma
3.4.3, page 324), using Lemmas 7.3 and 7.4 above.
Proposition 7.6.
E
ˆ
˙
sup |pMn ´ Mqpg ´ g0 q|
À
gPGIKδ
where
φn pδq “
Kδ
d
assuming that δ ď mintc2 e´1 , 1uK.
κ log
ˆ
φn pδq
?
n
c
¨
´ ? ¯˛
K κ
? ˙
K
κ
log
δ
‹
K κ ˚
˚1 `
‹
?
˝
‚
δ
δ n
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
30
7.2.1. Proofs
Proof of Proposition 7.6. The result follows from an application of Lemma 3.4.3 in van der Vaart and Wellner
(1996, page 324), along with our obtained entropy bounds. We write Mn pgq “ Pn pf q, where
f px, yq “ ygpxq ´ g 2 pxq{2. Then, we have
ˆ
˙
1
E
sup |pMn ´ Mqpg ´ g0 q|
“ ? Er}Gn pf ´ f0 q}FIKδ s
n
gPGIKδ
By (7.43) of Lemma 7.4 we have that if ||g ´ g0 ||2,X ă δ then }pf ´ f0 q{p12M K 2 q}B,P ă cK ´1 δ.
Applying Lemma 3.4.3 of van der Vaart and Wellner (1996)
˙
ˆ
f ´ f0
2
Er}Gn pf ´f0 q}FIKδ s “ 12M K Er}Gn
}FIKδ s
12M K 2
˙
ˆ
Jpδ{cKq
2
?
À 12M K Jpδ{cKq 1 `
pδ{cKq2 n
ˆ
˙
Jpδ{cKq
À K 2 Jpδ{cKq 1 ` K 2 2 ?
.
(7.44)
δ n
where
ˆ
żδd
1 ` HB ε,
Jpδq “
0
˙
1
pFIKδ ´ f0 q, } ¨ }B,P dε.
12M K 2
Applying the first part of Lemma 7.4 and Lemma 7.3, we find that
ˆ
˙
żδd
Kε
1 ` HB
, GIKpKδ{cq , } ¨ }2,X dε
Jpδq À
c
0
ˆ ? ˙
żδd
δ
cK κ
1 ` κ log
À
dε
ε
Kε
0
d
ˆ ? ˙
? żδ 1
c κ
log
dε
À δ ` κδ
ε
ε
0
ˆ ˙
ż δ{?c2 κ d
?
1
1
2 1{4
log
À δ ` κδpc κq
dε.
ε
ε
0
Next, it follows from Lemma 7.8 that
Jpδq
À
À
δ ď ce´1 , κ ě 1
À
δ ď c´1 , κ ě 1
À
g
˜?
¸
f
2κ
f δ
?
c
δ ` κδpc2 κq1{4 e ?
log
δ
c2 κ
g
˜?
¸
f
f
c2 κ
e
δ ` δ κ log
δ
g
˜?
¸
f
f
c2 κ
e
δ κ log
δ
d
ˆ? ˙
κ
,
δ κ log
δ
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
31
where the conditions put together imply that δ ď mintce´1 , c´1 u.
This, using (7.44), in turn implies that
ˆ
˙
2
2 Jpδ{cKq
?
Er}Gn pf ´f0 q}FIKδ s À K Jpδ{cKq 1 ` K
δ2 n
c
¨
´ ? ¯˛
d
K κ
ˆ ? ˙
Kδ
κ
log
δ
‹
K κ ˚
˚1 `
‹
?
À Kδ κ log
˝
‚
2
δ
δ n
ď
Kδ
d
κ log
ˆ
c
¨
´ ? ¯˛
K κ
? ˙
K
κ
log
δ
‹
K κ ˚
˚1 `
‹
?
˝
‚
δ
δ n
as required, for δ{cK ď mintce´1 , c´1 u.
Proof of Theorem 7.5. Fix η ą 0. By Lemma 4.4, we know that supxPX |p
gn pxq| “ Op plog nq.
Therefore, we can find a C1 “ C1,η such that P psupxPX |p
gn pxq| ą C1 log nq ă η{2. This allows us
to consider the previous entropy bounds with K “ C1 log n. Then, for any constant C2
P prn }p
gn ´ g0 }2,X ą C2 q
ˆ
˙
`
˘
ď P rn }p
gn ´ g0 } ą C2 , p
gn P GIpC1 log nq ` P sup |p
gn pxq| ą C1 log n .
xPX
Recall the results of Proposition 7.6. By van der Vaart and Wellner (1996, Theorem 3.2.5, page
289),
if for some α ă 2, φn pδq{δ α is decreasing in δ, and we can find an rn such that rn2 φn p1{rn q À
?
g0
q ă η{2, proving
n, then there exists a C2 “ C2,η such that P prn }p
gn ´ g0 } ą C2 , gpn P GIpC
1 log nq
the result. We now check these conditions. First, φn pδq{δ 3{2 is decreasing in δ. We also have,
c
¨
´ ? ¯˛
d
ˆ ? ˙
K κ log K δ κ ‹
K κ ˚
˚
‹,
?
φn pδq “ Kδ κ log
˝1 `
‚
δ
δ n
with K “ C1 log n. Choosing rn “ pn{κq1{2 plog nq´3{2 , it follows that rn2 φn p1{rn q À
required. Note also that rn´1 ď mintc2 e´1 , 1upC2 log nq for large enough n.
?
n, as
7.3. Proofs of various entropy bounds
In this Section we provide proofs of the required entropy bounds. We begin, however, with two
small technical lemmas which will be used throughout.
Lemma 7.7.
1. For 0 ă ε ď η ď e´1 and for K ě 1, we have
ˆ ˙
ˆ ˙
K
1
η
log
À K log
.
ε
ε
η
2. For 0 ă ε ď η ď 1{K and with K, κ ě 1 we have
ˆ ˙
ˆ ˙
ˆ? ˙
η
η
K
1
κ
log
À
ď
ď
log
log
ε
ε
η
ε
η
η
log
ε
ˆ
? ˙
K κ
η
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
Lemma 7.8. Suppose that δ ď e´1 . Then
żδc
1
1
log dx
x
x
0
À
32
c
1
δ log .
δ
Proof of Lemma 7.7. For the first part of the lemma, consider the function hpxq “ x´1 log x.
Since h1 pxq “ p1 ´ log xq{x2 we know that h is a decreasing function on re, 8q, and since ε ď η,
we have that
ˆ ˙
ˆ ˙
1
1
η
log
1 ď log
ď
.
ε
ε
η
Therefore,
log
ˆ
˙
K
ε
ˆ ˙
1
“ log K ` log
ε
ˆ ˙
1
η
ď K ` log
ε
η
ˆ ˙
η
1
ď 2K log
.
ε
η
For the second part, we proceed similarly, but first note that log K ď logp1{εq under the condition
that ε ď η ď 1{K.
0
Proof of Lemma 7.1. Recall that MΨ
is the space of increasing functions
r0,1sp0.5q pδq
th : r0, 1s ÞÑ r´0.5, 0.5s, }h}2,r0,1s ă δu. In van de Geer (1993, Lemma 4.7), it is shown that
ˆ ˙
ˆ ˙
δ
1
Ψ0
HB pε, Mr0,1sp0.5q pδq, } ¨ }2,r0,1s q ď A
log
,
ε
ε
for some constant A.
0
Now, write I “ ra, bs and consider any Ψ P MΨ
IK pδq. Then, hpuq “ Ψpa ` pb ´ aq ˆ uq{2K
U
is also increasing and satisfies h : r0, 1s ÞÑ r´0.5, 0.5s. Let thL
i , hi u denote lower and upper
Ψ0
ε1 -brackets for the space Mr0,1sp0.5q pδ1 q, where ε1 ą 0 and δ1 ą 0 will be chosen later. Define
Ψ0
L
U
ΨL
i pxq “ 2Khi ppx ´ aq{pb ´ aqq, and similarly for Ψi . Then for each Ψ P MIK pδq, there exists
L
U
an i such that Ψi ď Ψ ď Ψi and we have
ż
ż
` U
˘2
` U
˘2
L
2
Ψi pxq ´ Ψi pxq dx “ 4K
hi ppx ´ aq{pb ´ aqq ´ hL
dx
i ppx ´ aq{pb ´ aqq
I
I
2
“ 4K |I|
Furthermore, for Ψ0 “ 0, we have
żb
2
pΨpxq ´ Ψ0 pxqq dx
a
ż1
0
` U
˘2
du ď 4K 2 |I|ε21 .
hi puq ´ hL
i puq
“ 4K 2 |I|
ż1
0
h2 puqdu ď 4K 2 |I|δ12 .
Therefore, choosing ε21 “ ε2 {4K 2 |I| and δ12 “ δ 2 {4K 2 |I| we obtain
˙
˙
ˆ
ˆ
δ
ε
Ψ0
0
,
M
,
}
¨
}
HB pε, MΨ
pδq,
}
¨
}
q
ď
H
B
2,r0,1s
2,ra,bs
r0,1sp0.5q
ra,bsK
2K|I|1{2
2K|I|1{2
ˆ ˙
ˆ
˙
δ
2K|I|1{2
ď A
log
.
ε
ε
This completes the proof.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
33
řκ
Proof of Corollary 7.2. Write Ψ0 “ i“1 ci 1Ii , where Ii are disjoint intervals with union equal to
0
I, and c1 ă c2 ă ¨ ¨ ¨ ă cκ . For any f P MΨ
IK pδq,
řκwrite f “ f1 `. . .`fκ , where fi pxq “ f pxq1Ii pxq.
¯
Now, write fi “ fi ´ ci 1Ii . Then, f ´ Ψ0 “ i“1 f¯i and f¯i is bounded by 2K. By Minkowski’s
inequality, we have that
"ż
*1{2
*1{2
κ "ż
ÿ
2
2
¯
ď
pf psq ´ Ψ0 psqq ds
.
(7.45)
fi psqds
i“1
Ii
“0
pδi q where εi ą 0 and
Next, let fjLi , fjUi for ji “ 1, . . . , Ni denote an εi -bracket of the class MIΨi0p2Kq
“0
δi ą 0 will be chosen later. Then, for any f¯i P MIΨi0p2Kq
pδi q there exists a bracket such that fjLi ď
řκ ¯
řκ
řκ
fi ď
f U , so brackets for the
fL ď
f¯i ď f U , for each i. It therefore follows that
ji
i“1
ji
i“1
i“1
ji
class of interest can be formed by combining the brackets tfjLi , fjUi uji “1,...,Ni ,i“1,...,κ . Furthermore,
řκ
śκ
the number of brackets required is at most i“1 Ni . Note that, for example, i“1 fjLi is not
0
necessarily an increasing function, however, brackets of MΨ
IK pδq are not required to be in the
Ψ0
space MIK pδq (van der Vaart and Wellner, 1996, Definition 2.1.6, page 83). Let f U , f L denote
one such bracket, where we suppress its dependence on tji uκi“1 . By Minkowski’s inequality again,
we have
$ ˜
¸2 ,1{2
"ż
*1{2
κ
κ
&ż
.
ÿ
ÿ
“
pf U psq ´ f L psqq2 ds
ď
εi .
pfjUi ´ fjLi qpsq ds
% I i“1
i“1
Let Li denote the length of the interval Ii . It therefore follows from the above along with
(7.45) that
0
HB pε, MΨ
IK pδq, } ¨ }2,I q
ď
κ
ÿ
i“1
¯
´
“0
pδi q, } ¨ }2,Ii ,
HB εi , MIΨi0p2Kq
řκ
řκ
provided that we choose εi “ ε{κ and δi “ δ{κ, so that i“1 εi “ ε, i“1 δi “ δ, and δi {εi “ δ{ε.
Applying Lemma 7.1 several times, we obtain that there exists an A such that
˜
¸
κ ˆ ˙
1{2
ÿ
δi
4KLi
Ψ0
HB pε, MIK pδq, } ¨ }2,I q ď A
log
εi
εi
i“1
˜
¸
ˆ ˙ÿ
1{2
κ
δ
4KLi κ
“ A
log
ε i“1
ε
˜
¸
ř
ˆ ˙
1{2
4K κi“1 Li
δ
κ log
ď A
ε
ε
˜
? ¸
ˆ ˙
δ
4K κL
ď A
κ log
,
ε
ε
using Jensen’s inequality for the two last inequalities. This proves the result.
Proof of Lemma 7.3. Recall (4.19) and let I “ r´R, Rs. Let ε1 , ε2 P p0, 1q and δ1 , δ2 P p0, 1q
with ε1 ď δ1 and ε2 ď δ2 . Our goal here is to cover GIKδ with ε-envelopes created from ε1 0
envelopes of the class of increasing functions MΨ
p2IqK pδ1 q and ε2 -neighbourhoods covering of
tα P Sd´1 : }α ´ α0 } ď δ2 u “ Sd´1 pδ2 q Ă Sd´1 . The sizes ε1 , ε2 , δ1 , δ2 will be chosen appropriately
later on.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
34
To this end, let N “ N pε2 , δ2 q denote the number of neighbourhoods with diameter at most
ε2 required to cover Sd´1 pδ2 q. Let tα1 , . . . , αN u denote elements of each of these neighbourhoods.
For any g P GIKδ , with gpxq “ ΨpαT xq for some monotone Ψ and α P Sd´1 pδ2 q, we can find an
i P t1, . . . , N u such that }α ´ αi } ď ε2 . Then, using the monotonicity of Ψ together with the
Cauchy-Schwarz inequality we can write
ΨpαTi x ´ ε2 Rq ď gpxq
“ ΨpαTi x ` pα ´ αi qT xq ď ΨpαTi x ` ε2 Rq
U
for all x P X . Note that both αTi x´ε2 R and αTi x`ε2 R belong to 2I for ε2 ď 1. Next, let rΨL
j , Ψj s
denote the ε1 -brackets of the increasing functions Ψ : 2I ÞÑ r´K, Ks, choosing the interval where
the functions are defined as 2I “ 2r´R, Rs. These functions are defined on a larger interval then
needed, and to take care of this, we extend the original functions Ψ if necessary. Then, we have
U
that if Ψ P rΨL
j , Ψj s and }α ´ αi } ď ε2 then
T
U
T
T
ΨL
j pαi x ´ ε2 Rq ď Ψpα xq ď Ψj pαi x ` ε2 Rq
for all x P X , and hence these pairs pi, jq form brackets on the larger family ΨpαT xq. We now
need to obtain a bound for the new brackets
¸1{2
˜ż
´
¯2
U
T
L
T
.
(7.46)
Ψj pαi x ` ε2 Rq ´ Ψj pαi x ´ ε2 Rq dx
X
It follows from the Minkowski inequality that an upper bound for (7.46) is
˜ż
X
¸1{2
´
¯2
T
L
T
Ψpαi x ´ ε2 Rq ´ Ψj pαi x ´ ε2 Rq dx
`
˜ż
X
`
¸1{2
´
¯2
U
T
T
Ψj pαi x ` ε2 Rq ´ Ψpαi x ` ε2 Rq dx
˜ż
X
´
ΨpαTi x
` ε2 Rq ´
ΨpαTi x
¸1{2
¯2
.
´ ε2 Rq dx
(7.47)
Using the fact that for ?
any α P Sd´1 there exists a j P t1, . . . , du such that the jth component
of α is greater than 1{ d in absolute value, we can assume without generality that the first
component of αi satisfies this property. Hence, using the change of variable t1 “ αTi x ´ ε2 R and
tj “ xj , j “ 2, . . . , d as well as t1 P r´2R, 2Rs and tj P r´R, Rs for j “ 2, . . . , d, we get
ż ´
X
¯2
T
ΨpαTi x ´ ε2 Rq ´ ΨL
j pαi x ´ ε2 Rq dx
ď
ď
ż 2R ´
¯2
?
dp2Rqd´1
Ψptq ´ ΨL
j ptq dt
´2R
?
dp2Rqd´1 ε21 ,
and the same bound is found for the square of the second integral in (7.47). Similarly,
ż ´
¯2
ΨpαTi x ` ε2 Rq ´ ΨpαTi x ´ ε2 Rq dx
X
ď
ż 2R ´
¯2
?
dp2Rqd´1
Ψptq ´ Ψpt ´ 2ε2 Rq dt.
´2R
Finally, using the monotonicity of Ψ (extended to a larger interval) and the fact that it is bounded
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
by K, we have
ż 2R ´
¯2
Ψptq ´ Ψpt ´ 2ε2 Rq dt
´2R
ď 2K
“ 2K
˜ż
2R
´2R
˜ż
Ψptqdt ´
ż 2R´2ε2 R
2R
´2R´2ε2 R
Ψptqdt
´2R´2ε2 R
Ψptqdt ´
ż ´2R
¸
Ψptqdt
´2R´2ε2 R
35
¸
ď 8K 2 ε2 R.
Combining the above results yields
¸1{2
˜ż
´
¯2
U
T
L
T
Ψj pαi x ` ε2 Rq ´ Ψj pαi x ´ ε2 Rq dx
X
ď CR,d ε21 ` K 2 ε2
(1{2
2
where CR,d ą 0 depends only on R and d. We therefore set ε21 “ K 2 ε2 and ε2 “ ε2 p2CR,d
K 2 q´1
so that the bound in the previous inequality is equal to ε. This specifies the choice of ε1 , ε2 in
the coverings, and we also need to worry about the overall size of the space δ. Since 0 ă ε ď δ,
2
we will choose ε21 ď δ12 “ K 2 δ2 and ε2 ď δ2 “ δ 2 p2CR,d
K 2 q´1 . We now check that this choice is
appropriate. Recall the meaning of δ: we have that
*1{2
"ż
`
˘2
T
T
ď δ.
Ψpα xq ´ Ψ0 pα0 xq dx
X
Similarly to our calculations above, we can show that (changed here to squared version from
previous calculations)
ż
`
˘2
ΨpαT xq ´ Ψ0 pαT0 xq dx
X
ż
ż
`
˘2
`
˘2
T
T
Ψ0 pαT xq ´ Ψ0 pαT0 xq dx
Ψpα xq ´ Ψ0 pα xq dx ` 2
ď 2
X
X
where for all x,
`
˘2
Ψ0 pαT xq ´ Ψ0 pαT0 xq
`
˘2 `
˘2
ď Ψ0 pαT0 x ` δ2 Rq ´ Ψ0 pαT0 xq ` Ψ0 pαT0 x ´ δ2 Rq ´ Ψ0 pαT0 xq .
Using the same changes of variables as above, we conclude that
ż
(
`
˘2
2
δ12 ` K 2 δ2
“ δ2.
ΨpαT xq ´ Ψ0 pαT0 xq dx ď CR,d
X
It therefore follows that
HB pε, GIKδ , } ¨ }2,X q ď
HB
ˆ
˙
˙
ˆ
ε
δ
0
?
?
, MΨ
,
}
¨
}
2,2I
p2IqK
2CR,d
2CR,d
˜
¸
¸
˜
2
δ2
ε
,} ¨}
`HB
2 , Sd´1
2 K2
2K 2 CR,d
2CR,d
where the second entropy on the right hand side is at most
¸
˜
ε2
HB
2 , Sd´1 , } ¨ } ,
2K 2 CR,d
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
36
since Sd´1 pδ2 q Ă Sd´1 . Next, we use the results of Corollary 7.2 and Lemma 4.8 to explicitly
calculate these bounds:
˜ a
¸
?
ˆ ˙
ˆ ˙
2 κ|I|K 2CR,d
K
δ
log
` log
.
HB pε, GIKδ , } ¨ }2,X q À κ
ε
ε
ε
By the assumptions of the Lemma, we also have that κ ě 1, ε ď δ ď K ´1 ď e´1 . Hence, it
follows from Lemma 7.7 that
˜
¸ ˆ ˙
a
ˆ ˙
ˆ ? ˙
23{2 CR,d K κ|I|
δ
δ
K κ
HB pε, Gr´R,RsKδ , } ¨ }2,X q À κ
log
`
log
,
ε
ε
ε
δ
ˆ ˙
ˆ 3{2
? ˙
δ
p2 CR,d |I|1{2 ` 1qpK κq
À κ
log
ε
ε
ˆ ? ˙
ˆ ˙
K κ
δ
log
,
À κ
ε
ε
using that κ ě 1 and δ ď 1{p23{2 CR,d |I|1{2 ` 1q for the two last inequalities.
Proof of Lemma 7.8. Using change of variable with s “ logp1{xq first and integration by parts
second, we obtain
ż logp1{δq
żδc
1
1
s1{2 e´s{2 ds
log dx “ ´
x
x
8
0
ż8
“
s1{2 e´s{2 ds
logp1{δq
ż8
“ p´2q
s
1{2
logp1{δq
"
*
´1 ´s{2
e
ds
2
ż8
´
+
1 ´1{2 ´s{2
s
e
ds
“ p´2q s e
logp1{δq
logp1{δq 2
#
+
ż8
a
1 ´1{2 ´s{2
“ p´2q 0 ´ δ logp1{δq ´
s
e
ds
logp1{δq 2
ż8
a
“ 2 δ logp1{δq `
s´1{2 e´s{2 ds.
#
ˇ8
ˇ
1{2 ´s{2 ˇ
logp1{δq
2
We now use change of variable again (u “ s), to find
ż8
ż8
s´1{2 e´s{2 ds “ 2 ?
2
e´u
{2
du
logp1{δq
logp1{δq
“
a
?
2 2πP pZ ą logp1{δqq
where Z has a standard Gaussian distribution. Therefore, using Chernoff’s inequality and the
above calculations,
żδc
a
a
?
1
1
log dx
“
2 δ logp1{δq ` 2 2πP pZ ą logp1{δqq
x
x
0
!
)
a
a
?
ď
2 δ logp1{δq ` 2 2π exp ´0.5p logp1{δqq2
a
? ?
“
2 δ logp1{δq ` 2 2π δ
? a
δ ď e´1
ď
2p1 ` 2πq δ logp1{δq,
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
37
completing the proof.
8. Appendix B: Additional proofs
8.1. Proof of Proposition 2.1
For an arbitrary pair pi, jq with i ‰ j, the total space Rd is separated into three disjoint parts: (1)
`
the hyperplane Hij of all vectors α P Rd such that αT pXi ´ Xj q “ 0, (2) the open half-space Sij
`
´
T
T
T
T
such that α Xi ą α Xj for all α P Sij , and (3) the open half-space Sij such that α Xi ă α Xj
´
for all α P Sij
. There are finitely many such hyperplanes. We call a subset R Ă Rd a maximal
region if R X Hij is empty for all i ‰ j, but with R the closure of R, RzR is included in the
union of all hyperplanes Hij . Note that a maximal region is an open convex cone of Rd . Then,
S X is the intersection of Sd´1 with the (finite) union of all possible maximal regions in Rd .
We will prove that for arbitrary i ‰ j and α P Hij X Sd´1 , we can find αk P S X with
p α , αq ě hn pΨ
p αk , αk q.
hn pΨ
n
n
(8.48)
We can find a maximal region R and a sequence pαk qkPN such that αk P R X Sd´1 for all k
and αk Ñ α as k Ñ 8. We have mαk “ m since αk P S X , for all k. Moreover, there exists a
(unique) permutation πR such that
αTk XπR p1q ă αTk XπR p2q ă ¨ ¨ ¨ ă αTk XπR pmq .
Denote by l1 and l2 the indices such that i “ πR pl1 q and j “ πR pl2 q. We have αT Xi “ αT Xj
since α P Hij and therefore, letting k Ñ 8 yields
αT XπR p1q ď ¨ ¨ ¨ ď αT XπR pl1 q “ ¨ ¨ ¨ “ αT XσR pl2 q ď ¨ ¨ ¨ ď αT XπR pmq .
p αk minimizes hn pΨ, αk q over Ψ P M, rearranging the terms in the sum yields
Since Ψ
n
p αk , αk q “
hn pΨ
n
inf
η1 﨨¨ďηm
m
ÿ
i“1
nπR piq YπR piq ´ ηi
(2
(8.49)
(8.50)
for all k whereas because of (8.49),
p α , αq “
hn pΨ
n
inf
η1 﨨¨ďηl1 “¨¨¨“ηl2 﨨¨ďηm
m
ÿ
i“1
nπR piq YπR piq ´ ηi
(2
.
The above infimum is taken over a restricted set as compared to (8.50), so we conclude that
l
(8.48) holds for all k. This completes the proof of Proposition 2.1.
8.2. Proof of Proposition 3.1
Using the notation of Section 2, we have
ℓn pΨ, αq “ ´
m
ÿα
i“1
α
α
α
nα
i tYi ℓ rΨpZi qs ´ B ˝ ℓ rΨpZi qsu
(8.51)
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
38
for all α and Ψ. For a fixed α P Sd´1 , setting ηi “ ΨpZiα q for all i “ 1, . . . , mα , we define
Θ
“
tη “ pη1 , . . . , ηmα qT P Rmα | a ă η1 ď η2 ď ¨ ¨ ¨ ď ηmα ă bu
and we rewrite the optimisation problem over M as maximizing Lpηq over η P Θ where
Lpηq “
m
ÿα
i“1
α
nα
i tYi ℓpηi q ´ B ˝ ℓ pηi qu .
Recall that B 1 “ ℓ´1 (see Proposition 8.1 below) is increasing. Hence, for all i such that Yiα P
pa, bq, the function x ÞÑ Yiα x ´ B pxq is concave on pℓpaq, ℓpbqq with a first derivative that vanishes
at the point x “ ℓpYiα q. This means that the function η ÞÑ Yiα ℓpηq ´ B ˝ ℓ pηq achieves it maximal
value at the point η “ Yiα . Now, for all i such that Yiα “ a (which could happen only if a is
finite), the function η ÞÑ Yiα ℓpηq ´ B ˝ ℓ pηq is decreasing on pa, bq and its maximal value is
limtaℓpηq ´ B ˝ ℓ pηqu,
ηÓa
which is finite by assumption. Hence, the maximum is achieved as η Ó a “ Yiα . Similarly, for all
i such that Yiα “ b, the maximum of function η ÞÑ Yiα ℓpηq ´ B ˝ ℓ pηq is achieved as η Ò b “ Yiα .
We conclude from the preceding arguments that for all η P Θ, we have Lpηq ď Lpηq where η P Θ
is the truncated version of η defined by
η i “ minpmaxpηi , cq, Cq,
where c “ min1ďiďmα tYiα u and C “ max1ďiďmα tYiα u. Hence, with
Θ1
“ tη “ pη1 , . . . , ηmα qT P Rmα | c ď η1 ď η2 ď ¨ ¨ ¨ ď ηmα ď Cu
we conclude that
sup Lpηq “ sup Lpηq.
ηPΘ
(8.52)
ηPΘ1
But Θ1 is a non-empty compact set in Rmα since c and C are finite, so the above supremum is
pα P M
achieved on Θ1 . Using the connection between L and ℓn , this means that there exists Ψ
that takes values in rc, Cs such that
p α , αq.
inf ℓn pΨ, αq “ ℓn pΨ
ΨPM
p α (which is not unique). Note that all functions in M
Now, we characterize the solution Ψ
α
α
p
that agree with Ψ at Zi for all i “ 1, . . . , mα are minimizers of ℓn pΨ, αq over Ψ P M. For
convenience, we consider in the sequel the stepwise solution which takes a constant value between
α
, 8q. To
the jump points, is left-continuous and takes constants values on p´8, Z1α s and rZm
α
α
α
α
p
p
characterize Ψ (i.e., to characterize Ψ pZi q for all i “ 1, . . . , mα ), fix ε and consider for x P R
the perturbation function
¯
¯
´ ´
p α pzq ´ εIp´8,xs pzq ,
Ψε pzq “ ℓ´1 ℓ Ψ
`
˘
p α , αq {ε ě 0 which yields when ε Ó 0
for all z. If ε ą 0 then Ψε P M and hence ℓn pΨε , αq ´ ℓn pΨ
m
ÿα
i“1
¯)
!
´
α
α
α
1
p
Ψ
pZ
q
Ip´8,xs pZiα q ě 0.
nα
Y
´
B
˝
ℓ
i
i
i
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
39
Because B 1 “ ℓ´1 , this can be rewritten as
m
ÿα
!
)
α
α
α
p
nα
Y
´
Ψ
pZ
q
Ip´8,xs pZiα q ě 0.
i
i
i
m
ÿα
!
)
α
α
α
p
nα
Y
´
Ψ
pZ
q
Ip´8,xs pZiα q “ 0.
i
i
i
i“1
p α , we have Ψε P M provided that |ε| is very small. Passing to the
In case x is a jump point of Ψ
limit, we obtain for such an x that
i“1
α
p α pZ1α q ď ¨ ¨ ¨ ď Ψ
p α pZm
q and
Hence, Ψ
α
j
ÿ
i“1
!
)
α
α
α
p
nα
Y
´
Ψ
pZ
q
ě0
i
i
i
p α pZ α q ă Ψ
p α pZ α q. The equality condition comes
for all j P t1, . . . , mα u, with an equality if Ψ
j
j`1
p α , a given x is a jump point if and
from the fact that, as we consider a left-continuous solution Ψ
p α pyq. This means that pΨ
p α pZ α q, . . . Ψ
p α pZ α qq are the left derivatives of
p α pxq ă limyÓx Ψ
only if Ψ
1
mα
xα .
Y
p α P M satisfies (2.5), then it agrees with a minimizer at Z α for all
Conversely, if a function Ψ
i
i “ 1, . . . , mα , which implies that it is itself a minimizer. This completes the proof of Proposition
3.1.
l
8.3. Proof of Proposition 5.1
Consider vectors α and β in Sd´1 , and non-constant functions f P Mα and h P Mβ that satisfy
f pαT Xq “ hpβ T Xq a.s. and are both left-continuous (the right-continuous case can be treated
likewise) with no discontinuity point at the boundary of their domain. To prove Proposition 5.1,
it suffices to show that in such a case, we necessarily have α “ β and f “ h on Cα “ Cβ . We
prove below that we indeed have α “ β and f “ h.
By assumption we have f pαT xq “ hpβ T xq for almost all x P X in the Lebesgue sense. Using
left-continuity of both f and h, we conclude that the equality in the previous display holds for all
x in the interior of X . If we could prove that α “ β, this would imply that f “ h on the interior
of Cα “ Cβ . By continuity of both f and h at the boundaries of their domain, this would imply
that f “ h on Cα “ Cβ .
Hence, it suffices to show that α “ β. To show that α “ β, we first notice that because of the
convexity of X , for small enough L ą 0 we can find an open ball B with radius L included in X
on which x ÞÑ f pαT xq is not constant. We then have
f pαT xq “ hpβ T xq for all x P B.
(8.53)
Without loss of generality (possibly replacing f pzq by f pz ´ αT x0 q and hpzq by hpz ´ β T x0 q with
x0 being the center of the ball), we assume that B is the open ball with center x0 “ 0 and radius
L.
Assume β R tα, ´αu (which implicitly assumes that d ě 2). We will show that this yields
a contradiction. The vectors β and α are linearly independent so it follows from the CauchySwcharz inequality, where the equality case is excluded, that αT β ă 1. With v “ β ´ α, we then
have v T α “ β T α ´ 1 ă 0. Hence, for all a P r0, Lq we have
f paq “ f pαT paαqq “ hpβ T paαqq “ hpαT paαq ` v T paαqq ď hpaq,
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
40
using (8.53) combined to the monotonicity of h. Likewise, v T β ą 0 and therefore,
hpaq “ hpβ T paβqq “ f pαT paβqq “ f pβ T paβq ´ v T paβqq ď f paq
for all a P r0, Lq, whence hpaq “ f paq for all a P r0, Lq. Similarly, f p´aq “ hp´aq for all a P r0, Lq,
whence f “ h on p´L, Lq. Combining this with (8.53) we arrive at
f pαT xq “ f pβ T xq for all x P B.
(8.54)
Since x ÞÑ f pαT xq is not constant on B, there exists a point b P p´L, Lq of strict increase of
f . The ball B can be chosen in such a way that b ‰ 0. Either we have f pb ` εq ą f pbq for all
ε P p0, L ´ bq, or we have f pb ´ εq ă f pbq for all ε P p0, L ` bq. We assume that f pb ` εq ą f pbq
for all ε P p0, L ´ bq. The other case can be handled likewise. In the case where b ą 0, with
x “ pb ` εqβ, we have x P B and αT x ď b for sufficiently small ε, using that αT β ă 1. Hence,
f pαT xq ď f pbq ă f pb ` εq “ f pβ T xq by monotonicity of f , which yields a contradiction with
(8.54). In the case b ă 0, consider ε sufficiently small so that b ` ε ă 0. Then, with x “ bα we
have x P B and β T x ě b ` ε for sufficiently small ε. Hence, f pαT xq “ f pbq ă f pb ` εq ď f pβ T xq
which again, yields a contradiction. This means that β P tα, ´αu.
Assume β “ ´α. We will show that this yields again a contradiction. For all a P r0, Lq we have
f paq “ f pαT paαqq “ hpβ T paαqq “ hp´aq,
using (8.53) with β “ ´α. This means that f paq “ hp´aq. Likewise, hpaq “ f p´aq. By monotonicity of h we then have
f paq “ hp´aq ď hpaq “ f p´aq.
As f is non-decreasing, this means that f paq “ f p´aq for all a P r0, Lq. Hence, f is constant on
p´L, Lq, which yields a contradiction. This means that β ‰ ´α. We have proved that β P tα, ´αu,
l
hence α “ β. This completes the proof of Proposition 5.1.
8.4. Proof of Theorem 5.2
Assumption A4 together with Theorem 4.1 implies that
ż ´
¯2
p n pp
Ψ
αTn xq ´ Ψ0 pαT0 xq dQpxq “ op p1q.
Since convergence in probability is equivalent to the property that each subsequence has a further
subsequence along which the convergence holds with probability one, we assume in the sequel
without loss of generality (possibly arguing along subsequences) that
ż ´
¯2
p n pp
(8.55)
Ψ
αTn xq ´ Ψ0 pαT0 xq dQpxq “ 0
lim
nÑ8
with probability one. We will show that (8.55) implies (5.22) and (5.23) with an almost sure
convergence, whence the theorem follows.
p n , where we recall
In order to use compactness arguments, we consider a truncated version of Ψ
p
that Ψn denotes the LSE extended monotonically to the whole real line: we consider Ψ̄n such
that
$
p
p
’
&Ψn ptq if Ψn ptq P pK´ , K` q
p n ptq ě K`
Ψ̄n ptq “ K`
.
if Ψ
’
%
p
K´
if Ψn ptq ď K´
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
41
where K` denotes the largest value of Ψ0 whereas K´ denotes the smallest value of Ψ0 .
We argue along paths, which means that ω is considered as fixed here and such that (8.55)
p n itself is left-continuous)
holds. For all n, the function Ψ̄n is monotone, left-continuous (because Ψ
and bounded by maxtK` , ´K´ u in sup-norm. Hence, similar to Lemma 2.5 in van der Vaart
(1998), each subsequence of pΨ̄n qně0 possesses a further subsequence along which Ψ̄n converges
pointwise to a left-continuous monotone function m0 say, at each of its point of continuity. The
function m0 possesses right-limits at every point. Because α
pn belongs to the compact set Sd´1
for all n, we can extract a further subsequence along which α
pn converges in the Rd -Euclidean
distance to some vector a0 P Sd´1 . For simplicity, we denote the general term of the subsequence
by pΨ̄n , α
pn q. We aim to show that m0 “ Ψ0 and that a0 “ α0 . For this task, consider the
L2 -distance
ż
˘2
`
m0 paT0 xq ´ Ψ0 pαT0 xq dQpxq ď 3In,1 ` 3In,2 ` 3In,3
(8.56)
where
In,1 “
ż
˘2
`
αTn xq dQpxq
m0 paT0 xq ´ m0 pp
In,3
,
“
ż
˘2
`
Ψ̄n pp
αTn xq ´ Ψ0 pαT0 xq dQpxq
In,2 “
ż
`
˘2
m0 pp
αTn xq ´ Ψ̄n pp
αTn xq dQpxq.
We will show that In,j tends to zero as n Ñ 8 for j “ 1, 2, 3 to conclude that the L2 -distance
on the left-hand side of (8.56) equals zero. To deal with In,1 , we use the fact that because
m0 is monotone, the set of its discontinuity points is countable and hence has Q-measure zero.
This means that In,1 can be viewed as an integral over the set of continuity points of m0 . At the
continuity points of m0 we have m0 pp
αTn xq Ñ m0 paT0 xq as n Ñ 8 so it follows from the dominated
convergence theorem that In,1 converges to zero as n Ñ 8. Next, it follows from the definition
of Ψ̄n that
ż ´
¯2
p n pp
Ψ
αTn xq ´ Ψ0 pαT0 xq dQpxq.
In,2 ď
Hence, with (8.55) we conclude that In,2 Ñ 0. Finally, with R taken from (4.19) and Qn the
distribution of α
pTn X, where X is independent of the data points pX1 , Y1 q, . . . , pXn , Yn q, we have
żR
ż
`
˘2
`
˘2
m0 ptq ´ Ψ̄n ptq dt
m0 ptq ´ Ψ̄n ptq dQn ptq ď A
In,3 “
´R
for some A ą 0, using that X has a bounded density function. Because m0 is monotone, the
set of its discontinuity points is countable and hence has Lebesgue measure zero. Because of the
convergence of Ψ̄n at each continuity point of m0 we conclude from the dominated convergence
theorem that In,3 converges to zero as n Ñ 8. This means that the three terms on the right-hand
side of (8.56) tends to zero as n Ñ 8 and therefore,
ż
`
˘2
m0 paT0 xq ´ Ψ0 pαT0 xq dQpxq “ 0.
Possibly modifying m0 so that its restriction to Cα0 has no discontinuity point at the boundaries
of Cα0 , which does not modify the value of the above integral, we conclude from Proposition
5.1 that a0 “ α0 and m0 “ Ψ0 with possible exception at the boundaries of Cα0 . From the
calculations above, it follows that α
pn converges to α0 and Ψ̄n converges pointwise to Ψ0 on the
interior of Cα0 with probability one. Hence with probability one,
lim sup |Ψ̄n ptq ´ Ψ0 ptq| “ 0
nÑ8 tPI
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
42
for all compact interval I included in the interior of Cα0 . Uniformity of the convergence follows
from continuity of Ψ0 together with monotonicity of the functions involved. Let I be such that
p n on I for all large enough
K´ ă Ψ0 ptq ă K` for all t P I. For almost all ω, we then have Ψ̄n “ Ψ
n and therefore,
p n ptq ´ Ψ0 ptq| “ 0
lim sup |Ψ
nÑ8 tPI
with probability one. Theorem 5.2 follows.
l
8.5. Proof of Corollary 5.3
Proof. Borrowing, from (Murphy et al., 1999, Lemma 5.7, page 418) we have that for any random
variable X, if pErg1 pXqg2 pXqsq2 ď cErg12 pXqsErg22 pXqs for c ď 1, then
?
Erpg1 pXq ` g2 pXqq2 s ě p1 ´ cqpErg1 pXq2 s ` Erg2 pXq2 sq.
(8.57)
ş
Note that in the expectations, with PX denoting the distribution of X, Erg1 pXqs “ g1 pxqdPX pxq
and so on - that is, the expectations should be viewed as short-hand notation for the integral in
the x-variable.
Let A denote a subset of X to be chosen later such that QpAq ą 0. In order to adapt the
above result to our problem, we write
ż
ż
2
pp
gn pxq ´ g0 pxqq2 dQpxq “
pg1 pxq ` g2 pxqq dQpxq
A
A
“
Erpg1 pXA q ` g2 pXA qq2 s ˆ QpAq
p n pp
where g1 pxq “ Ψ
αTn xq ´ Ψ0 pp
αTn xq “ gr1 pp
αTn xq, g2 pxq “ Ψ0 pp
αTn xq ´ g0 pxq and XA denotes here
a random variable with density function x ÞÑ qpxqIA pxq{QpAq. To alleviate the notation, in the
sequel we simply write X instead of XA . We then have
Erg1 pXqg2 pXqs2
“
“
ď
by the Cauchy-Schwarz inequality. Hence,
Erg1 pXqg2 pXqs2
where
E
cn
“
Err
g1 pp
αTn Xqg2 pXqs2
Err
g1 pp
αTn XqErg2 pXq|p
αTn Xss2
‰
“
αTn Xs2 ,
αTn XqsE Erg2 pXq|p
Err
g12 pp
ď
cn Erg12 pXqsErg22pXqs.
”`
˘2 ı
Ψ0 pp
αTn Xq ´ ErΨ0 pαT0 Xq|p
αTn Xs
”`
.
˘2 ı
αTn Xq ´ Ψ0 pαT0 Xq
E Ψ0 pp
Using (8.57) with c “ cn , we conclude that
ż
! ”
ı
?
p n pp
pp
gn pxq ´ g0 pxqq2 dQpxq ě p1 ´ cn q E pΨ
αTn Xq ´ Ψ0 pp
αTn Xqq2
A
“
‰(
`E pΨ0 pp
αTn Xq ´ Ψ0 pαT0 Xqq2 ˆ QpAq.
(8.58)
Note that cn depends on α
pn and is therefore random in the data. We will prove below that there
exists some real number c P p0, 1q such that from any subsequence, we can extract a further
subsequence along which lim supnÑ8 cn ď c ă 1 with probability one. This means that
?
(8.59)
p1 ´ cn q´1 “ Op p1q.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
43
since convergence in probability is equivalent to the property that each subsequence has a further
subsequence along which the convergence holds with probability one.
Consider an arbitrary subsequence. Define rn “ }p
αn ´ α0 } and γn “ pp
αn ´ α0 q{rn . Because
γn belongs to the compact set Sd´1 , we can extract a subsequence along which γn converges in
probability to a limit, γ say. By Theorem 5.2, we can extract a further subsequence (that we still
index by n to alleviate the notation) along which α
pn , γn converges to α0 , γ with probability one.
This means that we can extract a subsequence along which both α
pn pωq Ñ α0 and γn pωq Ñ γpωq
for almost all paths ω. In the sequel, we argue with such a path ω fixed. Hence, α
pn and γn can
be considered as non random and we search for bounds that do not depend on the chosen path
ω. We denote by x0 a point in X such that z0 “ αT0 x0 where z0 is taken from Assumption A5
and we consider ε ą 0 such that Ψ0 is continuously differentiable over V :“ rz0 ´ 2ε, z0 ` 2εs with
a derivative that is bounded both from above and away from zero on V. Note that the derivative
is uniformly continuous on the compact set V. Furthermore, we denote by A the Euclidean ball
with center x0 and radius ε. Note that for large enough n, we then have αT0 x P V and α
pTn x P V
for all x P A, by the Cauchy-Schwarz inequality.
We have that
Ψ0 pαT0 xq “ Ψ0 pp
αTn xq ` Ψ10 pp
αTn xqpα0 ´ α
pn qT x ` oprn q
(8.60)
uniformly for all x P A since |pp
αn ´ α0 qT x| ď rn }x} where }x} ď }x0 } ` ε. Hence,
”`
˘2 ı
αTn Xq ´ ErΨ0 pαT0 Xq|p
αTn Xs
E Ψ0 pp
”` “
(ˇ T ‰˘2 ı
pn X
αTn Xqpp
αn ´ α0 qT X ` oprn q ˇ α
“ E E Ψ10 pp
ı
”` “
ˇ
‰˘
2
αn ´ α0 qT X ˇ α
pTn X
αTn Xqpp
` oprn2 q ` En,1
“ E E Ψ10 pp
where we recall that X denotes here XA , a random variable supported on A, and
“
‰
|En,1 | “ 2oprn q|E Ψ10 pp
αTn Xqpp
αn ´ α0 qT Xq | “ oprn2 q,
again by (A5). Similarly, we have
”`
”`
˘2 ı
˘2 ı
αTn Xq ´ Ψ0 pαT0 Xq
αTn Xqpp
αn ´ α0 qT X
E Ψ0 pp
“ E Ψ10 pp
` oprn2 q.
Combining these calculations, we arrive at
”`
“
‰˘2 ı
αTn XqγnT E X|p
αTn X
` op1q
E Ψ10 pp
”
ı
.
cn “
2
E pΨ10 pp
αTn XqγnT Xq ` op1q
αTn XsÝÑErX|αT0 Xs almost surely. This, along with
Lemma 8.2 in the Appendix shows that ErX|p
continuity of Ψ10 and the Lebesgue dominated convergence theorem (since |X| ď }x0 } ` ε almost
surely), implies that cn converges with
”
ı
E pΨ10 pαT0 Xqγ T ErX|αT0 Xsq2
”`
lim cn “
˘2 ı
nÑ8
E Ψ10 pαT0 Xqγ T X
“
‰
γ T E pΨ10 pαT0 Xqq2 ErX|αT0 XsErX|αT0 XsT γ
“ 1 T
‰
.
“
γ T E pΨ0 pα0 Xqq2 XX T γ
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
44
Now, }p
αn }2 “ }α0 ` rn γn }2 and therefore, 1 “ ||α0 ||2 ` rn2 ` 2rn xα0 , γn y, which implies that
xα0 , γn y “ ´rn {2 Ñ 0 since }α0 }2 “ 1. Since γ “ limnÑ8 γn , this implies that xα0 , γy “ 0 and
therefore, limnÑ8 cn ď c where
“
‰
γ T E pΨ10 pαT0 Xqq2 ErX|αT0 XsErX|αT0 XsT γ
‰
“
c“
sup
γ T E pΨ10 pαT0 Xqq2 XX T γ
}γ}“1,xγ,α0 y“0
does not depend on the chosen path ω. It remains to prove that c ă 1. Now, note that
“
‰
“
‰
E pΨ10 pαT0 Xqq2 XX T
“ E pΨ10 pαT0 Xqq2 ErX|αT0 XsErX|αT0 XsT
“
‰
`E pΨ10 pαT0 Xqq2 pX ´ ErX|αT0 XsqpX ´ ErX|αT0 XsqT .
1
T
Since
our ‰goal is now to show that the null-space of the matrix
“ Ψ0 pα0 XqTis bounded below,
E pX ´ ErX|α0 XsqpX ´ ErX|αT0 XsqT is spanned by α0 only, as this will imply that c ă 1.
To this end, consider any γ0 perpendicular to α0 with }γ0 } “ 1. Let A0 denote the matrix with
first row αT0 and second row γ0T and let Z “ A0 X. Since X has an everywhere-positive density,
so does Z. To see this, take Z 1 “ A10 X where A10 is a d ˆ d invertible matrix such that A10 has its
first two rows equal to A0 . (Such a matrix exists since γ0 and α0 are perpendicular and hence
linearly independent.) Then Z 1 has a density with respect to the Lebesgue measure which can
be explicitly calculated using the Jacobian formula and its marginal gives the density of Z, fZ
say. Now,
‰
‰
“
“
γ0T E pX ´ ErX|αT0 XsqpX ´ ErX|αT0 XsqT γ0 “ E pγ0T X ´ Erγ0T X|αT0 Xsq2 .
This equals zero iff γ0T X “ Erγ0T X|αT0 Xs almost surely, or, Z2 “ ErZ2 |Z1 s almost surely.
However, this means that the distribution of Z is concentrated on a one-dimensional subspace
of Z, which cannot hold since Z has an everywhere-positive Lebesgue density. This finally shows
that limnÑ8 cn ď c ă 1.
Then, if follows from (8.58) that we can find c1 ą 0 and c2 ą 0 independent on n and ω such
that for large n,
ż
ż
pp
gn pxq ´ g0 pxqq2 dQpxq ě c1 pΨ0 pp
αTn xq ´ Ψ0 pαT0 xqq2 dQpxq
A
A
ż
2
ě c
ppp
αn ´ α0 qT xq2 dQpxq
A
ż
ě c2 }p
αn ´ α0 }2 inf
pβ T xq2 dQpxq,
βPSd´1 A
by definition of A. The infimum above does not depend
ş on the chosen path ω and is achieved
at a point β0 , say, by continuity of the function β ÞÑ A pβ T xq2 dQpxq on the compact set Sd´1 .
Hence,
ż
ż
2
2
2
pp
gn pxq ´ g0 pxqq dQpxq ě c }p
αn ´ α0 }
pβ0T xq2 dQpxq.
A
A
The integral on the right-hand side is strictly positive since Q has a density function that is
everywhere-positive positive on A. This means that there exists K ą 0 such that from each
subsequence, we can extract a further subsequence along which
ż
gn pxq ´ g0 pxqq2 dQpxq
}p
αn ´ α0 }2 ď K pp
A
ż
gn pxq ´ g0 pxqq2 dQpxq
ď K pp
X
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
45
for large n, with probability one. Since the right-hand side is of order Op pn´2{3 q by Theorem 4.1,
this implies that }p
αn ´ α0 } “ Op pn´1{3 q.
In the sequel, we assume that Ψ0 has a derivative bounded from above on Cα0 and we extend
Ψ0 to the whole real line in such a way that the extension has a bounded derivative on R.
We assume without loss of generality that |α0,1 | ď d´1{2 and therefore, |p
αn,1 | ď 2d´1{2 with
probability that tends to one. Then, with q from (4.10) we have
ż
ż
T
T
2
p
p n pp
pΨn pp
αn xq ´ Ψ0 pp
αn xqq dQpxq ě q pΨ
αTn xq ´ Ψ0 pp
αTn xqq2 dx
X
X
ż
q
p n pt1 q ´ Ψ0 pt1 qq2 dt1 ¨ ¨ ¨ dtn ,
pΨ
“
α
pn,1 Cαp n ˆXd´1
where Xd´1 “ tpx2 , . . . , xd q : x P X u and Cαp n “ tp
αTn x : x P X u, using the change of variable
T
t1 “ α
pn x and tj “ xj for j “ 2, . . . , d. Since
ż
ż
d1{2
α
p´1
dt
¨
¨
¨
dt
ě
dt2 ¨ ¨ ¨ dtn ą 0,
2
n
n,1
2 Xd´1
Xd´1
there exists another constant c ą 0 such that
ż
p n pp
pΨ
αTn xq ´ Ψ0 pp
αTn xqq2 dQpxq
ě
X
ě
c
c
ż
Cαp n
p n ptq ´ Ψ0 ptqq2 dt
pΨ
ż c´vn
c`vn
p n ptq ´ Ψ0 ptqq2 dt
pΨ
(8.61)
with probability that tends to one, using the definition of vn and α
pn ´ α0 “ Op pn´1{3 q. On the
other hand,
ż
ż
T
T
2
p n pp
p
αTn xq ´ Ψ0 pαT0 xqq2 dQpxq
pΨn pp
αn xq ´ Ψ0 pp
αn xqq dQpxq ď 2 pΨ
X
X
ż
αTn xqq2 dQpxq
`2 pΨ0 pαT0 xq ´ Ψ0 pp
X
where by Theorem 4.1, the first integral on the right hand side is of order Op pn´2{3 q. For the
second integral, denoting by K an upper bound for the sup-norm of Ψ10 on R we have
ż
ż
T
T
2
pn qT xq2 dQpxq
pΨ0 pα0 xq ´ Ψ0 pp
αn xqq dQpxq ď K ppα0 ´ α
X
X
ď
K}α0 ´ αn }2 R2 “ Op pn´2{3 q
where R is taken from (4.19). Combining, we conclude that (5.24) holds. This proves the second
result.
8.6. Some properties of exponential families
Proposition 8.1. Let Y be an integrable random variable having a density with respect to a
dominating measure λ on R of the form
*
"
yℓpµq ´ Bpℓpµqq
(8.62)
y ÞÑ hpy, φq exp
φ
where µ is the mean, φ is a dispersion parameter, ℓ is a real valued function with a strictly positive
first derivative on a non void open interval pa, bq Ă R, and h is a normalizing function. We then
have
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
46
1. B is infinitely differentiable with B 1 “ ℓ´1 on pℓpaq, ℓpbqq, and ℓ is infinitely differentiable.
2. If ℓpµq belongs to a compact interval wich is stricly included in the domain of B, then we
can find positive numbers a0 and M that depend only on that compact interval such that
E|Y |s ď a0 s!M s´2 for all integers s ě 1.
Proof. Setting η “ ℓpµq, the density takes the form
*
"
yη ´ Bpηq
.
y ÞÑ hpy, φq exp
φ
Since it integrates to one with respect to the dominating measure λ, we have
"
* ż
" *
Bpηq
yη
exp
“ hpy, φq exp
dλpyq
φ
φ
for all η P pℓpaq, ℓpbqq. It follows from standard properties of the Laplace transform that the righthand side of the previous display is infinitely differentiable as a function of η on R, so we conclude
that B is infinitely differentiable on pℓpaq, ℓpbqq. Moreover, we can differentiate and interchange
derivation and integration to obtain that on pℓpaq, ℓpbqq,
"
" *
*
ż
Bpηq
yη
B
B
exp
“
hpy, φq exp
dλpyq
Bη
φ
Bη
φ
" *
ż
y
yη
“
hpy, φq exp
dλpyq
φ
φ
"
*
Bpηq
EpY q
exp
.
“
φ
φ
Hence, EpY q “ B 1 pηq. Going back to the parameter µ “ EpY q, we conclude that µ “ B 1 pℓpµqq
for all µ P pa, bq, whence B 1 “ ℓ´1 . This proves the first assertion.
To prove the second assertion, note that with again η “ ℓpµq we have
*
"
ż
yη ´ Bpηq
dλpyq
E rexpttY us “
hpy, φq exp ty `
φ
for all t P R. Now, denote by rc, ds the compact interval to which η is assumed to belong. Because
this interval is strictly included in the domain of B, there exists ε ą 0 such that rc ´ ε, d ` εs is
included in the domain of B. With t “ ˘ε{φ, using the fact that the density in the exponential
family with natural parameter η ` φt integrates to one, we obtain
"
*
Bpη ` φtq ´ Bpηq
E rexpttY us “ exp
.
(8.63)
φ
Choosing t “ ε{φ we conclude that
E rexptt|Y |us
ď
"
Bpη ` φtq ´ Bpηq
exp
φ
*
"
Bpη ´ φtq ´ Bpηq
` exp
φ
*
where the left-hand side is bounded uniformly in η since B is continuous on rc ´ ε, d ` εs. In
the sequel, we denote by C a positive number that is greater than the right-hand side for all
η P rc, ds. Since
ÿ tk E|Y |k
ts E|Y |s
ě
E rexptt|Y |us “
k!
s!
kě0
for all s ě 0, we conclude that E|Y |s ď a0 s!M s´2 for all integers s ě 1, where a0 “ C{t2 and
M “ 1{t. This concludes the proof of Proposition 8.1.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
Balabdaoui, F., Durot, C. and Jankowski, H./LSE in the monotone single index model
47
8.7. Proofs for Subsection 5.2
Lemma 8.2. Let X be a random variable having a density function q with respect to Lebesgue
measure on a bounded subset of Rd , and let αn be a non-random sequence in Sd´1 that converges
to some α0 as n Ñ 8. If q is continuous and bounded on its domain, we then have
ErX|αTn Xs ÝÑ ErX|αT0 Xs
with probability one, as n Ñ 8.
Proof. In the following, α0,j and αn,j denotes the j-th component of α0 and αn respectively.
Without loss of generality, we can assume that α0,1 ‰ 0, and hence αn,1 ‰ 0 provided that
n is large enough. Consider the transformation Z1 “ αTn X and Zj “ Xj , j P t2, . . . , du. Let
Z “ pZ1 , . . . , Zd q. Simple calculations show that the density of Z is given by
ˆ
˙
z1 ´ αn,2 z2 ´ . . . ´ αn,d zd
1
fZ pzq “ q
, z2 , . . . , zd
.
αn,1
αn,1
This yields the density of the conditional distribution of Xj given that αTn X “ t, hnj pxj |tq, for
j P t2, . . . , du:
¯
ş ´ t´αn,2 z2 ´...´αn,d zd
,
z
,
.
.
.
,
z
,
x
,
z
.
.
.
,
z
q
2
j´1
j j`1
d dz2 . . . dzj´1 dzj`1 . . . dzd
αn,1
¯
.
hnj pxj |tq “
ş ´ t´αn,2 z2 ´...´αn,d zd
q
, z2 , . . . , zd dz2 . . . dzd
αn,1
where the domains of integrations in the numerator and denominator are
tpx2 , . . . , xj´1 , xj`1 , . . . , xd q : x P X u and tpx2 , . . . , xd q : x P X u respectively. This in turn
gives
ż
T
ErXj |αn Xs “ xj hnj pxj |αTn Xqdxj .
Using convergence of αn to α0 and the assumptions on q, it follows that hnj pxj |tq converges to
h0j pxj |tq for all t, where h0j p¨|tq is defined in the same manner as hnj p¨|tq with αn replaced by α0 .
Hence, h0j p¨|tq is the conditional density of Xj given that αT0 X “ t. Using again the Lebesgue
dominated convergence theorem, since X is supported on a bounded subset of Rd , we then have
that
ErXj |αTn Xs Ñ ErXj |αT0 Xs
almost surely for all j P t2, . . . , du. For j “ 1, note that
¸
˜
d
ÿ
1
αn,j ErXj |αTn Xs
αTn X ´
ErX1 |αTn Xs “
αn,1
j“2
˜
¸
d
ÿ
1
T
T
Ñ
α0 X ´
α0,j ErXj |α0 Xs “ ErX1 |αT0 Xs
α0,1
j“2
which completes the proof.
imsart-generic ver. 2011/11/15 file: IsoSIM_ArXiv.tex date: January 10, 2018
| 10math.ST
|
arXiv:1606.04380v2 [math.AC] 27 Feb 2017
On the generators of the canonical module of a
Hibi ring: a criterion of level property and the
degrees of generators
Mitsuhiro MIYAZAKI∗
Department of Mathematics, Kyoto University of Education, 1
Fujinomori, Fukakusa, Fushimi-ku, Kyoto, 612-8522, Japan
Abstract
In this paper, we study the minimal generating system of the
canonical module of a Hibi ring. Using the results, we state a characterization of a Hibi ring to be level. We also give a characterization of
a Hibi ring to be of type 2. Further, we show that the degrees of the
elements of the minimal generating system of the canonical module of
a Hibi ring form a set of consecutive integers.
Hibi ring, canonical module, level ring, Cohen-Macaulay type
MSC: 13F50, 13H10, 06A07
1
Introduction
A Hibi ring is an algebra with straightening law on a distributive lattice with
many good properties. It appears in many scenes of the study of rings with
combinatorial structure. In many important cases, the initial subalgebra of
a subalgebra of a polynomial ring has a structure of a Hibi ring. A Hibi ring
is a normal affine semigroup ring and is therefore Cohen-Macaulay by the
result of Hochster [2]. Moreover there is a good description of the canonical
module of it [1]. However, as the Example of [1, §1 e)] shows, a Hibi ring is
not necessarily level.
The present author gave a sufficient condition for a Hibi ring to be level [3]
and showed that the homogeneous coordinate ring of a Schubert subvariety
∗
The author is supported partially by JSPS KAKENHI Grant Number 15K04818.
1
of a Grassmannian is level. However, the sufficient condition given in [3] is
far from necessary.
In this paper we analyze the generating system of the canonical module
of a Hibi ring by using the description of the canonical module of a normal
affine semigroup ring by Stanley [5]. Using this, we give a combinatorial
necessary and sufficient condition for a Hibi ring to be level. We also show
that the set of degrees of the generators of the canonical module of a Hibi
ring form a set of consecutive integers. Here we call a member of the minimal
generating system a generator for simplicity.
Further, we give a combinatorial criterion that the Hibi ring under consideration has Cohen-Macaulay type 2. Recall that a Hibi ring is Gorenstein
i.e., type of it is 1 if and only if the set of join-irreducible elements of the
distributive lattice defining that Hibi ring is pure [1, §3 d)]. Since type 2 is
next to type 1, we think it worth to give a criterion of a Hibi ring has type
2.
This paper is organized as follows. Section 2 is a preparation for the main
argument. In Section 2, we recall some basic facts on Hibi rings, normal affine
semigroup rings and levelness of standard graded algebras. Also, we state
some basic facts on posets.
In Section 3, we state a necessary and sufficient condition that a Hibi
ring is level by considering the degrees of the generators of the canonical
module of the Hibi ring. The key notion is a sequence with “condition N”
(see Definition 3.1). We named the condition as condition N from the shape
of the Hasse diagram related to the elements in the sequence. We show that
from a generator of the canonical module of a Hibi ring, one can construct
a sequence with condition N with additional condition in the poset of joinirreducible elements corresponding to the Hibi ring. We also show a kind of
converse of this fact. Thus, the range of the degrees of the generators of the
canonical module of a Hibi ring can be described by sequences with condition
N. Using this, we obtain a combinatorial criterion for a Hibi ring to be level.
Further, we show that the degrees of the elements of the minimal generating system of the canonical module of a Hibi ring form a consecutive
integers, i.e., if there are elements of degrees d1 and d2 in the minimal system of generators of the canonical module of a Hibi ring with d1 < d2 , then
there exists an element with degree d in the minimal system of the generators
of the canonical module of the Hibi ring for any integer d with d1 ≤ d ≤ d2 .
In Sections 4 and 5, we state a combinatorial criterion that a Hibi ring is
of type 2. In Section 4, we state a necessary and sufficient condition that a
Hibi ring is level and has type 2 and in Section 5, we state a necessary and
sufficient condition that a Hibi ring is not level and has type 2.
2
2
Preliminaries
In this paper, all rings and algebras are assumed to be commutative with
identity element. We denote by N the set of non-negative integers, by Z
the set of integers, by R the set of real numbers and by R≥0 the set of
non-negative real numbers.
First we recall some definitions concerning finite partially ordered sets
(poset for short).
Definition Let Q be a finite poset.
• A chain in Q is a totally ordered subset of Q.
• For a chain X in Q, we define the length of X as #X − 1, where #X
denotes the cardinality of X.
• The maximum length of chains in Q is called the rank of Q and denoted
as rankQ.
• If every maximal chain of Q has the same length, we say that Q is pure.
• If x, y ∈ Q, x < y and there is no z ∈ Q with x < z < y, we say that
y covers x and denote x <· y or y ·> x.
• For x, y ∈ Q with x ≤ y, we set [x, y]Q := {z ∈ Q | x ≤ z ≤ y},
and for x, y ∈ Q with x < y, we set [x, y)Q := {z ∈ Q | x ≤ z < y},
(x, y]Q := {z ∈ Q | x < z ≤ y} and (x, y)Q := {z ∈ Q | x < z < y}.
We denote [x, y]Q ([x, y)Q, (x, y]Q (x, y)Q resp.) as [x, y] ([x, y), (x, y],
(x, y) resp.) if there is no fear of confusion.
• Let ∞ be a new element which is not contained in Q. We set Q+ :=
Q ∪ {∞} with the order x < ∞ for any x ∈ Q.
• If I ⊂ Q and x ∈ I, y ∈ Q, y ≤ x ⇒ y ∈ I, then we say that I is a
poset ideal of Q.
Next we recall the definition and basic facts of Hibi rings. Let H be a finite
distributive lattice and x0 the unique minimal element of H. Let P be the
set of join-irreducible elements of H, i.e., P = {α ∈ H | α = β ∨ γ ⇒ α = β
or α = γ}. Note that we treat x0 as a join-irreducible element of H.
Then it is known that H is isomorphic to J(P )\{∅}, ordered by inclusion,
by the correspondence
α 7→ _
{x ∈ P | x ≤ α in H} for α ∈ H and
I 7→
x
for I ∈ J(P ) \ {∅},
x∈I
3
where J(P ) is the set of poset ideals of P . In particular, if P is a poset with
unique minimal element, then there is a finite distributive lattice whose set
of join-irreducible elements is P .
Let {Tx }x∈P be a family of indeterminates indexed by P and K a field.
Q
Definition ([1]) RK (H) := K[ x≤α Tx | α ∈ H].
RK (H) is called the Hibi ring over K on H. We set deg Tx0 = 1 and deg Tx =
0 for any x ∈ P \ {x0 }. Then, RK (H) is a standard graded K-algebra, i.e., a
non-negatively graded K-algebra whose degree 0 part is K and is generated
over K by the degree 1 elements.
Q
ν(x)
Definition ([1]) For a map ν : P → N , we set T ν := x∈P Tx . We set
T (P ) := {ν : P → N | x ≤ y ⇒ ν(x) ≥ ν(y)} and T (P ) := {ν : P → N |
ν(z) > 0 for any z ∈ P and x < y ⇒ ν(x) > ν(y)}. For ν ∈ T (P ), we extend
ν as a map from P + to N by setting ν(∞) = 0.
With this notation,
L
ν
Theorem 2.1 ([1]) RK (H) =
ν∈T (P ) KT . In particular, by the result
of Hochster [2], RK (H) is a normal affine semigroup ring and is CohenMacaulay.
Note that deg T ν = ν(x0 ).
Here we recall the description of the canonical module of a normal affine
semigroup ring by Stanley.
Theorem 2.2 ([5, p. 81]) Let S be a finitely generated additive submonoid
n
of
X1 , . . . , Xn indeterminates. If the affine semigroup ring
L N and
s
KX
in
Ls∈S
LK[X1 , . . . , Xn ] is s normal, then the canonical module of
s
s∈S KX is
s∈S∩relintR≥0 S KX , where relintQ denotes the interior of Q
in the affine space spanned by Q.
By this result, the canonical module of a normal affine semigroup ring has the
unique minimal fine graded generating system up to non-zero scalar multiplication. Therefore, we call a member of the minimal fine graded generating
system a generator for simplicity.
By applying this theorem to RK (H), we see the following
L
Corollary 2.3 The canonical module of RK (H) is ν∈T (P ) KT ν .
In order to describe the generators of the canonical module of RK (H), we
state the following
4
def
Definition We define the order on T (P ) by setting ν ≤ ν ′ ⇐⇒ ν ′ − ν ∈
T (P ) for ν, ν ′ ∈ T (P ), where (ν ′ − ν)(x) := ν ′ (x) − ν(x).
Then the following fact is easily verified.
Corollary 2.4 Let ν be an element of T (P ). Then T ν is a generator of the
canonical module of RK (H) if and only if ν is a minimal element of T (P ).
Finally, we recall the following characterization of Gorenstein property of
RK (H) by Hibi.
Theorem 2.5 ([1, §3 d)]) RK (H) is Gorenstein if and only if P is pure.
Here we recall that Stanley [4] defined the level property for a standard
graded Cohen-Macaulay algebra over a field and showed that a standard
graded Cohen-Macaulay algebra A over a field is level if and only if the
degree of the generators of the canonical module of A is constant.
In the following, let H be a finite distributive lattice with unique minimal
element x0 and let P be the set of join-irreducible elements of H. We set
r := rankP + . Since deg T ν = ν(x0 ) for ν ∈ T (P ), and ν0 : P → N , x 7→
rank[x, ∞] is a minimal element of T (P ), we see by Corollary 2.4 the following
Corollary 2.6 RK (H) is level if and only if ν(x0 ) = r for any minimal
element ν of T (P ).
Next we state some basic facts on T (P ).
Lemma 2.7 Let I be a non-empty poset ideal of P and ν ∈ T (P ). Suppose
that ν(x) − ν(y) ≥ 2 for any x ∈ I and y ∈ P + \ I with x < y. If we set
ν(x)
if x 6∈ I and
′
ν (x) =
ν(x) − 1
if x ∈ I,
then ν ′ ∈ T (P ). In particular, ν is not a minimal element of T (P ).
As a corollary, we have the following
Lemma 2.8 Let ν ∈ T (P ). If {z ∈ P + | ν(z) > rank[z, ∞]} is a non-empty
poset ideal of P + , then ν is not a minimal element of T (P ).
We also state the following fact.
5
Lemma 2.9 Let Q be a poset and s a positive integer. If for any z0 , z1 , z2
and z3 ∈ Q such that z0 ≤ z1 ≤ z2 ≤ z3 , z0 is a minimal element of Q and
z3 is a maximal element of Q, it holds
rank[z0 , z1 ] + rank[z1 , z2 ] + rank[z2 , z3 ] = s,
then Q is pure of rank s.
Proof Let x0 < x1 < · · · < xt be an arbitrary maximal chain of Q. We shall
show that t = s.
Since
rank[x0 , xt ] = rank[x0 , x0 ] + rank[x0 , x0 ] + rank[x0 , xt ] = s
by assumption, we see that
t ≤ rank[x0 , xt ] = s.
Assume t < s. Then {i | i < rank[x0 , xi ]} 6= ∅. Set j = min{i | i <
rank[x0 , xi ]}. Then j ≥ 1. Further, we see by assumption, that
s = rank[x0 , xj−1 ] + rank[xj−1 , xj ] + rank[xj , xt ]
= (j − 1) + 1 + rank[xj , xt ].
Since rank[x0 , xj ] + rank[xj , xt ] = rank[x0 , xj ] + rank[xj , xj ] + rank[xj , xt ] = s,
we see that rank[xj , xt ] = s − rank[x0 , xj ]. Thus,
s = j + (s − rank[x0 , xj ]) < j + s − j = s,
by the definition of j. This is a contradiction.
3
A criterion of levelness of a Hibi ring
In this section, we give a necessary and sufficient condition for a Hibi ring
to be level. Recall that H is a finite distributive lattice with unique minimal
element x0 , P is the set of join-irreducible elements of H and r = rankP + .
First we make the following
Definition 3.1 Let y1 , x1 , y2 , x2 , . . . , yt , xt be a (possibly empty) sequence
of elements in P . We say the sequence y1 , x1 , y2 , x2 , . . . , yt , xt satisfies the
condition N if
(1) x1 6= x0 ,
6
(2) y1 > x1 < y2 > x2 < · · · < yt > xt and
(3) for any i, j with 1 ≤ i < j ≤ t, yi 6≥ xj .
Remark A sequence with condition N may be an empty sequence, i.e., t
may be 0.
For a sequence with condtion N, we make the following
Definition Let y1 , x1 , y2 , x2 , . . . , yt , xt be a sequence with condition N.
We set
r(y1 , x1 , . . . , xt , yt ) :=
t
X
(rank[xi−1 , yi ] − rank[xi , yi ]) + rank[xt , ∞],
i=1
where we set an empty sum to be 0.
Remark For an empty sequence, we set r() = rank[x0 , ∞] = r.
Next we state a useful criterion of minimal property of an element of T (P ).
First we state the following
Lemma 3.2 Supposet that ν ∈ T (P ). If there are elements z1 , w1 , z2 , w2 ,
. . . , zs , ws ∈ P such that
(1) z1 > w1 < z2 > w2 < · · · < zs > ws and
(2) ν(wi ) − ν(zi+1 ) = rank[wi , zi+1 ] for 0 ≤ i ≤ s, where we set w0 = x0
and zs+1 = ∞,
then ν is a minimal element of T (P ).
Remark s may be 0 in Lemma 3.2, i.e., if ν(x0 ) = r, then ν is a minimal
element of T (P ).
Proof of Lemma 3.2 The case where s = 0 is obvious. Thus, we consider
the case where s > 0. Assume the contrary and suppose that there is ν ′ ∈
T (P ) with ν ′ < ν. Then ν(w0 )−ν ′ (w0 ) = (ν−ν ′ )(w0 ) > 0 since ν−ν ′ ∈ T (P )
and ν 6= ν ′ . Set j = max{i | ν ′ (wi ) < ν(wi )}. Since ν(wj ) − ν(zj+1 ) =
rank[wj , zj+1 ] ≤ ν ′ (wj ) − ν ′ (zj+1 ) by assumption, we see that
ν ′ (zj+1 ) < ν(zj+1 ).
In particular, j < s. Since ν − ν ′ ∈ T (P ), (ν − ν ′ )(wj+1) ≥ (ν − ν ′ )(zj+1 ),
i.e., ν(wj+1 ) − ν ′ (wj+1 ) ≥ ν(zj+1 ) − ν ′ (zj+1 ) > 0. This contradicts to the
definition of j.
Next we show a kind of converse of this lemma.
7
Lemma 3.3 Let ν be a minimal element of T (P ). Then there exists a sequence of elements y1 , x1 , . . . , yt , xt with condition N such that
ν(xi ) − ν(yi+1 ) = rank[xi , yi+1 ]
for 0 ≤ i ≤ t, where we set yt+1 = ∞.
Remark y1 , x1 , . . . , yt , xt may be an empty sequence, i.e., t may be 0.
Proof of Lemma 3.3 Set U1 := {y ∈ P + | rank[x0 , y]P + = ν(x0 ) − ν(y)},
D1 := {x ∈ P + \ U1 | ∃y ∈ U1 such that y > x}, U2 := {y ∈ P + \ (U1 ∪ D1 ) |
∃x ∈ D1 such that x < y and rank[x, y]P + = ν(x) − ν(y)}, D2 := {x ∈
P + \ (U1 ∪ D1 ∪ U2 ) | ∃y ∈ U2 such that y > x} and so on. Since P is a finite
set, the procedure stops after finite steps, i.e., Ut = ∅ or Dt = ∅ for some t.
It is proved by the induction on s that
U1 ∪ D1 ∪ U2 ∪ D2 ∪ · · · ∪ Us ∪ Ds
is a poset ideal of P + for any s. Therefore,
I = U1 ∪ D1 ∪ U2 ∪ D2 ∪ · · ·
is a poset ideal of P + .
Suppose that ∞ 6∈ I. Then P + \ I 6= ∅. Moreover, if z ∈ I, z ′ ∈ P + \ I
and z < z ′ , then ν(z) ≥ ν(z ′ ) + 2. In fact, if z ∈ Ds for some s, then
ν(z) − ν(z ′ ) > rank[z, z ′ ] ≥ 1, since z ′ 6∈ Us+1 ∪ Ds ∪ Us ∪ · · · ∪ U1 . If z ∈ Us
for some s, then there is x ∈ Ds−1 such that ν(x) − ν(z) = rank[x, z], where
we set D0 = {x0 }. Since z ′ 6∈ Us ∪ Ds−1 ∪ Us−1 ∪ · · · ∪ U1 , we see that
ν(x) − ν(z ′ ) > rank[x, z ′ ]. Therefore,
ν(z) − ν(z ′ ) = (ν(x) − ν(z ′ )) − (ν(x) − ν(z))
> rank[x, z ′ ] − rank[x, z] ≥ rank[z, z ′ ] ≥ 1.
Thus, by Lemma 2.7, we see that ν is not a minimal element of T (P ),
contradicts the assumption.
Therefore, ∞ ∈ I. Take t with ∞ ∈ Ut+1 . Set yt+1 = ∞, take xt ∈ Dt
such that xt < yt+1 and rank[xt , yt+1 ] = ν(xt ) − ν(yt+1 ), take yt ∈ Ut such
that yt > xt , take xt−1 ∈ Dt−1 such that xt−1 < yt and rank[xt−1 , yt ] =
ν(xt−1 ) − ν(yt ), take yt−1 ∈ Ut−1 such that yt−1 > xt−1 and so on. Then it is
easily verified that y1 , x1 , . . . , yt , xt is a sequence with condition N and
ν(xi ) − ν(yi+1 ) = rank[xi , yi+1 ]
for 0 ≤ i ≤ t.
Here we state the following fact.
8
Lemma 3.4 There are only finitely many sequences with condition N.
Proof If y1 , x1 , . . . , yt , xt is a sequence with condition N, then xi ≤ yi and
xj 6≤ yi for j > i. Therefore, we see that x1 , . . . , xt are distinct elements of
P . Since P is a finite set, we see the result.
Now we make the following
Definition We set rmax := max{r(y1, x1 , . . . , yt, xt ) | y1 , x1 , . . . , yt , xt is a
sequence with condition N}.
We note here the following fact.
Remark r is not necessarily equal to min{r(y1 , x1 , . . . , yt , xt ) | y1 , x1 , . . . ,
yt , xt is a sequence with condition N}.
As a corollary of Lemma 3.3, we see the following fact.
Corollary 3.5 If ν is a minimal element of T (P ), then r ≤ ν(x0 ) ≤ rmax .
Proof Since ν ∈ T (P ), we see that r ≤ ν(x0 ). On the other hand, by
Lemma 3.3, we see that there is a sequence y1 , x1 , . . . , yt , xt with condition
N such that ν(xi−1 ) − ν(yi ) = rank[xi−1 , yi ] for 1 ≤ i ≤ t + 1, where we set
yt+1 = ∞. Thus,
ν(x0 ) =
=
≤
t
X
i=1
t
X
i=1
t
X
(ν(xi−1 ) − ν(yi ) + ν(yi ) − ν(xi )) + ν(xt ) − ν(yt+1 )
(rank[xi−1 , yi ] − (ν(xi ) − ν(yi ))) + rank[xt , yt+1 ]
(rank[xi−1 , yi ] − rank[xi , yi ]) + rank[xt , yt+1 ]
i=1
= r(y1 , x1 , . . . , yt , xt )
≤ rmax .
Next we define two elements of T (P ) defined by a sequence of elements
with condition N.
9
Definition 3.6 Let y1 , x1 , . . . , yt , xt be a sequence of elements with condition N. Set yt+1 = ∞. We define
µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ) :=
t
X
(−rank[xk , yk ] + rank[xk , yk+1])
k=i
for 1 ≤ i ≤ t + 1 and
ν ↓ (y1 ,x1 ,...,yt ,xt ) (z) := max{rank[z, yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ) | z ≤ yi }
for z ∈ P + . We also define
µ↑ (y1 ,x1 ,...,yt ,xt ) (xi ) := rmax +
i
X
(−rank[xk−1 , yk ] + rank[xk , yk ])
k=1
for 0 ≤ i ≤ t and
ν ↑ (y1 ,x1,...,yt ,xt ) (z) := min{−rank[xi , z] + µ↑ (y1 ,x1 ,...,yt ,xt ) (xi ) | xi ≤ z}
for z ∈ P + .
Here we note the following fact.
Lemma 3.7 Let y1 , x1 , . . . , yt , xt be a sequence of elements with condition
N. Then ν ↓ (y1 ,x1 ,...,yt ,xt ) , ν ↑ (y1 ,x1,...,yt ,xt ) ∈ T (P ).
Proof Set ν ↓ = ν ↓ (y1 ,x1 ,...,yt ,xt ) , ν ↑ = ν ↑ (y1 ,x1 ,...,yt ,xt ) and yt+1 = ∞. It is easily
verified by the definition that if z1 , z2 ∈ P + and z1 < z2 , then
ν ↓ (z1 ) > ν ↓ (z2 ) and ν ↑ (z1 ) > ν ↑ (z2 ).
Further, since ν ↓ (∞) = max{rank[∞, yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ) | ∞ ≤ yi } = 0,
we see by the above inequality that ν ↓ (z) > 0 for any z ∈ P .
Now we prove that ν ↑ (z) > 0 for any z ∈ P . It is enough to show that
ν ↑ (∞) = min{−rank[xi , ∞] + µ↑ (y1 ,x1,...,yt ,xt ) (xi ) | xi ≤ ∞} ≥ 0. Assume the
contrary and take i with −rank[xi , ∞] + µ↑ (y1 ,x1 ,...,yt ,xt ) (xi ) < 0. Then y1 , x1 ,
. . . , yi , xi is a sequence with condition N and rmax − r(y1 , x1 , . . . , yi , xi ) =
−rank[xi , ∞] + µ↑ (y1 ,x1 ,...,yt ,xt ) (xi ) < 0. This contradicts to the definition of
rmax .
We state the following important properties of ν ↓ and ν ↑ .
10
Lemma 3.8 Let y1 , x1 , . . . yt , xt be a sequence of elements in P with condition N. Suppose that r(y1, x1 , . . . , yt, xt ) = rmax . Then ν ↓ (y1 ,x1 ,...,yt ,xt ) and
ν ↑ (y1 ,x1 ,...,yt ,xt ) are minimal elements of T (P ). Further,
ν ↓ (y1 ,x1 ,...,yt ,xt ) (yi ) = µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ),
(3.1)
ν ↓ (y1 ,x1 ,...,yt ,xt ) (xi−1 ) = rank[xi−1 , yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ),
(3.2)
ν ↑ (y1 ,x1 ,...,yt ,xt ) (xi−1 ) = µ↑ (y1 ,x1 ,...,yt ,xt ) (xi−1 ),
(3.3)
ν ↑ (y1 ,x1 ,...,yt ,xt ) (yi ) = −rank[xi−1 , yi ] + µ↑ (y1 ,x1,...,yt ,xt ) (xi−1 )
(3.4)
for 1 ≤ i ≤ t + 1, where we set yt+1 = ∞. In particular, ν ↓ (y1 ,x1 ,...,yt ,xt ) (x0 ) =
ν ↑ (y1 ,x1 ,...,yt ,xt ) (x0 ) = rmax .
Proof By Lemma 3.7, we see that ν ↓ (y1 ,x1 ,...,yt ,xt ) ∈ T (P ). First we
show (3.1). Assume the contrary and take j with ν ↓ (y1 ,x1 ,...,yt ,xt ) (yj ) 6=
µ↓ (y1 ,x1 ,...,yt ,xt ) (yj ). Then j ≤ t, ν ↓ (y1 ,x1 ,...,yt ,xt ) (yj ) > µ↓ (y1 ,x1,...,yt ,xt ) (yj ) and
there exists yi such that yj ≤ yi and
ν ↓ (y1 ,x1,...,yt ,xt ) (yj ) = rank[yj , yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ).
Since y1 , x1 , . . . , yt , xt is a sequence with condition N, yk 6≥ xj for any k
with 1 ≤ k ≤ j − 1. Thus, i ≥ j, since xj < yj ≤ yi . Since xj−1 < yj ≤ yi , we
see that y1 , x1 , . . . , yj−1 , xj−1 , yi , xi , . . . , yt , xt is a sequence with condition
N. Further, since
rank[xj−1 , yi ] ≥ rank[xj−1 , yj ] + rank[yj , yi ],
we see that
r(y1 , x1 , . . . , yj−1, xj−1 , yi , xi , . . . , yt , xt )
j−1
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + rank[xj−1 , yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi )
=
k=1
≥
j−1
X
(rank[xk−1 , yk ] − rank[xk , yk ])
k=1
+rank[xj−1 , yj ] + rank[yj , yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi )
=
j−1
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + rank[xj−1 , yj ] + ν ↓ (y1 ,x1 ,...,yt ,xt ) (yj )
k=1
j−1
>
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + rank[xj−1 , yj ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yj )
k=1
= r(y1 , x1 , . . . , yt , xt )
= rmax .
11
This contradicts to the definition of rmax . Therefore, we see (3.1).
Next we show (3.2).
Assume the contrary and take j with
↓
ν (y1 ,x1 ,...,yt ,xt ) (xj )
6=
rank[xj , yj+1] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yj+1).
Then
↓
↓
ν (y1 ,x1 ,...,yt ,xt ) (xj ) > rank[xj , yj+1] + µ (y1 ,x1,...,yt ,xt ) (yj+1) and there is yi
such that xj ≤ yi and
ν ↓ (y1 ,x1 ,...,yt ,xt ) (xj ) = rank[xj , yi ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yi ).
Since y1 , x1 , . . . , yt , xt is a sequence with condition N, we see that i ≥ j.
Moreover, since
rank[xj , yj ] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yj )
= rank[xj , yj+1] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yj+1)
6= ν ↓ (y1 ,x1 ,...,yt ,xt ) (xj ),
we see that i ≥ j + 2. Therefore, y1 , x1 , . . . , yj , xj , yi , xi , . . . , yt , xt is a
sequence with condition N and
r(y1 , x1 , . . . , yj , xj , yi , xi , . . . , yt , xt )
j
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + rank[xj , yi ] + µ↓ (y1 ,x1,...,yt ,xt ) (yi )
=
k=1
j
=
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + ν ↓ (y1 ,x1 ,...,yt ,xt ) (xj )
X
(rank[xk−1 , yk ] − rank[xk , yk ]) + rank[xj , yj+1] + µ↓ (y1 ,x1 ,...,yt ,xt ) (yj+1)
k=1
j
>
k=1
= r(y1 , x1 , . . . , yt , xt )
= rmax ,
contradicting the definition of rmax . Therefore, we see (3.2).
By (3.1), (3.2) and Lemma 3.2, we see that ν ↓ (y1 ,x1 ,...,yt ,xt ) is a minimal element of T (P ). Further, we see that ν ↓ (y1 ,x1 ,...,yt ,xt ) (x0 ) = r(y1 , x1 , . . . , yt , xt ) =
rmax .
By considering the poset Q whose base set is P + and
z < w in Q ⇐⇒ z > w in P + ,
we see that ν ↑ (y1 ,x1 ,...,yt ,xt ) is also a minimal element of T (P ) and (3.3) and
(3.4) hold.
12
Since deg T ν = ν(x0 ) for ν ∈ T (P ), we see by Corollaries 2.6 and 3.5 and
Lemma 3.8, the following
Theorem 3.9 RK (H) is level if and only if r(y1 , x1 , . . . , yt , xt ) ≤ r for any
sequence of elements in P with condition N, i.e., rmax = r.
As a corollary, we can reprove our previous result.
Corollary 3.10 ([3]) If [x, ∞] is pure for any x ∈ P \ {x0 }, then RK (H)
is level.
Proof By assumption,
rank[x, y] = rank[x, ∞] − rank[y, ∞]
for any x, y ∈ P \ {x0 } with x ≤ y. Therefore, for any sequence y1 , x1 , . . . ,
yt , xt with condition N,
r(y1 , x1 , . . . , yt , xt )
t
X
=
(rank[xi−1 , yi ] − rank[xi , yi ]) + rank[xt , ∞]
i=1
= rank[x0 , y1 ] + rank[y1 , ∞]
≤ r.
Next we show that for any integer d with r ≤ d ≤ rmax , there is a
generator of the canonical module of RK (H) with degree d. First we state
the following
Lemma 3.11 Let ν be a minimal element of T (P ) and k a positive integer.
Set νk (x) = max{ν(x) − k, rank[x, ∞]} for x ∈ P + . Then νk is a minimal
element of T (P ).
Proof It is clear that νk ∈ T (P ). By Lemma 3.3, we see that there is a
sequence y1 , x1 , . . . , yt , xt of elements of P with condition N such that
ν(xi ) − ν(yi+1 ) = rank[xi , yi+1 ]
for 0 ≤ i ≤ t, where we set yt+1 = ∞.
First consider the case where νk (xi ) − νk (yi+1) = rank[xi , yi+1] for 0 ≤ i ≤
t. Then, by Lemma 3.2, we see that νk is a minimal element of T (P ). Next
13
consider the case that there exists i with νk (xi ) − νk (yi+1) 6= rank[xi , yi+1 ].
Set
j = min{i | νk (xi ) − νk (yi+1 ) > rank[xi , yi+1 ]}.
If νk (xj ) = ν(xj ) − k, then
νk (xj ) − νk (yj+1)
= ν(xj ) − k − max{ν(yj+1) − k, rank[yj+1 , ∞]}
≤ ν(xj ) − ν(yj+1 )
= rank[xj , yj+1].
This contradicts to the definition of j.
Thus, νk (xj ) = rank[xj , ∞]. Since νk (xi ) − νk (yi+1 ) = rank[xi , yi+1] for
0 ≤ i ≤ j − 1 by the definition of j, by applying Lemma 3.2 to y1 , x1 , . . . ,
yj , xj , we see that νk is a minimal element of T (P ).
As a corollary, we see the following fact.
Theorem 3.12 Let d be an integer with r ≤ d ≤ rmax . Then there exists a
generator of the canonical module of RK (H) with degree d.
Proof Set k = rmax − d. Take a sequence y1 , x1 , . . . , yt , xt with condition
N such that r(y1 , x1 , . . . , yt , xt ) = rmax and put
ν(z) = max{ν ↓ (y1 ,x1 ,...,yt ,xt ) (z) − k, rank[z, ∞]}
for z ∈ P . Then ν(x0 ) = max{ν ↓ (y1 ,x1 ,...,yt ,xt ) (x0 )−k, r} = max{rmax −k, r} =
max{d, r} = d. Since ν is a minimal element of T (P ) by Lemmas 3.11 and
3.8, we see by Corollary 2.4 that T ν is a generator of the canonical module
of RK (H) with degree d.
Finally in this section, we make a remark on P when RK (H) is level.
First we state the following
Lemma 3.13 Set F = {x ∈ P | rank[x0 , x]P + + rank[x, ∞]P + < r}. Suppose
that RK (H) is level. Then for any z1 , z2 ∈ P + \ F with z1 ≤ z2 ,
rank[x0 , z1 ]P + + rank[z1 , z2 ]P + + rank[z2 , ∞]P + = r.
(3.5)
rank[z1 , z2 ]P + \F = rank[z1 , z2 ]P + .
(3.6)
and
14
Proof We first show (3.5). The cases where z1 = x0 , z2 = ∞ or z1 = z2 are
clear from the definition of F . Suppose that z1 , z2 ∈ P \ F and x0 < z1 < z2 .
Then z2 , z1 is a sequence with condition N. Since RK (H) is level, we see by
Theorem 3.9 that
rank[x0 , z2 ]P + − rank[z1 , z2 ]P + + rank[z1 , ∞]P + ≤ r.
Since z1 , z2 6∈ F , we see that rank[x0 , z2 ]P + = r − rank[z2 , ∞]P + and
rank[z1 , ∞]P + = r − rank[x0 , z1 ]P + . Thus,
2r − rank[x0 , z1 ]P + − rank[z1 , z2 ]P + − rank[z2 , ∞]P + ≤ r,
i.e.,
rank[x0 , z1 ]P + + rank[z1 , z2 ]P + + rank[z2 , ∞]P + ≥ r.
The opposite inequality is obvious. Thus, we see (3.5).
Next we prove (3.6). Assume the contrary. Take a maximal chain
z1 = w0 < w1 < · · · < wt = z2
of [z1 , z2 ]P + such that t = rank[z1 , z2 ]P + . Since
rank[x0 , wi ]P + + rank[wi , ∞]P +
≥ rank[x0 , z1 ]P + + rank[z1 , wi ]P + + rank[wi , z2 ]P + + rank[z2 , ∞]P +
= rank[x0 , z1 ]P + + rank[z1 , z2 ]P + + rank[z2 , ∞]P +
= r,
we see that wi 6∈ F for any i = 1, . . . , t − 1. Therefore, rank[z1 , z2 ]P + \F ≥
t = rank[z1 , z2 ]P + . The opposite inequality is obvious.
By the above lemma and Lemma 2.9, we see the following fact.
Proposition 3.14 Let F be as in Lemma 3.13. If RK (H) is level, then
P + \ F is pure of rank r. In particular, if RK (H) is level and F = ∅, then
RK (H) is Gorenstein.
15
Example 3.15 Let
t
❇
❇
t❇
❇
t ❇
t
❇
❇
❇t
t
t
❅
❅
❅t
t
P1 =
t
t
❇
❇
t❇
t
and P2 =
t
t
t
t
t
✁
✁
❇
✁ t
t ❇
✟t
✁
✟
❇
✟
❇ ✁✟✟
❇✁
t
t
t✟
t
t
❅
.
t
t
✟
✟
✟
✟
❅
❅✟
t ✟
Then rankPi+ = 6 for i = 1, 2. Let Fi be the subset of Pi defined as in
Lemma 3.13 and let Hi be the distributive lattice corresponding to Pi for
i = 1, 2. Then Pi+ \ Fi is pure but RK (Hi ) is not level for i = 1, 2. Thus,
the converse of Proposition 3.14 does not hold. There are 2 generators with
degree 6, 3 generators with degree 7 and 6 generators with degree 8 of the
canonical module of RK (H1 ) and there are 2 generators with degree 6, 48
generators with degree 7 and 108 generators with degree 8 of the canonical
module of RK (H2 ).
4
Characterization of level type 2 Hibi rings
As Example 3.15 shows, it is very hard to describe Cohen-Macaulay type
of the Hibi ring in terms of the combinatorial structure of P . However,
we can characterize Hibi ring RK (H) to be of type 2 with respect to the
combinatorial property of P . Recall that, by Corollary 2.4, we see that
typeRK (H) is the number of minimal elements of T (P ).
In this section, we state a characterization of a Hibi ring to be level and
of type 2. First we make the following
Definition Let y1 , x1 , y2 , x2 , . . . , yt , xt be a sequence of elements in P . If
the following 4 conditions are satisfied, we say that y1 , x1 , y2 , x2 , . . . , yt , xt
is an irredundant sequence.
(1) y1 , x1 , y2 , x2 , . . . , yt , xt satisfies condition N.
(2) r(y1 , x1 , . . . , yt , xt ) = rmax .
16
(3) If y1′ , x′1 , y2′ , x′2 , . . . , yt′′ , x′t′ is a sequence of elements in P with condition
N and r(y1′ , x′1 , . . . , yt′′ , x′t′ ) = rmax , then t ≤ t′ .
(4) For any i with 1 ≤ i ≤ t, rank[z, yi+1 ] − rank[z, yi ] < rank[xi , yi+1 ] −
rank[xi , yi ] for any z ∈ (xi , yi+1]∩(xi , yi ] and rank[xi−1 , z]−rank[xi , z] <
rank[xi−1 , yi ] − rank[xi , yi ] for any z ∈ [xi−1 , yi ) ∩ [xi , yi ), where we set
yt+1 = ∞.
It is clear that there exists an irredundant sequence. Further, by Theorem
3.9, RK (H) is level if and only if the empty sequence is an irredundant
sequence.
Next we state the following
Lemma 4.1 Set F = {x ∈ P | rank[x0 , x] + rank[x, ∞] < r}. Then the
number of generators of the canonical module of degree r is greater than #F .
In particular, typeRK (H) > #F .
Proof Set F = {f1 , f2 , . . . , fu } and i < j if fi < fj . For t with 1 ≤ t ≤ u + 1,
set
rank[x, ∞]
if x 6∈ F or x ∈ F and x = fj with j ≥ t,
νt (x) =
r − rank[x0 , x]
otherwise.
Then it is easily verified that νt is an element of T (P ). Further, since νt (x0 ) =
r, we see that νt is a minimal element of T (P ).
Since for any t, t′ with t < t′ , νt (ft ) < νt′ (ft ), we see that νt 6= νt′ .
Therefore, we see that there are at least u + 1 minimal elements ν of T (P )
such that ν(x0 ) = r.
Now we state the main theorem of this section.
Theorem 4.2 RK (H) is level and typeRK (H) = 2 if and only if there exists
z ∈ P with the following conditions.
(1) rank[x0 , z] + rank[z, ∞] = r − 1.
(2) P + \ {z} is pure of rank r.
(3) [x0 , z] and [z, ∞] are pure.
Remark As Example 3.15 shows, (3) of Theorem 4.2 does not follow from
(1) and (2).
17
Proof of Theorem 4.2 We first assume that RK (H) is level and
typeRK (H) = 2 and prove that there exists z ∈ P with (1), (2) and (3). Set
F = {x ∈ P | rank[x0 , x] + rank[x, ∞] < r}. If F = ∅ then, by Proposition
3.14, we see that RK (H) is Gorenstein, contradicting the assumption. Therefore, F 6= ∅. If #F ≥ 2, then by Lemma 4.1, we see that typeRK (H) > 2,
again contradicting the assumption. Thus, #F = 1.
Set F = {z}. We show that z satisfies (1), (2) and (3). Suppose that
rank[x0 , z] + rank[z, ∞] ≤ r − 2.
Then if we set
µt (x) =
rank[x, ∞]
rank[x, ∞] + t
if x 6= z,
if x = z,
then µt is a minimal element of T (P ) for 0 ≤ t ≤ 2. Thus, typeRK (H) ≥ 3,
contradicting the assumption.
Therefore,
rank[x0 , z] + rank[z, ∞] = r − 1,
i.e., we see (1). Further, we see by Proposition 3.14, that P + \ {z} is pure of
rank r. Thus, we see (2).
Now let y be an arbitrary element of [z, ∞] such that z <· y. We shall show
that [y, ∞] is pure and rank[y, ∞] = rank[z, ∞] − 1. The case where y = ∞
is clear. Suppose y 6= ∞. Then y, z is a sequence with condition N. Since
RK (H) is level, we see by Theorem 3.9 that rank[x0 , y] − 1 + rank[z, ∞] =
r(y, z) ≤ r. Since rank[x0 , y] = r − rank[y, ∞], we see that rank[z, ∞] −
rank[y, ∞] ≤ 1. Therefore, rank[y, ∞] = rank[z, ∞] − 1. Further, [y, ∞] is
pure by Proposition 3.14.
Since y is an arbitrary element of [z, ∞] with z <· y, we see that [z, ∞] is
pure. We see that [x0 , z] is pure by the same way. Thus, we see (3).
Next we assume that there exists z ∈ P with (1), (2) and (3) and prove
that RK (H) is level and typeRK (H) = 2.
We first note that it follows from (2) that for any w1 , w2 ∈ P + \ {z},
rank[w1 , w2 ]P + \{z} = rank[w1 , w2 ]P + . In particular, rank[x, y] = rank[x0 , y] −
rank[x0 , x] for any x, y ∈ P + \ {z} with x < y. We also see that if x < z,
then rank[x, z] = rank[x0 , z] − rank[x0 , x] and if y > z, then rank[z, y] =
rank[z, ∞] − rank[y, ∞], since [x0 , z] and [z, ∞] are pure.
Now let y1 , x1 , . . . , yt , xt be an irredundant sequence and set yt+1 = ∞.
We shall show that t = 0. Assume the contrary. Then, since y1 , x1 ,
. . . , yt , xt is an irredundant sequence, we see that r(y2 , x2 , . . . , yt , xt ) <
r(y1 , x1 , . . . , yt , xt ), i.e.,
rank[x0 , y2 ] < rank[x0 , y1] − rank[x1 , y1] + rank[x1 , y2 ].
18
(4.1)
First consider the case where x1 6= z. Since rank[x1 , yi ] = rank[x0 , yi ] −
rank[x0 , x1 ] for i = 1, 2, we see that the right hand side of (4.1) is equal to
rank[x0 , y2]. This is a contradiction.
Next consider the case where x1 = z. Since rank[x1 , yi ] = rank[x1 , ∞] −
rank[yi , ∞] for i = 1, 2 we see that the right hand side of (4.1) is equal to
rank[x0 , y1 ] + rank[y1 , ∞] − rank[y2 , ∞].
Since yi 6= z for i = 1, 2, we see that rank[yi , ∞] = r −rank[x0 , yi ] for i = 1, 2.
Therefore, the right hand side of (4.1) is equal to rank[x0 , y2 ]. This is also a
contradiction. Thus, we see that t = 0 and RK (H) is level.
Let ν be an arbitrary minimal element of T (P ). Since RK (H) is level,
we see that ν(x0 ) = r. Thus, we see that ν(x) ≥ rank[x, ∞] and r − ν(x) =
ν(x0 ) − ν(x) ≥ rank[x0 , x] i.e., rank[x, ∞] ≤ ν(x) ≤ r − rank[x0 , x] for
any x ∈ P . In particular, we see by (2) that ν(x) = rank[x, ∞] for any
x ∈ P + \ {z}. We also see by (1) that
rank[z, ∞] ≤ ν(z) ≤ r − rank[x0 , z] = rank[z, ∞] + 1.
Thus typeRK (H) ≤ 2. Further, we see typeRK (H) ≥ 2 by Lemma 4.1.
5
Characterization of non-level type 2 Hibi
rings
In this final section, we give a characterization of a Hibi ring to be non-level
and of type 2.
Theorem 5.1 RK (H) is non-level and typeRK (H) = 2 if and only if there
exist x, y ∈ P \ {x0 } with the following conditions.
(1) x <· y.
(2) rank[x0 , y] + rank[x, ∞] = r + 2.
(3) P + = [x0 , y] ∪ [x, ∞].
(4) rank[x0 , z1 ] + rank[z1 , z2 ] + rank[z2 , ∞] = r for any z1 , z2 ∈ P + with
z1 ≤ z2 and (z1 , z2 ) 6= (x, y).
Proof Set F = {z ∈ P | rank[x0 , z] + rank[z, ∞] < r}.
First we assume that RK (H) is non-level and typeRK (H) = 2 and prove
that there exist x, y ∈ P with conditions (1) to (4). Since RK (H) is not
19
level, we see by Theorem 3.9 that rmax > r. Further, since typeRK (H) ≥
rmax − r + 1 by Theorem 3.12, we see that rmax = r + 1 and the generating
system of the canonical module of RK (H) consists of 2 elements: one has
degree r and the other one has degree r + 1. Thus, by Lemma 4.1, we see
that F = ∅.
Let y1 , x1 , . . . , yt , xt be an irredundant sequence and set yt+1 = ∞. First
we claim that t = 1. Since rmax > r, we see that t ≥ 1. Further, since y1 ,
x1 , . . . , yt , xt is an irredundant sequence, we see that r(y2, x2 , . . . , yt , xt ) <
r(y1 , x1 , . . . , yt , xt ), i.e.,
rank[x0 , y2 ] < rank[x0 , y1] − rank[x1 , y1] + rank[x1 , y2 ].
Moreover, since F = ∅, we see that
r =
<
≤
=
rank[x0 , y2 ] + rank[y2 , ∞]
rank[x0 , y1 ] − rank[x1 , y1 ] + rank[x1 , y2 ] + rank[y2 , ∞]
rank[x0 , y1 ] − rank[x1 , y1 ] + rank[x1 , ∞]
r(y1, x1 ).
Since rmax = r + 1 and y1 , x1 , . . . , yt , xt is an irredundant sequence, we see
that r(y1, x1 ) = r + 1 and t = 1.
Set x = x1 and y = y1 . We show that these x and y satisfy (1) to (4).
We first show that (1). Assume the contrary. Then there is w ∈ P such
that x < w < y and rank[x, y] = rank[x, w] + rank[w, y]. Since y, x is an
irredundant sequence, we see that
rank[x0 , w] − rank[x, w] ≤ rank[x0 , y] − rank[x, y] − 1
and
−rank[w, y] + rank[w, ∞] ≤ −rank[x, y] + rank[x, ∞] − 1.
Therefore,
rank[x0 , w] − rank[x, w] − rank[w, y] + rank[w, ∞]
≤ rank[x0 , y] − 2rank[x, y] + rank[x, ∞] − 2
= r(y, x) − rank[x, y] − 2.
On the other hand, since
rank[x, w] + rank[w, y] = rank[x, y] and rank[x0 , w] + rank[w, ∞] = r,
20
we see r ≤ r(y, x) − 2. This contradicts to the fact that r(y, x) = r + 1.
Therefore, we see (1). Moreover, since
rank[x0 , y] + rank[x, ∞] − 1
= rank[x0 , y] − rank[x, y] + rank[x, ∞]
= r(y, x)
= r + 1,
we see (2).
Next we prove (3). Assume the contrary and let z ∈ P + \([x0 , y]∪[x, ∞]).
Since z 6≤ y, we see by the definition of ν ↓ (y,x) (see Definition 3.6) that
ν ↓ (y,x) (z) = rank[z, ∞]. On the other hand, since z 6≥ x, we see by the
definition of ν ↑ (y,x) that ν ↑ (y,x) (z) = r + 1 − rank[x0 , z]. Since the minimal
generating system of the canonical module of RK (H) contains only one element with degree r + 1 and ν ↓ (y,x) and ν ↑ (y,x) are minimal elements of T (P )
with degree r + 1 by Lemma 3.8, we see that ν ↓ (y,x) = ν ↑ (y,x) . Therefore,
rank[z, ∞] = ν ↓ (y,x) (z) = ν ↑ (y,x) (z) = r + 1 − rank[x0 , z].
This contradicts to the fact that rankP + = r. Therefore, we see (3).
Finally, we prove (4). The case where z1 = z2 , z1 = x0 or z2 = ∞ are
clear since F = ∅. Suppose that x0 < z1 < z2 < ∞. Then z2 , z1 is a sequence
with condition N. If r(z2 , z1 ) ≤ r, then
rank[x0 , z2 ] − rank[z1 , z2 ] + rank[z1 , ∞] ≤ r.
Since F = ∅, we see that
2r − rank[z2 , ∞] − rank[z1 , z2 ] − rank[x0 , z1 ] ≤ r.
Therefore,
rank[x0 , z1 ] + rank[z1 , z2 ] + rank[z2 , ∞] ≥ r.
Since the converse inequality holds in general, we see the result.
Now assume that r(z2 , z1 ) > r. Since rmax = r + 1, we see that r(z2 , z1 ) =
rmax = r + 1. Moreover, since the generating system of the canonical module
contains exactly one element with degree r + 1, we see that
ν ↓ (y,x) = ν ↑ (y,x) = ν ↓ (z2 ,z1 ) = ν ↑ (z2 ,z1 )
by Lemma 3.8. Since ν ↓ (z2 ,z1 ) (y) = ν ↓ (y,x) (y) > rank[y, ∞], we see that
y ≤ z2 . On the other hand, since ν ↓ (y,x) (z2 ) = ν ↓ (z2 ,z1 ) (z2 ) > rank[z2 , ∞], we
21
see that z2 ≤ y. Therefore, we see that z2 = y. We also see that z1 = x by
the same way. This proves (4).
Next we assume that there exist x, y ∈ P with conditions (1) to (4) and
prove that RK (H) is non-level and typeRK (H) = 2. First we see by (4) that
F = ∅. Further, we see that y, x is a sequence with condition N and by (1)
and (2) that
r(y, x) = rank[x0 , y] − rank[x, y] + rank[x, ∞] = r + 1.
Thus, we see that rmax > r. In particular, RK (H) is not level by Theorem
3.9.
Let y1 , x1 , . . . , yt , xt be an irredundant sequence. Since rmax > r,
we see that t ≥ 1. First consider the case where there is no i such
that (xi , yi ) = (x, y). Then r(y1 , x1 , . . . , yt , xt ) ≤ r by (1), (2) and (4),
contradicting the fact that y1 , x1 , . . . , yt , xt is an irredundant sequence.
Thus, there exists i such that (xi , yi ) = (x, y). Then by (4), we see that
r(y1 , x1 , . . . , yt , xt ) = r(yi , xi ) = r(y, x) = r + 1. Therefore, we see that y, x
is the unique irredundant sequence and rmax = r(y, x) = r + 1.
Let ν be a minimal element of T (P ) with ν(x0 ) = r. Since ν(z) ≥
rank[z, ∞] and r − ν(z) = ν(x0 ) − ν(z) ≥ rank[x0 , z], we see that ν(z) =
rank[z, ∞] for any z ∈ P + , since F = ∅. Therefore, in the generating system
of the canonical module of RK (H), there exists exactly one element with
degree r.
Let ν be a minimal element of T (P ) with ν(x0 ) = r + 1. It is enough
to show that ν = ν ↓ (y,x) . As in the proof of Lemma 3.3, set U1 := {w ∈
P + | rank[x0 , w]P + = ν(x0 ) − ν(w)}, D1 := {z ∈ P + \ U1 | ∃w ∈ U1 such
that w > z}, U2 := {w ∈ P + \ (U1 ∪ D1 ) | ∃z ∈ D1 such that z < w and
rank[z, w]P + = ν(z) − ν(w)}, D2 := {z ∈ P + \ (U1 ∪ D1 ∪ U2 ) | ∃w ∈ U2 such
that w > z} and so on. Since ν(x0 ) = r + 1, we see that ∞ 6∈ U1 . On the
other hand, since
∞ ∈ U1 ∪ D1 ∪ U2 ∪ D2 ∪ · · · ,
we see that D1 6= ∅.
Let z1 ∈ D1 and let z2 ∈ U1 such that z2 > z1 . Then
rank[x0 , z2 ]
= ν(x0 ) − ν(z2 )
= ν(x0 ) − ν(z1 ) + ν(z1 ) − ν(z2 )
≥ rank[x0 , z1 ] + 1 + rank[z1 , z2 ].
(5.1)
Therefore, we see by (4) that (z1 , z2 ) = (x, y). Further, by (1), (2) and the
fact that F = ∅, we see that equality holds for (5.1). Therefore, we see that
ν(x0 ) − ν(x) = rank[x0 , x] + 1
22
(5.2)
and
ν(x) − ν(y) = 1.
(5.3)
Let z be an arbitrary element of P . We shall show that ν(z) = ν ↓ (y,x) (z).
First consider the case where z = x. Since ν(x) = rank[x, ∞] by (5.2)
and the facts that F = ∅ and ν(x0 ) = r + 1, we see that ν(x) = ν ↓ (y,x) (x) by
Lemma 3.8.
Next consider the case where z = y. Since ν(x) = rank[x, ∞], we see by
(5.3) and (1) that ν(y) = rank[x, ∞] − rank[x, y] = ν ↓ (y,x) (y) by Lemma 3.8.
Now consider the case where z ≤ y and z 6= x. Since rank[x0 , y] =
rank[x0 , z]+rank[z, y] by (4), we see that z ∈ U1 and ν(z)−ν(y) = rank[z, y].
Since ν(x0 ) = r + 1, we see that ν(z) = ν(x0 ) − rank[x0 , z] > rank[z, ∞].
Therefore, we see by the definition of ν ↓ (y,x) that
ν(z) = ν(y) + rank[z, y] = ν ↓ (y,x) (y) + rank[z, y] = ν ↓ (y,x) (z).
Finally consider the case where z 6≤ y. By (3), we see that z > x.
Further, we see by (4) that rank[x, ∞] = rank[x, z] + rank[z, ∞]. Since
ν(x) = rank[x, ∞], we see that ν(z) = rank[z, ∞] = ν ↓ (y,x) (z).
References
[1] Hibi, T.: Distributive lattices, affine smigroup rings and algebras with
straightening laws. in “Commutative Algebra and Combinatorics” (M.
Nagata and H. Matsumura, ed.), Advanced Studies in Pure Math. 11
North-Holland, Amsterdam (1987), 93–109.
[2] Hochster, M.: Rings of invariants of tori, Cohen-Macaulay rings generated by monomials and polytopes. Ann. of Math. 96 (1972), 318-337.
[3] Miyazaki, M.: A sufficient condition for a Hibi ring to be level and levelness of Schubert cycles. Communications in Algebra. 35 (2007), 28942900.
[4] Stanley R. P.: Cohen-Macaulay Complexes. in Higher Combinatorics
(M.Aigner, ed.), Reidel, Dordrecht and Boston, (1977), 51–62.
[5] Stanley, R. P.: Hilbert Functions of Graded Algebras. Adv. Math. 28
(1978), 57–83.
23
| 0math.AC
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.